Sampling

18/1/2014 Methods of Randomization in Experimental Design : SAGE Research Methods
http://srmo.sagepub.com.proxy.grenoble-em.com/view/methods-of-randomization-in-experimental-design/SAGE.xml?rskey=ewhuX9&row=402 1/1
Table of contents
Front Matter
1. Randomized Experiments
2. Between-Subjects Designs Randomization
3. Within-Subjects Designs Randomization
4. Validity Issues, Analysis Guidelines, and Reporting Standards
Back Matter
Grenobl e Ecol e de Management
18/1/2014 Chapter 1: Randomized Experiments : SAGE Research Methods
http://srmo.sagepub.com.proxy.grenoble-em.com/view/methods-of-randomization-in-experimental-design/n1.xml?rskey=ewhuX9 1/11
Chapter 1: Randomized Experiments
1.1 Nature and Structure of Randomized Experiments
In broad terms, methods are the linking procedures between theory and data. They embody the theoretical hypothesis in
the research design, specifying the conditions and technical devices to collect, analyze, and interpret relevant basic
information (raw data). In the nomothetic or quantitative approach, as opposed to the idiographic or qualitative one,
methods of empirical investigation are usually classified on the basis of the structural properties of the underlying design
and the degree to which they allow valid causal inferences.
Leaving out theoretical innovation, research synthesis and evaluation (including meta-analysis and bibliometric surveys),
and documental studies focused on previously archived materials, all investigations based either on self-reports (via self-
administrated questionnaires or interviews) or on direct observation of research participants fall into one of the three main
categories: (1) experimental, (2) quasi-experimental, and (3) nonexperimental research (see Figure 1.1; see also Alferes,
2012, for a methodological classification of theoretical and empirical studies in psychology and related disciplines). As the
reader can verify from the decision chart shown in Figure 1.1, this classification schemeinspired by the Campbellian
approach to the validity of causal inferences (Campbell, 1957; Shadish et al., 2002)is organized around two key
attributes of empirical studies: (1) systematic variation of the presumed causes (independent variables manipulation) and
(2) use of randomization procedures. The presence of the first attribute separates experimental and quasi-experimental
research from nonexperimental research; the presence of the second one separates randomized experiments
(experimental designs) from nonrandomized experiments (quasi-experimental designs). The specific use of randomization
procedures in experimental design depends on the manipulation strategy adopted by the researcher: (a) each unit is only
exposed to one experimental condition and the randomization procedure is used to determine what exactly is that
condition (between-subjects designs), or (b) each unit is exposed to two or more experimental conditions and the
randomization procedure is used to determine the order in which the conditions will be presented (within-subjects
designs).
As stated in the opening paragraph of the Preface, this book is about experimental designs and the devices required for an
unbiased and efficient estimation of treatment causal effects. Formally speaking, a causal effect is
Figure 1.1 Decision chart for the classification of experimental, quasi-experimental, and nonexperimental designs according to
the Campbellian approach
the difference between what happens when an experimental unit is subjected to a treatment and what would happen if it
was not subjected to the same treatment or, which is equivalent, the difference between the responses of an experimental
unit when simultaneously subjected to two alternative treatments (differential causal effect). Stated in another way, the
inference of a causal effect requires counterfactual evidence. Yet, regarding a concrete experimental unit, it is impossible
to obtain such evidence: We are unable to apply and not apply a treatment (or apply two alternative treatments) to an
experimental unit at the same time.
A tentative solution for this problem could be either the comparison of the responses of two units (one receiving and one
not receiving the treatment or one receiving the treatment and the other the alternative treatment) or the comparison of the
responses of the same unit observed in two successive periods. However, this time, the counterfactual evidence is
equivocal: In the first case, the treatment effect is completely confounded with the intrinsic characteristics of the
experimental unit; in the second case, the treatment effect is completely confounded with any systematic or random
variation potentially associated with the different periods of observation. A better solution, and the only one really feasible,
is to replicate the experiment with other experimental units. Provided that certain assumptions are verified, having
observations from several units can allow us to separate the treatment effect from subjects and temporal sequence
effects.
What we have been saying is the core content of Rubin's causal model (Rubin, 1974, 2006, 2007; see also Rubin, 2004,
for a pedagogical introduction), which defines a causal effect in terms of the mean difference in the potential outcomes
between those who were submitted and those who were not submitted to a treatment, as long as the experimenter can
guarantee what Rubin calls stable-unit-treatment-value assumption (SUTVA). Among other things, this assumption states
that the potential outcome of one unit must not be affected by the actual assignment (to experimental conditions) of the
remaining units. More precisely, SUTVA is a twofold assumption. First, it implies that there are no hidden or different
versions of the treatments; that is, the selected treatment levels (or treatment level combinations) are administrated
without any modification from the beginning to the end of the experiment. Second, it implies no interference or interaction
between subjects who are receiving different treatment levels (or treatment level combinations). If we are dealing with a
randomized experiment, this means that the independence condition introduced by the initial randomization must be
preserved during the experiment to avoid potential contamination effects resulting from interactions between subjects
assigned to distinct experimental conditions. Some authors subsume randomization procedures under the SUTVA rubric,
despite the clear statement of Rubin (2007, 2010) that substantive assumptions must be distinguished from the underlying
assignment mechanism.
Rubin's causal model can be seen as an elaboration of the three fundamental principles of experimentation introduced by
Fisher (1935/1966): (1) replication, (2) randomization, and (3) local control. The first principle implies the recording of
observations from several units to estimate causal effects (mean differences). Randomization guarantees that the estimate
is a nonbiased one. Local control ensures more precision in the estimation (i.e., a more efficient nonbiased estimator; for
a synthesis of the properties of statistical estimators, see Fox, 2009). Stated in another way, randomization rules out
alternative explanations based on the intrinsic characteristics of experimental units, whereas local control reduces the
magnitude of random noise (residual variability) in the experiment.
In addition to the elaboration of Fisher's principles, a distinctive feature of Rubin's causal model is the replacement of the
observed outcomes notation with the potential outcomes notation underlying his contrafactual approach to
experimentation (randomized experiments) and quasi-experimentation (observational studies, according to the terminology
introduced by Cochran [1965, 1983] and popularized by Rosenbaum [2002, 2010] and Rubin himself [Cochran & Rubin,
1973; Rubin, 1973]). In the context of this introductory chapter, it is sufficient to remark that the potential outcomes
notation, initially proposed by Neyman (1923/1990), constitutes a coherent framework for the analysis of randomized and
nonrandomized experiments and is particularly relevant in cases where the conditions established by the initial
randomization are broken throughout the experiment and the SUTVA is violated. We will revisit this issue in the final
chapter, which is centered on practical matters and the guiding principles of data analysis.
For now, we will be returning to the basics of experimental design, reproducing an extended definition given by Kirk (1995),
in which the nature and the structure of randomized experiments are clearly detailed:
The term experimental design refers to a plan for assigning subjects to experimental conditions and the statistical
analysis associate with the plan. The design of an experiment to investigate a scientific or research hypothesis
involves a number of interrelated activities:
1. Formulation of statistical hypotheses that are germane to the scientific hypothesis. A statistical hypothesis is
a statement about (a) one or more parameters of a population or (b) the functional form of a population.
Statistical hypotheses are rarely identical to scientific hypotheses; they are testable formulations of scientific
hypotheses.
2. Determination of the experimental conditions (independent variable) to be used, the measurement
(dependent variable) to be recorded, and the extraneous conditions (nuisance variables) that must be
controlled.
3. Specification of the number of subjects (experimental units) required and the population from which they will
be sampled.
4. Specification of the procedure for assigning the subjects to the experimental conditions.
5. Determination of the statistical analysis that will be performed. (pp. 1-2)
In the next section, these interrelated activities involved in experimental design are discussed in the broader context of
the validity of causal inferences, and the main structural features of randomized experiments (experimental factors,
pseudofactors, classificatory factors, and outcome measures) are conveniently described. Section 1.3 gives an overview
of experimental designs in connection with methods of randomization and must be read as an advanced organizer for the
core contents of Chapters 2 and 3. This introductory chapter ends with a brief section devoted to important terminological
and notational issues.
1.2 Experimental Design and Validity of Scientific Inferences
Scientific hypotheses are conjectural statements about the relationships between theoretical constructs, empirically
represented by particular events or realizations (called operationalizations or measurements). In the nomothetic tradition,
relational or causal connections specified in the hypothesis are (ideally) theory driven. Inferences from data (particular
observables) to the hypothesis (corroboration, falsification) are the realm of the scientific enterprise and must be
distinguished from statistical inferences, which are about estimating population parameters from sampling particulars
(Meehl, 1990). Study designs are the embodiment of the theoretical hypothesis, and their structural features define the
conditions and constraints of scientific inference.
We can easily define the structural features of experimental designs by relying on the well-known Lewinian truism, which
states that in general terms, behavior (B) is a function (F) of the person (P) and of his environment (E), B = F(P, E)
(Lewin, 1946/1997, p. 337). First, behavioral measures (B) are dependent variables (outcome measures). Second,
environmental or situational variables (E) susceptible of being manipulated are experimental factors (independent
variables, treatments, or interventions). Third, personal or dispositional variables (P) are classificatory factors, which are
ideally controlled by random assignment of units to experimental conditions. Finally, pseudofactorsthat is,
environmental or situational variables other than the focal or primary experimental factorscan be incorporated in the
design (as secondary experimental factors), locally controlled (by holding them constant), or statistically handled (by
measurement and subsequent modeling as covariables). When classificatory factorsconceptualized either as
randomization restrictions (blocks or strata) or as measured covariables (substantive dispositional moderators)are
included in the design, they are occasionally labeled passive independent variables and contrasted with true (i.e.,
manipulated) or active independent variables.
A substantive classification of classificatory factors, experimental factors, outcome measures, and pseudofactors is given
in Table 1.1, where some disciplinary traditions in experimental research are also identified. Using the terminology
proposed by Cronbach (1982) to describe the basic
Table 1.1 Classification of Experimental Factors, Classificatory Factors, Pseudofactors, and Outcome Measures in Randomized
Experiments
elements of experimental designs (UTOS: units, treatments, observations, and settings), the relationships between these
elements are depicted in the lower left panel of Figure 1.2.
Figure 1.2 is a graphical representation of the widely known Campbellian approach to the validity of scientific inferences
(Campbell, 1957;
Figure 1.2 Components of the Campbellian validity typology
Campbell & Stanley, 1966; Cook & Campbell, 1979; Shadish & Cook, 2009; Shadish et al., 2002). To state it briefly,
scientific claims are evaluated by the degree to which the underlying design allows the researcher to make and to
generalize local causal inferences. Making valid local causal inferences is synonymous with giving unequivocal evidence of
covariance between the presumed causes and the expected effects and, simultaneously, ruling out alternative
explanations for the observed relationship. That is, the validity of local causal inferences depends on the statistical and
experimental components of the research design (see statistical conclusion validity and internal validity in the lower right
panel of Figure 1.2). Generalization of causal relationships requires a twofold approach: (1) generalizing from sampling
particulars to higher order constructs (causal generalization as representation, or construct validity) and (2) generalizing
the local causal relationship to other persons and instances not included in the study (causal generalization as
extrapolation, or external validity) (see the two upper panels of Figure 1.2).
The four dimensions of the validity of scientific inferences (external, construct, internal, and statistical conclusion validity)
can be thought of as the organizing principles for the main areas of the methodological field (sampling, measurement,
design, and analysis), matching, term by term, the critical challenges that all researchers must deal with: (1) to
extrapolate from observed samples to target populations, (2) to guarantee that their empirical realizations adequately
represent the theoretical constructs involved in the scientific hypothesis, (3) to establish causality, and (4) to model the
data ensuring that the underlying statistical relationships are not spurious. The first section of Chapter 4focusing on
practical matters related to planning and monitoring randomized experimentsis organized around the most common
strategies to overcome potential drawbacks in local (Subsection 4.1.2) and generalized (Subsection 4.1.1) causal
inferences.
1.3 Randomized Experiments and Methods of Randomization
In Section 1.1, experimental designs are contrasted with nonexperimental and quasi-experimental designs and classified
into two broad categories according to the researcher's manipulation strategy: (1) between-subjects versus (2) within-
subjects designs. Additionally, between-subjects designs are split into completely randomized designs and restrictedly
randomized designs, on the basis of the restrictions (blocking or stratifying) imposed on random assignment procedures
(see Figure 1.1). This classification scheme can be elaborated to account for other features of experimental design, such
as the number of treatments, the pattern of treatment level combinations, and the introduction of control devices other than
randomization and blocking (e.g., single and multiple pretest-posttest observations, covariates, or concomitant variables).
Relying on the elementary building blocks of randomized experiments (CRcompletely randomized, RBrandomized
block, and LSLatin square basic designs) and their possible combinations, Kirk (1995, 2003a) proposes a very useful
classification of factorial extensions of one-treatment designs. Excluding the examples of four systematic designs and
two miscellaneous designs, Kirk (2003a, p. 11) lists 34 randomized designs, classified on the basis of six criteria: (1) the
number of treatments (one vs. two or more treatments), (2) the absence/presence of covariates (analysis of variance
[ANOVA] vs. analysis of covariance [ANCOVA] designs), (3) the randomization procedures used (simple random
assignment vs. blocking prior to random assignment), (4) the structure of factorial designs (crossed vs. hierarchical
designs), (5) the absence/presence of confounding in crossed designs, and (6) the type of hierarchical designs (complete
vs. partial nesting). Ignoring ANCOVA designs, whose structure is similar to the equivalent ANOVA designs, and grouping
together the variants and extensions of incomplete block (Designs 4) and Latin square (Designs 5) designs, we get the 22
designs listed in Figure 1.3.
Adopting an analysis-centered perspective, Kirk (1995, 2003a) subsumes within-subjects or repeated measures designs
under the rubric cross-over design, a special type of the randomized block design where each block is formed by a single
experimental unit that is observed in all experimental conditions (see Chapter 3, Table 3.3, Design 6N). This classification
of within-subjects designs is consistent with the underlying statistical model of Design 6N, which for computational
purposes is precisely the same. However, as Kirk (1995) observes, compared with designs containing homogeneous but
different units per block, the cross-over design has distinct interpretations and generalizes to different target populations.
More important in the context of this book, the counterbalancing procedures applying to Design 6N and to other types of
cross-over designs are quite different from the random assignment procedures used in the between-subjects designs, and
therefore, we have chosen to handle within-subjects designs randomization in a separate chapter (Chapter 3). For the rest,
Figure 1.3 can be taken as an outline for the description of random assignment and blocking procedures in between-
subjects designs (Chapter 2). The distinct types of the one-factor cross-over design (Designs 6 in Figure 1.3) and their
factorial extensions are listed and labeled in Table 3.3 (Chapter 3). The main randomization methods described and
illustrated in Chapters 2 and 3 are named and sequentially numerated in Tables 1.2 and 1.3.
This arrangement of the core contents of this monograph fits a widely used organization scheme in statistical analysis
and research
Figure 1.3 Classification of randomized experiments (based on Kirk, 1995, 2003a)
Table 1.2 Methods of Randomization of Between-Subjects Experimental Designs
Table 1.3 Methods of Randomization of within-Subjects (Cross-Over)
design reference books (e.g., Anderson, 2001; Keppel & Wickens, 2007; Maxwell & Delaney, 2004; Tabachnick & Fidell,
2007; Winer, Brown, & Michels, 1991). Shadish et al. (2002, p. 258), adopting a more methodological approach and
omitting blocking and stratifying procedures, present diagrams for nine randomized experiments classified according to
three criteria: (1) the inclusion of pretests (posttest-only designs vs. pretest-posttest designs), (2) the number of
experimental factors (one-factor vs. two-factor designs), and (3) the manipulation strategies used (between-subjects vs.
within-subjects designs). From the randomization perspective adopted here, eight designs are completely randomized
designs (three variations of the pretest-posttest design, three variations of the posttest-only design, the longitudinal
design, and the factorial design), and the remaining one (cross-over design) is a within-subjects or repeated measures
design.
1.4 Terminological and Notational Issues
The designation under which randomized experiments are known varies according to the dominant traditions in each
scientific discipline (e.g., randomized controlled trials or randomized clinical trials in medicine and the health sciences)
and, even within the same discipline, researchers use different labels, as is the case in some areas of psychology, where
true experiments, defined as the gold standard of experimentation, are contrasted with quasi-experiments, which are
done without prior random assignment of subjects to experimental conditions (see Figure 1.1). The same goes for the
labels currently applied to the major categories of experimental designs: (a) between-subjects designs are also called
independent group designs and (b) within-subjects designs are known under the interchangeable names of repeated
measures designs and cross-over designs (see Figure 1.1). The terminology is even more diverse when we consider the
designations given to the manipulated, controlled, and measured variables in a randomized experiment, as the reader can
notice on a careful inspection of Table 1.4.
Finally, a word of caution concerning the precise meaning of the key terms treatment, treatment level (also called
treatment arm in some areas of medicine and the health sciences), treatment level combination, and experimental
condition. Experimental conditions correspond to the differential features of the independent variable(s) manipulation or
treatment(s) implementation. In one-factor experiments, the number of experimental conditions is identical to the levels of
the treatment, while in (multi)factorial experiments, this number equals the number of treatment level combinations
included in the design. To make a fair trade-off between clarity of exposition and economy of words, avoiding
misunderstandings and giving the reader a consistent frame of reference, we have adopted in this monograph the
notational system depicted in Table 1.5.
Table 1.4 Common Designations Given to Manipulated, Controlled, and Measured Variables in Randomized Experiments
Table 1.5 Notational System for Treatments and Treatment Levels in Between-Subjects and Within-Subjects Experimental
Designs
http://dx.doi.org/10.4135/9781452270012.n1
18/1/2014 Chapter 2: Between-Subjects Designs Randomization : SAGE Research Methods
Chapter 2: Between-Subjects Designs Randomization
2.1 Randomization and Local Control
In spite of recent advances in the design and analysis of nonrandomized experiments (e.g., Rosenbaum, 2002, 2010;
Rubin, 2006) and innovative approaches to the inference of causality in nonexperimental studies (e.g., Pearl, 2009), the
randomized experiment has remained the best strategy when the researcher's main objective is to establish causal
relationships. Needless to say, a large number of interesting and crucial problems cannot (yet?) be approached with
experimental methods, namely, complex cognitive and affective meditational process in the behavioral sciences (Kenny,
2008) or social phenomena whose nature and structure inhibit the deliberate and systematic variation of their causal
conditions and the effective control of contextual features. In any case, it is not by virtue of hidden or transcendental
properties but for supplying a coherent and unified framework to infer causality that the randomized experiment plays a
central role in the scientific enterprise. This framework, with roots in David Hume's conceptualization of causal process
and John Stuart Mill's methods of inductive logical reasoning, includes the demonstration of temporal precedence, the
assessment of covariation, and the exclusion of alternative explanations for the observed phenomena.
This chapter is focused on the third distinctive feature of randomized experiments: exclusion of alternative explanations by
random assignment and local control of the heterogeneity of experimental units. These principles of experimentation
introduced by Ronald Fisher (1925/1970, 1935/1966) in the beginning of the 20th centurycan be thought of as central
devices for making local molar causal inferences, as they have been conceptualized by Donald Campbell and colleagues
(Campbell, 1957, 1986; Campbell & Stanley, 1966; Cook & Campbell, 1979; Shadish & Cook, 2009; Shadish et al., 2002)
for more than 50 years.
Drawing causal inferences by excluding plausible rival explanations for experimental effects is an unfinished task that we
are obliged to pursue, and the aid of some guidelines may make this task less painful. It is exactly with this purpose, and
not as a kind of an experimental design recipe book, that the so-called threats approach to scientific inferences
developed by Campbell and colleagues is introduced in Figure 2.1, as a frame of reference for understanding the role of
randomization and local control in experimental design. Figure 2.1 reproduces the lower panel of Figure 1.2, where the
experimental and statistical components of valid local inferences are schematized, and adds an organized list of internal
and statistical conclusion validity threats to the picture (for the evolution and revisions of the initial version of validity
threats presented by Campbell in 1957, we recommend the firsthand reports given in Campbell, 1986; Campbell &
Stanley, 1966; Cook & Campbell, 1979; Shadish et al. 2002).
Figure 2.1 Local molar causal inferences: Threats to internal validity (Threats 19) and to statistical conclusion validity (Threats a-
i)
The group of validity threats shown at the top of Figure 2.1 is a generic one and concerns the central aim of
experimentation: estimation and evaluation of the magnitude of causal effects. We shall return to this topic in Chapter 4.
The remaining four groups of threats are classified in terms of their proximity to the structural features of randomized
designs, as they were presented in Chapter 1. The first group (ambiguous temporal precedence and unreliability of
treatment implementation) is directly related to independent variable manipulation and is not a focal issue in the context of
this chapter. The same is true for the second group of threats subsumed under the dependent variables measurement
(testing, instrumentation, and regression for internal validity; unreliability of measures and restriction of range for statistical
conclusion validity). Note that rigorously speaking regression artifacts are relational threats underlying all structural
features (for an outstanding review and conceptualization, see Campbell & Kenny, 1999).
The last two groups of threats are more directly related to our purposes. Unwanted but systematic variations of
environmental or situational factors other than experimental factorsthreats to internal validity generically labeled as
historyas well as contextual random noise (extraneous variance) can confound the effect of the independent variable.
The environmental or situational factors can be controlled by holding them constant or by their inclusion in the design as
blocking variables (e.g., time blocking), covariables, or secondary experimental factors. Random noise is inevitably a
source of the residual (error) variance and must be considered in the data analysis stage.
Finally, the threats grouped under the Units box are the main concern of randomization and local control procedures.
Technically speaking, the aim of randomization is to prevent the systematic bias associated with the intrinsic
characteristics of experimental units in case they have been assigned to experimental conditions on the basis of some
nonrandom criteria, hence the designation selection or selection bias for this threat to internal validity. In addition to the
systematic bias potentially introduced by selection, the overall heterogeneity of experimental units, as is the case for
situational random noise, increases the residual (error) variance, and therefore the net effect of the experimental treatment
vanishes. Local control is the main strategy to deal with the overall heterogeneity of experimental units. Note that
randomization also plays a role in the control of additive and interactive effects of selection with other threats to internal
validity, namely, attrition (also called experimental mortality) and maturation (psychobiological changes in experimental
units that may be confounded with the effects of the independent variables manipulation). Also observe that randomization
per se doesn't rule out the main effects of attrition and maturation, and therefore other control devices (pre- and posttest
measures for attrition and multiple-pretest observations for maturation) must be added to the experimental design (see
Shadish et al., 2002). In the final chapter, these threats to internal and statistical conclusion validity, as well as some
threats classified under the rubrics construct and external validity in the Campbellian framework, will be discussed in
connection with the potential outcomes approach to experimentation proposed by Rubin (1974, 2006, 2007).
From a more statistical point of view, random assignment is the fundamental device to obtain an unbiased estimate of the
experimental effects by ensuring the independence of observations and making the error terms in the statistical model
uncorrelated with the terms standing for the experimental effects (for a more comprehensive discussion of statistical
issues, see Chapter 4). On the other hand, local control or blocking can be seen as a complementary technique for
eliminating or reducing the heterogeneity of units in the experiment and, this way, allowing a more precise (efficient) and
unbiased estimate of causal effects. To speak about reducing or eliminating the heterogeneity of experimental units is a
somewhat misleading shortcut to refer to the role of local control in experimental design. Rigorously speaking, local
control, or blocking, doesn't reduce any heterogeneity of experimental units. What blocking actually does is to distribute
the units homogeneously across the experimental conditions on the basis of their scores (values) on the blocking variable
(theoretically related with the outcome measure), which isin the data analysis stageincorporated in the statistical
model in such a way that the variability in the outcome measure associated with it is subtracted from the residual
variability (variability on the outcome measures within conditions). Saying it in other words, in the zero-order elimination of
heterogeneity of experimental units strategy (completely randomized designs) the total variance in the outcome measure
is divided into systematic variance (theoretically explained by the systematic variation of the experimental factors) and
random or residual variance (theoretically explained by all insufficiently controlled sources of error). In blocking (one or
more ways of elimination of heterogeneity), the variability attributable to the blocking variables is considered as systematic
variability and subtracted from the residual or error variability, which becomes smaller, thereby allowing a more efficient
estimation of the causal effect of the experimental treatment.
In conclusion, the main reason for randomization is to prevent threats to internal validity, namely, selection and related
threats. The reason for choosing restrictedly randomized designs, instead of completely randomized designs, is primarily
a statistical conclusion validity issue: to circumvent the potential heterogeneity of research units. In this sense,
restrictedly randomized designs can be thought of as error-purifying designs (Weiss, 2006). For the same reason, a
within-subjects (repeated measures or cross-over) design constitutes an ideal error-purifying design, in which the
heterogeneity of units is theoretically zero. Unfortunately, there are other problems underlying within-subjects designs
(see Sections 3.1 and 3.2).
Before proceeding with the practical features of randomization and local control, it is convenient to distinguish between a
randomization plan and a design layout (i.e., the specification of design structure, namely, the nature and number of
treatments and treatment levels and the desired pattern of treatment level combinations). Federer and Balaam (1973), in a
systematic bibliography of theoretical, methodological, and statistical literature on experimental investigations, adopt the
expressions experiment design and treatment design for these two complementary features of experimentation. The
former label applies to the way treatments are arranged in an experiment, which is the basis for the classification of
experimental designs according to the strategies of randomization and elimination of the heterogeneity of experimental
units: zero-way elimination (e.g., completely randomized designs), one-way elimination (e.g., randomized block designs),
two-way elimination (e.g., Latin square designs), three-way elimination (e.g., Graeco-Latin square designs), and four- and
higher-way elimination (e.g., Hyper-Graeco-Latin square and other complex designs). In contrast, treatment design refers
to the actual selection of treatments, including the selection of treatment levels (e.g., selection on the basis of previous
regression analysis, dose-response studies, fixed vs. random factors, etc.), handling more than one treatment (e.g.,
factorial designs), confounding, nesting, and fractional replication.
Hinkelmann and Kempthorne (2008) make a similar distinction between treatment design (the same label given by Federer
& Balaam, 1973) and error control design (their label for experiment design) and present sampling and observation design
as a third component of experimental design. In light of these distinctions, this chapter focuses primarily on experiment or
error control design. However, instead of presenting randomization methods as stand-alone tools, emphasis is placed on
the interdependence between random assignment strategies and the structural features of experimental designs
(treatment design). Sampling and observational design issues will be briefly discussed in the final chapter.
Taking a practical approach, random assignment or random allocation is the process of associating, term by term, a
systematically ordered list of treatment levels or treatment level combinations with a randomly ordered list of available
experimental units (simple random assignment, SRA) or with multiple randomly ordered lists of available experimental
units within homogeneous blocks (blocked random assignment, BRA) or strata (stratified random assignment, StrRA). It
is formally equivalent to matching a randomly ordered list of treatment level replications to a systematically ordered list of
available units (e.g., names in alphabetical order, order of enrollment in the experiment, etc.). For example, in captive or
nonsequential assignment (i.e., in situations where all experimental units are available at the onset of the experiment;
Sections 2.2-2.4), we associate a systematically ordered list of treatment replications to a randomly permuted list of units
(i.e., identification numbers [IDs] in the allocation frame are randomly ordered), so that the reader can easily grasp the
structure of the experiment in conjunction with the final allocation of units. In sequential assignment (Section 2.5),
considering the allocation procedural constraints, we associate a randomly permuted list of treatment replications to a
systematically ordered list of IDs (1-n) corresponding to the order of the units enrollment in the experiment.
When properly donethat is, when the systematically ordered list is exhaustive and unambiguous, the correspondence
between ID numbers and units is kept unchangeable through the experiment, and the random mechanism underlying the
randomly ordered list is a nonfaulty onethis association process ensures that each experimental unit has a nonzero
probability of being allocated to a particular treatment level (or treatment level combination) or, which is the same, that
each treatment level (or treatment level combination) has a nonzero probability of being assigned to a particular
experimental unit. Ideally, this probability should be equal for all units or treatment levels. Local and procedural
constraints, substantive restrictions, and the practical or theoretical relevance of specific intended comparisons can result
in or justify unequal probabilities of assignment. In any case, theoretical probabilities of assignment must be specified in
advance (see Subsection 4.2.1 for a discussion of assignment mechanisms and underlying probabilities).
The remaining sections of this chapter are organized around the type of limitations imposed on random assignment: (a)
procedural constraints (number of available units per experiment and number of intended units in each experimental
condition; equal-size homogeneous blocks of experimental units) and (b) substantive restrictions (restrictions based on
the intrinsic characteristics of experimental unitsclassificatory factors). Section 2.2 deals with simple random
assignment in completely randomized designs, that is, designs that contain only procedural constraints. Sections 2.3 and
2.4 deal with restrictedly randomized designs, that is, designs involving substantive restrictions, in which experimental
units are grouped in homogeneous blocks (randomized block and Latin square designs) or strata prior to random
assignment. Randomization and local control procedures are first described and illustrated with one-treatment designs
and then extended to factorial designs. From Section 2.2 to Section 2.4, all procedures are presented on the assumption
that the allocation frame (i.e., a systematically ordered list of all units numerated from 1 to n) is available at the onset of
the experiment (captive or nonsequential assignment). Adjustments of these procedures to sequential assignment and
adaptive randomization strategies (e.g., minimization) are described in Section 2.5.
2.2 Completely Randomized Designs
2.2.1 Basic Design
When the researcher follows the old standard rule of one variable, [holding] all the conditions constant except for one
factor which is his experimental factor or his independent variable (Woodworth, 1938, p. 2) and he or she has no
compelling reasons to engage in local control of classificatory factors, the randomization plan consists in assigning at
chance the t treatment levels to the N available experimental units (or, which is the same, assigning at chance the N
experimental units to the t treatment levels). The resulting design is called Completely Randomized Design (Design 1; see
Figure 1.3), and the randomization procedure is known as SRA. This is the most basic experimental design and one of
the three elementary building blocks of complex experimental designs (Kirk, 1995; the remaining two elementary building
blocks are the one-factor randomized block design and the Latin square designsee Section 1.3 and Subsections 2.3.1
and 2.3.4).
Rigorously speaking, SRA implies equal probabilities of assignment, which means that each experimental unit has an
equal probability of being assigned to each of the treatment levels included in the design (Method 1: Equal Probabilities of
Assignment). With a finite and small number of available experimental units, Method 1 seldom conduces to a balanced
design, that is, a design in which all treatment levels are equally replicated. The search for balanced designs, whose
properties express an ideal trade-off between costs and power of experiments and allow simplicity in the data analysis
stage, often justifies the introduction of procedural constraints to ensure that all experimental conditions have the same
number of experimental units (Method 3: Forced Equal Sizes). On the other hand, Method 4Forced Unequal Sizescan
be used on the basis of the theoretical or practical relevance of specific comparisons, which are made with greater power
than less relevant comparisons. Finally, Method 2Unequal Probabilities of Assignment, also known as Biased Coin
Method (Efron, 1971)is of special interest in the context of the adaptive randomization procedures discussed in
Subsection 2.5.5 (for a more detailed discussion of the probabilistic nature of the basic SRA methods, see Subsection
4.2.1).
To carry out SRA, we begin by associating, term by term, a series of random numbers to the entries of an allocation
frame. Units in the allocation frame must be unambiguously identified and sequentially numerated (1-n). Table 2.1 shows
such an allocation frame with 20 experimental units, ordered alphabetically by name and numerated from 1 to 20 (note
that any other systematic criterion could be used to order the units, as long as the correspondence between ID numbers
and units is kept unchangeable from the beginning to the end of the study). The associated series of random numbers
could be produced by any random numbers generating mechanism (see Appendix 1) or simply taken from an existing
table of random numbers. Throughout this monograph, we have used Table A1.1 to get series of random numbers drawn
from a uniform distribution in the interval [.000, .999]. For example, having randomly picked number 6 in the intersection of
Row 27 and Column 13 of Table A1.1 as the starting point and having decided in advance to read the table vertically (down
columns instead of across rows or diagonals, from top to bottom and from left to right), to select numbers with three digits
(digits occurring in the random selected column and in the two immediately adjacent columns), and to divide these
numbers by 1,000 (in order to obtain a random uniform series between .000 and .999), we got the series of 20 random
numbers shown in Table 2.1, beginning in .621 and ending in .143 (for a detailed description of this procedure and
alternative procedures for using tables of random numbers, see Appendix 1).
Table 2.1 Completely Randomized Design: Allocation Frame and Random Numbers (RN) (Design 1; Methods 1, 2, 3, and 4)
In the next step, we sort the ID numbers (units) by the random series (see the last three columns of Table 2.1; also
repeated in Table 2.2) and assign the treatments levels according to the intended SRA method (see Table 2.2). For forced
sizes methods (Methods 3 and 4), we simply associate the randomly ordered list of units with the systematically ordered
list of treatment level replications. For instance, in Method 3Forced Equal Sizesthe randomization plan contains five
replications for each treatment level (T1, T2, T3, and T4), which are systematically ordered in the sixth column of Table
2.2. According to the random permutation of their ID numbers, Ted, Gene, Walt, Ron, and Chris are assigned to
Treatment Level T1; Mike, , and Pete to T2, and so forth. With the understanding that a different number of replications
is systematically listed for each treatment level (in our example, 7, 3, 3, and 7 replications, respectively), random
assignment in Method 4Forced Unequal Sizesis analogous to that in Method 3 (see last column of Table 2.2).
For a priori probabilities methods (Methods 1 and 2), the randomization strategy capitalizes on the procedure we have
used to draw random numbers. More specifically, in Method 1 Equal Probabilitiesthe probability that a three-digit
random number drawn from a uniform distribution within the interval [.000, .999] falls in one of the four equal-length
intervals ([.000, .250[, [.250, .500[, [.500, .750[, and [.750, .999]) is exactly p = 1/4 = .25. As a result, all the units (Ted,
Gene, Walt, Ron, Chris, and Mike) whose associated random numbers fall in the interval [.000, .250[are assigned to
Treatment Level T1; those in the interval [.250, .500[are assigned to Treatment Level T2, and so on (see fourth column of
Table 2.2). Likewise, in Method 2Unequal Probabilitiesthe probabilities for the intervals [.000, .400[, [.400, .600[,
[.600, .800[, and [.800, .999] are p1 = .40, p2 = .20, p3 = .20,
Table 2.2 Completely Randomized Design: Procedures for Allocation of 20 Units (See Table 2.1) to Four Treatment Levels (T1,
T2, T3, and T4) (Design 1; Methods 1, 2, 3, and 4)
and p4 = .20, respectively. The random assignment procedure is similar to the one for Method 1 (see fifth column of Table
2.2). Note that in a priori probabilities methods there is no need to sort the allocation frame by random numbers: Once the
random series is drawn, the assignment is automatically done. In the example, we have sorted the allocation frame to
permit a quick contrast between the four basic methods of SRA.
Although Method 1Equal Probabilities of Assignmentcorresponds to the formal definition of SRA, Method 3Forced
Equal Sizesin view of the aforementioned reasons (optimal trade-off between the costs and power of experiments and
data analysis facilities) is the standard choice in completely randomized designs. However, when the number of units per
experiment is not a multiple of the number of treatments, Method 3 cannot conduce to an equal number of subjects (ns)
per experimental condition. In this situation, the reader may adopt one of the following procedures for treatment exclusion
of the last replication: (a) elimination of treatments based on theoretical/practical relevance, (b) elimination of the last
treatment(s), and (c) random elimination of treatments. This may appear to be a nonissue, because when applying the
forced equal sizes strategy in a completely randomized design the difference between the number of experimental units
assigned to different treatment levels cannot be greater than 1. Nevertheless, when Method 3 is applied within strata in the
stratified random assignment (see Section 2.4) and there is no compelling reason to adopt procedure (a), the choice of
procedure (b) instead of procedure (c) can result in severe imbalance.
To carry out procedure (c) we draw a random permutation of treatments order (see Appendix 2), and we apply this
permutation to the last treatment replication. For instance, if only 19 units are available in the allocation frame for Design 1
(see Table 2.2), all treatment levels will be replicated four times but only three of them will be replicated five times.
Supposing we have picked from Table A1.1 the random permutation of four digits: 2, 4, 1, 3 (Rows: 3550; Column: 19);
then treatment T3's fifth replication will be excluded from the systematically ordered list of treatments included in Table
2.2: T1, T1, T1, T1, T1, T2, T2, T2, T2, T2, T3, T3, T3, T3, T4, T4, T4, T4, T4. A generalization of procedure (c)which we
have called last replication correction (LRC)to stratified random assignment is presented in Section 2.4.
2.2.2 Factorial Designs
The old standard rule of one variable is no longer the dominant strategy in experimental design. For theoretical or
economical reasons, researchers in the social, behavioral, educational, or health sciences run experiments with two or
more independent variables hoping to analyze the patterns of interaction (moderation effects) in the determination of the
outcome response.
According to the pattern of treatment level combinations, completely randomized factorial designs can be classified into
crossed or hierarchical (nested) designs. The former category applies to designs where all possible combinations of
treatment levels are included (see Design 7: Completely Randomized Factorial Design; Table 2.3). The latter one refers to
designs where at least the levels of one treatment are not completely crossed with the levels of a second treatment (see
Design 18: Completely Randomized Hierarchical Design; Table 2.3). With more than two treatments, it is possible to have
simultaneously crossed and nested factors (partial nesting). For instance, in Design 20Completely Randomized Partial
Hierarchical Design (see Table 2.3)Treatment C is crossed with Treatments A and B, but Treatment B is nested in
Treatment A. Finally, when some possible treatment level combinations are omitted from crossed designs, at the expense
of confounding treatment and interaction effects, the designs are called completely randomized fractional factorial designs
(e.g., Design 14 in Table 2.3 is a one-half fraction of the corresponding completely randomized factorial design with four
treatmentsA, B, C, and Dand two levels per treatment).
Nesting structures and fractional replication are complex topics in treatment design (design layout) and are out of the
scope of the present monograph (for trustworthy presentations and discussions, see Hinkelmann & Kempthorne, 2005,
2008; Kirk, 1995). However, from the error control design perspective, the randomization of completely randomized
factorial designs is a mere extension of the single-factor completely randomized design. More precisely, instead of
treatment levels, treatment level combinations included in the layout are assigned to experimental units. For instance, in
the four designs depicted in Table 2.3, the available experimental units would be randomly assigned to eight (Design 7),
four (Design 18), eight (Design 20), and eight (Design 14) experimental conditions or treatment level combinations. Any of
the four basic SRA methods could be used, but Method 3Forced Equal Sizesremains the standard choice.
2.3 Restrictedly Randomized Designs: Blocking
Restricted randomization is the combination of SRA with local control procedures. On the basis of their intrinsic
characteristics, experimental units are arranged in homogeneous groups (blocks or strata) prior to random assignment.
Random assignment within homogeneous groups rules out alternative explanations to treatment effects based on the
individual attributes of experimental units (classificatory factors). Blocking and stratifying reduce the overall heterogeneity
of units or, which is the same, minimize the residual variability or error term in ANOVA models, allowing a more precise
estimation of treatment effects.
Table 2.3 Layouts of Completely Randomized Factorial Designs (Designs 7, 14, 18, and 20)
Procedural constraints make the distinction between blocking and stratifying. In blocking, the number of units per block
must be equal to or a multiple of the number of experimental conditions, with the exception of incomplete blocks, in which
this number is smaller than the number of experimental conditions, but it is the same for all blocks. In stratifying, the
number of experimental units varies across strata, and treatments are unequally replicated within each stratum. In this
section, we describe randomization methods applying to homogeneous blocks of experimental units. The next section
deals with stratified random assignment.
2.3.1 Randomized Block Design
Using Federer and Balaam's (1973) terminology, randomized block designs illustrate the one-way elimination of the
heterogeneity of the experimental units. In randomized block designs, the researcher begins by arranging the available
experimental units in equal-size homogeneous groups (blocks), on the basis of individual scores (values) on the blocking
variable. The blocking variable can be any classificatory factor (see Section 1.2), as long as the information it conveys is
theoretically relevant for dependent variable prediction. Saying it in another way, the choice of the blocking variable must
be carefully made, with the understanding that its role in the minimization of residual variability is a function of the
expected correlation with the outcome measure. Vis--vis the measurement scale, the blocking variable can be a
nonmetric (nonordered categories and ordered categories; nominal and ordinal scales) or a metric (continuous variable;
interval and ratio scales) variable.
Suppose you are studying the impact of extrinsic motivation (a treatment with three levels: high, medium, and low
incentives) on academic achievement (outcome measure) and you have good reasons to suspect that students cognitive
abilities (classificatory factor) could diminish the expected relationship Higher incentives Better achievement. Instead
of assigning at random the available experimental units to the three treatment levels (SRA; completely randomized
design), you could assess students IQ (intelligence quotient) prior to randomization, place them in homogeneous blocks
on the basis of their IQ level, and finally proceed to random assignment of treatment levels within each block. If the
number of units per block is the same as the number of treatment levels, so that in each block only one experimental unit
is assigned to each treatment level, your randomization strategy is called BRA (Method 7) and the resulting design is a
Randomized Block Design (Design 2). With this strategy, you have controlled the variability of cognitive abilities within
experimental conditions (local control) and therefore created the optimal circumstances for a more precise estimation of
the effect of motivation on academic achievement.
Table 2.4 Randomized Block Design: Allocation Frame and Random Numbers
The adopted blocking and randomization procedures are illustrated in Tables 2.4 and 2.5, assuming that 18 experimental
units (students) have been enrolled and are available at the onset of the experiment. For example, Gabby, Deb, and
Nancy, according to their IQ scores, have been placed in Block 6 (see Table 2.4); in line with the associated random
numbers, they have been allocated to Treatment Levels T1, T3, and T2, respectively (see Table 2.5). Note that in the
allocation frame for BRA, the IDs are sequentially assigned within blocks, which themselves are also sequentially
numerated (1-n). Note also that Method 7Blocked Random Assignmentis simply the application of SRA (Method 3:
Forced Equal Sizes) within the blocks included in the allocation frame.
Table 2.5 Randomized Block Design: Allocation of 18 Units (Grouped in Six Blocks of 3 UnitsSee Table 2.4) to Three
Treatment Levels (T1, T2, and T3) (Design 2; Method 7)
2.3.2 Generalized Randomized Block Design
When the number of experimental units per block is greater than the number of treatment levels (t) but is fixed at any
multiple of this number (kt, where k is any integer 2), using Method 7Blocked Random Assignmentproduces a
Generalized Randomized Block Design (Design 3). The randomization procedure is identical to the procedure for the
Randomized Block Design (Design 2), with the exception that in each block two or more experimental units are assigned
to each treatment level. Tables 2.6 and 2.7 show the allocation frame, the randomization procedure, and the final
allocation for a generalized randomized block design with three treatment levels (T1, T2, and T3) and a nominal blocking
variable: gender of the experimental units.
Table 2.6 Generalized Randomized Block Design: Allocation Frame and Random Numbers (RN) (Design 3; Method 7)
Table 2.7 Generalized Randomized Block Design: Allocation of 18 Units (Grouped in Two Blocks of 9 UnitsSee Table 2.6) to
Three Treatment Levels (T1, T2, and T3) (Design 3; Method 7)
The option between Design 2 (Randomized Block Design, also called Complete Randomized Block Design; see
Hinkelmann & Kempthorne, 2008) and Design 3 (Generalized Randomized Block Design) is simultaneously a procedural
and a substantive (theoretical) issue, as it depends on having enough homogeneous experimental units to form larger size
blocks and the expected relationship between the experimental factor and the blocking variable. More precisely, in
randomized block designs, the interaction between the blocking variable and the experimental factor is completely
confounded with the residual variance (crossing both variables in the design layout gives rise to one experimental unit per
cell), and therefore only the main effects for both variables can be estimated. In contrast, generalized randomized block
designs, in which each cell has at least two experimental units, allow the estimation of the interaction effect besides the
main effects of the experimental factor and the blocking variable (see Subsection 4.2.2). Briefly stated, when larger size
blocks are possible and some kind of interaction between the independent and the blocking variable can be anticipated,
the best option is the generalized randomized block design.
As Kirk (1995) remarks, this design is often misspelled in empirical research reports and confounded with the two-factor
Completely Randomized Factorial Design (Design 7; see Table 2.3). The confusion comes from the fact that both designs
share the same underlying statistical model; however the Completely Randomized Factorial Design has two true
experimental factors (active independent variables), whereas the Generalized Randomized Block Design is limited to one
active independent variable (the experimental factor) and one passive independent variable (the blocking variable; see
Table 1.4 for terminological details).
2.3.3 Incomplete Randomized Block Designs
When the number of treatment levels is large and the distribution of observed scores (values) on the blocking variable
doesn't recommend the arrangement of available experimental units in blocks of size equal to or a multiple of the number
of experimental conditions, a strategy to circumvent the problem consists in the formation of a homogeneous block whose
size is smaller than the number of treatment levels. In this situation, the design layout must include a fixed number of
treatment level arrangements, selected in such a way that all pairwise associations between treatment levels within
blocks are equally replicated.
Design 4A in Table 2.8 corresponds to this requirement in the sense that every pairwise association of the six treatment
levels (1, 2, 3, 4, 5, and 6) occurs exactly twice across the 10 blocks (e.g., the association of Treatment Levels 1 and 2 is
presented in the first two blocks, whereas the association of Treatment Levels 3 and 5 is displayed in Blocks 7 and 9).
Design 4A is called a Balanced Incomplete Randomized Block Design, with t (treatment levels) = 6, b (blocks) = 10, k
(experimental units per block) = 3, r (treatment level replications per experiment) = 5, and (number of times each pair of
treatment levels is replicated within blocks) = 2.
There are no balanced designs for certain combinations of the above parameters. When the relationship = [r(k 1)]/(t
1) gives a noninteger solution for , one can be sure that no balanced design exists. When the solution is an integer, one
can look for the corresponding design in authoritative sources (e.g., Cochran & Cox, 1957; Federer, 1955; Hinkelmann &
Kempthorne, 2005, 2008; Kempthorne, 1952), in spite of knowing that even in this case not all balanced designs can be
built. A less satisfactory solution is to rely on what has been called a Partially Balanced Incomplete Randomized Block
Design, a design in which some pairwise associations between treatment levels occur 1 times while other associations
happen 2 times (e.g., in Design 4C/Table 2.8, the associations 14 and 25 occur twice, whereas the associations 12
and 23 occur only once). The layouts for other alternative incomplete randomized block designs (e.g., balanced
treatment and extended designs) are shown in the lower left part of Table 2.8.
Table 2.8 Treatment Levels Arrangements in Complete (Designs 2 and 3) and Incomplete (Designs 4A, 4B, 4C, 4D, and 4E)
Randomized Block Designs
Designing of incomplete block experiments is a complex matter that we cannot cover in this monograph, and therefore the
main utilization of these designs is found in agricultural studies, engineering, and technological research and not in the
social, behavioral, educational, or health sciences. However, the randomization and local control of these designs supply
the prototype for other designs covered in this chapter. Regardless of the layout of the incomplete block design,
randomization is always identical and consists of a two-step procedure: (a) random assignment of treatment level
arrangements to blocks and (b) random assignment of treatment levels to experimental units within each block.
Method 8Two-Step Blocked Random Assignmentis illustrated in Table 2.9 with Design 4C (six treatment levels and
12 experimental units, grouped in three blocks of 4 units). In the first step, incomplete blocks of experimental units are
randomly allocated to the treatment level arrangements included in the design layout. The procedure is similar to SRA
(Method 3: Forced Equal Sizes), with the exception that the units of assignment are the blocks themselves, instead of
individual experimental units, and the single treatment levels are replaced with treatment level arrangements. In the
second step, we make use of Method 7Blocked Random Assignmentalready employed in the randomization of
(Complete) Randomized Block (Design 2) and Generalized Randomized Block (Design 3) designs. The final allocation of
the 12 experimental units is shown at the bottom of Table 2.9.
Note that the same randomization procedure could be used with any multiple of the number of blocks. If we formed six
blocks of four units, instead of the three blocks used in the illustrative example, in the first step, two homogeneous blocks
of experimental units would be randomly allocated to each of the three arrangements of treatment levels included in the
design.
2.3.4 Latin Square Design
Methods 7 (Blocked Random Assignment) and 8 (Two-Step Blocked Random Assignment) are applied when the strategy
of elimination of the heterogeneity of experimental units is confined to a single classificatory factor (blocking variable). To
pursue with the example given for the randomized block design (see Subsection 2.3.1), suppose that the researcher is
manipulating material incentives and he or she has sound reasons to expect that, in addition to the cognitive abilities, the
upper-class students performance, as compared with the lower-class students, is less responsive to the nature of the
incentives. One possible strategy is to classify the students in
Table 2.9 Randomization of a Partially Balanced Incomplete Block Design: Allocation of 12 Units (Grouped in Three Blocks of 4
Units) to Three Arrangements of Six Treatment Levels (T1 T4 T2 T5; T2 T5 T3 T6; and T3 T6 T1 T4) (Design 4C; See Table 2.8;
Method 8)
homogeneous blocks defined simultaneously by two blocking variables: cognitive abilities and socioeconomic status
(SES). With the procedural constraints that the number of students should be the same in all blocks and that the number
of levels of both blocking variables should equal the number of treatment levels included in the design layout, this strategy
of local control gives rise to a Latin square design (Design 5A).
The design receives its name from an ancient Roman puzzle, which consists in finding out how many possible
arrangements of the first n letters (A, B, C, etc.) in n rows and n columns could be done with the restriction that each
letter appears once, and only once, in each row and in each column of the square. Square 1 in Figure 2.2 is an example
of a Latin square of order 5 (an array in five rows and five columns of the first five Latin letters), having the special property
that the letters in the first row and in the first column are arranged in alphabetical order (standard or normalized Latin
square).
Figure 2.2 Random selection of a Latin square
In treatment and error control design, the letters of the square represent the treatment levels, whereas the rows and the
columns are filled with the blocking variables (for variations and extensions of this basic design, see the next subsection).
Once the blocks (n
2
blocks; numbers of rows times number of columns) have been formed, the randomization consists in
choosing at random one of the possible arrangements of letters (symbols) to assign the treatment levels to the
experimental units (Method 9: Two-Way Blocked Random Assignment). If the experimenter had to find out all the possible
arrangements (their number increases dramatically with the order of squares; see Appendix 3 for enumeration of Latin
squares) for a given square, the research enterprise would be converted into a never-ending task. An easy way to obtain a
random square from a relatively large collection of all available squares (see Appendix 3 for details of and simplified
procedures for squares of orders 24) is to begin with the standard Latin square and to permute, randomly and
independently, the order of the rows, the order of the columns, and the labels (letters/symbols) of the treatment levels.
This procedure is illustrated in Figure 2.2 for a square of order 5 (the three random permutations were picked from Table
A1.1, according to the guidelines given in Appendix 2).
Returning to our example, suppose that the experimenter has managed to form the nine blocks depicted in the allocation
frame of Table 2.10 (the levels of the row variable correspond to the levels of cognitive abilities, defined by two cutoff points
in IQ scores, and the levels of the column variable represent three SES-ordered categories). In the next step, adopting the
aforementioned procedure (actually, with an order 3 Latin square it is sufficient to permute randomly the three rows and
the last two columns of the standard Latin square in order to obtain one of the 12 possible different squares; see Appendix
3), the experimenter gets the randomized Latin square shown in the upper right side of Table 2.10 (the replacement of
letters A, B, and C with T1, T2, and T3 is only a matter of consistence with the notation employed in the preceding
designs). As a result, the final allocation of the 18 experimental units is the one shown in the lower portion of Table 2.10
(e.g., Daysie and Nattie, whose attributes have placed them in Block 4, would be tested in Treatment Level T3).
Despite its advantages in simultaneously handling two classificatory factors, the associated procedural constraints (the
same number of levels needed for the experimental factor and both blocking variables and the requirement of equal
numbers of experimental units per block) have made the use of Latin square designs relatively rare in the behavioral and
related sciences (for a comprehensive approach, see, among others, Cochran & Cox, 1957; Federer, 1955; Fisher &
Yates, 1963; Hinkelmann & Kempthorne, 2005; Kirk, 1995). The main role of Latin squares in experimental design still lies
in their utilization in the randomization of within-subjects designs (see Chapter 3). The extension of Latin square designs
to the control of more than two classificatory factors (e.g., Graeco-Latin squares) is discussed in the next subsection in
conjunction with blocked random assignment in factorial designs.
Table 2.10 Randomization of a Latin Square Design: Allocation of 18 Units (Grouped in Nine Blocks of 2 Units) to Three of
Treatment Levels (T1, T2, and T3) (Design 5A; Method 9)
2.3.5 Factorial Designs
With the randomization of the Latin square design, we have described the basic methods of blocking and randomization of
the one-factor restrictedly randomized designs: (a) Blocked Random Assignment (complete and generalized randomized
block designs: Designs 2 and 3Method 7); (b) Two-Step Blocked Random Assignment (balanced, partially balanced,
and extended incomplete randomized block designs: Designs 4A to 4EMethod 8); and (c) Two-Way Blocked Random
Assignment (Latin square design: Design 5AMethod 9). In Subsection 2.2.2, we have generalized the basic SRA
methods, used in the one-factor completely randomized design (Design 1Methods 1 to 4), to their factorial applications
(Completely Randomized Factorial Design, Completely Randomized Fractional Factorial Design, Completely Randomized
Hierarchical Design, and Completely Randomized Partial Hierarchical DesignDesigns 7, 14, 18, and 20). In the current
subsection, Methods 7, 8, 9, and 10 (Blocked Random Assignment via Graeco-Latin Squares, an extension of Method 9)
are applied to the remaining factorial designs classified in Figure 1.3, beginning with the factorial randomized block and
related designs and concluding with variations and extensions of the Latin square design.
2.3.5.1 Factorial Randomized Block and Related Designs
Broadly stated, factorial randomized block designs are randomized designs with two or more treatments and one blocking
variable. Similar to factorial completely randomized designs (see Subsection 2.2.2), factorial randomized block designs
are classified into crossed versus hierarchical designs, and within the former category, the effects of treatments and their
interactions may be either unambiguously estimated (designs without confounding) or confounded, as a result of some
specific procedural constraint underlying the randomization plan. In the latter category, with three or more treatments, the
hierarchical structure can accommodate complete crossing between two or more treatments.
2.3.5.1.1 Crossed Designs without Confounding
The two basic factorial designs in which all treatment effects and their interactions can be unambiguously estimated are
the Randomized Block Factorial Design (Design 8) and the Generalized Randomized Block Factorial Design (Design 9).
From the perspective of treatment design, both layouts are identical to the corresponding Completely Randomized
Factorial Design (Design 7; see Table 2.3). From the perspective of error control design, the randomization is done via
Method 7 (Blocked Random Assignment) and is similar to the randomization of their one-factor counterparts (Designs 2
and 3; see Subsections 2.3.1 and 2.3.2), with the understanding that the experimental units within blocks are assigned to
treatment level combinations, instead of being allocated to a single treatment level.
2.3.5.1.2 Crossed Designs with Confounding
The Completely Randomized Fractional Factorial Design (Design 14; see Table 2.3), as compared with the Completely
Randomized Factorial Design (Design 7; see Table 2.3), achieves a reduction in the total number of experimental units at
the cost of confounding the treatment main effects with the interaction effect. The same is true for Design 15
Randomized Block Fractional Factorial Designwhose randomization plan relies on Method 7 (Blocked Random
Assignmentthe same method used for Designs 2 and 3, assuming that we are dealing with treatment level combinations
and not with treatment levels themselves).
In Design 15, the confounding is between treatments and their interaction. Two additional types of confounding can result
from the procedural constraints imposed on the randomization plan. The first one gives rise to what has been called a
Randomized Block Completely Confounded Factorial Design (Design 11). In this design (see Table 2.13), all the
combinations between treatment levels are present in the global layout, but they are broken into groups of incomplete
combinations that are themselves randomly assigned to homogeneous blocks of experimental units. This procedure
achieves the reduction of block size with the correlative price of confounding the interaction between treatments with
groups of blocks. The second additional type of confounding reduces the size of the blocks by confining the blocked
random assignment to one of the treatments included in the design. More precisely, entire blocks are randomly assigned
to the levels of the first treatment, and the experimental units within blocks are subsequently assigned at random to the
levels of the second treatment. From a theoretical or substantive point of view, this means that the estimation of the
second treatment effect is more important than the estimation of the first treatment effect (Kirk, 1995). This design,
combining features of the Completely Randomized Design and the Block Randomized Design, is commonly labeled Split-
Plot Factorial Design (Design 10; when the second treatment is a within-subjects experimental factor, the design is known
as a Mixed Design, a somewhat misleading designationsee Subsection 3.7.2).
The randomization plan for Designs 10 (Split-Plot Factorial Design) and 11 (Randomized Block Completely Confounded
Factorial Design) is an extension of the strategy described for the Incomplete Randomized Block Designs (see
Subsection 2.3.3). In Design 10, the experimenter begins by randomly assigning the available blocks to the levels of the
first treatment; in the second step, units within blocks are themselves assigned at random to the levels of the second
treatment. For instance, in Table 2.11, Block 2 is assigned to Treatment Level A1 (Step 1); according to the intrablock
randomization, Treatment Levels B2, B1, and B3 are assigned to Della, Freda, and Sally (Step 2). The randomization plan
for Design11Randomized Block Completely Confounded Factorial Designis similar, with the understanding that in the
first step blocks are randomly assigned to groups of treatment level combinations instead of treatment levels of the
second independent variable (see Table 2.12 for the randomization procedure and Table 2.13 for the final allocation of
experimental units).
Table 2.11 Randomization of a Split-Plot Design: Allocation of 12 Units (Grouped in Four Blocks of 3 Units) to Six Combinations
of Treatment Levels (A1B1, A1B2, A1B3, A2B1, A2B2, and A2B3) (Design 10; Method 8)
Note that there is an additional difference in the underlying layout of Designs 10 and 11. In the first design, no constraint is
imposed on the number of levels of the independent variables; in the second design, the number
Table 2.12 Randomized Block Completely Confounded Factorial Design: Randomization Procedures for Allocation of 18 Units
(Grouped in Six Blocks of 3 Units) to Three Groups of Treatment Level Combinations (Design 11; Method 8)
Table 2.13 Randomized Block Completely Confounded Factorial Design: Final Allocation of 18 Units (Grouped in Six Blocks of 3
Units) to Three Groups of Treatment Level Combinations (See Table 2.12) (Design 11; Method 8)
of levels must be equal for both treatments. This requirement also holds for Design 12Randomized Block Partially
Confounded Factorial Design (see Table 2.14)in which there are at least three treatments with two levels and the
confounding scheme consists in confounding different interactions with different groups of blocks (see Kirk, 1995, chaps.
1214, for a complete discussion of crossed designs with confounding). The randomization of Design 12 is analogous to
the randomization of Design 11.
2.3.5.1.3 Hierarchical Designs
The layout of randomized block factorial designs with complete or partial nesting between experimental factors
Table 2.14 Layout of a Randomized Block Partially Confounded Factorial Design with Three Two-Level Treatments (A2B2C2) and
Eight Blocks of 4 Units (Grouped in Four Groups of Treatment Level Combinations) (Design 12)
(Designs 19 and 21) is similar to the layout of their equivalent completely randomized designs (Designs 18 and 20; see
Table 2.3), but the randomization plan, instead of Method 3, relies on Method 7Blocked Random Assignment (see
Subsection 2.3.1). Finally, there is also a hierarchical version of the split-plot design with more than two treatments
(Design 22Split-Plot Partial Hierarchical Design; see Table 2.15 for an illustrative design layout). The randomization plan
for this design, based on Method 8 (Two-Step Blocked Random Assignment), is the same as what we have described for
Design 10 (Split-Plot Factorial Design; see Table 2.11), with the exception that in the second step the intrablock
randomization is done in reference to the treatment level combinations (in the example, combinations of Treatments B and
C; see Table 2.15) and not in connection with the levels of a single treatment.
2.3.5.2 Latin Square and Related Designs
In Subsection 2.3.4, the Latin square design (Design 5A; see Table 2.10 and also the layout in Table 2.16) was introduced
as the basic two-way elimination strategy of the heterogeneity of experimental units, in the sense that it allows the
systematic variation of one experimental factor (whose levels are represented by the symbols or by Latin letters in the
cells of the square) and the simultaneous control of two classificatory factors (whose levels are appointed to the rows and
columns of the square and must be equal to the number of treatment levels). With the limitation that all factors must have
the same number of levels, this basic design can be used in situations where one or two additional experimental factors
(active independent variables)or even some kind of a pseudofactor (e.g., time sequence) replace the blocking variables
in the rows and/or columns of the square.
Table 2.15 Layout of a Split-Plot Partial Hierarchical Design with Three Treatments (A, B, and C) and Four Blocks (Design 22)
When both rows and columns represent two experimental factors, the design becomes a special type of factorial fractional
design (Design 16: Latin Square Fractional Factorial Design; see Table 2.16), more precisely a 1/t fractional replication of
the corresponding completely randomized factorial design (e.g., Design 16 includes nine combinations of treatment levels,
while the equivalent completely randomized design A3B3C3 contains 27 experimental conditions). When the columns are
appointed to a second experimental factor and the rows continue to represent a blocking variable, the design receives the
label Latin Square Confounded Factorial Design (Design 13) and is structurally identical to a randomized block completely
confounded factorial design (Design 11; see Table 2.12), where the number of blocks (levels of the blocking variable) is the
same as the number of groups of treatment combinations included in the design layout. The randomization of Designs 13
and 16 is done with Method 9Two-Way Blocked Random Assignment and is analogous to the randomization of Design
5A (see Table 2.10; in the examples in Table 2.16, a random Latin square should be selected for the assignment of the
levels of Treatment B, in Design 13, and Treatment C, in Design 16).
Table 2.16 Layout (Before Randomization) of Latin and Graeco-Latin Square Designs (Designs 5A, 5B, 13, 16, and 17)
Designs 13 and 16 modify the logic of local control (two-way elimination of the heterogeneity of experimental units)
enclosed in the classic Latin square design (Design 5A), by reducing the number of blocking variables to handle additional
experimental factors. On the contrary, Design 5BGraeco-Latin Square Design (see Table 2.16) extends this logic by
accommodating a third blocking variable. This is done by juxtaposing two orthogonal Latin squares (see Table A3.2,
Appendix 3) in the layout cells. The symbols of the first Latin square (Vcell1, Vcell2, and Vcell3, in the example of Design
5B; see Table 2.16) are appointed to the third blocking variable, whereas the symbols of the juxtaposed orthogonal square
(A1, A2, and A3, in the same example) continue to represent the experimental treatment (compared with the 3 3
standard Graeco-Latin square depicted in Table A3.2, Vcell1, Vcell2, and Vcell3 correspond to the first Latin square
defined by the letters A, B, and Cand A1, A2, and A3 correspond to the second Latin squaredefined by the numbers
1, 2, and 3). Provided that the experimenter is able to find equal-sized homogeneous blocks of experimental units
categorized simultaneously by the three blocking variables (Vrow, Vcol, and Vcell) and the corresponding Graeco-Latin
square, the randomization of Design 5B consists simply in taking a random permutation for the symbols of the second
Latin square that represent the levels (A1, A2, and A3) of the experimental treatment (in this example, Treatment A). This
procedure is identical to the third random permutation illustrated in Figure 2.2. Note that this design should have been
described in Subsection 2.3.4, since it is a one-factor design and a natural extension of the Latin square design. We
have postponed its presentation so that we will be able to compare its structural features with the remaining factorial Latin
and Graeco-Latin square designs described in this section.
Finally, Design 17Graeco-Latin Square Fractional Factorial Designis a variation of Design 5B, in which the blocking
variables appointed to the columns and cells have been replaced by two experimental treatments. The randomization of
Design 17 is analogous to that of Design 5A, with two exceptions: (1) the initial standard square is one of the Graeco-
Latin squares shown in Table A3.2 and (2) a fourth independent random permutation must be selected for the symbols
(subscripts) of the second orthogonal Latin square. In our example (see Design 17 in Table 2.16), the first and second
random permutations illustrated in Figure 2.2 would apply to the rows and the columns of the Graeco-Latin square, the
third random permutation would define the symbols for levels of Treatment B, and the fourth random permutation would
define the symbols for levels of Treatment C. We have named this randomization method Blocked Random Assignment
via Graeco-Latin Squares (Method 10). Note that there are no order 2 or 6 Graeco-Latin squares (see Table A3.2).
2.4 Restrictedly Randomized Designs: Stratifying
Blocking is the optimal strategy of local control and therefore must be used whenever possible. Nevertheless, as the
reader has already realized, blocking methods have intrinsic procedural constraints, namely, equal-size homogeneous
blocks and an equal number of levels for treatments and blocking variables in the two- and three-way blocking techniques,
which inhibit their utilization in a considerable number of real-word randomized experiments. An alternative approach to
local control of classificatory factors is the stratification of available experimental units prior to random assignment. More
precisely, at the onset of the experiment the units are grouped in homogeneous groups (called strata) on the basis of their
scores (values) in one or more stratifying variables.
The nature of the stratifying variables is the same as that of the blocking variables mentioned in the previous section (see
also the typology of classificatory factors in Table 1.1), and their choice must be made according to the same criteria:
known or expected correlations with the outcome measures. Briefly stated, stratified random assignment is a method of
local control similar to blocked random assignment (in the sense that experimental units are previously assembled in
homogeneous groups) but without the requirements of equal-size groups and an equal number of levels across
classificatory factors, as is the case in two- and three-way blocking techniques. For example, if the researcher has sound
reasons to suspect that gender (two categories), age (three age intervals), and SES (four ordered categories) can make
the expected relationship between his or her experimental factor and the outcome measure disappear, he or she begins
by grouping the available experimental units in 24 (2 3 4) strata and subsequently randomly assigns the units to
treatment levels within each stratum. Any of the four basic SRA methods can be used (see Section 2.1), but the ideal of
balanced designs recommends the forced equal sizes strategy (Method 3). Method 11 (Stratified Random Assignment:
Captive Procedure) is simply the application of Method 3 to all of the unequal-size strata included in the randomization
plan.
In the example of Tables 2.17 and 2.18, the researcher has grouped 60 experimental units (available at the onset of the
experiment) in 12 strata, according to the observed scores (values) in three stratifying variables (A, B, and C). For each
stratum, a series of random numbers is picked from Table A1.1 (the actual random series are shown in Table 2.17). In the
next step (see Table 2.18), Method 3 (SRSForced Equal Sizes), coupled with the LRC, explained in Subsection 2.2.1,
is applied within each stratum. More precisely, when the number of units in a stratum is not equal to or is
Table 2.17 Stratified Random Assignment (Nonsequential Procedure): Allocation Frame and Random Numbers (RN) for
Allocation of 60 Units (Grouped in 12 Strata Defined by Three Stratifying Variables: A, B, and C) to Three Treatment Levels (T1,
T2, and T3) (Method 11)
Table 2.18 Stratified Random Assignment (Nonsequential Procedure): Final Allocation of 60 Units (Grouped in 12 Strata Defined
by Three Stratifying Variables: A, B, and CSee Table 2.17) to Three Treatment Levels (T1, T2, and T3) (Method 11)
not a multiple of the number of treatment levels (in the example, three treatment levels: T1, T2, and T3), the last t
replications in the systematically ordered list of treatment levels shown in Table 2.18 are randomly selected. For instance,
in Stratum 8, the systematically ordered list of treatment levels is perfectly balanced until the sixth replication (T1, T2, T3,
T1, T2, T3). For the last two replications, Treatments T1 and T3 are selected, whereas Treatment T2 is excluded, on the
basis of the corresponding random permutation (1, 3, 2) picked from Table A1.1 according to the guidelines given in
Appendix 2. Once the systematically ordered lists of treatment replications have been established with the LRC
procedure, random assignment consists simply in the association of these lists with the randomly ordered lists of units
within strata (in Stratum 8, Adela, Magda, and Brenda have been assigned to Treatment Level T1; Nelly and Ina to T2;
and, last, Stella, Dina, and Claire to T3). Finally, note that the application of Method 11 to factorial designs is similar to
the one-factor example given in this section, with the difference that treatment level combinations, instead of the treatment
levels themselves, are assigned to experimental units within strata.
With the stratified random assignment, we have concluded the description of the two broad strategies (blocking and
stratifying) for local control of classificatory factors in restrictedly randomized designs. We have not yet mentioned a
restrictedly randomization strategy that bears some analogy with cluster random sampling and is often described in the
health sciences literature as cluster-randomized trials. The strategy mainly consists in having experimental units nested
in higher level entities or clusters (e.g., hospitals) and randomly assigning all the units grouped in one cluster to one of the
experimental conditions included in the design layout. From a practical perspective, this strategy is an application of SRA
to all available clusters.
2.5 Sequential Assignment and Adaptive Randomization Methods
Until now (Section 2.2-2.4), as stated in the opening section of this chapter, all the randomization and local control
procedures have been described with the understanding that the experimental units are available at the onset of the study.
In technical jargon, this is called captive or nonsequential assignment (Underwood, 1966) and can be seen as the ideal
mode of conducting randomized experiments. Unfortunately, a substantial number of experimental studies must be carried
out with virtual allocation frames, in the sense that the experimental units are individually enrolled on successive
occasions, which can be distributed for considerable time periods (months, or even years). This is the typical situation in
some disciplines, namely, in the health sciences, where the experimental units are generally patients with specific
diseases and prognostic characteristics and the treatment delivery must be scheduled according to the flow of
participants from the beginning to the end of the study.
In any case, the sequential random assignment is analogous to the nonsequential random assignment, with some
procedural adaptations to compensate for the lost degrees of freedom. As already stated, in nonsequential assignment,
it is an arbitrary and personal choice either to associate a systematically ordered list of available units to a randomly
ordered list of treatment level replications or to match a randomly ordered list of units to a systematically ordered list of
treatments. For the reasons mentioned in the introductory section (i.e., to give the reader a more comprehensive picture of
the randomization procedure in conjunction with the design layout), we have chosen the second association procedure.
The first degree of freedom we lost in the sequential assignment is the arbitrariness of this choice. Saying it in other
words, temporal allocation frames must necessarily include a randomly ordered list of treatment level replications that is
subsequently associated with a systematically ordered list of enrolled experimental units (sequential ID numbers from 1 to
n).
Before proceeding with the description of sequential random assignment methods, we must address a preliminary issue
concerning what is frequently called batch and trickle randomization, that is, situations where the experimental units
are enrolled in small groups in successive time periods. Batch and trickle randomization can be more accurately
described as special problems in the organization of temporal allocation frames (e.g., attribution of sequential ID numbers)
than as randomization methods in their own right. More precisely, a temporal allocation frame is a randomly ordered list of
treatment replications coupled with a systematically ordered list of ID numbers (1-n), ascribed to the experimental units as
they are enrolled in the experiment. If the experimental units are sequentially and individually enrolled, the attribution of ID
numbers poses no problem. However, if the units come in groups, the researcher must decide which ID numbers are
attributed to the units within enrollment groups (batches). This can be done by establishing temporal precedence based
on systematic criteria (e.g., alphabetic order, date of birth) or by random determination of precedence (e.g., the
experimental units draw lots), as shown in Table 2.19. In any case, the temporal precedence in the attribution of ID
numbers must be independent of the randomly ordered list of treatment replications previously recorded in the temporal
allocation frame. In the remaining subsections, the procedural adaptations of simple, blocked, and stratified randomization
to the sequential assignment are conveniently described, as well as minimizationan alternative method to sequential
stratified random assignment.
Table 2.19 Attribution of Sequential ID Numbers When the Order of Enrollment in the Experiment Is Unclear
2.5.1 Sequential Simple Random Assignment
With a priori probabilities methods (Methods 1 and 2), the application of sequential SRA in completely randomized
designs is straightforward. The only task the researcher must perform is to have an adequately large series of random
numbers and assign the treatment levels or treatment level combinations to the arriving experimental units (ID numbers 1-
n, ascribed by order of enrollment) according to the probabilistic scheme previously defined (see Table 2.20 and
Subsection 2.2.1).
With the forced-size strategies, the application of Methods 3 and 4 must begin necessarily with the determination of the
number of desired replications of treatment levels by means of educated guessing and formal power analysis (note that
this strategy would be the standard approach routinely adopted in all randomization methods described in this monograph;
see Subsection 4.1.2). Once this number has been identified, a randomly ordered list of treatment replications, containing
the corrected number of replications for each treatment level, is generated and associated, step by step, with the
sequential ID numbers. This is an easy task when there is no temporal or budget-fixed limit for the conclusion of the
experiment and the treatments can be delivered to all the virtual experimental units listed in the temporal allocation
frame.
Table 2.20 Sequential Simple Random Assignment: Assignment of 18 Experimental Units to Three Treatment Levels (T1, T2, and
T3) With Equal a Priori Probabilities (Method 1Sequential Version) and Forced Equal Sizes (Method 5Time-Blocking/Two
Randomization Fractions) Procedures (Design 1)
In situations where the study can be abruptly interrupted by previously fixed temporal limits or other external constraints,
the best solution to approach the forced-size strategy is to break the total number of experimental units down into a
reasonable number of equal-size fractions (usually two to four fractions) and to apply SRA within each fraction. This is
similar to the random permuted blocks technique (see Appendix 1 and Table A1.3), but we avoid using the term block in
this context, because we have been using it so far to identify homogeneous groups of experimental units formed on the
basis of substantive restrictions (scores or values in classificatory factors). A more appropriate designation could be time
blocking, with the understanding that each consecutive randomization fraction corresponds to a level of the blocking
variable. In the data analysis stage, the effect of this blocking variable and its interaction with the experimental factor(s)
may be analyzed, hoping that (needless to say!) no interaction or main effect will be detected, or else the design will
become a special type of randomized block design, and sound explanations for the impact of time of enrollment in the
experiment would be required. To account for this variation of the forced-size strategy in sequential assignment, we give
the designations Method 5 (Forced Equal Sizes Sequential Assignment) and Method 6 (Forced Unequal Sizes Sequential
Assignment) to the basic SRA Methods 3 and 4. The illustration of Method 5 is given on the right side of Table 2.20;
Method 6 is similar, with the exception that the equal number of treatment level replications within each randomization
fraction is replaced with the desired pattern of unequal replications.
2.5.2 Sequential Blocked Random Assignment
The distinction between situations where there is no temporal limit for the conclusion of the study and situations where an
ending point is imposed by extrinsic time schedules or any other constraints (e.g., institutional resources, research funds)
is even more crucial in experimental studies relying on blocked random assignment. If there is no temporal limit, it is
possible to carry out all the experimental designs and randomization methods described in Section 2.3, with two
additional adaptations: (1) the allocation frames within blocks start with the randomly ordered list of treatment replications
and (b) the experimenter waits until all associated (virtual) ID numbers are ascribed to the successively enrolled
experimental units, with the understanding that once a block has been filled up, no more units owning the block-defining
characteristic(s) are enrolled in the experiment. When the experimenter's degrees of freedom don't include the decision
about the study end, sequential blocked random assignment is simply not feasible, and the alternatives are sequential
simple or stratified random assignment.
Note that there is an additional restraint on blocked random assignment in the no temporal limit situation: When the
cutoff points for continuous blocking variables are arbitrary and are chosen on the basis of the observed distribution of
scores (as in the example in Subsection 2.3.1 with IQ), the blocks become undefinable at the beginning of the
experiment. A way to circumvent this obstacle is to establish in advance fixed cutoff points (within reasonable ranges) for
the blocking variable. By definition, with nominal (nonordered categories) and ordinal (ordered categories) blocking
variables, no such obstacles exist. The randomization methods applying to sequential blocked random assignment are
listed in Table 2.21, as well as the methods appropriate to sequential simple (previous subsection) and stratified (next
subsection) random assignment.
2.5.3 Sequential Stratified Random Assignment
In the same sense that Nonsequential Stratified Random Assignment (Method 11) is the application of Method 3 (SRA
with forced equal sizes, including last permutation correction) within each stratum integrated in the design layout,
Sequential Stratified Random Assignment (Method 12) is the use of Method 5 (sequential SRA, including time-blocking
strategy) in equivalent circumstances. To carry out Method 12 is an easy, but time-consuming, task. Once the strata have
been defined, a randomly ordered list of treatment level replications is attributed to each randomization fraction within
strata. The intrastratum randomization is identical to that with Method 5. However, in this case, the size of the
randomization fractions must take account of the (intended) total number of experimental units and the expected numbers
for each stratum. Usually, the size of the randomization fractions is smaller than in sequential SRA, and in some
circumstances it is advisable to let the randomization fraction size vary across strata. Table 2.22 shows an example with
four strata, three treatment levels, and randomization fractions (time blocking) with six experimental units.
Table 2.21 Randomization Procedures in Sequential Assignment
Table 2.22 Stratified Random Assignment (Sequential Procedure): Assignment of N Experimental Units (Grouped in Four Strata
Defined by Two Stratifying Variables: A and B) to Three Treatment Levels (T1, T2, and T3) With Time Blocking Within Strata
(Method 12)
2.5.4 Minimization and Related Techniques
Minimization, or more rigorously some techniques of deterministic assignment and adaptive randomization procedures,
was initially proposed by Taves (1974) and Pocock and Simon (1975; for a recent review, see Scott, McPherson, Ramsay,
& Campbell, 2002) as an alternative to sequential stratified random assignment, and its status among randomization
methods is still a controversial issue. Even though we agree with authors who see no recommendable use for
minimization (e.g., Berger, 2005; Fleiss, 1986), we will describe and illustrate the main features of the procedure in this
final subsection, postponing the substantive discussion to Chapter 4.
The difficulties in achieving balanced designs with sequential stratified random assignment and related problems with
overstratification set the stage for minimization and adaptive randomization strategies. Basically, the procedure consists
in skipping the stratification phase and combining sequential SRAMethod 1 (equal a priori probabilities of assignment)
either with deterministic (nonrandom) assignment or with sequential SRAMethod 2 (unequal a priori probabilities of
assignment). The final goal is to have a balanced design in which the levels of classificatory or prognostic factors are
homogeneously distributed across the experimental conditions included in the design layout, hence the expression
minimizing treatment imbalance. The main steps of Minimization (Method 13) are summarized in Table 2.23, and the
procedure is illustrated in Table 2.24.
Table 2.23 Major Steps in the Minimization Method (Method 13)
Suppose that the researcher is carrying out a randomized experiment with three treatment levels and 50 units have
already been assigned. When the 51st unit joins the experiment, it must be previously classified according to the relevant
prognostic factors (in the example, the structural characteristic of the 51st unit puts it on Levels 1, 2, and 2 of the 3
prognostic factors considered by the researcher). Treatment imbalanceif the unit is assigned to each of the treatment
levels included in the designis then assessed on the basis of the previously chosen measure of treatment imbalance
(calculus with range and variance measures is illustrated in Section B of Table 2.24). Finally, if the observed imbalance is
greater than the maximum tolerate imbalance, also defined in advance at the beginning of the study, the unit is assigned
to the treatment level that minimizes the overall imbalance (in the example, the 51st unit is assigned to Treatment 3, for
which the range or the variance measure gives the smallest value: range = 4; variance = 1.66). This is a deterministic
(nonrandom) assignment. A variation of the procedure uses the biased coin method (Efron, 1971), our Method 2 (unequal
a priori probabilities of assignment), with probabilities of assignment inversely proportional to the amount of imbalance for
each treatment.
Please note that the unit is not assigned to the treatment level with the lowest number of experimental units, as a less
attentive inspection of the data in Section A of Table 2.24 would suggest (our example is the
Table 2.24 Minimization: Assessment Treatment Imbalance Within the Levels of Classificatory (Prognostic) Factors (Method 13)
original example of Pocock & Simon, 1975). For instance, if the levels of prognostic factors for the 51st unit were Level 2
(Factor 1), Level 1 (Factor 2), and Level 1 (Factor 3), the range measures (6, 3, and 4, respectively) or the variance
measures (4.00, 1.00, and 2.00, respectively) would imply the assignment of Treatment T2, in the case of deterministic
assignment, or greater probabilities of assignment to Treatment T2, in the case of the application of Method 2. We have
also assumed that all prognostic factors were equally important; therefore, no weighting scheme has been applied to the
calculus of the amounts of imbalance. Finally, note that when the probabilistic scheme used in Method 2 approaches the
equal probabilities of assignment procedure, minimization becomes SRA and then the baby is lost with the bath water;
minimization no longer does what it purports to do (Berger, 2010, p. 406).
http://dx.doi.org/10.4135/9781452270012.n2
18/1/2014 Chapter 3. Within-Subjects Designs Randomization : SAGE Research Methods
Chapter 3. Within-Subjects Designs Randomization
3.1 Basic Assumptions and Specific Threats to Validity
A within-subjects design is an experimental design in which two or more units (subjects) are exposed to two or more
treatment levels (or treatment level combinations) over two or more periods and outcome measures are recorded after each
exposure. As such, within-subjects designs contrast with between-subjects designs, wherein each experimental unit is
only exposed to one treatment level (or treatment level combination). Figure 3.1 outlines a within-subjects design in which
every experimental unit receives three treatment levels (A, B, and C, not necessarily in this order) in three consecutive
periods (Periods 1, 2, and 3). Briefly stated, within-subjects designs concern experimental situations where a number of
subjects receive a sequence of treatments over a number of time periods (Bate & Jones, 2006, p. 3249).
Figure 3.1 Direct and carryover effects in within-subjects designs (the first-order residual effect of A on O2 is omitted)
In methodological and statistical literature, within-subjects designs are also labeled repeated measures designs or
repeated measurement designs, emphasizing the idea that several measures are taken on the same subjects. However, it
is convenient to make a clear distinction between what has been called a longitudinal data structure (i.e., collecting one or
more measures on the same units over successive periods) and an experimental repeated measures design. In a
longitudinal data structure, the interval between the first and the last period of measurement can vary from a few seconds
to months or even years, and the intervals between adjacent periods can have the same or different durations. The same
is true in relation to a repeated measures design, given that from a measurement perspective the design is a special kind
of longitudinal data structure. However, what makes the specificity of within-subjects or repeated measures experimental
designs is the administration of different treatment levels to the same units and the recording of outcome measures
immediately after each administration, not the longitudinal data structure in itself. Longitudinal data structures can arise
from experimental, quasi-experimental, and nonexperimental studies (see Figure 1.1). More specifically, they can take the
form of either before- and after-treatment measures in interrupted time series quasi-experimental designs or
single/multiple pretest-posttest measures in between-subjects experimental (pretest-posttest designs and longitudinal
designs) and quasi-experimental (regression discontinuity designs and nonequivalent comparison group designs; see
Shadish et al., 2002) designs. Finally, longitudinal data structures can be obtained in nonexperimetal research, as it is the
typical case of panel surveys. An additional distinction must also be made between repeated measures designs and
designs with multiple measures, these being true between-subjects designs in which two or more different outcome
measures have been collected after the treatment administration (multivariate analysis of variance designs).
Moreover, in some disciplines (e.g., experimental psychology) the expression within-subjects designs applies to
completely different situations: subjects-by-trials and subjects-by-treatments designs (see Kirk, 1995). The first
situation corresponds to studies in which the same participants are submitted to several trials to quantify learning
progress in specific tasks. Rigorously speaking, these are not randomized experiments, since all participants are
measured in the same series of ordered trials, and the aim of the study is to establish a functional relationship (learning
curve) between the amount of training (number of trials) and the quality of the performance (indexed by the number of
errors or the time required for accomplishing the learning task). Subjects-by-trials experiments will not be covered in this
monograph, but we should note that comparing the learning curves of two or more groups of participants tested under
different learning conditions could be an example of a between-subjects design if the participants had been randomly
assigned to distinct learning groups. The second situation, subjects-by-treatments designs, corresponds exactly to our
initial definition of within-subjects experimental designs and is the main topic of this chapter. When moving from one
period to the next, subjects are exposed to a different treatment level; thus, they are said to cross over treatments, hence
the designation cross-over designs, under which these designs are also known.
As compared with between-subjects designs, within-subjects designs have two main advantages. First, for the same
number of observations, they require fewer experimental units (NW = NB/t; compare the structural features of basic one-
factor designs in Table 4.2, Chapter 4). Second, as each experimental unit acts as its own control by being observed in
different experimental conditions, error variability is expected to be substantially reduced. Taken together, these
advantages add up to make within-subjects designs more powerful than between-subjects designs. Unfortunately, the
advantages can be overcome by serious drawbacks.
First of all, there are research problems and theoretical hypotheses that by their very nature and content don't allow the
use of cross-over designs. For example, when social or educational programs are known for having lasting and
nonreversible effects (hopefully!), it makes no sense to deliver alternative programs to the same subjects (e.g., assessing
the differential effectiveness of two methods of teaching reading skills to the same subjects in two consecutive periods).
By the same token, when manipulations rely on the cognitive definition of situations, as in the case of many experiments
in social and cognitive psychology, once a cover story is created to allow the implementation of a treatment, there is no
possibility of constructing a plausible scenario to introduce an alternative treatment. In brief, before proceeding with a
cross-over design, the researcher must always take into account the possibility that his or her manipulations may be
meaningless to the study participants. This is basically a problem of construct validity of the presumed causes (see
Subsection 4.1.1).
Second, in a cross-over experiment there is always the very possibility that the effect of a treatment level will be
confounded with the effects of the treatment levels and the measurement operations that took place in the previous
periods (see Figure 3.1). This kind of confounding is in essence a matter of internal and statistical conclusion validity of
local causal inferences that is known under several designations in methodological textbooks and the experimental design
literature.
By definition, the measurement operationsthat is, the operationalization of the dependent variable but not, as it is
obvious, the outcome actually recordedare identical for all subjects in all periods. As a result, no matter which treatment
level has been administered in the previous period, the potential confounding effect of the associated measurement
operation is indistinguishable from the confounding effects of the measurement operations associated with the remaining
treatment levels. In other words, the confounding effects of previous measurement operations are intrinsically position
effects: How many measurement operations have preceded the current observation?
Also by definition, treatment levels are always distinct operationalizations of the independent variable, therefore the
potential confounding effect of the treatment level administrated in the previous period is unlikely to be identical to the
confounding effects of the other treatment levels. In other words, these effects are essentially sequence effects: Which is
precisely the treatment level administered in the previous period, or more generally, which is the sequence of treatment
levels administered in all previous periods?
Position effects and sequence effects in cross-over designs have been contrasted under different labels: rank order effects
versus sequence effects (Lindquist, 1953); practice, order, time-related, or progressive-error effects versus treatment-
carryover or treatment-transfer effects (Kiess & Bloomquist, 1985); practice or testing position effects versus carryover
effects (Keppel, 1991); testing position effects versus sequence effects (Reese, 1997); and order main effects versus
carryover effects (Maxwell & Delaney, 2004)the list is not exhaustive. Generic labels for both effects (some of them
rather ambiguous as order effects) have also been used, such as multiple-treatment effects (Kiess & Bloomquist, 1985)
or context effects (Greenwald, 1976). In specialized cross-over designs literature (e.g., Bose & Dey, 2009; Jones &
Kenward, 2003), sequence effects are interchangeably named carryover effects or residual effects and are distinguished
from treatment direct effects (see Figure 3.1).
Taken all together, the effects of previous treatments and the effects of previous measurement operations can be
conceptualized as external events or situational nuisance variables, and in this sense, they can be thought of as a threat
to internal validity called history. Taken alone, the effects of previous measurement operations correspond exactly to the
Campbellian definition of testing. In practice, given that each treatment is always coupled with the corresponding outcome
measure, we cannot disentangle the testing effects of previous measurement operations from the residual effects of
previous treatments. So we think that the carryover effects of previous periods can be best described as additive and
interactive effects of history and testing. From a substantive point of view, these effects can be explained by perceptive
(e.g., assimilation/contrast), learning (e.g., transfer, fatigue, boredom), motivational, and more general cognitive and
affective processes involved in the definition/interpretation of the experimental situation.
In a very clear presentation of Rubin's causal model, Holland (1986) puts into words the fundamental problem of causal
inferenceIt is impossible to observe the value of Yt(u) and Yc(u) on the same unit, and, therefore, it is impossible to
observe the effect of t on u (p. 947)and elaborates on two general solutions, which he calls the scientific solution and
the statistical solution. The latter solution is the one adopted in between-subjects designs and relies on the
randomization methods described in the preceding chapter, on Rubin's SUTVA, and on the estimation of average causal
effects; the former solution is the solution achieved by cross-over designs if the strong assumption of homogeneity or
invariance is tenable. To be more precise, an ideal cross-over design is carried out by presupposing that the units general
behavior is constant across experimental conditions (temporal stability) and there are no lasting carryover effects of
previous periods (causal transience).
In practice, there are two basic and complementary strategies to approach this ideal: (1) to insert sufficiently wide time
intervals (washout or rest periods; see Figure 3.1) between adjacent experimental periods, hoping that potential carryover
effects will dissipate, and (2) to compensate for carryover effects by counterbalancing the order of presentation of the
different treatments. The first strategy is always recommendable, despite knowing that the longer the duration of the
experiment is, the lower the probability of ensuring the constancy of units general behavior across experimental
conditions. The implementation of the second strategy is the main subject of Sections 3.3 through 3.7. However, before
proceeding with the description and the illustration of counterbalancing schemes, we must classify the different kinds of
cross-over designs and summarize their structural properties. This will be done in the next section.
3.2 Treatment Design and Methods of Randomization
Formally speaking, a cross-over designdenoted by CO(t, n, p)is an array of n rows and p columns, defining np
experimental sites (cells), coupled with a plan for the random assignment of t treatment levels (or treatment level
combinations) to the np experimental sites. By convention, each row corresponds to one experimental unit (subject) that
receives two or more treatment levels in successive periods (columns). Also by convention, the different treatment levels
assigned to the np experimental sites are denoted by Latin letters (A, B, C, ).
There are two basic structural properties of cross-over designs: uniformity and balance. Uniformity concerns the number of
replications of treatment levels within rows (subjects) and columns (periods). Balance relates to the sequence of treatment
levels across the rows.
A cross-over design is said to be uniform on rows when, for each experimental unit (subject), the same number of periods
is assigned to each treatment level; likewise, a design is said to be uniform on columns when, in each period, the same
number of experimental units (subjects) is assigned to each treatment level. For example, Designs 6B2p (see Table 3.1)
and 6N4 (see Table 3.2) are uniform on rows but not on columns; similarly, Designs 6D2e (see Table 3.1) and 6R4 (see
Table 3.2) are uniform on columns but not on rows. It is evident that the necessary condition for uniformity on rows is that
p = rt and the necessary condition for uniformity on columns is that n = ct, where r and c are integers greater than or
equal to 1.
Table 3.1 Uniformity and Balance in One-Treatment Cross-Over Designs: Examples for Two Treatment Levels (t = 2)
Table 3.2 Uniformity and Balance (First-Order Carryover Effects) in One-Treatment Cross-Over Designs: Examples for Three and
Four Treatment Levels (t = 3 and t = 4)
A cross-over design is said to be a uniform design when it is simultaneously uniform on rows and columns (i.e., p = rt
and n = ct). For example, Designs 6A2 (r = 1, c = 1), 6B2e (r = 2, c = 1), 6H3 (r = 1, c = 2), and 6M4 (r = 1, c
= 1) are uniform designs (see Tables 3.1 and 3.2).
A cross-over design is said to be balanced with respect to the first-order carryover effects when each treatment level is
not immediately preceded by itself but is immediately preceded by all the remaining treatment levels the same number of
times. For example, in Design 6G4, Treatment Level A is immediately preceded once (and only once) by Treatment Levels
B, C, and D; the same is true for B, C, and D regarding the remaining treatments. This pattern doesn't occur in Design
6M4, wherein A is immediately preceded twice by B, once by C, and never by D (see Table 3.2).
The necessary condition to achieve balance for the first-order carryover effects is that the equation [n(p 1)]/[t(t 1)] = 1
gives an integer value (1) for 1. In this caseand when all additional conditions are fulfilled the parameter 1 can be
read as the number of times a treatment level is immediately preceded by each of the remaining treatment levels in
balanced designs (see Tables 3.1 and 3.2).
A cross-over design is said to be strongly balanced with respect to the first-order carryover effects when each treatment
level is immediately preceded by all treatment levels (including the treatment level itself) the same number of times. For
example, in Design 6R4, A is immediately preceded by A, B, C, and D only once; the same is true for B, C, and D (see
Table 3.2).
The necessary condition to have a strongly balanced design for the first-order carryover effects is that the equation [n(p
1)]/[t
2
] = 2 gives an integer value (1) for 2. If this is trueand if all additional conditions are satisfiedthe parameter 2
is the number of times a treatment level is immediately preceded by all treatment levels (including the treatment level
itself) in strongly balanced designs (see Tables 3.1 and 3.2).
For a given number of treatment levels and periods and an anticipated degree of balance and/or uniformity, a standard
design is the design that accomplishes the intended structural properties with the minimum possible number of
experimental units. By convention and to avoid misidentification or giving different names to the same design, standard
designs are always written with the first row and the first column in alphabetical order (standard form). So a standard
design is a minimal design in its standard form. With the exception of nonutilizable designs and Designs 6N3 and 6N4, all
designs presented in Tables 3.1 and 3.2 are in their standard form (please note that in a standard design there are no
duplicated sequences). In Section 3.5, we illustrate the procedure for reducing a given design to its standard form (see
Table 3.11, Steps 2 and 3).
Cross-over designs can be classified on the basis of the following criteria: (a) the number of treatment and treatment
levels (one vs. two or more treatments, two vs. three or more treatment levels); (b) the combination of within-subjects
treatments with between-subjects treatments (designs involving only within-subjects treatments vs. designs including at
least one between-subjects treatment); (c) the relationship between the number of periods and the number of treatment
levelsor treatment level combinationsincluded in the study (complete, incomplete, and extended sequences); (d) the
structural properties of the standard design, namely, uniformity (nonuniform vs. uniform designs) and balance
(nonbalanced, partially balanced, nearly balanced, balanced, nearly strongly balanced, and strongly balanced designs);
(e) the number of admissible partitions of the standard design (designs consisting of a single set of sequences vs.
designs organized into different subsets of sequences; one-brick designs vs. multiple-bricks designs); and (f) the number
of replications of the standard design (one vs. two or more replications).
Table 3.3 shows a classification of cross-over designs based on the first four criteria. Please note that criterion (c) and
criterion (d) are not strictly independent, since the extension of sequences (i.e., the number of periods) is directly related
to the structural properties of the intended design. The last two criteria are especially relevant when within-subjects
treatments are combined with between-subjects treatments and/or between-subjects blocking structures. In addition to
power-related issues, the replication of standard designs creates the conditions to give identical sets of sequences to
groups of experimental units either defined by between-subjects blocking structures or generated by between-subjects
randomization. Standard designs consisting of multiple bricks can also be broken into different but structurally equivalent
subsets of sequences, which can be randomly assigned to distinct groups of experimental units.
The randomization methods relevant to cross-over designs can be categorized in terms of the intended structural
properties and the presence/absence
Table 3.3 Classification of Cross-Over Experimental Designs
of restrictions imposed by concomitant consideration of between-subjects treatments or blocking structures. In other
words, if neither uniformity nor balance is wanted for the design, the proper method is to randomly order the treatment
levels separately for each unit (Method 14: Random Counterbalancing); if uniformity, but not balance, is a requisite,
randomly selected Latin squares are the suitable choice (Method 15: Positional Counterbalancing). Finally, if some
degree of balance is required, the researcher must select the appropriate standard design and proceed to the
randomization following the major steps for Sequential Counterbalancing (Methods 16-18) described in Table 3.4. The
choice of the sequential counterbalancing method depends on the restrictions imposed by between-subjects blocking
structures and/or between-subjects randomization.
Having in mind that in practice it is impossible to disentangle the testing effects of previous measurement operations and
the residual effects of previous treatments (see Section 3.1), the main theoretical assumptions about carryover effects and
the distinctive features of counterbalancing methods are summarized in Table 3.5. The reader should be careful about the
clear delimitation between the so-called symmetrical and asymmetrical residual effects of previous treatments. As we
noted in Section 3.1, by definition the residual effects of different treatment levels are expected to be dissimilar. Yet,
depending on the very nature of the treatment and the magnitude of the differential features of treatment levels (i.e.,
experimental conditions), it can be conceivable to handle all carryover effects as symmetrical effects. These matters are
not consensual, and they are intertwined with the general approaches to cross-over
Table 3.4 Major Steps in the Randomization of Cross-Over Experimental Designs
designs. In broad terms, we can identify two main approaches. The first one remains on theoretical ground, and in the
presence of asymmetrical residual effects, the researcher is recommended to leave out cross-over designs and opt for
between-subjects designs (e.g., Maxwell & Delaney, 2004). When no asymmetrical residual effects are expected, random
or positional counterbalancing could be employed, depending on the assumptions about testing effects: Random
counterbalancing is sufficient when no testing effects are expected, and positional counterbalancing is the suitable
strategy for control cumulative testing effects. The second approach takes a more practical point of view and considers
that the issue is essentially an empirical one (e.g., Jones & Kenward, 2003). Sequential counterbalancing could and
should be used as a device to control asymmetrical effects. In any case, carryover effects must be assessed and
statistically modeled in conjunction with direct effects in the data analysis stage. It should also be noted that there are
studies in which the assessment of carryover effects is the experimenter's main goal.
Table 3.5 Assumptions about Carryover Effects of Previous Periods and Distinctive Control Features of Counterbalancing
Methods
In Sections 3.3 to 3.6, the five randomization methods of cross-over designs are described and illustrated with one-
treatment designs (Designs 6A-6T; see Table 3.3). The standard designs used for the illustration of randomization
methods are identified with the labels given in Table 3.3, followed by the number of treatment levels included in the design;
for example, Design 6G4 in Table 3.15 means that we have used a balanced Williams's square for an even number of
treatment levels (Design 6G; see Table 3.3) with four treatment levels.
To make the text less cumbersome, we will use the term treatment(s) instead of treatment level(s), with the understanding
that we are referring to the different levels of a treatment and not to the treatment itself (i.e., the within-subjects
experimental factor; see Table 1.5 for the notational system used throughout the book). Please note that this is the
current practice in the within-subjects designs literature.
The final section deals with the adjustment of counterbalancing schemes to factorial cross-over designs (Designs 6U and
6V; see Table 3.3). As we did in Chapter 2, instead of presenting randomization methods as stand-alone tools, emphasis
is placed on the interdependence between counterbalancing strategies and the structural features of cross-over
experimental designs. This is particularly relevant for the construction and selection of appropriate standard designs in the
case of sequential counterbalancing.
3.3 Random Counterbalancing
When no asymmetrical residual effects of previous treatments are expected and the testing effects of previous
measurement operations can be dismissed, the only requisite for the cross-over design is uniformity on rows (i.e., all
treatments are assigned to each subject the same number of times), and the counterbalancing strategy consists in
randomly ordering the treatments separately for each subject (Method 14: Random Counterbalancing). This is easily done
by taking as many random permutations of t integers (t = number of treatments) as are necessary to randomize the order
of the treatments for all participants in the study. In Table 3.6, the procedure is illustrated with seven subjects and four
treatments: Seven random permutations of the first four integers were drawn from Table A1.1 according to the procedure
explained in Appendix 2. The treatments are ordered in consonance with the random permutations adopting the system of
correspondence, previously established by the researcher, between integer numbers and treatment labels (e.g., the
random permutation taken for Brent4, 3, 1, 2leads to the sequence D, C, A, B; see Table 3.6).
The design produced by this method of counterbalancing is called a Repeated Measures Randomized Block Design
(Design 6N; see Table 3.3) and is structurally equivalent to the between-subjects Randomized Block Design (Design 2;
see Subsection 2.3.1). In the within-subjects version, subjects are thought of as distinct blocks, and the p experimental
sites within each row act as units of randomization. So Method 14Random Counterbalancingis a simple variation of
Method 7Blocked Random Assignment With One Blocking Variable.
3.4 Positional Counterbalancing
If no asymmetrical residual effects of previous treatments are anticipated but the cumulative testing effects of previous
measurement operations are a matter of concern, the cross-over design should guarantee, in addition to uniformity on
rows, that each treatment occurs equally often in all periods (uniformity on columns). This is done by selecting a standard
Latin square (see Table A3.1, Appendix 3) and by permuting, randomly and independently, the order of the rows, the order
of the columns, and the labels (letters/symbols) of the treatments.
Table 3.6 Random Counterbalancing: Independent Random Ordering of Four Treatments (A, B, C, and D) for Seven Units
(Design 6N4; Method 14)
The procedure of selection and randomization of the Latin square is the same as the one we illustrated in Chapter 2 (see
Figure 2.2) along with the randomization of the Between-Subjects Latin Square Design (Design 5A). However, although in
the between-subjects design the randomization is limited to a single Latin square, in the case of the Repeated Measures
Latin Square Design (Design 6M; see Table 3.3) a new randomized Latin square is required for each consecutive group of
n experimental units. Thus, Method 15Positional Counterbalancingcan be seen as a variation of Method 9Two-Way
Blocked Random Assignment: Latin Squares (see Subsection 2.3.4). More precisely, for each group of n (n = p = t)
experimental units, the double-blocking structure of Latin squares defines n
2
experimental sites (units of randomization),
and the random assignment of treatments to experimental sites is done by the randomization of the proper standard Latin
square.
Method 15 is illustrated in Table 3.7 with a five-treatment design and 15 experimental units. Please note that in captive
assignment situations the names of experimental units should be listed in a systematic order (alphabetical or other)
before the randomization. In noncaptive assignment circumstances, the names are replaced with sequential ID numbers
(1-n), which are ascribed to experimental units as long as they are enrolled in the experiment. In any case, the number of
experimental units must be a multiple of the number of treatments. Standard Latin squares of orders 2 to 12 are given in
Table A3.1 (Appendix 3). For orders higher than 5, the randomization is the very same as that given in Table 3.7. For
orders 2 to 4, shortcuts and details are given in Appendix 3.
Table 3.7 Positional Counterbalancing: Random Assignment of Sequences of Five Treatments (A, B, C, D, and E) to 15 Units
(Three Latin Squares) (Design 6M5; Method 15)
In introductory methodological textbooks, positional counterbalancing is frequently called partial counterbalancing and is
contrasted with situations where all possible sequences of t treatments are included in the design (complete
counterbalancing). Both designations are somehow ambiguous and should be avoided. Partial counterbalancing can be
confused with partially balanced designs, in which a precise degree of balance is achieved (see Subsection 3.5.2). This is
not the case with repeated measures Latin square designs: With the exception of the 2 2 Cross-Over Design (Design
6A2; see Table 3.1), which is a balanced design (position and sequence are redundant terms), all designs including three
or more treatments are nonbalanced designs (e.g., in the first square of Table 3.7, Treatment A is preceded twice by C,
once by B and E, and never by D; in the whole design, Treatment A is preceded once by B, twice by D, three times by E,
and six times by C).
What is achieved by repeated measures Latin square designs is that all treatments are administered the same number of
times (in the example of Table 3.7, c = 3) in all positions (periods); that is the reason why we adopt the designation
positional counterbalancing, instead of partial counterbalancing (for misunderstandings underlying the use of the
expression complete counterbalancing, see Subsection 3.5.1).
3.5 Nonrestricted Sequential Counterbalancing
When the potential occurrence of asymmetrical carryover effects of previous periods is a matter of concern, the selection
of a balanced design achieves an ideal sequencing of treatments and, more important, allows the simultaneous
assessment of direct and carryover effects. In the analysis stage, statistical adjustment for carryover effects provides the
researcher with a more accurate estimation of direct effects.
In sequential counterbalancing, the critical feature is the selection of a standard design with the desired structural
properties. Once the design has been chosen, the randomization is straightforward and consists in simple variations of the
methods described in Chapter 2, with the understanding that sequences of treatments, instead of the treatments
themselves, are randomly assigned to experimental units.
The major steps of sequential counterbalancing methods are summarized in Table 3.4. This section deals with Method 16
Nonrestricted Sequential Counterbalancingthat is, with situations where the random assignment of sequences is
done at once for all experimental units, not being restricted by between-subjects blocking structures and/or random
assignment of units to the levels of between-subjects treatments included in the design, as is the case with Methods 17
and 18, discussed in the next section.
Since the implementation of Method 16 is identical for all designs described in this section, we will give the details in one
case (Design 6F: Complete Sets of Mutually Orthogonal Latin Squares), devoting the bulk of the text to the selection and
construction of the standard designs to be randomized, with special emphasis on uniform balanced designs. In
consonance with the classification of cross-over designs presented in Table 3.3, designs consisting of complete (p = t),
incomplete (p < t), and extended (p > t) sequences are presented in the three main subsections. Within each subsection,
designs are organized according to their degree of balance.
3.5.1 Complete Sequences (p = t)
In complete-sequences designs, the number of periods equals the number of treatments (p = t), and all standard designs
are uniform designs; that is, the number of replications per row (subject) and per column (period) is the same for all
treatments.
3.5.1.1 Balanced Designs
The most basic cross-over design is the design with two treatments (A and B) and two periods involving only two possible
sequences: AB and BA (Design 6A2; see Table 3.1). With this design, both positional and sequential counterbalancing
produce a balanced design, hence the researcher can adopt either Method 15 or Method 16.
3.5.1.1.1 All Possible Sequences
At first glance, the inclusion of all possible sequences of treatments would be the ideal procedure for all cross-over
designs. However, the minimum number (n = t!) of experimental units required for the standard design precludes its use in
experimental studies with a high number of treatments. For example, with 2, 3, 4, 5, and 6 treatments, the numbers of
different sequences of the standard designs are 2, 6, 24, 120, and 720, respectively (see Table 3.12).
When applying Method 16Nonrestricted Sequential Counterbalancingto cross-over designs including all possible
sequences of treatments (Designs 6A and 6E), Step 1 (random assignment of treatments to symbols of the standard
design; see Table 3.9) is unnecessary, because changing the symbols of treatments would result in the same set of
sequences.
One important structural property of cross-over designs including all possible sequences of treatments is that they are
balanced not only for the first-order carryover effects but also for higher order carryover effects (see Table A2.1, Appendix
2, where the entire set of sequences for two up to five treatments is given). This is the reason why sequential
counterbalancing with all possible sequences is frequently called complete counterbalancing. As we have already noted,
this is an ambiguous designation, because in cross-over designs literature complete balance is synonymous with strong
balance (Jones & Kenward, 2003), as is the case with Designs 6P, 6R, and 6S (see Table 3.3 and Subsections 3.5.2 and
3.5.3).
3.5.1.1.2 Sets of Mutually Orthogonal Latin Squares
A realistic alternative to all possible sequences designs, which preserves balance for the first- and higher order carryover
effects and requires a substantially smaller number of experimental units (see Table 3.12), is the use of complete sets of
mutually orthogonal Latin squares (MOLS). Two Latin squares of the same order are said to be orthogonal if, when
superimposed, each symbol (e.g., Latin letter) of the first square occurs once and only once in conjunction with each
symbol (e.g., Greek letter or number) of the second square. For a given number of t treatments, complete sets of t 1
MOLS can be constructed when t is a prime or a power of a prime number; that is, we can have complete sets of t 1
MOLS for 3, 4 (=2
2
), 5, 7, 8 (=2
3
), 9 (=3
2
), 11, 13, () treatments.
When a complete set of MOLS exists and the squares are odd sided, the t 1 squares can be easily constructed by
cyclically moving the rows or the columns of the initial square (see illustrations in Table 3.8 for 3- and 5-sided squares).
For even-sided squares, the procedure is a little more complicated and is not illustrated here (see Fisher & Yates, 1963).
However, standard forms of complete sets of 4- and 8-sided MOLS, as well as odd-sided squares up to 11, are shown in
Tables A3.3 and A3.4 (Appendix 3).
Table 3.8 Construction of Complete Sets of Mutually Orthogonal Latin Squares (MOLS): Procedure and Examples for Odd-Sided
Squares
Once the complete set of MOLS (standard form) has been chosen, the researcher can proceed to Nonrestricted
Sequential Counterbalancing (Method 16) according to the steps summarized in Table 3.4 and illustrated in Table 3.9 for
the case of four treatments (A, B, C, and D). Given that the
Table 3.9 Nonrestricted Sequential Counterbalancing: Random Assignment of Sequences of Four Treatments (A, B, C, and D;
Three Mutually Orthogonal Latin SquaresMOLS 4; See Table A3.3) to 12 Units (Design 6F4; Method 16)
standard form of MOLS doesn't constitute all possible sequences of treatments, Step 1 ensures that the sequences to be
included in the cross-over design are randomly selected within the range of possibilities allowed by the combinatorial
structure of the design. In Step 2, selected sequences are randomly assigned to experimental units. Please note that
Step 2 is a simple variation of Method 3Simple Random Assignment With Forced Equal Sizesdescribed in
Subsection 2.2.1, in which sequences of treatments, instead of the treatments themselves, are assigned to experimental
units.
In Table 3.9, we have associated a randomly ordered list of sequences to a systematically ordered list of experimental
units. As we noted in Chapter 2, it is formally equivalent to associate a randomly ordered list of experimental units to a
systematically ordered list of sequences. We have chosen the first procedure because it applies to both captive and
noncaptive assignment circumstances. In the captive assignment, subjects names are listed in some systematic order
(e.g., alphabetical) prior to the randomization; in the noncaptive assignment, the names are replaced with sequential ID
numbers (1-n), which are ascribed to experimental units as they are enrolled in the experiment. In the example of Table
3.9, there is only one replication of Design 6F4; therefore, the number of experimental units is equal to the number of
sequences; if the final design comprises two or more replications of the standard design, the number of experimental units
must be a multiple of the number of sequences.
As stated in Table 3.4, regardless of the number of constitutive bricks (in the example, three squares) and the number of
replications of the standard design, all the sequences included in the final design are seen as a single set of sequences,
and the only procedural constraint is the assignment of the same number of experimental units to each sequence. It is
always a good practice to keep track of the location of sequences in the standard design (e.g., ascribing them a
sequence number-SN in Table 3.9). In case of experimental attrition in multiple-bricks designs, this will allow an accurate
statistical analysis of intact subsets of the whole design.
3.5.1.1.3 Williams's Designs
All possible sequences designs and MOLS designs achieve structural balance for first- and higher order carryover effects
by using a considerable number of experimental units per treatment. In the former designs, this number quickly becomes
unreasonable, whereas in the latter ones it is kept in acceptable sizes. However, there are situations where a limited
number of available subjects and/or a restrictive budget force the experimenter to work with fewer subjects than those
required to implement a MOLS design. In these situations, it would be highly desirable to know what is the minimum
number of subjects required to achieve a given degree of balance. With respect to the first-order carryover effects, the
answer to this question was given by Williams (1949, 1950) about 60 years ago: For an even number of treatments, it is
possible to have a balanced design in that the number of experimental units equals the number of treatments; for an odd
number of treatments, the number of subjects should double the number of treatments. Williams's designs are special
cases of Latin squares in which each treatment is immediately preceded equally often by all the remaining treatments
(see Designs 6H3 and 6G4 in Table 3.2 and Designs 6G4 and 6H5 in Table 3.11).
There are several procedures for constructing Williams's designs (see, Namboodiri, 1972), including the original ones given
by Williams (1949). In Table 3.10, we illustrate the procedure devised by Sheehe and Bross (1961), which applies to both
even and odd treatments and has the advantage of being easy to remember. Williams's Latin squares are the most
frequently used designs in sequential counterbalancing and are also known under the names single carryover design (even
number of treatments) and double carryover design (odd number of treatments).
Like any other cross-over design, the designs obtained by the procedure explained in Table 3.10 can and should be
reduced to their standard form. In Table 3.11, we show how this is done by renaming the treatments so that the first row is
written in alphabetical order and by reordering the rows so that the first column is also written in alphabetical order.
Standard forms of Williams's squares of 2 up to 12 orders are given in Table A3.5 (Appendix 3). Starting with these
standard forms, the sequential counterbalancing of William's designs is the same as what we have illustrated in Table 3.9
with MOLS. Please note that the Williams's design for two treatments is the 2-sided standard Latin square (see Table
A3.1) and the Williams's design for three treatments is identical to the MOLS design (see Table A3.3).
3.5.1.1.4 Other Balanced Designs
For some cross-over designs with an odd number of treatments, Williams's designs are no longer minimal designs, since
Hedayat and Afsarinejad (1975, 1978) have devised balanced squares with n = t for 9, 15, 21, and 27 treatments (the
standard design for t = 9 is given in Table 3.13, Design 6I9).
3.5.1.2 Nearly Balanced Designs
In the cases of three, five, and seven treatments, Williams's designs (i.e., double carryover designs requiring 6, 10, and 14
experimental units, respectively) remain the minimal balanced designs. When the number of available experimental units
is not enough to comply with the requirements of Williams's designs, alternative nearly balanced designs (with n = t) can
Table 3.10 Construction of Balanced Latin Squares for the First-Order Carryover Effects in Cross-Over Designs (Williams's
Squares)
be used with five and seven treatments. These designs, developed by Russell (1991), approach the combinatorial
properties of balanced designs in the sense that all treatments are equally often immediately preceded by the remaining
treatments, with the exception of two of them. For example, in Design 6L7 (see Table 3.13), Treatment A is immediately
preceded once
Table 3.11 Reduction of Williams's Squares to the Standard Form
by Treatments B, C, D, and G, but it is preceded twice by E and never by F (a similar pattern occurs for treatments B, C,
D, E, F, and G).
3.5.1.3 Choosing a Uniform Balanced Design
Putting aside all possible sequences designs, the choice of a uniform balanced standard design in sequential
counterbalancing is between MOLS (Designs 6F) and minimal designs (Designs 6G, 6H, and 6I). As Jones and Kenward
(2003) remarked, when the main concern is balance with respect to the first-order carryover effects, which is often the
case, minimal designs are the obvious choice, having the practical advantage of making the experimental mise en place
less expensive and less prone to procedural errors, given that a smaller number of distinct sequences is involved in the
research protocol. On the other hand, MOLS designs allow, in the analysis stage, the assessment and modeling of higher
order carryover effects that, in certain circumstances, can be, in their own right, the aim of the experiment. Please note
that in this choice we are not talking about power: Either standard MOLS or standard minimal designs can be replicated
to meet the power requirements. Needless to say, with a limited number of available experimental units, minimal designs
are the obvious choice.
In Table 3.12, we give the minimum number of experimental units required for the sequential counterbalancing with the
uniform balanced designs reviewed in this subsection. In Table 3.13, we show Newcombe and Prescott's Triples
(Designs 6J5 and 6J7), which are balanced designs requiring n = 3t subjects and can be used when a complete
replication of a double carryover design, calling for n = 4t subjects, is not feasible. For a thorough review of uniform
balanced designs, see Bate and Jones (2008). For more advanced topics in the choice of designs when few subjects are
available and the implementation of complex algorithms for the construction of cyclic cross-over designs, see Bate,
Godolphin, and Godolphin (2008).
3.5.2 Incomplete Sequences (p < t)
There are situations in which researchers carry out designs with a relatively high number of treatments, but because of the
time needed for the implementation of each treatment and/or for some technical feature of the design, it is not
recommended that the subjects be observed in all the experimental conditions. In these situations, one possible strategy
is the administration of incomplete sequences of treatments with a prespecified pattern. When the number of p periods (p
< t) is equal for all subjects and all treatments are equally replicated within each period (uniformity on periods), the
resulting design is called an incomplete cross-over design. Depending on the pattern of treatment replication, the design
can be balanced, strongly balanced, or partially balanced.
Table 3.12 Sequential Counterbalancing: Minimum Number of Experimental Units Required in Uniform Balanced Cross-Over
Designs
3.5.2.1 Balanced Designs
There are several ways to build an incomplete balanced design, but here we only give two of them, referring the reader to
the text of Jones and Kenward (2003) for a systematic presentation. The first and the easiest way to get an incomplete
balanced design for t treatments and p < t periods is to rely on MOLS designs and to delete one or more periods of the
whole design. For example, if an incomplete design with t = 5 and p = 3 is needed, the researcher may go to Table A3.3
(Appendix 3), select the 5 5 standard MOLS, and delete the last two columns. The resulting design (Design 6O5), with
five treatments and 20 sequences of three different treatments (ABC, BCD, CDE, DEA, EAB, ACE, BDA, CEB, DAC,
EBD, ADB, BEC, CAD, DBE, ECA, AED, BAE, CBA, DCB, and EDC), requiring a minimum of 20 subjects, includes four
replications of each treatment per period, and, as the reader could easily verify, is a balanced design for the first-order
carryover effects: All treatments are immediately preceded equally often by the remaining treatments (1 = 2). Note that
the designs obtained by this procedure are also balanced for higher order carryover effects.
Table 3.13 Sequential Counterbalancing: Complementary Balanced (Designs 6I9, 6J5, and 6J7) and Nearly Balanced (Designs
6L5 and 6L7) Cross-Over Designs
The second way to obtain an incomplete balanced design is to pick the corresponding between-subjects balanced
incomplete block design and replace each row with the possible permutations of the treatments involved. For example, if
Design 4A (6 treatments, 10 blocks, and 3 units per block; see Table 2.8) is used as a template, each of the 10 rows
would give rise to 6 permutations of 3 treatments. In the case of the first row, the set of Treatments 1, 2, and 5 would
generate the following permutations: 1, 2, 5; 1, 5, 2; 2, 1, 5; 2, 5, 1; 5, 1, 2; and 5, 2, 1. Repeating the procedure for the
remaining rows and replacing treatment numbers with letters, to be consonant with the notation adopted for cross-over
designs, the final design (Design 6O6) would consist of 60 sequences (the last 6 sequences would be DEF, DFE, EDF,
EFD, FDE, and FED). Like designs based on MOLS, designs based on balanced incomplete block designs are balanced
for carryover effects of all orders.
3.5.2.2 Strongly Balanced Designs (Extra-Period Designs)
If the reader adds an extra period to a balanced incomplete cross-over design containing p < t 1 periods and the
treatments in the last period of the original design are exactly replicated in the extra period, he or she gets a strongly
balanced design (see Section 3.2 for the formal definition of strong balance), in which the number of periods remains
smaller than the number of treatments (p < t). For example, if the design described in the previous paragraph (Design
6O6) is converted into an extra-period design (Design 6P6) with four periods, the last six sequences would be read as
DEFF, DFEE, EDFF, EFDD, FDEE, and FEDD.
3.5.2.3 Partially Balanced Designs
When the number of available experimental units and/or the values of t and p wanted by the researcher don't permit a
balanced incomplete cross-over design, he or she can adopt a special type of incomplete-sequences designs (p < t) in
which all treatments are equally replicated in each period (uniformity on periods) but not immediately preceded by the
remaining treatments the same number of times. More specifically, in a partially balanced design (Design 6Q), a
treatment is immediately preceded 1 times by some treatments and 2 times by the remaining different treatments. For
example, in Design 6Q4 (t = 4, n = 8, p = 3), including the sequences ABC, ABD, BAC, BAD, CDA, CDB, DCA, and
DCB, Treatment C is immediately preceded twice by Treatment D and only once by Treatments A and B (please note that
all treatments are replicated twice within each of the three periods).
There is an abundant literature on partially balanced cross-over designs and a great diversity of methods of design
construction, including cyclic designs (e.g., Davis & Hall, 1969; Iqbal & Jones, 1994) and the partially balanced designs of
Patterson and Lucas (1962). These and other methods, along with an extensive enumeration of designs, are
systematically presented in Jones and Kenward's (2003) book, to which the reader is referred. From the randomization
perspective, both partially balanced designs and balanced and strongly balanced designs with incomplete sequences (p <
t) are handled in the same way as the complete-sequences designs (p = t) described at length in Subsection 3.5.1. The
same is true for the extended-sequences designs (p > t), briefly presented in the next subsection.
3.5.3 Extended Sequences (p > t)
When the number of treatments is relatively low, it is possible to replicate one or more treatments within the same
sequence. Designs 6B and 6D, depicted in Table 3.1, are special cases of extended sequences with two treatments (A
and B), and some of them are known under different labels: reversal, switch-back, or alternating-treatment designs (e.g.,
Designs 6B2f, 6B2m, and 6B2l); switching treatments in blocks designs (e.g., Design 6D2l); and multiple baseline
designs (e.g., Design 6D2f; see Cotton, 1998, for designs historically focused on the point at which treatment change is
introduced). Multiple baseline designs and some unique sequences of treatments (e.g., ABA, ABAB) have also been used
in the so-called single-case (quasi)experimental studies (see Anderson, 2002; Barlow & Hersen, 1984). However, from the
standpoint of treatment design, the main aim of extended sequences (p > t) with two or more treatments is to obtain
strongly balanced designs.
3.5.3.1 Extra-Period Strongly Balanced Designs
The easiest way to have a strongly balanced design is to add an extra period to a uniform balanced design, as illustrated
in the right columns of Table 3.2 with Williams's Designs: Design 6H3 is converted into Design 6R3, and Design 6G4 gives
rise to Design 6R4. However, adding an extra period (identical to the last one of the original design) breaks the uniformity
of the p = t + 1 Extra-Period Strongly Balanced Designs (Designs 6R). The choice between the two designs lies in
theoretical grounds and statistical trade-offs in the estimation of direct and first-order carryover effects: In uniform balanced
designs (e.g., Designs 6H3 and 6G4), the direct effects of treatments are orthogonal to the subjects (rows) and to the
periods (columns) but not to the first-order carryover effects; in nonuniform extra-period strongly balanced designs (e.g.,
Designs 6R3 and 6R4), the direct effects of treatments are orthogonal to the first-order carryover effects and also to the
periods (columns), but the orthogonality relative to the subjects (rows) is lost.
3.5.3.2 Many-Periods Strongly Balanced Designs
A second type of extended-sequences design, known as Many-Periods Designs (Design 6S), makes use of more periods
than the extra-period designs and can simultaneously achieve uniformity and strong balance, making the direct effects of
treatments orthogonal to the first-order carryover effects, to the subjects (rows), and to the periods (columns). An example
for two treatments is given in Table 3.1 (see Design 6D2k). For examples of designs with more than two treatments (e.g.,
Quenouille's designs, Berenblut's designs) and for bibliographic sources on the methods of construction, as well as for the
illustration of Nearly Strongly Balanced Designs (Designs 6T), we refer the reader to Jones and Kenward (2003) once
again.
3.6 Restricted Sequential Counterbalancing
Method 16Nonrestricted Sequential Counterbalancingis applicable to all designs reviewed in the previous section,
regardless of their structural properties or the relationship between the number of treatments and the number of periods
per sequence. In Table 3.9, we used a single replication of a standard design with three bricks (MOLS for four treatments)
to illustrate Method 16. Method 16 could also be applied to single-brick designs or to several replications of the same
standard design consisting of either one or multiple bricks, with the understanding that each different sequence included in
the whole design is assigned to the same number of subjects. Method 16 is the default method for sequential
counterbalancing in cross-over designs, and its distinctive feature is that all sequences are considered a unique set of
sequences. There is no place to create groups of experimental units on the basis of their structural characteristics (i.e.,
classificatory factors; see Table 1.1), and the only procedural constraint is that the number of experimental units equals
the number of sequences in the whole design. It is for this reasonthe absence of restrictions to the randomization
imposed by between-subjects blocking structuresthat we call Method 16 Nonrestricted Sequential Counterbalancing.
In some circumstances, either for theoretical reasons (e.g., studying the moderator effects of classificatory factors) or for
local constraints (e.g., patients coming from different hospitals), it is advisable or even required to split the available
experimental units into groups defined by one or more structural characteristics (e.g., gender, educational level,
personality traits, diagnostic categories, etc.). In these circumstances, the random assignment of sequences of
treatments is done within each group of experimental units, that is, the randomization is restricted by between-subjects
blocking structures. Ideally, the same sequences of treatments should be administered to all groups (Method 17:
Restricted Sequential Counterbalancing: The Same Sequences per Group). This is easily done by replicating the standard
design and giving one or more replications to each group of experimental units. When this is not feasible, different subsets
of sequences can be randomly assigned to different groups of experimental units (Method 18: Restricted Sequential
Counterbalancing: Different Sequences per Group), keeping the desired structural properties in the whole design but
missing these properties within each subset of sequences. In any case, the number of groups must be consonant with the
admissible partitions of the whole cross-over design, and the number of experimental units in each group must equal the
number of sequences per partition. Table 3.14 summarizes the acceptable methods of sequential counterbalancing
according to the number of replications and the possible partitions of the standard design.
3.6.1 The Same Sequences per Group
Table 3.14 Replications and Partitions of Cross-Over Standard Designs: Acceptable Methods of Sequential Counterbalancing
As the reader must have already realized, Method 17Restricted Sequential Counterbalancing: The Same Sequences
per Groupis a simple extension of Method 16 intended to accommodate between-subjects blocking structures
introduced by some meaningful classificatory factor (see Table 1.1). Method 17 is fully illustrated in Table 3.15, in which a
four-sided Williams's square is replicated three times (Step 2), and the random assignment of sequences is done within
each group of four experimental units (Step 3). The randomly selected sequences in Step 1 are identical for the three
groups (Step 3), and the whole design is a
Table 3.15 Restricted Sequential CounterbalancingThe Same Sequences per Group: Random Assignment of Sequences of
Four Treatments (A, B, C, and D; Three Replications of the Williams's Design; See Table A3.5) to Three Groups of Four Units
(Design 6G4; Method 17)
balanced design (Design 6G4) with 1 = 3. For each group taken alone, the design remains a balanced design with 1 =
1. This is particularly relevant when experimental attrition or some unexpected technical mistake prevents the
experimenter from obtaining complete data for one or
Table 3.16 Restricted Sequential CounterbalancingDifferent Sequences per Group: Random Assignment of Sequences of Four
Treatments (A, B, C, and D; Three Mutually Orthogonal Latin SquaresMOLS 4; See Table A3.3) to Three Groups of Four Units
(Design 6F4; Method 18)
more groups. In these cases, the data of intact groups can be analyzed separately, ensuring the degree of balance
initially intended.
3.6.2 Different Sequences per Group
Contrary to Method 17, which requires at least two replications of a single-brick or a multiple-bricks standard design (see
Table 3.14), Method 18Restricted Sequential Counterbalancing: Different Sequences per Groupapplies only to
standard designs consisting of two or more distinct subsets of sequences (bricks) and can be used with a single
replication of the selected design. In this case, as illustrated in Table 3.16 with a standard design with three bricks (MOLS
for four treatments), each different brick is randomly assigned to a different group of experimental units categorized by
some important classificatory factor. The whole design, with 12 sequences, is a balanced design (1 = 3), but no balance
is achieved within each brick (e.g., in Brick 3, randomly assigned to Group 2, Treatment A is immediately preceded twice
by C, once by B, and never by D; see Table 3.16). The best one can say is that each brick (Latin square) taken by itself is
a uniform design that could be obtained by Method 15Positional Counterbalancing (see Section 3.4).
In practice, the main difference between Methods 17 and 18 (see Tables 3.4, 3.15, and 3.16) is that Method 18 includes a
specific step (Step 3 in Table 3.16) to randomly assign the different subsets of sequences to the different groups defined
by the between-subjects blocking structures. Note that Method 18 could also be applied to several replications of multiple-
bricks standard designs. In any case, once the user specifies the number of replications of the standard design, the
number of admissible partitions in Step 2 of Table 3.16 as well as the number of groups of experimental units required for
the whole design are automatically defined: (a) they equal the number of bricks in the case of a single replication and (b)
they correspond to the product number of replications times number of bricks in the case of two or more replications (in
this case, some groups are administered identical subsets of sequences). Obviously, the number of experimental units
per group must equal the number of sequences per partition of the whole design.
As we have already stated, when the procedural constraints (number of groups created by blocking variables and number
of experimental units per group) are compatible with the whole design structure (number of subsets of sequences resulting
from replications and/or partitions of the standard design and number of sequences per subset), Method 17 is always the
best choice, and Method 18 should be used parsimoniously in restricted sequential counterbalancing. Nevertheless, there
are some important balanced designs whose randomization requires the mise en place of Method 18. To be more specific,
when the number of treatments is an odd number and some additional procedural constraints and randomization
restrictions are assembled (i.e., the total number of available experimental units is n = t(t 1), and these units can be
meaningfully categorized by some classificatory factor into (t 1)/2 equal-size groups), locally balanced designs
(Anderson & Preece, 2002) are particularly useful.
Locally balanced designs are built by adding vertically (t 1)/2 different pairs of Williams's squares (double carryover
designs; see Subsection 3.5.1). In the randomization phase (Step 3 of Method 18; see Table 3.16), each pair of Williams's
squares is randomly assigned to one of the previously formed groups of homogeneous experimental units. Besides the
achieved local balance within each pair of Williams's squares and, obviously, on the whole design, locally balanced
designs have the additional structural properties summarized in Table 3.17, in which the standard forms of designs with
five (Design 6K5) and seven (Design 6K7) treatments are depicted. Property (b) is specially relevant in the situations
where subjects unexpectedly drop out from the experiment in the last period(s), because the data for the first (t + 1)/2
periods can be analyzed as if they had been obtained in an Incomplete Balanced Design (Design 6O; see Subsection
3.5.2). Please note that the locally balanced design for t = 3 is simply the Williams's design shown in Table 3.2 (Design
6H3). For an odd number of treatments higher than seven and for details on the construction of locally balanced designs,
refer to the original work of Anderson and Preece (2002).
3.7 Factorial Designs
From Section 3.3 to Section 3.6, we explained and illustrated randomization methods applied to cross-over designs in
which only one experimental factor is manipulated (Designs 6A-6T, according to the classification scheme presented in
Table 3.3). As the reader has been warned at the end of Section 3.2, for the sake of simplicity and in consonance with the
current practice in cross-over designs literature, we have used the term treatment(s) instead of treatment level(s), with the
understanding that we have been referring to the different levels of a treatment and not to the treatment itself. In this final
section, the randomization of true factorial cross-over designs (Designs 6U and 6 V in Table 3.3) is discussed, and we
return to the terminology used throughout Chapter 2 and Sections 3.1 and 3.2. Designs consisting of multiple within-
subjects factors (treatments), each with two or more treatment levels, are covered in the first subsection; designs
containing simultaneously between-subjects and within-subjects treatments are addressed in the second subsection.
Table 3.17 Standard Forms of Locally Balanced Designs for Five and Seven Treatments
3.7.1 Only Within-Subjects Treatments
In some disciplines, it is not uncommon to see researchers manipulating simultaneously two or more within-subjects
experimental factors. For example, in classical experimental psychology, subjects can be faced with stimuli differing
consistently in two independent dimensions; in cognitive social psychology, manipulations of physiological arousal can be
coupled with distinct instructional definitions of situations or social stimuli; in the health and pharmacological sciences,
the conjoint effects of two drugs, each one given at different doses, can be the main aim of the experiment; finally and
among other possible examples, in market research, consumers may be asked to evaluate similar products coming from
different brands with alternative forms.
The examples given above are focused on the simplest case of factorial cross-over designs, that is, a design including
only two treatments: W (e.g., Drug A in the third example and Brand Name in the fourth example) and X (e.g., Drug B in
the third example and Form of Product in the fourth example). If each treatment has only two treatment levels (e.g., low
and high doses for Drugs A and B in the third example, established vs. new for Brand Name, and solid vs. liquid for Form
of Product in the fourth example), we get a factorial cross-over design with four experimental conditions or treatment level
combinations: W1X1, W1X2, W2X1, and W2X2 (see Table 1.5 for the notational system used throughout this book). In this
situation, it is tempting to rely on one of the four periods designs presented in the previous sections and, according to the
assumptions about carryover effects, choose a suitable design for Random (Method 14), Positional (Method 15), or
Sequential (Methods 16, 17, or 18) Counterbalancing, depending on the characteristics of the selected design and/or the
absence/presence of between-subjects blocking structures (see Table 3.14). Randomization could be easily done by
replacing the symbols A, B, C, and D of the final design with treatment level combinations W1X1, W1X2, W2X1, and
W2X2, respectively. This procedure can actually be done, and it is absolutely correct (see, e.g., Fletcher & John's, 1985,
utilization of 8 8 Latin squares in the construction of a 2 4 factorial cross-over design). However, as the number of
treatments or the number of levels per treatment increases, the number of required periods (extension of sequences) for
accommodating all the possible treatment level combinations quickly becomes excessive. At this point, the best solution
is to skip complete-sequences designs and rely on designs with fewer periods than treatment level combinations but
having a well-known factorial structure and a specific degree of balance.
Generalized cyclic designs (GC/n), introduced by John (1973) and developed in subsequent articles (e.g., Fletcher, 1987;
Fletcher & John, 1985), are special cases of incomplete block designs (see Chapter 2, Subsection 2.3.3) exhibiting what
has been called a factorial (cross-over) structure: The different treatments (factors) are orthogonal to each other and each
treatment (factor) is orthogonal to the carryover effects belonging to the remaining treatments (factors) (see Fletcher &
John, 1985, for a more formal definition of factorial structure, and Jones & Kenward, 2003, for a brief review of the literature
on GC/n and similar designs). In Table 3.18, a simple procedure for the construction of a generalized cyclic design is
illustrated with the original example given by John (1973) for a 4 3 factorial design, using p = 4 periods (Design 6U12). In
this design, there are two within-subjects treatments (W and X), the first one with four levels (W1, W2, W3, and W4) and
the second with three levels (X1, X2, and X3), giving rise to 12 treatment level combinations (W1X1, W1X2, W1X3, ,
W4X1, W4X2, and W4X3). Each different combination is administered in all periods (uniformity on periods), and a block of
four treatment level combinations is given to each subject (see Form B of Design 6U12 in Table 3.18).
For design construction purposes, it is more convenient to rely on a slightly different notational system (see Form A of
Design 6U12 in Table 3.18), in which the sequences of treatment level combinations are denoted by series of p pairs of
digits (in each pair, the first digit indicates the levels of the first treatment, varying from 0 to tW-1, and the second digit
represents the levels of the second treatment, varying from 0 to tX-1). Beginning with a predefined generating sequence (in
the example, the generating sequence is the sequence 00, 10, 21, and 32 in the first block), in each period the levels for
Treatment W are obtained by cycling the first digit of the corresponding pair of digits in the generating sequence one step
(i.e., adding 1 to the previous treatment level and reducing modulus tW = 4 when necessary). Also in each period, the
levels for Treatment X are obtained by repeating tW times the second digit of the pair and by cycling this digit one step
the next tW times, reducing modulus tX = 3 when necessary. Adopting our notational system, the 0 to tW-1 levels for
Treatment W are converted into 1 to tW levels (0 = W1, 1 = W2, 2 = W3, and 3 = W4), and the 0 to tX-1 levels for
Treatment X are converted into 1 to tX levels (0 = X1, 1 = X2, and 2 = X3). As the reader could verify by himself, the design
obtained has a factorial structure, in which (a) Treatments W and X are orthogonal, (b) Treatment W is orthogonal to the
carryover effects of Treatment X, and (c) Treatment X is orthogonal to the carryover effects of Treatment W (see Table
3.18).
Once Design 6U12 has been generated, 12 experimental units are required, and the randomization is done via Method 16
Nonrestricted Sequential Counterbalancing. If the whole design is replicated and there are meaningful between-subjects
blocking structures, Method 17Restricted Sequential Counterbalancing: The Same Sequences per Groupcould also
be applied. When the whole design includes multiple bricks (as in the case of Design 6U6, whose Form A is also depicted
in Table 3.18) and a between-subjects blocking structure is present, the researcher must apply Method 18Restricted
Sequential Counterbalancing: Different Sequences per Group (see Tables 3.4 and 3.14).
Table 3.18 Construction of Generalized Cyclic Designs With Two Treatments and Four Periods: Examples of One-Brick (Design
6U12) and Two-
Fletcher (1987, pp. 653-654) gives an extended list of generating sequences for two (2 2, 2 3, 2 4, 3 3, and 4 4)
and three (2 2 2 and 3 3 3) treatments generalized cyclic designs, including, for each design, variants with
different numbers of periods (p) and experimental units (n). The reader can take these generating sequences to build a
specific design, relying on the procedure explained in the previous paragraphs and illustrated in Table 3.18 with two
treatments generalized cyclic designs. In the case of the three treatments (e.g., W, X, and Y) designs, the procedure for
design construction can be easily extended. For example, suppose you have a 3 2 2 design, with p = 6 and n = 12,
and the generating sequence is 001, 201, 010, 100, 110, and 211. Taking the 3-tuple in the second period (201), the
treatment level combinations for the 12 blocks in this period would be obtained: (a) for Treatment W levels, by cycling the
first digit one step (reducing modulus tW = 3 when necessary); (b) for Treatment X levels, by repeating tW = 3 times the
second digit and cycling this digit one step the next tW times (reducing modulus tX = 2 when necessary); and (c) for
Treatment Y levels, by repeating tWtX = 6 times the third digit and cycling this digit one step the next tWtX times
(reducing modulus tY = 2 when necessary). The 12 treatment level combinations for the second period would be 201, 001,
101, 211, 011, 111, 200, 000, 100, 210, 010, and 110. Replacing this notation with our notational system, these
combinations would be read as W3X1Y2, W1X1Y2, W2X1Y2, , W3X2Y1, W1X2Y1, and W2X2Y1. Of course, the
procedure should be applied to the generating 3-tuples for the remaining five periods.
Generalized cyclic designs can be used in studies that have treatments with either an equal or a different number of
treatment levels. When the number of treatment levels is the same for all the treatments involved in the study and the
number of periods equals the number of treatment levels, factorial cross-over designs can be alternatively built on the
basis of special types of MOLS, consisting of pairs of nearly column-complete Latin squares. Bate and Boxall (2008),
using Williams's (1949) approach to MOLS construction, give a catalog of factorial cross-over designs for three up to nine
treatment levels per experimental factor, including the particular case of six treatment levels for which no MOLS exist. In
Table 3.19, we give an example of a 5 x 5 design (Design 6U25), referring the reader to the original source for theoretical
and technical details on the construction and utilization of these and other designs.
3.7.2 At Least One between-Subjects Treatment
To conclude our presentation on randomization methods in experimental design, we must consider designs having both
between and within-subjects manipulations. The simplest design of this kind would be a design with two treatments: one
between-subjects factor (TA) and one within-subjects factor (W). Design randomization would be a two-step procedure. In
the first step, the available experimental units would be randomly assigned to the levels of TA (TA1, TA2, TA3, ) using
one of the simple random assignment (SRA) methods described in Subsection 2.2.1, even though Method 3Forced
Equal Sizeswould be the most appropriate choice. In the second step, a counterbalancing scheme would be applied for
determination of the order in which the levels of W (W1, W2, W3, ) would be administered to each experimental unit.
This design has been popularized under the name of
Table 3.19 Construction of Factorial Cross-Over Designs From NC MOLS and Random Assignment of Treatment Levels to
Experimental Conditions: Example of a Design (Design 6U25) With Two Treatments (W and X) and Five Levels per Treatment
mixed design (e.g., Lindquist, 1953), but we avoid this designation because in statistics the term mixed design refers
specifically to any design that includes both random- and fixed-effects factors, regardless of whether they are between-
subjects or within-subjects (Maxwell & Delaney, 2004, p. 593). From a structural perspective, the design is similar to the
Split-Plot Factorial Design (Design 10) discussed in Chapter 2 (see Table 2.11), with the understanding that each block is
formed by a single experimental unit.
Theoretically it is possible to combine the structural features of many between-subjects designs classified in Figure 1.3
with counterbalancing schemes for one or more within-subjects factors (a very early systematization of these designs is
given by Lindquist, 1953). However, with the increase in the number of factors and factor levels, the designs quickly
become statistically intractable, not to mention the procedural constraints imposed on randomization.
Except for the combination of Random Counterbalancing (Method 14) with one of the four SRA methods (Methods 1-4),
the two-step procedure described above must be reversed; that is, the optimal randomization strategy has to begin by
choosing the counterbalancing scheme (positional or sequential counterbalancing) for the within-subjects factor and to
replicate the standard design one or more times within each level (or treatment level combination) of the between-subjects
factor(s). On the understanding that Method 3 is used to randomly assign the participants to the experimental conditions
defined by the between-subjects treatment design, the number of experimental units required for the whole design must
equal the number of sequences in the standard cross-over design times the number of replications of this design per
between-subjects experimental condition times the number of between-subjects experimental conditions. If Positional
Counterbalancing (Method 15) is chosen for the within-subjects factor, the number of sequences in the standard cross-
over design will equal the order of the selected standard Latin square. If sequential counterbalancing is the researcher's
option, what we said about between-subjects blocking structures in Section 3.6 can then be applied, mutatis mutandis, to
the experimental groups created via Method 3. To be more precise, the randomization of designs with both between- and
within-subjects factors is a special case of Methods 17 and 18, bearing in mind that the restrictive features of these
methods apply to randomized groups of participants and not to groups defined by the intrinsic characteristics of available
experimental units. Whenever possible, Method 17 must be used instead of Method 18, which is to say that the
sequences for the levels of the within-subjects factor are the same in all randomized groups.
http://dx.doi.org/10.4135/9781452270012.n3
18/1/2014 Chapter 4: Validity Issues, Analysis Guidelines, and Reporting Standards : SAGE Research Methods
Chapter 4: Validity Issues, Analysis Guidelines, and
Reporting Standards
In the previous chapters, the methods of randomization in between-subjects and within-subjects designs were intentionally
presented without any mention of statistical software for random assignment or counterbalancing procedures. As the
reader was informed in the Preface, the 18 methods described and illustrated throughout Chapters 2 and 3 could be
rigorously implemented by relying on the materials given in the appendices and using the common features (e.g., copy
and paste, replace, tables, and sort commands) included in any word processor software. This deliberate pedagogical
strategy was planned to be in agreement with the user-friendly philosophy underlying the SAGE Quantitative
Applications in the Social Sciences Series since the very first title in the mid-1970s, as Lewis-Beck (2008) has recently
reminded us.
However, once the reader has mastered the critical features of local control and randomization procedures (i.e., the error
control design) and their intrinsic relationships with the associated design layout (i.e., the treatment design), either the
use of basic random number generators or the adoption of specialized software is, obviously, the ideal approach to
randomization in real-world experimental research. Therefore, in this final chapter and whenever it is convenient,
references to particular programs, computational routines, and randomization tools will be given to the reader. Needless to
say, statistical software is being continuously updated, and exact citations of program releases are condemned to perish
in a few months. So we invite the reader to regularly consult the software providers websites. A good place to begin with
is the directory of statistical software providers offered by STATA (http://www.stata.com/links/stat_software.html). For an
excellent overview of free online software, we recommend John C. Pezzullo's Interactive Statistics Pages
(http://statpages.org/). Finally, in the companion website of this monograph (http://www.sagepub.com/alferes), we will try
to keep the reader informed about both specific random assignment software and the randomization features included in
general-use statistical packages (e.g., BMDP, GenStat, IBM SPSS, R, SAS, STATA, STATISTICA, SYSTAT). As
mentioned in the Preface, IBM SPSS and R versions of the author's SCRAED program are free and available from the
companion website of this monograph, in which catalogs (e.g., TEXT and EXCEL files) of between-subjects and within-
subjects designs and statistical code for running the associated analysis in different statistical packages (e.g., IBM
SPSS) and computing environments (e.g., R) are also available.
In addition to these potentially useful tips on the selection and use of computer programs, this final chapter has three
main objectives: (1) to restate the crucial role of randomization procedures in the broader context of the validity of
scientific claims, (2) to establish the relevant connections between design and data analysis issues, and (3) to give the
reader a clear picture of the standards required for truthful communication of research findings, with special emphasis on
randomized experiments. A guide to the chapter contents is given in Figure 4.1, where the main features of randomized
experiments, from the planning stage to the final report, are summarized. Figure 4.1 could be misread as a tentative plan
for writing a comprehensive book on experimental design and data analysis. Yet there are excellent and authoritative
reference works in this field, so that the reader will be referred to them whenever it is opportune. Our goal is obviously
more modest, and this chapter is focused on the links between the core contents of the previous chapters and the
practice of conducting randomized experiments, with special attention to the inescapable connections between local and
generalized causal inferences and to the intersection points between experimental design features and data analysis
strategies. More specifically, Chapters 2 and 3 were centered in error control design (randomization plans) and its
interface with treatment design (design layouts), leaving out the interface with sampling and observational design (see
Figure 4.1). This interface is the main concern of Section 4.1. On the other hand, some theoretical and technical details of
randomization methods omitted in the two previous chapters are discussed in Section 4.2 along with general guidelines for
data analysis. The concluding section is a brief guide for reporting randomized experiments.
4.1 Planning and Monitoring Randomized Experiments
As it was asserted in Section 1.2, scientific hypotheses are conjectural statements about the relationships between
theoretical constructs, empirically represented by particular events or realizations. At the most abstract level, constructs
can be thought of as the main nodes in the nomological network specified by theoretical conjectures. The translation of
theoretical constructs into particular events (manipulations and measurement operations) is the first task an experimenter
must deal with by clearly formulating the research hypothesis. In the methodological literature, there are some nuances in
the use of the term hypothesis, ranging from an abstract proposition to a statistical hypothesis statement, as it can be
read in Kirk's (1995) definition of experimental design, quoted in Section 1.1. A good strategy to formulate a research
hypothesis is the perspectivist approach suggested by McGuire (2004b), in which the researcher begins to express the
hypothesis in six different modes: (1) natural language (e.g., frustration, operationalized by [], leads to aggression,
measured by []); (2) abstract symbolic
Figure 4.1 Conducting and reporting a randomized experiment
language (e.g., DV = f (IV1, IV2, IV1*IV2); DV = dependent variable, IV = independent variable); (3) pictorial representation
(e.g., graphical anticipation of the expected main results); (4) tabular depiction (e.g., the planned layout, cases/rows by
variables/columns in a statistical package data editor); (5) descriptive statistics (e.g., means, mean differences,
standardized effect sizes); and (6) inferential statistics (e.g., t tests, F tests, confidence intervals). The worksheets for
generating a program of research given in McGuire (2004a) can be an invaluable tool for beginners and even for
experienced researchers.
Once the research hypothesis has been formulated, the next step is to build the research design connecting the three
main components shown in Figure 4.1 and to anticipate all possible drawbacks in the empirical realization of the study by
a tentative answer to the following questions:
1. a. Who (the experimenter) will deliver the treatments and make the observations and/or collect the self-reports
(measurement operations), and which exactly will be the treatment levels or treatment level combinations to be
administered and the behavioral responses to be recorded (outcome measures)?
2. b. Where (laboratory or field settings), when (study temporal structure), and under what conditions (previous local
control and randomization, constancy of pseudofactors, treatment implementation, and manipulation check) will these
responses be recorded?
3. c. How many participants will be involved in the study (power analysis and study size), how will they be recruited and
enrolled (sampling strategy), and how will their intrinsic characteristics (classificatory factors/blocking
structures/dispositional moderators/covariates) be identified and/or measured?
4. d. How will participants be informed of the study objectives in advance (informed consent), and what will be the
protocol and the timing for the covariates measurement, local control, randomization, follow-up of treatments
implementation, data collection, and debriefing interview?
5. e. What theoretical and practical implications of the study can be anticipated?
In the next two subsections, we give the reader an overview of the interrelated activities involved in planning and
monitoring a randomized experiment addressing the questions above in the context of Campbell's threats approach to
the validity of causal inferences and the substantive assumptions underlying Rubin's causal model. Obviously, this is a
selective overview centered on the interfaces of error control designthe main subject of this monographand cannot be
read as an art of experimentation exposition. In the social sciences, one of the pioneering efforts to write essays on this
subject is documented by the chapters on field experiments (French, 1953) and laboratory experiments (Festinger, 1953)
included in the remarkable edited volume Research Methods in the Behavioral Sciences (Festinger & Katz, 1953),
associated with the creation of the Institute for Social Research at the University of Michigan in 1949. Aronson, one the
most prominent students of Festinger, has continued this tradition in his contributions to the Experimental Method
chapters (e.g., Aronson & Carlsmith, 1968; Wilson, Aronson, & Carlsmith, 2010) included in the successive editions of
the Handbook of Social Psychology, to which the reader is referred. An outstanding presentation and discussion of 28
classic social psychology experimentspublished between 1959 and 2001is given in Abelson, Frey, and Gregg
(2004).
4.1.1 External and Construct Validity Issues
Although Chapters 2 and 3 were essentially focused on strategies to ensure valid local inferences, in the introductory
chapter, we have conveniently highlighted the connections between the experimental and statistical components of
causality and the extrapolation (external validity) and representation (construct validity) features underlying generalized
causal inferences (see Figure 1.2). This subsection gives a brief review of these connections, with special emphasis on
the sampling strategies behind randomized experiments and some specific threats to the construct validity of
experimental treatments.
4.1.1.1 Sampling Strategies and External Validity
If there were no budgetary constraints or practical limitations in recruiting participants and obtaining their informed consent
and if the experimenter's goal is to simultaneously ensure the optimal conditions for both local inferences of causality and
generalization of causal relationships, randomized experiments would always be preceded by probabilistic sampling from
target populations. The structural characteristics of the target population would be synonymous with the eligibility criteria
for the randomized experiment, and the preliminary sampling procedure would be done by trying to minimize coverage,
sampling, and nonresponse errors (Groves, 1989; Groves et al., 2004).
4.1.1.1.1 The Ideal Model for Generalized Causal Inferences
This ideal model of generalized causal inference (see Figure 4.2) has been applied in some areas of research (e.g.,
split-ballot experiments in cross-sectional survey research, some of them focusing on the study of question formats and
the cognitive processes underlying response editing, or repeated measures designs embedded in longitudinal survey
research), but it is far from being the norm in social and behavioral experimental studies. In addition to financial matters
and practical reasons, there are other sound arguments in favor of experimentation without previous probabilistic sampling.
One extreme position defending external invalidity was taken by Mook (1983) on the grounds that the bulk of
experimental research is centered on the identification of causal processes and the appraisal of competing theories, and
therefore generalization to the real world is not the main concern. A classic example is Harlow's (1958) study on the
nature of love, in which psychoanalytic and behaviorist hypotheses about the secondary nature of the infant-mother
attachment were experimentally disconfirmed. Harlow's subjects were rhesus monkeys, and the experimental settings
comprised, among other devices, wire and cloth surrogate mothers and artificial living cages. There was no formal
sampling of the participants, and we could not easily imagine a situation more distant from the daily existence of normal
savage monkeys.
This lack of ecological validity is frequently a source of misunderstandings about randomized experiments among the
general public and even in some academic forums. As Aronson, Wilson, and Brewer (1998) consistently argue, the basic
requirements for an experiment are experimental realism (i.e., an impact and meaningful manipulation) and psychological
realism (i.e., a manipulation that triggers the relevant psychological processes) and not naive or mundane realism (i.e., a
true but nonpertinent similarity with people's everyday life).
4.1.1.1.2 Realistic Models for Generalized Causal Inferences
Having said this, we are not suggesting that external validity issues can be dismissed. On the contrary, the generalization
of causal relationships can and must be the central aim of the scientific enterprise. However, there are more realistic
ways to achieve this desideratum by circumventing the difficulties inherent to probabilistic sampling strategies. This is not
the place to describe these alternative ways, but we refer the reader to Shadish et al. (2002), where Cook's (1993, 2000)
model of generalized causal inference is fully explained. This approach, relying on special strategies of purposive or
intentional sampling (e.g., deliberately heterogeneous sampling, modal instance sampling), makes simultaneous use of
different principles of generalization (e.g., assessing superficial and deep similarities, making discriminations,
interpolations, and extrapolations) and is applied to both external and construct validity issues. In brief, realistic models
for generalized causal inference encapsulate the often quoted reply of Sir Ronald Fisher when asked how many
experiments would be necessary to make a certain point: Elaborate! was the elliptic answer.
Figure 4.2 Random sampling, random assignment, and validity of scientific inferences
4.1.1.1.3 Specific Threats to External Validity and the No Different Versions of Treatments Assumption
In the last update of Campbell's threats approach to the validity of scientific inferences (Shadish et al., 2002), external
validity is equated not just with the extrapolation of causal relationships to target populations but also with the
extrapolation of these relationships to other operationalizations of treatments and outcome measures, as well as to
different physical and social settings (see Figure 2.1). Among the potential threats to external validity (interactions of the
causal relationship with units, treatment variations, outcomes, and settings, and context-dependent mediation), the
second one is particularly relevant in the context of this monograph given that the variations in treatment implementation
or partial implementation of treatments encompass to some extent the first component of Rubin's SUTVA, which is the
no different versions of treatments assumption. We will return to this point in the discussion of construct validity issues.
4.1.1.1.4 Assessing Eligibility and Participants Enrollment
With non-probabilistic intentional or convenience sampling, the first activity an experimenter is obliged to do is to clearly
define and check the eligibility criteria for the available participants. This is also the initial step in the flowchart embedded
in Figure 4.1, which is an adaptation of the diagrams inserted in the successive versions of the Consort Statement (see
Section 4.3). This chart will be used in Subsection 4.1.2 to organize the review of statistical conclusion and internal
validity issues. The panel regarding classificatory factors in Table 1.1 can be seen as a reminder of the typical eligibility
criteria used in experimental research. Informed consent matters will be briefly discussed in connection with the
postexperimental interview in the final paragraph of Subsection 4.1.2.
4.1.1.1.5 Random Assignment and Random Sampling
We should conclude these short notes on external validity issues by noting that SRA can be thought of as a special case
of simple random sampling, wherein multiple random samples (e.g., groups of experimental units assigned to different
treatments) are drawn without reposition from a population consisting of all available experimental units. The number of
samples equals the number of treatment levels (or treatment level combinations), and the sampling procedure ends when
there are no more units to sample. In forced-size methods (Methods 3 and 4), the sample size (i.e., the number of
experimental units per treatment) is fixed by the researcher; in a priori probabilities methods (Methods 1 and 2), the
sample size varies around the theoretical probabilities of assignment previously specified by the experimenter. By the
same token, blocked random assignment (Methods 7-10 and the two variants of SRAMethods 5 and 6) and stratified
random assignment (Methods 11 and 12) are special cases of stratified random sampling, in which multiple random
samples are drawn without reposition within each stratum of the population (i.e., all the available experimental units). To
extend the analogy, the so-called cluster assignmentwhere entire groups of subjects are randomly assigned to
experimental conditionsis a special case of cluster sampling, in which all the identified clusters are assigned to
experimental treatments. A very readable introduction to random sampling is Kalton (1983). Dattalo (2010) gives an
overview of the methodological and statistical alternatives to random sampling and random assignment. For reference
works on sampling strategies, the reader may consult the classics of Cochran (1977) and Kish (1965) or more recent
sources, such as Lohr (2010) or Thompson (2002).
4.1.1.2 Treatments, Outcome Measures, and Construct Validity
Ensuring the construct validity of causal inferences is perhaps the most demanding task a researcher must deal with. In a
certain sense, definitions of constructs can be always replaced with definitions of definitions, in a never-ending regressive
process. Putting aside the neo-positivist illusion of single and unequivocal operational definitions, Cook and Campbell
(1979; see also Campbell & Fiske, 1959; Cook, 1985) defend the path of multiple operationalizations and the assessment
of convergent and discriminant validity of the different measures and manipulations of scientific constructs. In the
systematization presented by Shadish et al. (2002), the threats to construct validity can be divided into two broad
categories: (1) general threats, which affect simultaneously all the structural elements of a research design (see Threats
1-5 in Figure 4.3), and (2) specific threats, which can be more directly related to treatments, settings, units, or
observations (see Threats 6-14 in Figure 4.3).
In Figure 4.3, we also sketch out the main contributions of Rosenthal (1976) to the study of experimental artifacts, with
particular emphasis on experimenter effects and the psychological processes underlying participants behavior in an
experimental situation. It is not our intention to make a detailed explanation of the very nature of all these artifacts, and for
that the reader is referred to the classic volumes edited by Rosenthal and Rosnow (1969) and Miller (1972) and to the
more recent synthesis by Rosnow and Rosenthal (1997). However, some specific threats and artifacts strictly concerning
the main subject of this monograph will be considered in the next few paragraphs.
4.1.1.2.1 Specific Threats to Construct Validity and the No Different Versions of Treatments Assumption
First, the no different versions of treatments component of Rubin's SUTVA can be more clearly understood in the context
of the multiple events that may introduce some kind of confounding in the treatment implementation. More precisely, and
particularly in the social and behavioral sciences, in which the treatments may consist of complex and prolonged
manipulations, the experimenter's expectancies, as well as his or her more or less unconscious actions intended to
compensate for the less desirable effects, can result in different operationalizations of the same treatment. This is the
reason why blind studies and randomization concealment are highly recommended standards in all best practices
sources (see Section 4.3). The first expression refers to the practice of masking the research hypothesis and the very
nature of the experimental conditions from those who administer the treatments and/or collect the outcome measures,
and obviously also from the participants. The second one concerns the procedure of keeping the actual random
assignment unknown until the moment of treatment implementation.
4.1.1.2.2 Specific Threats to Construct Validity and the No Interference between Units Assumption
Second, and as explicitly shown in Figure 4.3, participants behavior in a randomized experiment must be conceived as a
special instance of general behavior, in which the outcome responses are also determined by situational demand
characteristics and the participants can
Figure 4.3 Threats to construct validity: Synthesis of Shadish, Cook, and Campbell's (2002) approach (Threats 1-14; Threats 8a
and 8b are generically named reactivity to the experimental situation) and Rosnow and Rosenthal's (1997) framework (Artifacts A-
M)
actively (re)define the experimental situations in a manner that is not coincident with the intended operational definition of
the focal treatment. In addition to the cognitive and affective processes underlying participants decisions and actions in a
randomized experiment, the possibility of social interactions between participants assigned to different treatments should
also be taken into account, especially when the settings, the nature, and the duration of the manipulation favor the
occurrence of these interactions, as is the case with social and educational programs and prolonged medical or
psychological interventions. The component of Rubin's SUTVA labeled no interference between experimental units is a
general assumption that encompasses some of the threats to construct validity, namely, treatment diffusion and the
opposite behaviors encapsulated in the Campbellian expressions compensatory rivalry and resentful demoralization (see
Threats 12-14 in Figure 4.3; for a detailed explanation of the many procedures to prevent artifacts and potential threats to
the construct validity of causes and effects in experimental research, the reader is referred to Rosnow & Rosenthal, 1997;
Shadish et al., 2002).
4.1.2 Statistical Conclusion and Internal Validity Issues
Given that local control and randomization procedures were conceptualized as the essential devices to ensure statistical
conclusion and internal validity of causal inferences, we have thoroughly presented their rationale in the introductory
sections of Chapters 2 and 3. Consequently, here we will be very briefly focusing on some issues omitted in Sections 2.1
and 3.1. In passing, it should be noted that in the very first systematizations of the Cambpellian approach (Campbell,
1957, 1969; Campbell & Stanley, 1966), statistical conclusion and internal validity were subsumed under the label internal
validity, and the main threats to internal validity could be broadly defined as the shortcomings ruled out by a proper
randomization. We echoed this initial conceptualization when representing the complementarity of the statistical and
experimental components of valid local molar causal inferences in Figure 1.2.
4.1.2.1 Power Analysis, Covariance Assessment and Testing, and Statistical Conclusion Validity
Demonstrating covariance between treatments and outcome measures is the preliminary condition to infer causality.
Besides the unreliability of measures or treatment implementations and the random variance associated with settings and
experimental units, covariance estimation and testing could be affected by the group of threats to statistical conclusion
validity depicted at the top of Figure 2.1. Here, we will resume these threats with a special emphasis on power issues and
the strategies for study size planning.
4.1.2.1.1 Study Size and Power Analysis
A recurring question among students and novice researchers is the famous N question: Now that the experimental
design has been built and the variables operationalized, how many participants do I need to test my research hypothesis?
To frame the question more correctly, how many participants (or observations) per experimental condition are needed?
Unfortunately, this question cannot be answered directly, since the solution requires that the researcher is able to give the
response to three preliminary questions: (1) What is the maximum risk or margin of error that he or she is willing to incur
by rejecting the null hypothesis when it is true? (2) What is the admissible margin of error if he or she does not reject the
null hypothesis when it is false? (3) Is he or she competent enough to provide an estimate of the expected experimental
effect?
With the help of relatively simple formulas and tables (Cohen, 1988), an adequate answer to these preliminary questions
allows the researcher to determine the required number of participants (or observations) per experimental condition. The
answer to Question 1 is to establish the probability for the so-called Type I error (a), which is the highest margin of error
that the research can accept when declaring that the statistical hypothesis under investigation (i.e., the null hypothesis)
has been rejected. Please note that in statistical terms, the hypothesis under investigation is the null hypothesis (there
is no covariation between the treatment and the outcome response, which is to say that there are no differences between
the outcome means of different experimental conditions). Stating things this way may confuse the less informed
researchers, who have come to believe that their aim is to prove the alternative hypothesis (i.e., there are differences
between the outcome means of different experimental conditions) and not to reject the null hypothesis. A substantive
response to Question 1 can be given by setting a at the conventional level of .05.
The answer to Question 2 connects directly with the so-called Type II error (b): the probability of declaring that there are
no differences between the outcome means of experimental conditions, when actually these differences exist (not
rejecting the null hypothesis when it is false). The complementary probability (1 ) of Type II error is called power or
sensibility and corresponds to the correct decision of rejecting the null hypothesis when it is false. In other words, power
is the probability of detecting differences between conditions if such differences really exist. In a similar way to what is
done with Type I error, Cohen (1988) proposes to set at the conventional level of .20 as the highest margin for Type II
error, which would mean that the researcher has at least 80 chances in 100 (1 .80) to obtain significant results if
they actually exist.
Finally, to answer correctly Question 3, the researcher must rely on the results of previously published studies or define in
advance what is the minimal acceptable difference between experimental conditions (i.e., the magnitude or size of the
experimental effect), based on criteria of theoretical or practical relevance. However, the absence of studies involving the
same variables or insufficient information to clearly define those criteria makes the answer to Question 3 problematic.
We are not giving here the formulas to carry out prospective or a priori power analysis, but the reader can consult the
classic source already cited (Cohen, 1988) or some more recent reviews of power issues in experimental design (e.g.,
Kelley & Rausch, 2006; Murphy, 2002). The main general-use statistical packages (e.g., R, SAS, STATISTICA, SYSTAT)
include specific modules and routines for doing prospective power analysis for t tests, and ANOVA and ANCOVA designs.
A very recommendable stand-alone commercial package is BioStat's Power and Precision (a similar version, named
Sample Power, is sold in conjunction with the IBM SPSS statistical program). Among the free software, our choice goes
to G*Power (Faul, Erdfelder, Lang, & Buchner, 2007) and to the Java Applets offered by Lenth (2006-2009).
Using these packages and statistical tools to calculate the number of experimental units per experimental condition in
advance is a relatively simple task, particularly with regard to the specification of and 1 values. For the specification
of effect sizes, the user typically has at his disposal two main options: (1) to provide independent values for expected
means (or mean differences) and residual variability or (2) to directly introduce the expected standardized effect size (e.g.,
Cohen's d for t tests or Cohen's f for effects in ANOVA and ANCOVA designs).
We conclude these notes on power with three final remarks. First, we side with Lenth (2001) when he advocates that in
prospective power analysis it is better to gather sound information from pilot studies than to specify more or less arbitrary
small, medium, or large effect sizes (i.e., values of .20, .50, and .80 for the d statistic or values of .10, .25, and .40 for
the f statistic). Second, when using prospective power analysis in conjunction with Methods 7 to 10 (see Table 1.2) or
with Methods 15 to 18 (see Tables 1.3 and 3.4), the number of experimental units (or observations) per experimental
condition resulting from the formal analysis should be replaced with the next integer compatible with the procedural
constraints of the error control design (e.g., the number of participants per block must equal or be a multiple of the number
of experimental conditions) or with the intended structural properties of the chosen cross-over design. Finally, although we
recommend the regular use of prospective power analysis, we warn the reader against a posteriori or retrospective
analysis (some statistical packages routinely give values for observed power) and against the practice of relying on
observed power values to make substantive interpretations of the study results. As Maxwell and Delaney (2004) note,
Saying p > .05 is tantamount to saying that observed power is less than .5 (p. 125).
4.1.2.1.2 Effect Size, Statistical Testing, and Confidence Intervals Estimation
In the preceding discussion of prospective power analysis, we implicitly relied on the traditional hybrid formulation of the
null hypothesis significance testing (NHST). In spite of its ubiquity in published research, the NHST has always been the
subject of strong controversies, with sound arguments for and against (see the volume edited by Harlow, Mulaik, &
Steiger, 1997, and the review of Nickerson, 2000).
This matter is outside the scope of this monograph, but we should advise the reader to be critical with regard to common
misunderstandings (e.g., p values don't inform the researcher about the probability of the null hypothesis being true given
the obtained data, but on the contrary they correspond to the conditional probability of obtaining these or more extreme
data given that the null hypothesis is true) and to take into account the more informative procedure of calculating
confidence limits for mean differences. This shift from the NHST to the confidence intervals estimation for theoretically
relevant statistical contrasts and to the appraisal of effect sizes is more visible in the health sciences than in the social
and behavioral sciences (Fidler, 2011), despite the institutional or quasi-institutional recommendations (e.g., Wilkinson &
Task Force on Statistical Inference, 1999). By saying this, we are not proposing that NHST should be banned, but we are
rather endorsing its proper use (see Abelson, 1997; Jones & Tukey, 2000) in conjunction with the substantive evaluation of
effect sizes.
4.1.2.2 Flow of Participants through the Experiment and Internal Validity
In the middle panel of Figure 4.1, on the left of the chart showing the flow of participants, we listed the main activities of
researchers when conducting a randomized experiment. Eligibility matters and sampling strategies were briefly discussed
in Subsection 4.1.1, and prospective power analysis was approached in the opening paragraphs of this subsection. Here,
we resume our selective overview of researchers activities, beginning with the collection of biographic information about
enrolled participants and ending with the ethical issues underlying randomized experiments.
4.1.2.2.1 Biographic Information, Classificatory Factors, and Covariates Measurement
Collecting all relevant biographic information and recording the values of major classificatory factors (e.g., age, gender, and
ethnicity) are tasks intertwined with the assessment of participants eligibility for a randomized experiment. Depending on
the specifications in the error control design, once the participants meeting the eligibility criteria are enrolled and their
informed consent obtained, the experimenters may have to get additional data about classificatory factors and to proceed
with covariates measurement. These covariates could be either used as blocking/stratifying variables in the context of
local control procedures or statistically handled in the data analysis stage.
From a substantive point of view, we refer the reader to Lazarsfeld (1959) on the conceptualization and classification of
dispositional variables (generality vs. specificity, degree of directiveness/dynamics, and time perspective) in social and
behavioral research and also to the discussion given in Rosenberg (1979); another important classic is Campbell's (1963)
systematic review of acquired behavioral dispositions. For more contemporary approaches to dispositional variables in
personality psychology, the reader may consult the synthesis presented in McCrae and Costa (1999).
4.1.2.2.2 Allocation Frames, Local Control, and Randomization
The next step is to build the allocation frame and to implement one of the methods of randomization or counterbalancing
fully described and illustrated in Chapters 2 and 3. As already stated, this can be done with the same resources we have
used throughout this monograph or with the help of RGN and specialized software.
A Google search with the expression random number generator gives more than 4,250,000 results, so the reader has
certainly many accessible choices. The vast majority of RGN, either available on the Internet or incorporated in statistical
spreadsheets and statistical software, produce pseudo-random numbers on the basis of some existing algorithm (see
Appendix 1). Provided that a uniform probability distribution is chosen, the starting point is arbitrarily set (e.g., current
date and time expressed in milliseconds), and the algorithm is accurate, these pseudo-RNG are good enough for
experimental design purposes. If the reader wants true random numbers, he or she can obtain them from websites such
as http://RANDOM.ORG (http://www.random.org/), from which we got the series included in Table A1.1. As
http://RANDOM.ORG states, the randomness of these numbers comes from atmospheric noise, and no predictable
pattern can be identified in the obtained series.
Virtually all statistical packages include RNG, and most of them also include the so-called design of experiments module
accessible from their graphic interfaces (e.g., SAS, STATISTICA, SYSTAT). One problem with these modules is that they
allow the randomization of many complex designs employed in engineering, industry, and technological research (e.g.,
fractional factorial designs with a high number of treatments, such as Taguchi designs, or response surface models, such
as Box-Behnken designs), but some basic designs used in the social, behavioral, and health sciences are not available,
as is the case with many restrictedly randomized between-subjects designs or counterbalancing schemes for cross-over
designs. A very recommendable choice that allows the implementation of most of the between-subjects and within-
subjects randomization methods described in this monograph is the Design module of GenStat. However, for some
designs the best solution is to rely on the programming features of available software and to build specific routines for the
intended randomization procedure. In fact, this is the solution we have adopted in SCRAED, a program that includes the
18 randomization methods explained in Chapters 2 and 3 and whose IBM SPSS and R versions can be freely downloaded
from the companion website of this monograph.
4.1.2.2.3 Treatment Implementation, Manipulation Check, and Compliance
As stated in all reporting standards (see Section 4.3), information about the settings, contents, and duration of
experimental manipulations is a crucial feature in the assessment of the internal validity of causal inferences. From a
practical point of view, it is convenient to make a clear distinction between randomized experiments in which the time
elapsed between the beginning of treatment delivery and the measurement of the dependent variable(s) is relatively short
(i.e., about 30-40 minutes to 2-3 hours) and randomized experiments in which the modus operandi is extended over
several days or even months or years. Although not necessarily so, experiments of the first type are generally conducted
in laboratory settings, and the researchers have a high degree of control over participants and nuisance situational
variables (pseudofactors). In these circumstances, pseudofactors are controlled by holding them constant in all
experimental conditions, and a relatively brief manipulation check (i.e., collecting a measure that is known to be sensitive
to the focal treatment) can be done immediately after the treatment delivery. Even in the absence of manipulation checks,
researchers can easily verify to what extent participants comply with the assigned experimental condition.
Experiments of the second type, that is, long-duration experiments, are typically (yet not inevitably) conducted in field
settings, and in many instances treatment exposition occurs without the researchers supervision, as is the case with
medical prescriptions or some important components of social and educational interventions. In these situations, a well-
planned follow-up is a critical device to evaluate whether participants comply with the assigned treatment. Treatment
noncompliance or compliance with an unassigned treatment breaks the very nature of a randomized experiment. It is also
a violation of Rubin's SUTVA and must be directly addressed in the data analysis stage (see Section 4.2).
4.1.2.2.4 Attrition, Measurement Operations, and Internal Validity
In the Campbellian framework, experimental mortality or attrition is a threat to the internal validity of causal inferences
and is broadly defined as differential loss of respondents from the comparison groups (Campbell & Stanley, 1966, p. 5) or
loss of respondents to treatment or to measurement (Shadish et al., 2002, p. 55). In Section 2.1, we explicitly
associated attrition with the Units box (see Figure 2.1), but it could as well be simultaneously associated with the
Treatment and the Observations boxes, implying both treatment noncompliance and the absence of outcome measures
from randomly assigned participants. It is also theoretically conceivable that a participant has fully complied with the
treatment, although he or she has been absent from the measurement session (see the hypothetical example in Table
4.1, in which the reverse situation is also shown: collecting outcome measures from noncompliers). In practice, given that
there is no record of the outcome measure, the information about compliance with treatment becomes irrelevant (i.e., the
participant must necessarily be ignored in the analysis stage).
A rather important issue is that attrition artifacts can more seriously affect between-subjects designs than within-subjects
or cross-over designs. In the first case, the comparability between experimental conditions introduced by random
assignment is derogated; in the second one, the worst scenario is the breaking of the intended degree of balance for the
chosen design.
4.1.2.2.5 Debriefing Interview and Ethical Issues
Ethical issues must be a major concern in all randomized experiments, beginning with obtaining participants informed
consent and ending with a warm and elucidative postexperimental or debriefing interview. In addition to the informative
features related to the current experiment, the postexperimental interview can be an excellent occasion to improve
participants knowledge about scientific methods of inquiry, with potential benefits to their personal lives. In the past four
decades, we have observed the emergence and consolidation of a powerful movement condemning all experiments
involving deceptive features. Providing that these features are really required for the appraisal of the research hypotheses
and don't cause irreversible damage to the participants, we don't side with this politically correct movement. Technical
illusions, to borrow Milgram's terminology (Evans, 1976, as cited in Korn, 1997, p. 104), are often irreplaceable devices,
and their use can be fully justified in the debriefing session. As a marginal note, we refer the reader to one of the rare
moments of self-disclosure in Festinger's (1980) writings, where he confesses that in many years of doing experiments
involving elaborate deceptions, I have never encountered an instance in which a subject felt harmed or ill used (p. 250).
For a well-informed history and appraisal of deceptive experimentation in social psychology, see Korn (1997).
As is obvious, ethical issues are not confined to deceptive experiments, and they are a matter of concern to all types of
scientific inquiry. For an extensive examination of the ethical aspects of quantitative research, we refer the reader to the
handbook recently edited by Panter and Sterba (2011), especially to the chapters focusing on randomized experiments
(Mark & Lenz-Watson, 2011), sample size planning (Maxwell & Kelley, 2011), and data analysis (Rosnow & Rosenthal,
2011). For a well-articulated defense of randomization procedures in scientific inquiry and a systematic discussion of
ethical and substantive objections to random assignment, the reader is referred to Cook and Payne (2002).
4.2 Analyzing Randomized Experiments
In a wonderful book about the proper use of statistics in scientific research, Abelson (1995) remarks that data analysis is
best described by the detective work, the narrative devices, and the rhetorical principles needed to organize a useful
argument from quantitative evidence (p. xiii). Relying on Tukey's (1977) philosophy of data analysis, Abelson elaborates
on the fundamental criteria for producing sound statistical arguments: indexing the magnitude (e.g., effect sizes) of
qualitative claims, articulating theoretically relevant comparisons (e.g., formulating specific directional hypotheses instead
of the general or omnibus null hypothesis and, whenever possible, specifying quantitatively the values to be nullified,
instead of the zero difference null hypothesis), and worrying about the generality, the interestingness, and the credibility
of research results.
As the reader was informed in the Preface and in the introductory paragraphs of this chapter, this is not a section on data
analysis but rather a brief guide to the relevant connections between design and statistical issues in randomized
experiments. The section begins with an overview of the probabilistic nature of random assignment mechanisms and ends
with a synthesis of the trade-offs between local and statistical control in experimental and quasi-experimental research. In
the two intermediate subsections, we give some guidelines for the analysis of intact and broken experiments.
4.2.1 Random Processes, Probabilities, and Assignment Mechanisms
At this point, we cannot resist quoting the best and shortest definition of randomization we have found: Randomization is
something that everybody says they do, but hardly anybody does properly (Crawley, 2005, p. 10). Perhaps because
reference to randomization procedures in published research is often limited to a single sentence, such as The
participants have been randomly assigned to the experimental conditions, students and novice researchers tend to
underestimate the true role of randomization in causal inference. We hope that the core contents of this monograph will
help the reader to improve this state of affairs.
From a theoretical point of view, the true role of randomization or probabilistic assignment in experimental design can be
best understood in the context of the missing data framework and the potential outcomes approach proposed by Rubin
(1974, 2006, 2007; see also Little & Rubin, 2002). Suppose we want to compare the effect of a new experimental
treatment (say T1) with an alternative treatment (say T2), and 20 participants are available. As explained in the
introductory chapter, it is impossible to obtain counterfactual evidence at an individual level: We are not able to apply two
different treatments to an experimental unit at the same time. So we are obliged to estimate the differential effect of the
experimental conditions T1 and T2 at a group level, and for that we randomly assign the 20 participants to both
conditions (10 participants per condition) using Method 3Forced Equal Sizes (see Subsection 2.2.1). The actual
assignment shown in the three left columns of Table 4.1 is one of the 184,756 equiprobable assignments under the
procedural constraints of Method 3. Suppose also that all participants have complied with the assigned treatment and that
we have collected the outcome measures depicted in the fourth and fifth columns of Table 4.1. Using Rubin's potential
outcomes notation, a rapid inspection of these columns shows that half of the data are missing: Participants exposed to
T1 (Units 11, 15, , 14) have no outcome measures under T2, and the same is true for participants exposed to T2 (Units
5, 6, , 2) in relation to the outcome measures under T1. In formal terms, what Rubin says is that the underlying
probabilistic assignment mechanism in itself makes the missing observations in the unassigned experimental conditions
into missing completely at random data, which are therefore nonimputable to any intrinsic characteristic of the
experimental units that have been randomized. Under these circumstances, the average effect size calculated on the
basis of the observed outcomes is a nonbiased estimator of the true effect size that would be obtained if the
counterfactual evidence could be assessed at an individual level. In this example of an intact or nonbroken experiment, the
nonbiased sample estimator of the population average effect size is g = 0.65 (Hedges's correction to Cohen's d = 0.68;
see Analysis 1 in Table 4.1).
The number of possible assignments (NPA) using Method 3 is given by the formula NPA = N!/(n!
t
), where N is the number
of available experimental units, n is the number of units per experimental condition, and t is the number of experimental
conditions. If we apply this formula to our example in
Table 4.1 Assignment Mechanisms, Outcomes Notation, and Statistical Analysis of Intact and Broken Experiments
Table 2.2 (Design 1: Completely Randomized Design, with 4 treatment levels, 20 experimental units, and 5 experimental
units per treatment level), we will find NPA = 20!/(5!
4
) = 11,732,745,024 possible assignments under Method 3 (for
calculation of factorials, see Appendix 2). This formula is a shortcut for the general formula applied to forced-sizes SRA
methods: NPA = N!/(n1! n2! nt!), where n1, n2, , nt are the intended number of units for the different
experimental conditions. Applying this formula to our illustration of Method 4 (Design 1: Completely Randomized Design,
with 4 treatment levels, 20 experimental units, and 7, 3, 3, and 7 units for Treatment Levels T1, T2, T3, and T4,
respectively) results in NPA = 20!/(7! 3! 3! 7!) = 2,660,486,400 possible assignments.
When dealing with a priori probabilities SRA methods (Methods 1 and 2), the number of possible assignments are given
by the so-called rule of product or fundamental principle of counting: NPA = t
N
, where t is the number of experimental
conditions and N the number of available experimental units. In the same example (see Table 2.2), the number of possible
assignments are NPA = 4
20
= 1,099,511,627,776. Please note that some of these assignments may correspond to
situations in which one or more experimental conditions are empty; that is, no experimental units have been assigned to
these conditions. In Method 1, all assignments are equiprobable. In Method 2, depending on the prespecified probabilities,
some assignments are more probable than others. In the companion website of this monograph, under the link
Probabilities and Random Assignment Mechanisms, the reader can access a fully worked example showing a physical
random assignment mechanism and containing all the details for counting the number of possible assignments in
Methods 1 to 4. To conclude our example, we should note that if the reader uses Method 1 instead of Method 3, the
probability of getting a random assignment with equal sizes (n = 5) per treatment level is p =
11732745024/1099511627776 = .0107. By the same token, if the reader uses Method 1 instead of Method 4, the
probability of getting a random assignment with treatment levels sizes n1 = 7, n2 = 3, n3 = 3, and n4 = 7 is p =
2660486400/1099511627776 = .0024.
4.2.2 Statistical Analysis of Intact Experiments
As the reader would have realized in the previous subsection, an intact experiment is a randomized experiment in which
the conditions introduced by the initial randomization are kept unchanged from the beginning to the end; that is, all
participants comply with the randomly assigned experimental condition, and the researcher is able to collect reliable
outcome measures from all of them. In these conditions, often assembled in short-duration laboratory experiments but
seldom met in long-duration field experiments, modeling data and estimating relevant parameters are relatively simple
tasks, providing that the statistical model is correctly specified and the computational algorithms are sufficiently accurate.
All truthful and pedagogically oriented statistical textbooks begin with one or more chapters focused on the proper use of
descriptive statistics and the associated graphical displays. In addition to the efforts involved in planning the study and
collecting reliable data, the first task a researcher is supposed to carry out before plunging into inferential statistics is a
careful inspection of his or her raw data. This preliminary work of scrutinizing data is at least as important as obtaining the
final results to be published and routinely takes much more time than the few seconds required to run a complex analysis
via any standard statistical software. So it is always a good idea to advise students and novice researchers to begin by
mapping raw data into convenient tabular and graphical displays (e.g., box-and-whisker, stem-and-leaf, and influence
plots; informative and sober bar and line charts), avoiding the beautiful but almost always deceptive features of colored 3-
D plots (see Wilkinson, 2005, for an outstanding discussion of graphical displays).
As indicated in Figure 4.1, the ultimate statistical analysis must also be preceded by relevant ancillary analysis (e.g., an
internal analysis based on manipulation check measures) and, when necessary, by (admissible) data transformations
required for meeting the assumptions underlying the statistical model. In addition to carefully checking these
assumptions, the reader is also advised against the practice of relying on the default statistical models available from the
GUI (graphical user interface) menus of standard statistical packages. In Table 4.2, we summarize the structural features
and the statistical models underlying the five most basic one-factor designs, whose randomization was explained in
Chapters 2 (Designs 1, 2, 3, and 5A) and 3 (Design 6N). If, for example, the reader uses IBM SPSS default options to run
the analysis of Designs 2 and 5A, he or she gets an ANOVA Table filled with blanks telling that some parameters in the
model cannot be estimated. The reason for this is trivial: By default, the program assumes interactive terms for all the
main factors included in the model, as is the case with Design 3, in which the interaction () between the true
experimental factor (a) and the blocking variable () can be estimated. To properly analyze Design 2 (one true
experimental factor and one blocking variable) and Design 5A (one true experimental factor and two blocking variables),
the user must delete the interactive terms from the default statistical models (see Table 4.2). Obviously, this can easily be
done within the IBM SPSS GUI, but note that we are talking about the most elementary statistical models. For running
complex models that include simultaneously fixed and random factors and have different error terms for relevant contrasts,
it is
Table 4.2 Comparison of Randomization Procedures, Structural Features, and Statistical Models of between-Subjects and Within-
Subjects Designs (Illustration With One Treatment With Three Levels and Six Observations per Treatment Level)
always a significant advantage to master the programming features of the statistical package (e.g., SAS and IBM SPSS
syntax code) or to rely on flexible computing environments, such as the free R software. Even if the user completely
ignores the statistical coding potentialities, there is no GUI that can substitute for having a proper knowledge of the
statistical model underlying the experimental design. Please note that what we are saying about IBM SPSS applies to the
GUI menus of similar statistical packages, and the main point is not whether the analysis is conducted via point-and-
click actions or via previously written code but rather the specification and estimation of the correct statistical model.
To express it in general terms, we can adopt the formulation given by Hinkelmann and Kempthorne (2008, p. 35), in which
the outcome responses (i.e., the data) are a function of explanatory variables plus random error: Response =
f(Explanatory variables) + Error. The model for explanatory variables includes treatment effects (i.e., the effects of the
factors incorporated in the treatment design) and blocking/stratifying effects (i.e., the effects of the variables used in the
error control design). Depending on both the nature of the treatment design and the nature of the error control design,
interactive terms between explanatory variables may be included in the statistical model. The error term corresponds to
the random variance that can be imputed to the unreliability of measures or of treatment implementation as well as to
unintended variations in settings, and obviously to the uncontrolled variability of experimental units. What randomization
effectively does is to make the error terms uncorrelated (i.e., independent) with the terms standing for the explanatory
variables.
As the reader has already been informed, this is not a section on data analysis. So the next paragraphs will be limited to
some guidelines and bibliographic orientations for the proper analysis of randomized experiments. Inevitably, these
bibliographic orientations reflect our own teaching and research practice, but we do not intend in any way to disqualify
other excellent books on experimental design and data analysis.
4.2.2.1 between-Subjects Designs
Before proceeding with these guidelines, it should be noted that t tests, such as those included in Table 4.1, and
ANOVA/ANCOVA models are special cases of the general linear model (GLM), which also includes correlation/regression
analysis, as well as some interdependence models such as principal components and factorial analysis. Note also that t
tests themselves are special cases (one experimental factor only with two levels) of the so-called one-way ANOVA
models. Running an ANOVA on the example analyzed in the previous subsection gives an observed F (1, 18) = 2.33, p =
.144 (i.e., t
2
= 1.526
2
= F = 2.33).
For an integrated approach to ANOVA and least squares regression under the GLM, the reader is referred to Cohen,
Cohen, West, and Aiken (2003), Fox (2008), and Kutner, Nachtsheim, Neter, and Li (2004). In particular, Fox's book also
covers generalized linear models (an extension of the GLM), which can be used with remarkable advantages over the
traditional nonparametric alternatives to ANOVA in situations where outcome measures are taken as unordered or ordered
categorical variables. Furthermore, it should also be stated that when the experimental factors are truly continuous
variables, the analysis can be done via regression, without losing information by their artificial reduction to the status of
nominal independent variables. As a matter of fact, statistical software packages can handle all ANOVA models via the
so-called regression approach, in which the nominal independent variables are previously recoded into dummy variables.
With intact experiments, provided that the statistical model is correctly specified and the error terms are appropriate for
the intended specific contrasts, all the general-use statistical packages mentioned in the introductory paragraphs of this
chapter can be used for doing a proper analysis of between-subjects randomized experiments. As the reader has already
been advised, the point-and-click actions and the help menus included in the chosen software do not excuse one from
having knowledge of the underlying statistical models, which can be obtained only from authoritative sources. In addition
to Fisher's (1925/1970, 1935/1966) seminal books, there are three classic sources, published in the 1950s, that remain
irreplaceable references in experimental design and data analysis: Cochran and Cox (1957), Federer (1955), and
Kempthorne (1952). Written without the assistance of present-day computers and statistical software, these books
include fully worked examples, formulas, and supplementary materials that continue to be invaluable tools for the
statistically skilled researcher. Among the updated editions of more recent classics, our choice goes to Hinkelmann and
Kempthorne (2005, 2008), Keppel and Wickens (2007), Kirk (1995, 2013), Maxwell and Delaney (2004), and Winer et al.
(1991). As the reader has certainly noticed, Kirk (1995) gave us the explicit layout for the presentation of randomization
methods in Sections 2.2 and 2.3. We finish this overview by recalling the experimental design and ANOVA QASS
monographs whose serial numbers are mentioned in the Preface (the titles are given in the list included in the first pages
of this monograph) and the new edition of Judd, McClelland, and Ryan (2009), focused on statistical models comparison,
as well as the book by Tabachnick and Fidell (2007), which makes a fair trade-off between hand calculations and
statistical packages procedures (IBM SPSS, SAS, and SYSTAT).
With specific regard to the estimation of effect sizes and the proper calculation of their confidence intervals, the reader can
choose among several excellent books (e.g., Cortina & Nouri, 2000; Grissom & Kim, 2005; Kline, 2004; Rosenthal,
Rosnow, & Rubin, 2000) and/or seek advice from systematic reviews such as Borenstein (2009), Fleiss and Berlin (2009),
Kirk (2003b), or Thompson (2006). Reichardt (2000) discusses effect sizes in the context of the Campbellian threats
approach to the validity of causal inferences.
4.2.2.2 Within-Subjects (Cross-Over) Designs
The considerations and the guidelines given regarding the statistical analysis of intact between-subjects designs are also
valid for the analysis of intact cross-over designs. However, we must add two complementary notes.
The first note concerns the preliminary check of the assumptions underlying the statistical model. In between-subjects
designs, these assumptions are independence of errors, normality of sampling distributions, and homogeneity of variance.
The reader must realize that the independence assumption introduced by random assignment in between-subjects
designs cannot be taken for granted in cross-over designs given that some degree of correlation between the measures
collected on the same units in successive periods is always expected. Standard statistical packages generally include
one or more routines (e.g., Mauchly's Test of Sphericity) to test if the magnitude of these correlations can compromise
further analysis. In formal terms, this additional assumption, which includes separated but related components (additivity
and homogeneity of covariance), is know as the sphericity assumption.
The second note concerns the choice of the statistical model itself. Even if the sphericity assumption holds, the reader
must be advised against statistical packages with default options for the analysis of cross-over or repeated measures
designs. These options are usually appropriate for the analysis of designs generated on the basis of Random
Counterbalancing (Method 14; Design 6N) and Positional Counterbalancing (Method 15; Design 6M) but not for modeling
direct and carryover effects simultaneously, as is the case with designs built by sequential counterbalancing (Methods 16-
18; see Table 3.3 for the designations of designs). So, in addition to the bibliographic orientations given for between-
subjects designs, in which Designs 6M and 6N are generally covered, we refer the reader to the sources where models for
direct and carryover effects are fully explained and illustrated (e.g., Cotton, 1998; Jones & Kenward, 2003; Senn, 2003). In
consonance with the presentation of cross-over designs randomization methods in Chapter 3, our first choice is Jones and
Kenward (2003; in the companion website of this book, the reader can access XOEFFICIENCY, a GenStat module to
calculate the efficiencies of direct and carryover contrasts in cross-over designs).
4.2.3 Statistical Analysis of Broken Experiments
As the expression suggests, broken experiments are randomized experiments in which the ideal conditions introduced
by random assignment or counterbalancing are more or less compromised by unexpected threats to internal validity (e.g.,
attrition) and/or the violation of SUTVA components (e.g., treatment noncompliance).
4.2.3.1 Between-Subjects Designs
In the last two columns of Table 4.1, we show a hypothetical example of a between-subjects broken experiment (to be
more precise, the same completely randomized design whose standard analysis is illustrated in Subsection 4.2.1). As the
reader can easily verify, attrition has resulted in the absence of outcome measures for Participants 18, 17, 13, 7, and 10,
and Participants 15, 1, 6, 20, and 2 and have complied with the unassigned treatment. Moreover, the follow-up on
treatment compliance is uninformative for Participants 17, 13, and 8. In this situation, the assessment of the causal effect
must be reframed to obtain the least biased estimator. One common procedure is to ignore treatment compliance
information and analyze all cases without missing values on the outcome measure as if they have complied with the
randomly assigned treatment (see Analysis 2 in Table 4.1). This analysis is called intention-to-treat (ITT) analysis, and the
obtained effect can be thought of as the best available estimator of the average treatment effect (ATE) that would be
obtained in an intact experiment (see Analysis 1). ITT analysis is always preferable to the temptation to estimate the ATE
on the basis of treatment compliance information, ignoring random assignment. This is what is illustrated in Analysis 3
(see Table 4.1), and the obtained effect is called naive ATE (for consistent alternatives, see Angrist, Imbens, & Rubin,
1996).
Broken experiments fall somewhere in a gray area between intact randomized experiments and nonrandomized
experiments (quasi-experiments or observational studies), and their statistical analysis could be improved by the
sophisticated techniques developed in the context of propensity score analysis. We refer the reader to several fundamental
sources on propensity score analysis (Imbens & Rubin, in press; Rosenbaum, 2002, 2010; Rosenbaum & Rubin, 1983;
Rubin, 2006) and also to the excellent book by Guo and Fraser (2010), in which Rubin's approach and related techniques
are fully explained and illustrated. For well-articulated presentations of Rubin's causal model, causal inference, and
experimental methodology in the social and political sciences, see Morgan and Winship (2007) and Morton and Williams
(2010).
4.2.3.2 Within-Subjects (Cross-Over) Designs
Treatment noncompliance issues and their implications for the validity of causal inferences are essentially the same in
cross-over and between-subjects designs. On the contrary, whereas attrition (especially differential attrition) is a serious
threat to the internal validity of causal inferences in between-subjects designs, its negative impact on cross-over designs
can be more easily overcome. As we have already stated, omitting data from one or more participants in a cross-over
design results in breaking the intended degree of balance for the whole design. If this happens, data coming from intact
subsets of sequences can also be analyzed with statistical models appropriate for the degree of balance achieved in
these intact subsets (e.g., subsets of Designs 6K: Locally Balanced Designs; more generally, all multiple-bricks designs
or designs having two or more replications of the standard cross-over design).
4.2.3.3 Trade-Offs between Design and Analysis
The core contents of this monograph (Chapters 2 and 3) relate directly to design features (random assignment, local
control, and counterbalancing schemes) intended to maximize the internal and the statistical conclusion validity of local
causal inferences. In this chapter and in the introductory chapter, the complementarity between Campbell's threats
approach to causal inference and Rubin's causal model was conveniently highlighted. While the first approach is more
centered in design issues, the second one is more focused on the probabilistic nature of assignment mechanisms and the
formal assumptions underlying the specification and estimation of statistical models; however, there is no remarkable
divergence on the very nature of randomized experiments and their central role in causal inference. This can be noticed in
a recently published special section of Psychological Methods (Vol. 15, No. 1), with the benefit of first-hand accounts
(Cook & Steiner, 2010; Rubin, 2010; Shadish, 2010). As stated in the introduction to the special section (Maxwell, 2010)
and documented in the main articles (see also Imbens, 2010; West & Thoemmes, 2010), the two approaches are more
dissimilar, yet not irreconcilable, with respect to causal inference in nonrandomized experiments, but this is not the place
to elaborate on quasi-experimentation (observational studies). To conclude this short note, we must restate the intrinsic
relationship between the randomization method used in the error control design and the specification of statistical models
and error terms in the data analysis stage.
4.2.4 Randomization and Statistical Control
When previously measured covariates are continuous variables, ANCOVA may be a consistent alternative to local control
(blocking and stratifying) in randomized experiments. By definition, random assignment procedures ensure the statistical
assumptions (e.g., noninteraction between the covariable and the experimental factor) underlying ANCOVA models.
Please note that we are not talking about the use of ANCOVA to infer causality in nonexperimental designs or about the
inclusion of single or multiple covariates in quasi-experimental designs (see, e.g., the contribution of Reichardt to Cook &
Campbell, 1979), in which propensity score analysis and related techniques can do a better job. For the same reasons,
we were reticent about minimization as an acceptable method of randomization (see Subsection 2.5.4), regardless of
whether and how prognostic factors are used in the data analysis stage (for consistent alternatives, see Berger, 2005).
4.3 Reporting Randomized Experiments
Even though the modes of production and transmission of academic knowledge in contemporary societies could depart
significantly from the idealized Merton's (1942/1973) scientific ethosin which organized skepticism is tied with the
normative values of universalism, communalism, and disinterestednessand can probably be best described by the
Campbellian tribal model of the scientific social system (Campbell 1979), transparent communication of research findings
continues to be a highly prized standard, as it is documented in the activities of organizations such as The Cochrane
Collaboration (http://www.cochrane.org/) in the area of health care or its counterpart The Campbell Collaboration
(http://www.campbellcollaboration.org/) in the domains of education, crime and justice, and social welfare. Both
organizations focus on producing and disseminating systematic reviews of scientific studies and therefore helping
policymakers and people in general make well-informed decisions on the basis of the best available research evidence.
This is also the most important mission of the American Evaluation Association (http://www.eval.org/) in the areas of
evaluation studies and social experimentation (for systematic reviews, see Cook & Shadish, 1994; Shadish & Cook,
2009).
In this context, randomized experiments play a central role, and the effort to promote clear and truthful communication of
their methodological details and results gave rise to the publication of well-systematized reporting standards, such as the
CONSORT Statement (Consolidated Standards of Reporting Trials; see Schulz, Altman, & Moher, 2010) in the health
sciences, The American Psychological Association Journal Article Reporting Standards (JARS, 2008) in psychology, and
the Standards for Reporting on Empirical Social Science Research in AERA Publications (American Educational
Research Association, 2006) in education. Boutron, John, and Torgerson (2010) have also extended these guidelines to
the political and social sciences in a recently published article in the official journal of the American Academy of Political
and Social Science. Of course, this is not entirely new, since more or less explicit guidelines for research reporting have
always been present in institutional publications (e.g., Standard 13 of the Code of Ethics and Policies and Procedures of
the American Sociological Association, 1999), editorial and substantive pieces in scientific journals (e.g., a seminal
version of the Publication Manual of the American Psychological Association was published as an article inserted in the
1929 volume of the Psychological Bulletin), and reference books that gave substance to and organized the various
disciplines.
With the exception of the diagram embedded in Figure 4.1 showing the flow of participants through the different stages of
a randomized experiment, we are not reproducing here the extended tables and checklists inserted in the aforementioned
reporting standards, but the reader is referred to the companion website of this monograph, where an organized listwith
direct links to the primary documentsis available. Warning against the mechanical use of these tables, checklists, and
flowcharts and the implicit reification of the methods behind them, we encourage researchers to carefully read the
explanations and elaborations given by the original authors (e.g., Cooper, 2011; Moher et al., 2010). An interesting
appraisal of the impact of the Consort Statementwhose initial version was published by Begg et al. (1996)on the
quality of reporting randomized experiments in the health sciences is given by Plint et al. (2006).
If we had to summarize the contents of the various reporting standards in a single paragraph, we would say that a
scientific report (e.g., a journal article) of a randomized experiment must comply with three basic requirements: (1) to
allow for an exact or a conceptual replication of the experiment; (2) to give the theoretical, the methodological, and the
contextual information needed for a fair evaluation of the main results, claims, and practical implications of the study; and
(3) to provide the summary statistics required for a reliable meta-analysis. Briefly stated, any competent researcher in the
field should be able to replicate, evaluate, or include the experiment in a meta-analysis relying exclusively on the
published contents (comprising precise bibliographic references for complementary technical details), without the need to
contact the author(s) of the study. The conflict between comprehensiveness and availability of space in scientific journals
(Cooper, 2011) could be easily reduced by the current practice of providing supplementary materials in the journals
websites.
To conclude, we should also say that the randomization and local control procedures described in this monograph, as well
as the standardized statistical procedures applying to data analysis, are instrumental and not terminal values. They
are just devices that supply the machinery for unambiguous interpretation:
A clear grasp of simple and standardised statistical procedures will, as the reader may satisfy himself, go far to
elucidate the principles of experimentation; but these procedures are themselves only the means to a more
important end. Their part is to satisfy the requirements of sound and intelligible experimental design, and to
supply the machinery for unambiguous interpretation [italics added]. To attain a clear grasp of these requirements
we need to study designs which have been widely successful in many fields, and to examine their structure in
relation to the requirements of valid inference. (Fisher, 1935/1966, p. vii)
http://dx.doi.org/10.4135/9781452270012.n4

Sampling

Uploaded by

Copyright:

Available Formats

Sampling

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sampling

Uploaded by

Copyright:

Available Formats

18/1/2014 Methods of Randomization in Experimental Design : SAGE Research Methods

You might also like