See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225208510

Linear versus logistic regression when the dependent variable is a dichotomy

Article  in  Quality & Quantity · January 2009

DOI: 10.1007/s11135-007-9077-3

CITATIONS READS

544 2,106

1 author:

Ottar Hellevik
University of Oslo
51 PUBLICATIONS   1,285 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Candidate selection in local elections in Norway View project

All content following this page was uploaded by Ottar Hellevik on 24 January 2017.

The user has requested enhancement of the downloaded file.

Qual Quant (2009) 43:59–74
DOI 10.1007/s11135-007-9077-3

O R I G I NA L PA P E R

Linear versus logistic regression when the dependent

variable is a dichotomy

Ottar Hellevik

Published online: 16 February 2007

© Springer Science+Business Media B.V. 2007

Abstract The article argues against the popular belief that linear regression should
not be used when the dependent variable is a dichotomy. The relevance of the statis-
tical arguments against linear analyses, that the tests of significance are inappropriate
and that one risk getting meaningless results, are disputed. Violating the homoscedas-
ticity assumption seems to be of little practical importance, as an empirical compari-
son of results shows nearly identical outcomes for the two kinds of significance tests.
When linear analysis of dichotomous dependent variables is seen as acceptable, there
in many situations exist compelling arguments of a substantive nature for preferring
this approach to logistic regression. Of special importance is the intuitive meaning-
fulness of the linear measures as differences in probabilities, and their applicability in
causal (path) analysis, in contrast to the logistic measures.

Keywords Logistic regression · Binary variables · Significance tests

1 Introduction

In analyses of survey data it is not unusual that the dependent variable is a dichot-
omy. When the research problem requires a multivariate solution, regression analysis
is very convenient for handling large numbers of independent variables.1 Today it
seems to be a common belief that with a binary dependent variable (dichotomy coded

1 This opportunity is sometimes overexploited, however. A regression analysis of binary variables

does not have access to information that is lacking in the corresponding tabular analysis. When more
variables may be included in the regression analysis, this is due to the distributional assumptions on
which the regression analysis is based. With a large number of variables one runs the risk that an
estimate reflects the model more than the data. (Rubin 1997; Rothman and Greenland 1998).

O. Hellevik (B)
Department of Political Science, University of Oslo, P.O. Box 1097, 0317, Blindern, Oslo, Norway
e-mail: [email protected]
60 O. Hellevik

0–1) ordinary linear regression cannot be used.2 Logistic regression is a prerequisite

for an article to be considered for publication in a serious scientific journal, one hears.
Two statistical arguments are given for this rejection of linear regression. One is
that with linear coefficients we risk meaningless results, since a predicted probability
may fall outside the range 0–1. The other, which seems to be the main objection, is
that the statistical tests for linear analyses are inappropriate with a binary dependent
variable. These arguments are not as decisive as many seem to think, however. And
when results from logistic regression are difficult to understand,3 and in many cases
may seem counterintuitive from a substantive point of view, there is every reason to
ask for a more open minded discussion of the fruitfulness of different approaches
to the statistical analysis of survey data. Hopefully the outcome of such a discussion
would be that the choice of technique is guided more by considerations of what is
meaningful in relation to the research problem, and less by a desire to demonstrate
mastery of complicated statistical tools, than what may sometimes seem to be the case
today.

2 Outline of the article

The article starts out by examining the two arguments of a statistical nature that are
given as reasons why one should avoid a linear model when the dependent variable is
a dichotomy. It is claimed that the risk for “impossible predictions” may be avoided by
better specifications of the model. In the case of causal (path) analysis this objection
is not relevant. With regard to inappropriate significance tests, I show by means of
a simulation that the theoretical objection against the linear test is of little practical
importance. The significance probabilities from linear and logistic regression analyses
turn out to be nearly identical.
Having thus established that we actually have a choice when deciding how to ana-
lyze an empirical problem involving a binary dependent variable, arguments relevant
to this choice is discussed. In order to reach also those unfamiliar with loglinear anal-
ysis, we start with a brief presentation of some of the central measures of statistical
association used within this approach. A comparison between the results of linear and
logistic measures is made, and reasons for preferring one or the other discussed.
The conclusion is that there in many situations exist compelling substantive rea-
sons for preferring linear measures over the logistical. It would therefore be most
unfortunate if we in fact were restricted to the latter measures when the dependent

2 “Use of logistic regression — a statistical technique suited for the analysis of dichotomous categor-
ical dependent variables — has become increasingly common in public opinion research. Outcomes
involving voting, internet use, survey non-response, and opinions on a variety of controversial issues
often take the form of yes/no, agree/disagree, favor/oppose, or present/absent. These outcomes do not
meet the requirement of linear regression for an interval or ratio scale and often warrant use of logistic
regression.” From page 30 in the program for the AAPOR/WAPOR (American/World Association
of Public Opinion Research) annual conference in May 2004, advertising the course “Making Sense
of Logistic Regression: An Introduction”.
3 We cite once more from the ad for a course in logistic regression mentioned in note 2: “… for those
who haven’t had a graduate-level statistics course on the topic, making sense of logistic regression
is daunting. The technique involves unfamiliar terms such as logged odds, odds, predicted proba-
bilities, marginal effects, log likelihood values, pseudo-variance explained, and maximum likelihood
estimation that can make meaningful interpretation of results difficult”.
Linear versus logistic regression when the dependent variable is a dichotomy 61

variable is a dichotomy, as it is often claimed. We will now take a closer look at the
validity of the statistical arguments against the use of linear measures.

3 “Impossible” results of linear analyses?

Results outside of the interval 0–1 may occur when we determine the probability for a
high value on the dependent variable for various combinations of values on the inde-
pendent variables by summing the products of variable values and coefficients in the
regression equation. The problem of “impossible results” cannot arise in a loglinear
analysis, since these measures do not have limits to their variation. With a range from
plus to minus infinity “anything goes”.
But are predictions outside the range of meaningful variation really such a seri-
ous problem? “In practice, estimates of cumulative incidence would seldom be much
below zero or above 1, and then only for extreme combinations of values of the
independent variables. In place of such inadmissible estimates, it is reasonable to
substitute the theoretical minimum or maximum value, as appropriate, for the inad-
missible value.” (Rothman 1986: 291).
In a linear regression analysis with binary independent variables “impossible”
results may only occur when there is interaction in the data. In case of interaction, we
will, by including the relevant interaction terms in the model, always be guaranteed
that the predicted result will lie within the interval 0–1.
If some of the independent variables are continuous instead of binary, predictions
outside the permitted range will still be possible, however. This may occur if the asso-
ciation between such a variable and the dependent variable is non-linear. When there
is reason to question the assumption of linearity, one alternative is to transform the
continuous variable into a set of dummy variables. Then we would still be protected
against predictions outside the 0–1 interval, by including the relevant interaction terms
in the analysis.
If the purpose of the analysis is not prediction but causal decomposition of bivar-
iate associations (as in path analysis), the problem of “impossible predictions” is no
longer relevant. Whether or not the conditional effects of an independent variable
are different (interaction) or similar, we may calculate an average which tells us what
impact on the dependent variable these conditional effects add up to within a popula-
tion with a given composition (the distributions for the independent variables, which
determine the weights in a tabular analysis (Hellevik 1983, 1984)). Such causal effects
should not be used to predict probabilities, which means that whether a prediction
falls within the range of 0–1 or not is irrelevant. What matters for the results of a causal
analysis is whether the sum of components of direct, indirect and spurious effects is
identical to the bivariate association. This requirement is met by the linear regression
coefficients, but not by the logistic (Hellevik 1984).

4 Inappropriate linear significance tests?

When the danger of misleading significance tests are used as a decisive argument
against linear regression with binary dependent variables, this first of all indicates that
uncertainty arising from sampling is made into a bigger problem than it usually is in
survey analysis. Random sampling error is just one of many possible methodological
62 O. Hellevik

explanations of the results from surveys. With a reasonably large sample, random
sampling error may often be rejected off hand as a possible cause of a tendency, as
long as we are not interested in very weak effects.
Secondly it is pertinent to ask how important the objection based on statistical
theory against the linear test is in practice. The problem with a binary dependent
variable is that the homoscedasticity assumption (similar variation on the dependent
variable for units with different values on the independent variable) is not satisfied.
This is of no consequence for the regression coefficient. But the uncertainty estimate
for the coefficient, and thus the test of significance, is affected. This seems to be the
main basis for the widely held opinion that linear regression is inappropriate with a
binary dependent variable.
How far are the results from linear tests from those of logistic tests? One possibility
is that the divergence between the two is systematic, in the sense that the significance
probability from one of the tests always is larger than the other. If it were the case
that the linear tests consistently showed larger P-values than the logistic ones for the
same empirical data, this would mean that these tests give a conservative estimate of
the significance of a regression coefficient.
To look at the correspondence between the two tests a series of parallel linear and
logistic regression analyses involving two independent variables (the binary variable
gender and the continuous variable age) and a series of binary dependent variables
(preferences for various parties and political attitudes) were carried out. The per-
centage with a high value on the dependent variable varied between 1 and 52, with
an approximately even distribution on the ranges 0–9, 10–19, 20–29, and 30+, within
each of four categories of random samples with approximately 200, 500, 1250 and
2500 respondents. A total of 320 comparisons of significance probabilities for the
corresponding linear and logistic regression coefficients were made, 20 in each cell of
the 4*4 matrix defined by sample size and distribution on the dependent variable.
It turned out that there was no systematic tendency for one of the two sets of
P-values to be larger than the other. Even so the results were very interesting with
regard to the applicability of linear tests of significance with a binary dependent
variable.
Quite surprisingly in view of the dire warnings against the use of linear tests,
the differences between the corresponding significance probabilities were extremely
modest. The correlation between the two sets of P-values was as high as 0.9998, which
means an explained variance of 99.96% when one P-value is regressed on the other.
The average difference in probability (disregarding the sign) is 0.002. In just one case
out of the 320 did a 5% level significance test result in conflicting conclusions for the
two tests.4
The correspondence between the results is illustrated by the diagonal pattern of
a scatterplot for the P-values from the two kinds of regression analysis. This is an
agreement so perfect that it borders on complete identity (Fig. 1).
When we look at the P-values from the linear and logistic significance tests, they in
20% of the 320 cases were completely identical with four digit probabilities. In 43%
of the comparisons the difference was less than 0.001, in 90% less than 0.005 (Fig. 2).
4 The P-values were 0.0497 for the linear and 0.0505 for the loglinear coefficient. Of the remaining
319 coefficients 136 (43%) were found significant by both tests, and 183 (57%) were found not signifi-
cant. The one case of disagreement constitutes 0.3% of the results. In comparison, we can in 16 of the
tests expect to make a type I error by rejecting a true null hypothesis of b = 0 in the population when
the significance level is 5%.
Linear versus logistic regression when the dependent variable is a dichotomy 63

1,00

0,90

0,80

0,70
Logistical regression

0,60

0,50

0,40

0,30

0,20

0,10

0,00
0,00 0,10 0,20 0,30 0,40 0,50 0,60 0,70 0,80 0,90 1,00
Linear regression

Fig. 1 The correspondence between significance probabilities for 320 linear and logistic regression
coefficients (sample sizes varying between 200 and 2500, percentage high on dependent binary variable
between 1 and 52)

50

45 43

40

35

30
24
25 23

20

15

10
5
5 2 1
0 1 1 0 0
0
-0.0349 -
-0.0250
-0.0249 -
-0.0150
-0.0149 -
-0.0050
-0.0049 -
-0.0005
-0.0004 -
0.0004
0.0001 -
0.0049
0.0050 -
0.0149
0.0150 -
0.0249
0.0250 -
0.0349
0.0350 -
0.0449
0.0450 -
0.0549

Fig. 2 The percent distribution of differences between significance probabilities for 320 correspond-
ing linear and logistic regression coefficients (linear–logistic)

The size of the difference between the corresponding P-values is primarily affected
by the distribution on the dependent variable in this series of tests. Size of sample also
matters, with the exception of tests with a very skewed distribution on the dependent
variable (percentages below 10). In these particular cases the pattern actually is in the
opposite direction of what is to be expected (Table 1). When we compare the results
for the binary and the continuous independent variables, there is a tendency towards
slightly larger differences in P-value for the binary variable (in average 0.002).
64 O. Hellevik

Table 1 Average difference in size of significance probability for linear and logistic test (320
comparisons, evenly distributed between the 16 cells of the table)

Sample size Percentage with high value on dependent variable All

0–9 10–19 20–29 30–

200 0.0034 0.0033 0.0020 0.0017 0.0027

500 0.0062 0.0024 0.0010 0.0005 0.0027
1200 0.0060 0.0013 0.0009 0.0004 0.0020
2500 0.0068 0.0004 0.0010 0.0001 0.0021
All 0.0055 0.0019 0.0011 0.0008 0.0024

Table 2 Experiencing similarities between own work in the government office and that of a judge,
depending on own education (Proportions)

Education Law Other PD

Feels that own role is similar to that of a judge 0.75 0.20 0.55
Does not feel own role as similar to that of a judge 0.25 0.80 −0.55
Sum 1.00 1.00 0.00

The high agreement in test results must come as a surprise for all those who firmly
believe that linear regression cannot be used with a binary dependent variable due to
inadequacy of the significance test. The outcome is, however, in line with the statement
we find in some textbooks on regression analysis that linear tests will give misleading
results for significance tests only with small samples (Fox 1997: 451; Skog 1998: 301).
There also have been similar results in empirical analyses reported earlier (Kanagy
et al. 1994). The simulation reported in this article indicates that the linear test is so
robust that even sample sizes unusually small for survey research and very skewed
distributions on the dependent variable do not create problems.
It is our claim that the statistical objections to applying linear regression analysis
with a dichotomous dependent variable may be put to rest. This means that we do
have a choice between this approach and logistic regression. The discussion now turns
to reasons for preferring one method over the other. But first a quick review of the
two families of statistical measures.

5 Linear and loglinear measures

Table 2 shows the bivariate association between education and how own occupational
role is experienced among government office employees (from Lægreid and Olsen
1978).
The role perception differs between those administrators with an degree in law
and those with a different educational background. The association may be described
by the proportion difference PD, which corresponds to the unstandardized linear
regression coefficient b when the two variables are coded 0 or 1 (binary variables).5

5 The average for a 0–1 variable is the same as the proportion of units with value 1. The constant is
the proportion with this value among units with value 0 on the binary independent variable, and b
the difference in this proportion when we move to units with value 1 on the independent variable.
For a discussion of how regression coefficients in an analysis with binary variables correspond to
proportions and proportion differences, see Hellevik 1984.
Linear versus logistic regression when the dependent variable is a dichotomy 65

PD = b = Plaw − Pother = 0.75 − 0.20 = 0.55

Besides looking at the absolute difference or gap between the proportions, one may
calculate the ratio between them. We can call this relative difference for the proportion
ratio (PR):

PR = Plaw /Pother = 0.75/0.20 = 3.75

Relative measures are rarely used in comparisons of groups, but are common in
analyses of change over time. If the proportion with a higher education increases from
40 to 60%, this absolute change of 20 percentage points may be described as a change
of 50% relative to the starting point.
In a loglinear analysis odds are calculated instead of proportions. The odds is a ratio
which equals the proportion (or number) of units with a certain value on a variable
divided by the proportion (number) of units without this value. Below the odds for
experiencing own role as similar to that of a judge is calculated for those educated
in law as well as for other educations among government employees. An odds larger
than 1 shows that in the first group the number who feel similarities is higher than the
number who does not. The odds for other educations is far below 1, telling us that
here only a minority feels like a judge in their work.

Odds: O = P/(1 − P)
Oddslaw : Ol = 75/25 = 3.00
Oddsother : Oo = 20/80 = 0.25

To measure how different the two groups of bureaucrats are, we calculate the odds
ratio (OR), which as the name suggests equals the ratio between the two odds, in
other words the ratio between two ratios. The result of the calculation tells us that
the odds for those educated in law is twelve times the odds for employees with other
kinds of educations.

Odds Ratio: OR = Ol /Oo = 3.00/0.25 = 12.0

From the OR other measures of statistical association are derived. One is the natural
logarithm of OR, called the logit, which corresponds to the logistic regression coeffi-
cient we find in a regression analysis of the variation in the odds for having a high
value on the dichotomous dependent variable (here denoted B to distinguish it from
the linear b).

B = ln(OR) = ln(12.0) = 2.48

Another loglinear measure is lambda, which is 1/4 of B.

Lambda = 1/4 B = 1/4 ln(OR) = 1/4 ln(12.0) = 0.62

Lambda in many cases will be identical to or only slightly larger than the correspond-
ing proportion difference (or b) for a given empirical association. In our example
lambda is 0.62 as compared to 0.55 for PD. In certain situations, however, lambda
will take on much larger values. While PD or b varies between plus or minus 1, B or
lambda has plus/minus infinity as extreme values.
66 O. Hellevik

6 What is measured?

The definition and calculation of the linear, relative and loglinear measures differ, and
the results may be contradictory with regard to whether concrete empirical associa-
tions are strong or weak, and whether there is statistical interaction or not. To choose
between them, one ought to have an idea of what aspects of reality the different
measures capture. Only then it is possible to decide what measure is relevant for the
purpose of the investigation.
The linear measures have interpretations that are intuitively easy to comprehend.
A proportion may be understood as an estimate for the probability that the units stud-
ied have a specific value on a variable. A proportion difference or linear regression
coefficient accordingly is the difference in this probability between groups of units
defined by an independent variable. In a causal analysis we may talk about how the
probability for having a high value on the effect variable is influenced by substituting
units with different values on the causal variable, e.g. how the outcome of a recruit-
ment process is influenced when we replace persons with a low education with persons
with a high education (Hellevik 1984).
The relative measure may correspondingly be interpreted as the ratio of probabili-
ties. In epidemiology the term relative risk is used when comparing the incidence of a
medical condition within two groups by calculating the ratio. In studies of recruitment
to attractive positions such as higher education the term participation ratio is some-
times used (Kivinen, Ahoa and Hedman 2001; Hellevik 2002). PR here is interpreted
as a measure of equality–inequality in the distribution of a good, with 1 (identical
probabilities) indicating complete equality.
When the largest proportion is used as numerator, as in our illustrative example,
PR approaches infinity as inequality increases, and may thus be called a measure of
inequality. If instead the largest proportion is used as denominator, we have 0 when
inequality is extreme (one group gets nothing of the good). In this case increasing
values for PR towards the maximum of 1 means increasing equality.
An odds is the ratio between the probability for having a certain value on a variable
and the probability of not having this value. It is easily grasped what is meant by for
instance an odds of 3 (or 3 to 1, as it is often said). This means that 75% of the units
have a certain characteristic, which 25% lack. It becomes more difficult, however,
when we look at the odds ratio. The OR is sometimes described as the ratio between
probabilities, and it is stated that something is so and so many times more common
in one group than in another.6 This is incorrect, what OR shows is the ratio between
odds, not between proportions as PR does. But when the proportions are small the
results for PR and OR are similar. Within epidemiological research one for this reason
sometimes uses such an interpretation of OR.7 In analyses of survey data, however,
6 A research report gave the following explanation for OR (translated from Norwegian): “the odds
ratio expresses how many times more often failing health occurs in the group in question as compared
to the reference group”. Comment to an OR of 3.8: “failing health occurs 3–4 times as frequently for
those who …”. The ratio between the proportions involved was much lower, however: 0.39/0.15 = 2.6.
Referred to in Hellevik (1996).
7 “The odds ratio … approximates how much more likely (or unlikely) it is for the outcome to be
present among those with x= 1 than among those with x= 0. For example, if y denotes the presence
or absence of lung cancer and x denotes whether or not the person is a smoker, then ψ [OR] = 2
indicates that lung cancer occurs twice as often among smokers than among nonsmokers in the study
population”. Hosmer and Lemeshow (1989), Applied Logistic Regression (New York: John Wiley &
Sons).
Linear versus logistic regression when the dependent variable is a dichotomy 67

where the phenomena studied usually occur more frequently, this interpretation often
becomes misleading.
It does not get any easier to give a meaningful description of what is captured by
a loglinear measure, on the contrary, when we turn our attention to B (the logistic
regression coefficient). “The natural logarithm of the ratio between two odds” is
probably not very helpful for many readers. This may explain why sometimes pre-
sentations of results from logistic regression analysis only comment on whether the
effects are significant or not, and not on their magnitude. Or effects are converted to
linear differences in probabilities given certain assumptions.

7 Differences in results

When the definition does not contribute much to an understanding of what aspects
of social reality is captured by a stastical measure, it may be useful to look at how
the measure “behaves” in concrete situations. We will mainly be comparing the more
commonly used linear and loglinear measures, and comment briefly on the relative
measures towards the end.
If as an example 60% of the men in a sample are in favor of Norway joining the
European Union, while the corresponding figure is 45% for the women, we have a
PD of 0.60 − 0.45 = 0.15, which is also the result for the linear regression coefficient b.
The odds is 60/40 = 1.5 for men, and 45/55 = 0.82 for women, rendering the OR
1.5/0.82 = 1.83. We find the logistic regression coefficient B as ln(1.83) = 0.60.
Lambda is 1/4 B = 1/4 0.60 = 0.15. The results in other words are identical for b and
lambda. In general it holds that within the interval from 0.30 to 0.70 lambda will be
identical to or just slightly larger than the PD, and the logistic B will be four times the
size of the linear b.
When we move outside of this range, and towards the extremes of 0 or 1 for the
proportion with a high value on the dependent variable, lambda will take on much
higher values than PD. If for instance 3% of all men are candidates in local elections,
as compared to 1.5% of all women, we have a modest PD of 0.015, one tenth of
the effect on EU opinion mentioned above. With odds of 3/97 = 0.031 for men and
1.5/98.5 = 0.015 for women, we have an OR of 0.031/0.015 = 2.1. This means that
the association according to lambda is higher in the second than in the first example:
1/4 ln(2.1) = 1/4 0.74 = 0.185. In this case B is not four but nearly fifty times larger than
b (B / b = 0.74 / 0.015 = 49.3).
The same contrast between the two kinds of measures is found at the opposite
extreme of the scale. If for instance 99.99% of the men and 99.90% of the women
took part in the “elections” of a dictator, the difference between the sexes according
to the linear regression coefficient b is microscopic: 0.0009. The logistic coefficient B
is 2.30, or 2559 times the size of b. According to the logistic coefficient this difference
in turnout between men and women of 0.09% actually is larger than what we have
when the turnout is 25% for women and 75% for men (B in this case is 2.20).
This shows that statistical association must mean quite different things within the
two approaches. In the last two examples the linear b tells us that the sex of a per-
son makes very little difference when it comes to being a candidate or vote. In the
first case because very few are nominated, irrespectively of whether they are men or
women, while in the second case almost all vote regardless of sex. Knowing the sex of
a person accordingly has little value when we want to predict his or her position on
68 O. Hellevik

1,4

Lambda
1,2 = 1/4 B
PR
1
PD = b

0,8

0,6

0,4

0,2

0
10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90
Percentage with high value on dependent variable in total

Fig. 3 Lambda, proportion difference (linear b) and proportion ratio for comparisons of two groups
of equal size (with proportions for the binary dependent variable 0.10 above and below the total level)

the dependent variable. In a causal perspective we would say that the sex of a person
has very little effect on the probability of being a candidate or a voter.
Figure 3 gives a more general picture of the difference between the two kinds of
measures by comparing two groups (which for simplicity of calculations are of the
same size, e.g. men and women), at various levels for the dependent variable. The
proportion is 0.10 higher than the total level for one of the groups, and 0.10 lower for
the other group, which means that we have a PD of 0.20 at all levels (0.20 vs. 0.00;
0.25 vs. 0.05, and so on). The linear regression coefficient b thus is 0.2 at all levels for
the distribution on the dependent variable. The loglinear lambda takes on the same
values at the medium levels, but becomes much higher near the extremes. The same
pattern holds for the logistic regression coefficient B (not shown), which is four times
the size of lambda.
Figure 3 also includes the relative measure PR, with the largest proportion used as
the denominator. To the left in the figure, where the proportions in the two groups are
0 and 0.2, respectively, we have the minimum value for equality (0). As the propor-
tion in each group rises, PR moves closer to maximum equality (1). In other words,
the higher the level for the dependent variable, the less inequality resulting from the
constant absolute difference, according to the results for the relative measure. This
pattern of monotonous increasing values for PR is different from that of both the
linear and the loglinear measures. For this reason it has been argued that neither of
these two classes of measures are suited for analyses of inequality (Hellevik 1997,
2000, 2002).
Some of the examples discussed above indicate that loglinear measures may give
results that seem counterintuitive. We shall now look at situations where the special
properties of loglinear measures appear substantively meaningful.

8 Variations in rare but important phenomena

When studying a dependent variable where the value of interest only rarely occurs, the
effects of independent variables may be so small that they are hardly noticeable when
Linear versus logistic regression when the dependent variable is a dichotomy 69

analyzed by means of linear measures. Even so the tiny differences between groups in
the incidence of a fatal illnesses, suicide rates, etc. will be important enough to warrant
closer scrutiny. For this reason logistic measures, which may manifest strong effects in
such cases, are often used in epidemiology. Another example is quantitative studies
of conflicts between states, where the probability for war fortunately is small.
In analyses of such rare occurrences also the linear effects (or the relevant propor-
tions) should be reported. Otherwise there is a danger that misleading impressions
of large absolute differences are created. A newspaper reporting from the results of
a skin cancer research project (Veierød et al. 2003) used the headline “55% higher
chance of getting skin cancer if you take solarium”.8 The headline for interviews
with persons who take solarium was “Frightening”. They would probably have been
less alarmed if the newspaper instead had reported that the absolute difference in
probability was less than 0.001.9

9 Changes over time

With trend data the loglinear measures may be interpreted in relation to a model of
logistical growth (Heath et al. 1992). We may take as an example a process of social
diffusion, such as the dissemination of a rumor by word of mouth within a population,
from the starting point where no one has heard the rumor to the end point where
all have been reached. At the outset the proportion having heard the rumor will
grow very slowly, since so few are spreading it. Then the dissemination speeds up as
more people become active, until the growth once more flattens out as the ceiling of
100% knowing the rumor approaches, making it difficult finding someone who has
not already heard it.
A logistical growth curve depicts such a development. In this case lambda will have
the same size for all time intervals of equal width, no matter if we look at the beginning,
the middle period or the end of the diffusion process. The corresponding proportion
differences for changes in knowledge will vary considerably, however. They are much
smaller at the outset (start effect), and towards the end (saturation or ceiling effect),
but similar to lambda in the middle of the process. A stable lambda over time thus
tells us that the growth follows a logistical pattern (Fig. 4).
A logistical growth model does not seem appropriate for all developmental pro-
cesses, however. With regard to for instance attainment of higher education, we have
the possibility of decline as well as growth. This creates a problem for the substantive
justification of the loglinear measures. In a situation of growth, it may with reference
to the ceiling effect seem reasonable to regard an increase from 85 to 95% as stronger
or more “demanding” than an increase from 45 to 55%. This is what lambda indicates,
taking on the value of 0.3 in the first and 0.1 in the last case.
According to the loglinear measures also the decrease from 95 to 85% should be
regarded as three times as strong as the decrease from 55 to 45, since lambda is −0, 3
in the first and −0, 1 in the second case. Intuitively it is not easy to understand why it
should be regarded as “harder” to drop 10 percentage points from 95 than from 55%.
8 Headline in the newspaper Aftenposten 17.10.2003, referring to those who take solarium more than
once a month in a study of more than 100,000 women in Norway and Sweden.
9 The probabilities were 0.0024 and 0.0017, giving an absolute difference of 0.0007. This can be
calculated from information in table 5 in Veierød et al. (2003: 1536). The relative risk according to
these figures is 1.41, which after multivariate control for relevant variables increases to 1.55.
70 O. Hellevik

0,9 Proportion
Result for statistical measure
0,8
High on
dependent
0,7 variable
0,6
Lambda
0,5 T(i+1)
0,4 compared
to T(i)
0,3

0,2 PD
T(i+1)
0,1
compared
0 to T(i)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Time

Fig. 4 The behavior of lambda and the proportion difference in a situation with logistical growth
over time

It no longer makes sense to talk about a ceiling effect, when we are moving away from
instead of closer to 100%.
If the variable has a set of values that may be regarded as symmetrical (e.g. being in
favor of or opposed to a certain policy), a decline in support for one alternative may
instead be seen as an increase for the other. In this way all changes may be defined as
increases. The above discussion thus applies to variables with asymmetrical value sets,
such as level of education, income, voting, holding political office, etc., where some
values are regarded as more attractive or rewarding than others.

10 Dependent variable with a fixed overall distribution

In processes of change the level of the dependent variable at a certain point in time
among other things may depend upon what the level was at prior points in the process.
This is an argument for letting the level influence the result, as is the case for loglinear
measures. What then when we compare groups at the same point in time? In some
cases one may argue that there exists some form of coupling between the levels of the
dependent variable within the various groups, namely when the total distribution for
the dependent variable is fixed.
An example of a fixed distribution might be the proportion of a population obtain-
ing higher education, in a situation where a limited number of positions is provided,
and this number does not meet the demand. If we were told the number of educational
positions obtained by men, we would also know the number available to women, given
that the total number is fixed. In this situation there will be limits to the size of the
difference in educational attainment between the sexes.
Let us assume that there are positions of higher education for 40% of those seeking
such a position. Then the maximum possible value for PD would be 0.80.10 This is

10 For simplicity we assume the distribution on sex is exactly 50–50 for those seeking a higher educa-
tion.
Linear versus logistic regression when the dependent variable is a dichotomy 71

1,4

1,2 Lambda
(when PD is
Result for statistical measure

0.2)
1

0,8 PD/MaxPD:
50-50 on
independent
0,6
variable

0,4 PD/MaxPD:
25-75 on
0,2 independent
variable

0
10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90
Percentage with high value on dependent variable in total

Fig. 5 Comparison of lambda and the proportion difference standardized by taking maximum value
into account (for two different distributions on the independent variable)

seen if we “give” all positions to one of the groups, for example men. The proportion
with a higher education then becomes 0.80 for men, and 0.00 for women.
If the provision of positions of higher education equals 70% of the applicants, the
maximum sex difference is 0.60 (in this case with 100% attainment among men, and
40% among women, when they are given the remaining positions).
In this situation it makes sense to see the actual difference in relation to the max-
imum possible, to obtain a measure standardized to vary between 0 and ±1. This
maximum-corrected PD is found by dividing the actual PD with the maximum possi-
ble given the fixed distribution on the dependent variable. If for instance we have an
actual sex difference of 0.40, the maximum-corrected PD would be 0.4/0.8 = 0.50 in
the first example above, and 0.4/0.6 = 0.67 in the second.
It has been suggested that loglinear measures of association provide such a cor-
rection for variations in the marginal distribution on the dependent variable (Heath
et al. 1987: 270). It turns out, however, that the results for lambda deviate somewhat
from the maximum-corrected PD (Hellevik 2002), as shown in Fig. 5.
The deviation is especially large when the distribution on the independent variable
is uneven. The size of lambda is not influenced by this distribution, while the maximum
possible difference is. If we want to correct our measure of association for limits to its
maximum value given the fixed distribution on a dependent variable, other measures
are better suited than the loglinear ones.
When the distribution on the dependent variable is not fixed, the proportions of
other groups do not restrict the proportion with a certain value on this variable within
a group. Whatever the percentage among men who are pro Norwegian EU member-
ship, this does not in any way restrict the level possible for such an attitude among
women. The maximum difference possible in this case would be 100%. In such a
situation it is hard to see what the substantive justification could be for regarding
the difference between 60 and 40% as a much smaller association than the differ-
ence between 90 and 70%, which is what the loglinear measures suggest (lambda
equal to 0.20 and 0.34, respectively). But if we choose to apply these measures,
72 O. Hellevik

including the logistic regression coefficient, we implicitly assume that it is mean-

ingful to distinguish between these situations that according to linear measures are
identical.
To accentuate this point: We would have to argue why it makes sense to regard
a difference in the tendency to vote between 35% for women and 65% for men as
a much smaller association than the difference between participations rates of 99.90
and 99.99%, since lambda is 0.31 in the first and 0.58 in the second case.

11 Other perspectives on the interpretation of loglinear measures

This discussion does not exhaust the list of possible substantive justifications for using
non-linear models in analyses with dichotomous dependent variables. In studies of
choice situations economists use the so-called probit model, which is similar to the
logistic model.11 The underlying assumption is that choices result from cost-benefit
calculations which are not immediately observable, but which leads to a shift in value
on the dependent variable when benefits surpasses costs. Using the probit model
implies the assumption that the unobserved variation in preferences is normally dis-
tributed (se for instance Greene 1993: 642).12
This perspective is applicable in other situations where change of value on a dichot-
omous dependent variable may be seen as the result of unobserved characteristics or
propensities of the units under study. If it seems reasonable to assume that these
variables are normally distributed, the independent variables will have a diminishing
effect on the probability for a change in the dependent variable the closer we move
towards the extreme values of 0 and 1.

12 The purpose of the analysis

When deciding which statistical technique to use the purpose of the analysis ought to
be an important consideration. A decisive distinction is the one between prediction
and causal analysis (Hellevik 1983, 1984). Table 3 gives a brief overview over the
differences between the two approaches to multivariate analysis.
The use of loglinear measures is restricted to prediction analysis; they cannot be
used for causal purposes (path analysis). The effects found will not give a sum of
components identical to the bivariate association one wishes to interpret causally
(Davis and Schooler 1974; Reynolds 1977). A correct decomposition in this sense
can only be obtained by using linear measures. With linear regression coefficients
the sum of components will always correspond to the bivariate association, whether
the variables are binary or continuous. Another possibility is to use weighted aver-
ages of proportion differences from tabular analyses (Boyle 1966; Hellevik 1983,
1984).13

11 While the logistic model is based on the logistical distribution the probit model is based on the
better-known normal distribution. Both distributions have approximately the same bell shaped form.
The logistical growth curve discussed earlier is a cumulative logistical distribution.
12 Point communicated to me by Axel West Pedersen.
13 With more than three variables in the model there is, however, a possibility for minor deviations
from the bivariate association of the sum of components (Hellevik 1984: 168–170).
Linear versus logistic regression when the dependent variable is a dichotomy 73

Table 3 Prediction and causal analysis

Prediction analysis Causal analysis

Purpose Predict the value on Y (dependent Explain bivariate association between

variable) from knowledge of value on pair of variables by decomposing it into
set of X’es (independent variables) direct, indirect and spurious effect
Model Two tiered: Hierarchical: Causal ordering
Two or more X, one Y X1 > X2 > X3 > X4
Design for Multivariate table / regression equation Set of tables / regression equations
analysis with all X’es and Y (two alternative designs)
Calculation of Variable and interaction effects. Variable effects.
effects (In tabular analysis: unweighted average (In tabular analysis: weighted average
of conditional effects and differences of conditional effects.
between conditional effects.
In regression analysis: equation with In regression analysis: equation without
interaction terms) interaction terms)
Requirement Predicted value close to actual value for Sum of components identical to
for result dependent variable bivariate association
Statistical Percentage/proportion table analysis Percentage/proportion table analysis
techniques Linear regression Linear regression
Loglinear analysis of tables.
Logistic regression

13 Conclusion

The statistical arguments against the use of linear regression with a binary depen-
dent variable are not as decisive as it is often claimed. Even if the homoscedasticity
assumption is violated, this in practice has little effect on the outcome of significance
tests. The results for linear and logistic significance probabilities as we have seen turn
out to be nearly identical, even with small samples and skewed distributions on the
dependent variable. The problem of impossible predictions may be avoided by includ-
ing interactions terms in the regression equation and/or by transforming continuous
variables into dummy variables.
This means that we are not restricted to use logistic regression with a binary depen-
dent variable. This is fortunate, since there in many cases are compelling substantive
arguments for preferring the linear approach.
Proportion differences and coefficients from a linear regression analysis of binary
variables may be interpreted as the difference in probability for having a certain value
on the dependent variable for units with different values on an independent variable
(in a multivariate analysis controlled for other independent variables). This is an
interpretation that is substantively meaningful and easy to comprehend. The same
cannot be said for the loglinear measures of association.
The intuitively meaningful interpretation makes it easier to communicate research
results to broader audiences of interested citizens. Even without statistical training
they may be able to comprehend and critically evaluate the conclusions drawn by
researchers. For the natural sciences it may not be a problem that results are acces-
sible only for specialists. But for the social sciences where the phenomena under
study often will be of great concern for many outside the scientific milieu, it must be
seen as an important consideration that the statistical techniques should not be more
complicated than absolutely necessary.
74 O. Hellevik

The linear measures can be used to answer causal questions, which often is what
the researcher wants to do. Loglinear measures are not applicable in causal analyses,
since they do not provide an accurate decomposition of bivariate associations.
With proportions near 0 or 1 on the dependent variable, where the two kinds of
measures show conflicting results, those from loglinear analyses in some cases appear
to give little substantive meaning.

References

Boyle, R.P.: Causal theory and statistical measures of effect: a convergence. Am. Sociol. Rev. 31, 843–
851 (1966)
Davis, J.A., Schooler, S.R.: Nonparametric path analysis — the multivariate structure of dichotomous
data when using the odds ratio or Yule’s Q. Soc. Sci. Res. 3, 267–297 (1974)
Fox, J.: Applied Regression Analysis, Linear Models, and Related Methods. Sage Publications, Thou-
sand Oaks, CA (1997)
Greene, W.H.: Econometric Analysis, 2nd edn. Macmillan Publishing Company, New York (1993)
Heath, A., Jowell, R., Curtice, J.: Trendless fluctuation: a reply to Crewe. Polit. Stud. 35, 256–277 (1987)
Heath A., Mills C., Roberts J.: Towards meritocracy? Recent evidence on an old problem. In: Crouch
C., Heath A. (eds) Social Research and Social Reform: Essays in Honour of A.H. Halsey. Claren-
don Press, Oxford (1992)
Hellevik, O.: Decomposing proportions and differences in proportions: approaches to contingency
table analysis. Qual. Quant. 40, 79–111 (1983)
Hellevik, O.: Introduction to Causal Analysis. Exploring Survey Data by Crosstabulation. George
Allen & Unwin, London (1988. Oslo: Norwegian University Press) (1984)
Hellevik, O.: Fagkritikk av oppdragsforskning. Sosiologisk tidsskrift 4, 219–228 (1996)
Hellevik, O.: Class inequality and egalitarian reform. Acta Sociologica 40, 377–397 (1997)
Hellevik, O.: A less biased allocation mechanism. Acta Sociologica 43, 81–83 (2000)
Hellevik, O.: Inequality versus association in educational attainment research:. comment on Kivinen,
Ahola and Hedman. Acta Sociologica 45, 151–158 (2002)
Hosmer, D.W., Lemeshow, S.: Applied Logistic Regression. John Wiley & Sons, New York (1989)
Kanagy, C.L., Humphrey, C.R., Firebaugh, G.: Surging environmentalism: changing public opinion or
changing public? Soc. Sci. Quart. 75, 804–819 (1994)
Kivinen, O., Ahola, S., Hedman, J.: Expanding education and improving odds. Participation in higher
education in Finland in the 1980s and 1990s. Acta Sociologica 44, 171–181 (2001)
Lægreid, P., Olsen, J.P.: Byråkrati og beslutninger: En studie av norske departement. Universitetsforla-
get, Oslo (1978)
Reynolds, H.T.: The Analysis of Cross-Classifications. Free Press, New York (1977)
Rothman, K.J.: Modern Epidemiology. Little, Brown and Company, Boston (1986)
Rothman, K.J., Greenland, S. (eds.): Modern Epidemiology, 2nd edn. Lippincott-Raven Publishers,
Philadelphia (1998)
Rubin, D.B.: Estimating causal effects from large data sets using propensity scores. Ann. Internal
Med. 127, 757–763 (1997)
Skog, O-J.: Å forklare sosiale fenomener. En regresjonsbasert tilnærming. Ad Notam Gylden-
dal, Oslo (1998)
Veierød, M.B., Weiderpass, E., Thörn, M., Hansson, J., Lund, E., Armstrong, B., Adami, H-O.: A
prospective study of pigmentation, sun exposure, and risk of cutaneous malignant melanoma in
women. J. Nat. Cancer Inst. 95, 1530–1538 (2003)

View publication stats