Concept

Ratio
Proportion
Rate
Risk
Prevalence

Cumulative/Crude
Incidence (CD)
Incidence Density (ID)
Morbidity rate
Mortality rate
Case-fatality
Attack rate
Years of Potential Life
Loss (YPLL)

E
UE

O
a
c

O
b
d

Equation
Event or People (A)/ Event of People (B)

= +

#

#
#
#
#
Point = specific time pt; period = time interval OR
Incidence x Disease Duration
#
( )
#

#

#

#
#
#

Age at death predetermined age at death
(predetermined standard = 65)

Units
None
%

Use
Descriptive

Personyrs
%

Frequency of
event

Burden disease
in popn

Personyrs
Personyears

years

Disease
causation

Premature
mortality index

Pros/Cons

Affected by survival; no measure

risk; mix chronic/acute cases
Assume entire popn followed
through; always smaller than ID
Not include time not followed up

Research/resource priorities,
surveillance trends/interventions

Epidemiological Study Designs

Descriptive Describe health events with attention to person, place, time. Generate hypothesis and resource allocation.
Analytic Examine associations with methodological rigor. Observational = cohort & case control. Experimental = RCT

Concept
Relative Risk (RR)
Attributable Risk/Risk
Difference (AR)
Attributable Fraction
Exposed (AR%)
Population
Attributable Risk
(PAR)
Attributable Fraction
population (PAR%)

Odds Ratio (OR)

Mantel-Haenszel
Summary Odds Ratio
(ODMH)

Equation

= +

+
RateE RateUE =

+ +

100

1
= 100

Incidencetotal
IncidenceUE
OR
AR x PrevelanceE

100

+
( 1)

=
+
( 1) + 1

=
/

Use
+
Likelihood devng Outcome (O ) in Exposed
(E) group relative to Unexposed (UE) group;
ratio incidence in E vs. incidence in UE
Rate of O that can be attributed to the E in
the E group

Interpretation
Those with [E] are [RR%] [more/less] likely
to develop [O] than those with [UE]

Proportion of the disease amongst the E that

is attributable to the E

[AR%] of [O] amongst the [E] group can be

attributed to [E]

Excess rate of O in the total population that

is attributable to the E

[PAR]excess cases of [O] per [x ppy] in the

population can be attributed to [E]

Proportion of O in total popn which is

attributable to E; assume causation

[PAR%] of [O] in the population can be

attributed to [E]

Ratio of odds of E in O to odds of E in O ; for

case-control studies. No change with O #; no
+
change if examine % E in O /O . Equals RR if
O rare (prevalence <5%)
Used for stratified case-control studies if OR
b/w strata are similar.

Causal Relationships
Direct: factor -> disease; no intermediates
Indirect: factor-> A -> disease; factor causes disease through step(s)
Necessary: no factor = no disease; w/o factor disease never develops
Sufficient: factor -> disease. w/ factor the disease always develops

[AR#] cases of [O] per [x ppy] among the

[E] group can be attributed to [E]

Cases were [OR] times more likely than

controls to have been exposed
Cases were [ORMH] times more likely than
controls to have been exposed after
adjusting for [confounder]

Hillis Causal Criteria

1. Strength large effect
2. Consistency repeated observations in different settings
3. Specificity cause leads to single effect
4. Temporality cause precedes effect
5. Biologic gradient dose response relationship
6. Plausibility biologic
7. Coherence no conflict natural history/biology
8. Analogy
Direct Standardization (Type 1)
Requires: (a) age specific rates from study population
(b) age distribution from standard population
Represents what crude rate would be if study population
had same age distribution as standard
Cons: adjusted rate not meaningful for population, not suitable for
resource allocation, choice of standard affects comparison
1. Calculate age-specific rates for each study population (a)
2. Choose a standard population
a. Reference Population
b. Average of Study Populations
3. Calculate % per age category (%) as decimal
4. Calculate age-standardized rates for each population
[(a1)(%1)+(a2)(%2)+(a3)(%3)]
Indirect Standardization (Type 2)
Requires: (a) age specific rates from standard population
(b) age distribution of study population
Outcome= expected events; represent # occurring if study
population had same rates as standard population & SMR
Standardized Mortality Ratio (SMR)
= observed deaths/expected deaths (= 1.71)
If [panama] had the same [age]-specific rates as [Sweden]
then wed expect [1.71] times the mortality
No need category specific rates in study population
Useful when dealing with small number events and
unstable rates in study population; when no internal
comparison group
Pro: summary measure, statistical stability, minimize confounder efx
Con: not useful for multiple comparisons due to differing popn
structures, rate no longer meaningful by itself
Proportionate Mortality
= proportion death due to factor/ total deaths in popn
Not useful because competing values (more deaths due to
factors means less death due to other factors)
When population structure unknown, info from death
certificates
Incidence Density Ratio
= ID(Exposed)/ ID(Unexposed)
For cohort studies with varying follow up with ppy in den.
Case-Controls Studies
Use When
Disease with very long latency periods
Rare diseases
Wide ranges of exposures in single study
Case Selection
Incidence, dont use prevalence b/c:
Difficult determine if factor related to disease
occurrence/duration
No temporality
Prevent longer time for recall bias
Control Selection
Select to represent popn which wouldve been included as
case had they developed disease
Represent frequency of exposure in underlying source
popn (may differ from general popn)
Characteristics and sources of cases (comparable)

Individual Matching
Each ctrl matched to case, matched analysis
For each case select one or more ctrls with same
characteristics on potential confounders
Pro: ctrls factors difficult to measure; easier obtain comparable ctrl
group, gain precision of OR estimate (tighter CI)
Cons: complexity in ctrl accrual; info from ctrls on matching variable
need obtain before study inclusion (need screen more), matching on
many variables difficult, cant study matched variable, decrease OR
precision if not true confounder; only small gain if factor not strong
for disease
Frequency Matching
Select controls with similar distribution of confounder
Control in analysis
Analysis of Individually Matched Data
Each pair (case-ctrl) contributes to one observation/count
OR = #(CasesE & ControlsUE)/ # (CasesUE & ControlsE)
Stratification
Stratify by confounder, examine OR within lvls of
confounder
Want summary estimate, estimate of risk adjusted for
effects of confounder
Use ORMH when strata OR similar
Factor is true confounder if adjusted OR and unadjusted
OR differ by greater than 10%
Cohort Studies
Uses
Rare exposures
Multiple effects of single exposure
Identify temporal sequence
Expense follow-up not an issue
Fixed Cohorts: identify popn at time and no include more eligible
Dynamic Cohorts: open popn and includes ppl who enter later
Changes in exposure over time: re-classify during study or allow
exposure status to vary in analysis
Internal Comparisons
Study gradient (D-R) of disease
Variation with amount of exposure (often no unexposed)
External Comparisons
Estimate disease incidence in exposed group in absence of
exposure
As similar as possible to exposed group
Follow-Up
1. Length needs to be considered in design
Base apriori knowledge of time needed for disease show
Induction time (to induce) & latency time (express/detect)
Usually no know induction/latency; try estimate
2. Attempt high levels of follow up (prevent loss)
Be persistent
Bias
1. Selection (systematic differences b/w E & UE groups)
2. Information (Misclassification, measurement error)
3. Non-participation (systematic reason for non-Ps?)
4. Attrition (differential loss to follow-up)
5. Healthy Worker Effect (decrease O+ in worker)
Use internal comp, external comp of workers, artificially
adjust risk (inflate)
Nested Case-Control Study
Conduct cohort but no full evaluation of exposure
After follow-up, conduct case-control within cohort
Cases = identified cases of disease in cohort
Ctrls= sample of cohort free of disease at time of cases
Analysis as case-control

Randomized Control Trials

Therapeutic: conducted with diseased ppl (diminish Sy, prevent
recurrence, decrease mortality risk)
Preventative: disease-free ppl (decrease risk, inds or communities)
Efficacy Trials (Explanatory)
Does Tx work under ideal circumstances?
Tx more harm than good?
Only patients who cooperate
Effectiveness Trials (Pragmatic)
Does Tx work in ordinary settings?
Offer Tx to subjects and let them reject or accept
Intention to treat analysis
Failure Tx effect may be due lack efficacy or subject accept
Blinding (Masking)
Observers/subjects kept ignorant of group subject
assigned to
Avoid bias
Single blind = subject
Double blind = subject and interviewer/evaluator
Triple blind = subject, evaluator and analyst
Noncompliance
Potentially due to randomization, drop out, stop following
prescribed Tx
Build in checks to ensure compliance
Phases of RCTs
1. Tx any effect (pharm/tox)? All get Tx (single arm)
2. What dose achieves effect (efficacy)? What toxic effects
observed with Tx (safety)? Single arm
3. Large scale RCT for effectiveness and safety (work in ideal?
Ordinary?) Usually 2 or more arms
4. Post-marketing surveillance, cost, benefit, etc
Randomized Community Trials
Why?
Public health: interventions at this level
Feasibility: individual trials expensive
Benefits: intervention under control of experimenter so
can adjust for differences in communities
Community Selection/Recruitment
Size (expense) vs. statistic, similarity of communities, favourable
community relations, accessibility, consent, communication plan
Baseline Surveillance
Selection of outcomes, key population characteristics, approaches to
data collection, comparability of baseline and follow-up measures
Development of Interventions
Protocol (intervention and ctrl arm), Options (education, municipal
policy), random assignment
Data Collection and Analysis
1. Periodic Surveillance (outcomes, intermediate outcomes,
potential side effects)
2. Evaluation (intervention effective? Adjust community
differences, assumption of independence)
Natural History of Diseases
Normal: No disease, before onset (1 prevention, remove cause)
Preclinical: B/w bio onset and Sy appearance (2prevention, screen)
Clinical: Sy to disease outcome (3 prevention, Tx)
Lead time: Detected by screening/Dx to usual time for Dx
Screening
Early detection disease
Not diagnosis; pos+ screen = diagnostic tests after
Often for disease with long latency periods
Improve outcome of disease amongst those affected
Conditions for Screening
1. Long detectable preclinical phase
2. High prevalence amongst screened popn (cost/benefit)

3.

Seriousness (cost effectiveness for reduce mortality;

consequences fail detect vs. risk/discomfort of screen)
Measurement Validity
Degree to which method used correctly categorizes
False Pos+: # screen is pos+ but diagnosis is negFalse Neg-: # screen is neg- but diagnosis is pos+
True Disease Present No Disease
Test Positive
a
b
Test Negative
c
d
Sensitivity: probability testing pos+ if disease truly present
= a/(a+c)
Specificity: probability testing neg- if disease truly absent
= d/(b+d)
Depending on cut off, increase sensitivity or specificity
Everything pos+ = 100% sensitivity, 0% specificity and v-v
High Consequences for both false pos+ and negMissing a case (false neg-): increase sensitivity (decrease false neg-)
Identifying non-case (false +): increase specificity (decrease false +)
Feasibility
Acceptable to popn being screened (uncomfortable)
Cost effectiveness (screen+ diagnostic tests, cost per case)
Yield of cases (predictive value of screen)
Predictive Value
Pos+ PV: probability person truly has disease given pos+ test
= a/ (a+b)
Neg- PV: probability person truly no disease given neg- test
= d/(c+d)
Measurement Reliability
Observed agreement (O) = (a+d)/N; no ctrl chance
2
Expected agreement (E)= [c(a+c) + (b+d)(c+d)]/N
Kappa = (O-E)/(1-E); perfect agreement K=1, only chance K=0
Fair agreement = 0.21-0.40
Evaluating Screening Programs
Effectiveness: screening effective to reduce morbidity/mortality
Short-term outcome: severity of disease at diagnosis
Mortality: compare screen and unscreened popns
Volunteer Bias: Systematic differences in comparison groups
Lead Time Bias: Increased time b/w diagnosis and death (survival)
purely due to earlier diagnosis; compare age-specific mortality
Length Bias: Amongst those w/ disease who are screened, may be
over-rep of those with long pre-clin phases (and maybe more
favourable prognosis/benign disease); may never have shown Sy
Prevalence and Screening Tests
Increase Prevalence:
No change in sensitivity and specificity
If rare disease: decrease PPV, and v-v
Measurement Error, Sensitivity & Specificity
+
O (Cases)
O (Controls)
Exposed
a
b
UnExposed
c
d
Non-differential error with respect to case-ctrl status with a
sensitivity of [80%] and a specificity of {90%}.
1. Calculate new cases and controls:
A=a[80%], B=b[80%],C= c{90%},D= d{90%}
2. Calculate differences between old and new numbers and
move these numbers to create the final measure:
Af=A+(c-C), Bf= B+ (d-D).
Hypothesis Testing
Null Hypothesis (H0): p0=p1 or OR=1.0
Alternative Hypothesis (HA): p0p1 or OR1.0
Assume null is true to begin with
2
Chi-Square (X ) = ( )2 /E

P-values
2
Square root of X to compare to standard distrubtn
Standard normal, p=0.06 (less than 0.05, reject null)
Represents probability of observing result at least as
extreme as that observed by chance alone
Interpretation of Significance Tests
Type 1 Error: Reject H0 when it is true
Area under 0.05 (alpha = 0.05 in 2-tailed test)
Type 2 Error: Fail to reject H0 when it is false
Significance Values
P-value function of sample size and effect magnitude
Confidence Intervals (CIs)
Range within which true effect magnitude lies
If CI includes 1.0, then p>0.05 (accept null)
If CI excludes 1.0, then p<0.05 (reject null)
Width of CI indicates variability of estimate (reflects n)
Power
Ability of study to demonstrate association if it exists
Probability of rejecting the null when it is false (1-Type 2 Er)
Estimating Power and Sample Size
Begin with assumption null is false
Increase OR = decrease n
Exposure proportion amongst controls (p0) move to
extreme = increase n
Decrease n = decrease power
P0 move to extreme = decrease power
Effect Modification (Interaction)
Change in magnitude (or direction) of measure of
association b/w E and disease according to value of X
Contrast confounding which distorts the measure of
association resulting from a mixing effect of E on disease
rd
with that of 3 factor
Confounding = bias in effect to be eliminated/avoided
Effect mod = description of effect to report
Deal with both using stratified analysis
Not similar OR b/w strata = effect mod (similar = pool)
Differences crude OR and adjusted OR suggests confound
If interaction between E and population subgroup exists,
results NOT generalizable
Regression Analysis
For prediction, confound control, and effect mod detect
General Linear Model GLM
Y=a0 + a1X1 + a2X2
Y= outcome
X1 = exposure
X2= confounder
Coefficients a1 and a2 provide estimate of effect of X1 and
X2 that are mutually unconfounded
Data limits determine number of variables included
All GLM have 3 components:
1. Random component (identified response variable)
Identifies response variable Y and selects probability distribn
Standard GLM treat all Y as independent
Binomial distribution: binary outcome (0 vs. 1)
Normal distribution: continuous outcome
2. Systematic component (specifies explanatory variable)
Linearly predictors on right hand side of model equation
Linear combination of explanatory variables = linear predictor
Some can be interaction terms (e.g. x3=x2x1)
2
Others indicate curvilinear effects (e.g. x2=x1 )
3. Link Function (specifies a function of expected value
[mean] of Y)
Connects random and systematic components

Order of Interaction Tests

Rothman
1. Stratified analysis
2. Determine which confounders to include
3. Determine shape of E-D relation
4. Evaluate Interaction
Kristman/Kleinbaum
1. Stratified analysis
2. Shape of relationship
3. Evaluate interaction
4. Evaluate confounding
Epidemics
The occurrence of health event cases in a population in
excess of normal expectancy
Relative to usual frequency of health event in same
population, area, at same time
Time may be short or long (epidemics can include acute
and chronic health events)
Approaches to Study
Outbreak prior to study: usually use retrospective
Case control
When do epidemics occur?
Something changes (enhanced exposure or transmission)
Common circumstances:
1. Those at risk travel to endemic area
2. Infected carrier makes contact with uninfected popn
3. Cultural changes (behrs)
4. Host susceptibility compromised
Types of Epidemics
Common source outbreaks:
Common pathogenic source, simultaneous E, varying time/place of E
Point Epidemics:
Common pathogenic source, simultaneous E, short/sharp epidemic
curve
Propagative Epidemics:
Host-to-host pathogen transfer, infectious agent grows/multiplies in
host, excretion of pathogen (physical or behr) by host
Mixed Epidemics:
Common pathogenic source, host-to-host transfer
Cessation of Outbreak
Source pathogen identified and eliminated, transmission prevented,
susceptible popn reduced/eliminated (vaccination), risk factor mod
Epidemic curves
Frequency distributions, number of cases vs. time
Common source, acute outbreaks
Sharp/defined curve, short duration, no secondary peaks
Exposure to common source
Drawn out curves, longer, no secondary peaks
Propagative, epidemic curve
Drawn-out, irregular peaks (1 &2), varies with host susceptibility,
depends presence of other risk factors
Mixed
Common followed by propagative
Steps in outbreak investigation
1. Define the problem
Compare past/present rates, list possible etiology, develop case def
2. Appraise existing data
Time: Meant intubation time (lit) time epidemic peak, subtract min
st
intubation time from 1 case and max from last case to get range
Place: spots on map. Person: identify role of host factors (strata)
3. Develop a hypothesis
4. Test hypothesis
Eliminate source, prevent transmission, eliminate popn (vac)
5. Conclusion and interventions