Principles of Epi-Exam 2

Observational-natural experiments

-Ecologic

-Cross-sectional

-Case-control

-Cohort

The hierarchy of experiments/evidence

1. Randomized Clinical Trials

2. Cohort

3. Case-control

4. Cross-sectional

5. Ecologic

6. Case series

Selection of study design

Dependent on:

- The research question

- Amount of information already known about the research question

- Practical considerations (feasibility)

- Ethical considerations

Ecological studies

Examine the exposure-disease association among aggregates of people – usually defined by geographic groupings

What is the unit of analysis for ecological studies

group

Groupings are defined by what

- Geographic area (e.g., county, state, country)

- Time (ecological trend study)

What are measures used in ecological studies:

- Summary measure of the frequency of exposure for each study unit (e.g., per capita cigarette consumption)

- Summary measure of the frequency of outcome for each study unit (e.g., CHD mortality)

- Joint distribution of exposure and disease within individuals residing in each study unit is unknown

2×2 table

DATA USED IN ECOLOGIC STUDIES OF EXPOSURE-DISEASE ASSOCIATIONS

The measure for ecological studies is a…

correlation coefficient represented by r

Correlation Coefficient

A measure of association often used in ecologic studies = r, the correlation coefficient

- Quantifies extent of linear relation between exposure and disease across units of analysis

- Reflects the direction (+ or -) and the strength of the group-level association between exposure and disease

Interpreting Correlation Coefficient

-For every unit change in level of exposure, the disease frequency

increases or decreases proportionately

-Values range between

+1 (perfect positive correlation) and

-1 (perfect negative/inverse correlation)

Pearson Correlation Coefficient

• The sign (+ or -) of the correlation coefficient provides important information.

• +

  • high values of one variable are associated with high values of the other variable.

• -

  • low values of one variable are associated with high values of the other variable.

Strengths of correlation coefficient

• Correlations from 0 to 0.3 indicate poor or no relationship

• Correlations from 0.30 to 0.50 reflects a fair degree of relationship

• Correlations from 0.6 to 0.8 a moderate degree of relationship

• Correlations at least 0.8 reflect a strong relationship

What are 3 useful scenarios for ecological study design?

1. As a first look at a possible association between exposure and disease at the group level to generate a hypothesis

2. Only aggregate info is available on the exposure of interest

3. Exposure of interest only varies at population level (no individual variation within group)

What are the 3 types of data used in ecologic studies?

1. Incidence

2. Prevalence

3. Mortality (used the most often)

What are some advantages to ecologic studies?

1. Makes use of existing data

e.g. vital statistics, cancer statistics, pollution level

2. Does not require direct contact with individual subjects

e.g. reduce the risk from performing test/giving treatment to study participants; expedited/exempt IRB

3. Can be conducted quickly and inexpensively

Through using existing data and no direct contact

4. Essential for studying ecologic associations/population-level relationships

e.g. large-scale comparisons, health policy

The ecologic fallacy entails the following:

• Occurs when one draws inappropriate conclusions regarding relationships at the individual level based on group level (ecologic) data

• Error in inference due to failure to distinguish between correlation of an exposure and disease within a population versus within individuals

• May be ascribing to members of a group characteristics they in fact do not possess as individuals

What are the limitations to ecologic study?

1. Group data, not individual data (cannot infer results to the individual)

2. Information on confounders generally not available (cannot adjust for confounding variables)

3. Do not demonstrate causal relationships (can only be consistent with or not consistent with a causal association)

Correlation DOES NOT IMPLY causation because why?

Because there is possible a third confounding variable that explains the associations (e.g. high ice cream sales are associated with high homicidal rates b/c the heat is a confounding variable)

True or false: A cross-sectional study can be analytical or descriptive

True

What is a cross-sectional study?

a.k.a. a Prevalence Study because it provides “snapshot” of population at a specific point in time and looks for the coexistence of exposure and outcome in individuals at the same point in time

What are some features or a cross-sectional study?

• Exposure and disease measures made at the individual level

• Exposure and disease status determined at the same point in time for each participant

• Cases identified are prevalent cases

• Often used for chronic diseases (extended pre-clinical period and long duration)

• Seroprevalence studies

• Not useful for rare diseases or those of very short duration

• Often sample general population – these are generalizable

What are the uses of a cross-sectional study?

• Estimate of the magnitude and distribution of a health problem – measures the prevalence proportion

• Hypothesis generation

• Intervention planning

Design of a cross-sectional study invovles:

DATA EMPLOYED FOR CROSS-SECTIONAL STUDIES OF EXPOSURE-DISEASE ASSOCIATIONS

Prevalence of disease among exposed: a/a+b

Prevalence among unexposed: c/c+d

DATA EMPLOYED FOR CROSS-SECTIONAL STUDIES OF EXPOSURE-DISEASE ASSOCIATIONS

Proportion of exposed who have the disease: a/a+c

Proportion of exposed who do not the disease: b/b+d

Selection of study population

1. Selection based on a particular characteristic (e.g., ethnicity)

Example: If you wanted to compare the prevalence of coronary

heart disease (CHD) and markers of inflammation in Native

Americans, you could sample this population and then do an

ECG and blood test on each person in the sample. Prevalence

proportions of CHD would be compared in those with high and

low CRP.

2. Selection without regard to characteristic or exposure.

Example: If one wanted to examine relation between prevalence of

hypertension and body weight in Oklahoma County, a random sample

of households would be chosen and household members would have

their blood pressure, height and weight measured and be interviewed

regarding treatment for hypertension.

Prevalence proportion = number of cases of hypertension divided by

the number of persons examined. Compare by BMI category.

Ways to assess for exposure

• Questionnaires – e.g., smoking status, employment, medication use

• Records – e.g., current medication use

• Lab tests – e.g., blood lead, serum cholesterol

• Physical measurements – e.g., blood pressure, body mass index

• Other special procedures – e.g., toenail clippings for selenium

Exposure issues in cross-sectional studies

• Date of onset of exposure – may not be captured

• Past exposure – may not be captured

• Duration of exposure – may not be captured

• Only exposures that do not change or those that “track” – e.g., genetic factors, serum levels of persistent pesticides - can be considered as preceding disease

Measurement of disease

Methods of measurement

• Questionnaire/medical records

• Physical exam

• Special procedures, e.g., lab tests

Defined diagnostic criteria

• Objective measures

• May divide cases into definite, probable and possible disease categories

Cross-sectional study: to compare prevalence proportion

Prevalence proportion ratio (PPR)

What are the advantages to cross-sectional studies

• If based on random sample of general population, results can provide prevalence estimates for the entire population

  • Can estimate population prevalence proportions of disease and exposure

  • e.g., what is the prevalence of NSAIDs use in residents of Oklahoma County?

• Provides preliminary examination of disease –» exposure association

• Relatively short time needed for study, lower costs

What are some limitations to cross-sectional studies?

• Temporal sequence not determined (exception = if exposure is permanent factor)

• Potential for biased inference in prevalent cases (are they like incident cases?; P ~ I x D) e.g., survival bias

• Cases in remission or treated may be missed

• Not appropriate for rare diseases or diseases with very short duration

How might bias be introduced?

• In cross-sectional studies, prior exposure status is usually not determined – only current status

• People may change exposure status over time, so current status may not reflect past exposures

• Change in exposure status may be related to the outcome you are studying

• This leads to bias in the observed association

More on cross-sectional studies:

• Cross-sectional studies can suggest associations between exposure and disease (hypothesis generation)

  • Cross-sectional studies cannot establish causal relationships

  • No temporal sequence

• Prevalence cases are a function of both disease incidence (etiology) and disease duration (survival)

• Special circumstance when cross-sectional study data can evaluate etiologic associations

  • Only when current values of exposure variables are unaltered over time

    • factors present at birth (eye color, blood group, HLA antigens)

What is the temporal sequence problem?

If you wanted to know whether or not obesity predisposes to osteoarthritis of the hip, in a cross-sectional study a positive association may mean:

1) that obesity does predispose to osteoarthritis,

2) or it may mean that people who develop arthritis exercise less and put on more weight.

A weight history is necessary to determine weight before arthritis developed.

What is the problem with prevalence/incidence bias?

If a cross-sectional study showed a lower prevalence of dementia in users of NSAIDs compared to non-users, it could mean:

1) that NSAIDs protect against developing dementia (influenced etiology),

2) or that persons with dementia who used NSAIDs were more likely to have died than persons with dementia who did not use NSAIDs (influenced survival), thus making NSAIDs appear protective

True or False: An ecological study is higher on the "Hierarchy of Evidence" list when compared to cohort studies.

False

What is the key-characteristic of a cross-sectional study design?

Data collected at a single point in time

How do you calculate prevalence in a cross-sectional study?

Number of individuals with condition divided by total in group

True or False: Cross sectional studies tend to identify cases of long duration since people who die or recover quickly aren't found

True

What is the primary limitation of ecological studies?

They are susceptible to ecological fallacy

What would be the best way to measure exposure to secondhand smoke in children in a cross sectional study?

Measuring a nicotine biomarker in a child's saliva

What are case-control studies?

• Selects the study population on the basis of disease status

• A case-control study begins with people who have the disease (cases) and compares them to people who don’t have the disease (controls)

• Compare the odds of past exposure to a suspected risk factor between cases and controls (retrospective/goes back in time)

An example of a case selection:

Suppose we are interested in investigating an association between childhood cataracts and exposure to rubella virus in utero.

• Cases would be children with cataracts

• Controls would be children without cataracts

• For each child we would determine whether or not their mother was exposed to rubella during her pregnancy with that child

Issues in case selection: definition

Diagnostic criteria – Clearly defined, objective, standardized criteria

• Objective: to produce a uniform, homogenous group of cases

Example: In a case-control study of preterm delivery, cases would be

identified as babies born < 37 weeks of gestational age as defined by first

trimester ultrasound measurement of crown-rump length.

Issues in case selection: definition

Criteria for eligibility – Clearly defined reasons to include/exclude cases

• E.g., by age, gender, potential for exposure

• Apply equally to cases and controls

Example: A study of recent oral contraceptive use and myocardial infarction would exclude males and post-menopausal and surgically sterilized women because they have no risk of recent exposure. Including them could bias the results towards the null.

Why do we study incident cases for case-control vs prevalence?

• Reduce potential for Incidence/Prevalence bias (P≈ IxD etiology v. survival/duration)

• Diagnoses more likely to be uniform, using same criteria

• Recall of exposure may be better

Case selection

Cases are selected without reference to (without knowledge of) their exposure history

Sources of cases include:

• Hospital-based

• Community or population sample

• Registries which are not population-based

• Large pre-paid insurance plans

• Retirement communities

Advantages vs disadvantages of hospital-based

Hospital-based: all cases admitted to a single hospital or

group of hospitals within a given time period who meet the eligibility criteria

Advantages:

• Easier

• Cheaper

Disadvantages:

• Potential for biased sample of cases

• Referral patterns

• Only suitable for diseases that are usually hospitalized

Advantages vs disadvantages of community/population samples

Community or population sample of cases: include all (or random sample) of cases in defined geographical area within specified time period; use registry if it exists

Advantages:

• Representative

• Case group

Disadvantages:

• Costly

• Time-consuming

• Difficult to do without registry

Selection of Controls

• Purpose: Controls provide a comparison group for cases

• Definition: Controls are intended to represent the frequency of exposure in the population from which the cases arose

• Controls are selected from the same source population as the cases were chosen

• Controls are free of disease under study

• Controls usually similar to cases with regard to past potential for exposure, during same period of risk under study

• Usually select controls in same manner as cases selected in order to select from the same source population

• Eligibility criteria – any exclusion criteria applied to the cases also applied to the controls

Example for control selection

A case-control study of tonsillectomy (exposure) and lymphoma (outcome) conducted in Iowa would select cancer cases from the statewide SEER cancer registry.

• Because cases are identified on a statewide basis, controls should also be chosen on statewide basis.

• The “source population” for cases is the State of Iowa.

True or False: Controls are selected without reference to (without knowledge of) their exposure history

True

Sources of controls include:

• Hospital-based

• Community/general population samples

• Neighborhood

• Multiple types

What are community/general population sample controls?

• Controls are selected from a random sample of the general population

• Random-digit dialing commonly used

• Appropriate if cases are population-based

What is hospital-based control

• People seeking medical care at same institution as cases for conditions unrelated to disease under study

• Exclude persons with diseases known or suspected to be related to exposure under study (ex: smoking and MI; emphysema)

• The illness of the controls should have the same referral patterns to the health care facility as that of cases

• May use multiple diagnoses

What are neighborhood controls?

Controls selected from same neighborhood as cases through canvassing or use of phone number

What are multiple type controls?

Used to assess potential biases (e.g., recall) or to compensate for the deficiencies of other types of controls

Advantages vs Disadvantages of Hospital-based controls sources:

• Advantages: Captive population, clearly identified, economical method, less recall bias

• Disadvantages: Potential for selection bias, less generalizable

Advantages vs Disadvantages to community/general population control sources:

• Advantages: Highly representative; calculate population frequency of exposure; appropriate for population-based cases

• Disadvantages: Costly, problems with refusal and phone coverage, increase in cell phone usage

Advantages vs Disadvantages to neighborhood control sources:

• Advantages: Provides controls of similar socioeconomic status and environment as cases

• Disadvantages: Overmatching possible, low response rates

Advantages vs Disadvantages of multiple type control sources:

• Advantages: internal replicate of study; assess certain types of bias

• Disadvantages: costly and time-consuming; if results differ, must be able to explain

An example of a case-control study:

A case-control study of brain tumors in children included two

different types of controls:

• Children with other solid tumors (“sick controls”)

• A random sample of healthy children from the same birth cohorts as cases (community controls)

Purpose: to assess results for potential recall bias

Validity

Validity – accuracy of the data

• Internal validity - the study was done properly and the results are valid in the study population

• External validity – findings of the study are generalizable (representative) from the study population to the target population

Representativeness

Relates to the generalizability of the study (external validity)

• Advantages: increase generalizability, reduces potential for certain types of selection bias

• Disadvantages: may be more difficult, time consuming, resource intensive

Comparability

Internal validity: cases and controls have equal probability of past

exposure if there is no association between exposure and disease.

• Advantages: easier to detect smaller differences; reduces chances for unmeasured confounding

• Disadvantages: may have select group of cases, potential for Berkson’s bias (hospital-based studies only)

Assessment of Exposure

• Exposure is prior to disease onset (or reference date for controls)

  • Techniques: Information on prior exposure in cases and controls may be ascertained through personal interview, hospital or medical records, employment records, pharmacy or lab records, or direct measurement

• Timing of exposure – when did it occur in relation to disease onset or index date?

• Quantification of exposure for dose-response analysis

  • Includes amount, duration and frequency of exposure.

Example of assessment of exposure for a case-control

Weinmann et al. (1994) conducted a case control study of renal cell cancer in relation to prior use of antihypertensive medications within the membership of Kaiser-Permanente Northwest.

For cases and their matched controls, outpatient and inpatient medical records were reviewed for information regarding medication use up to a date three months prior to the case’s diagnosis

Measure of Association for case-control study

Odds Ratio (OR)

• Ratio of two odds (the odds that cases were exposed ÷ the odds that controls were exposed)

Example of calculating odds:

• If there are 100 smokers and 60 develop a chronic cough, the probability of smokers developing a cough is 60 / 100 = 60%

• The probability of smokers not developing a cough is 100% - 60% (or 40 / 100) = 40%

• The odds of developing a cough are 60:40 or 1.5

• Note how the odds of developing a cough differs from the probability of developing a cough

DATA USED IN CASE-CONTROL STUDIES OF EXPOSURE-DISEASE ASSOCIATIONS

CASE-CONTROL STUDY 2×2 table

CASE-CONTROL STUDY 2×2 table OR calculations

How to interpret OR in the example used (remember its odds of exposure not disease)

The odds of having smoked cigarettes among cases of lung cancer are 14.0 times greater than the odds of having smoked cigarettes among controls.

Interpretations of OR

Risk Ratio

• You cannot, in most circumstances, use case-control data to calculate incidence and thus, you cannot directly calculate risk ratios from case-control data

• Study groups are chosen on the basis of presence of disease, not exposure

• Odds ratios are good approximations of the risk ratio if your study is properly designed

Matching of cases and controls

• Definition: The process of selecting controls so that they are similar to cases for characteristics, such as age, gender, race, or socioeconomic status, that might be confounding variables.

• Matching is one of the ways to deal with confounding, by making cases and controls similar across these characteristics.

Types of Matching

• Individual (pair) matching – selecting one or more controls for each individual case

  • e.g., for a case of MI who was a white male aged 67 years, select one or more controls without MI who are white, male and 62-72 years of age –repeat for each case in the study

• Frequency (group) matching – the control group is selected so that its distribution is similar to that of cases for potentially confounding variables such as age, gender and race. Requires that all of the cases be selected first or that the distribution is known.

  • e.g., if 68% of your MI cases are male, 68% of your controls are chosen to be male

Example of Pair Matching

Only match b/c for pair matching

Explanation of matched pair analysis

• To calculate an OR for a matched case-control study with 1 case matched to 1 control (1:1 match) on a factor (sex as in our example): OR = b/c

• We divide the number of pairs in which the case, but not the matched control, is exposed (Case +, control - for exposure) by the number of pairs in which the control, but not the case, is exposed (case -, control + for exposure)

• Each cell in the 2x2 table on the previous slide represents the number of pairs in that cell

  • A: 8 pairs where case and the matched control are both exposed

  • B: 18 pairs where the case is exposed (8+ lbs), but the matched control is unexposed (<8 lbs)

    • Supports our hypothesis that positive exposure is associated with disease (high birth weight associated with brain tumors)

  • C: 7 pairs where case is unexposed (<8 lbs), but matched control is exposed (+8 lbs)

    • Data do not support our hypothesis

  • D: 38 pairs where case and matched control are both unexposed

• A and D are non-informative since cases and controls are alike regarding exposure

What are the advantages of case-control design?

• May be quicker and less expensive compared to cohort studies.

• Well suited for rare disease and diseases with long latent periods.

• Fewer subjects required than for cohort studies.

• Multiple etiologic factors can be evaluated.

What are the disadvantages of case-control design?

• More potential sources of bias and error than cohort studies

• Temporal relationship – did exposure cause disease or did disease cause exposure?

• Appropriate control group may be difficult to identify.

What is a cohort?

group of people who share a common experience

• Persons born in same year – birth cohort

• Persons who share a common behavior – e.g., cohort of smokers; employees at a tire manufacturing plant

• Persons in the same class – e.g., MPH cohort of Fall 2019

True or False: Prospective cohort studies follow exposed and unexposed individuals forward in time and relative to exposure to determine the incidence of disease in each group

True

3 Types of cohort studies:

1) Prospective (concurrent/longitudinal)

2) Retrospective (non-concurrent/historical)

3) Ambi-directional (combined concurrent/non-concurrent)

Prospective Cohort Studies

Investigator collects information on the exposure status of study subjects at the time the study begins and identifies new cases of disease that develop from that time on, until the end of the follow-up interval.

Retrospective Cohort Study

Investigator determines exposure status from information recorded at some time in the past, and disease status is determined from that point in the past up until the present (i.e., the follow-up period has already occurred).

Ambi-directional Cohort Study

Cohort and exposure status identified from past records, followed into the present, then followed into the future. Most often used when additional follow-up time is needed.

Selection of cohort study population: General population

General population sample – Representative sample from the general population

• Makes results highly generalizable

• Expensive, labor intensive, and may have problem with loss to follow-up. E.g., Framingham; PATH study

Selection of cohort study population: Special cohort population

• Special cohort – Defined population based on membership in a particular subgroup of interest.

• Follow-up may be easier

• Results may not be as widely generalizable. E.g., Nurses Health Study; U.S. Veterans; Strong Heart study

Selection of cohort study population: known exposure cohort population

Presence of a Distinctive Exposure – Selected because they are known to be exposed to a certain factor

• Generally used for occupational cohorts.

• Cohorts may be stable and easy to follow.

• Results may not be generalizable. E.g., Atomic bomb survivors; persons exposed to chemicals or other hazards on the job

What is the reference group for the exposed?

Internal Comparison – from the same population as the exposed group e.g., unexposed in the population (Framingham, low cholesterol levels)

2. External Comparison – outside of the exposed group e.g., general population data; another study’s data

3. Combined – both internal and external comparison groups can be used

Cohort Charactersitics

• Population based vs. Non-population based

• Open cohort (persons enter and leave over the course of follow- up) vs. Closed cohort (begin with a fixed study group; persons may leave but no new members enter)

Assessments of exposure:

1. Definition of exposure – what will constitute “exposure”?

2. Sources of exposure data which can be found through interview/survey, existing records, physical exam, and direct measurement

3. Timing of exposure - onset

4. Quantifying exposure

• Frequency – how often? Continuous or intermittent?

• Intensity – how much?

• Duration – how long?

5. Changes in exposure status - how to deal with?

Measurement of Disease

• Determine disease-free status at start of study

• Outcomes must be clearly defined and measurable – who is an incident case?

• How will you determine whether or not disease has occurred?

• Procedures should be well-defined and the same for both exposed and unexposed groups

• Best if assessment of disease incidence is done blinded to exposure status

Outcome measures in cohort studies:

Cumulative Incidence

Incidence Density

Cumulative Incidence

CI: all cases known to have occurred in the baseline cohort during the follow-up time divided by the study population at baseline, per unit of time.

• Numerator is # of incident cases; Denominator is # people in study population

• Closed cohort

• CI is a risk measure