Friday, December 7, 2012

Study Design: Cohort, Case-Control, & Cross-Sectional

In the Ph.D. program in which I'm enrolled, every student must take (and pass) a comprehensive examination (Part 1 of the General Examination) consisting of one epidemiology exam and one biostatistics exam.  Fortunately I'm well past this milestone (I took and passed this exam in late summer 2009), but while studying for this exam, I created a matrix briefly describing the three major study designs (cohort, case-control, & cross-sectional), their advantages & disadvantages, primary biases, and measures of association.  I figured that in addition to needing to know these study designs inside-and-out for the exam,  recording them in a matrix would be helpful for later reference.  Most of the material for the matrix came from my Occupational Epidemiology course notes and text ("Research Methods in Occupational Epidemiology" by Checkoway, Pearce, and Kriebel), although I vaguely recall supplementing it with Rothman, et al.'s "Modern Epidemiology". 

Study Design
Description/Defining Characteristic(s)
Advantages/Intended Uses
Disadvantages/Discouraged Uses
Prominent Bias(es)
Measure of Association
Cohort:  Prospective (Concurrent) and Retrospective (Non-concurrent)
1)  A direct analogue of the experimental study in that the subjects are disease-free at baseline, followed-up over time, then the outcome is observed among the exposed and unexposed groups.   
2)  Includes all persons in the population giving rise to the study cases (rather than a sample). 
The prospective cohort study is ideal for studies where the outcome occurs w/in a relatively short period of time, given the resources necessary to conduct said study. 

The retrospective cohort study, however, is better suited, relative to prospective studies, for studying diseases with a long induction/latency period.

1)  Temporality:  Since the exposure is assessed prior to disease, there is more evidence that the exposure ‘causes’ the disease. 
2)  Exposure is ascertained prior to development of disease.
3)  Minimal selection bias, although if (loss-to) follow-up differs by outcome, then this bias could become problematic.  
4)  Assembly of a cohort can yield a broad picture of the health experience of the cohort.  Allows for investigation of rates for multiple health outcomes (Checkoway, pp.174).  .

The cohort study can be used to study common exposures in “general population” studies (e.g. smoking, oral contraceptives), insofar as the exposure in the exposed group is large enough (don’t want the unexposed to overwhelm the exposed). 
The prospective cohort study is not intended for studies of an outcome w/ a long induction/latency period, e.g. “rare” disease.  Note that a “rare” disease is customarily one where 5% or less of the population will develop the disease. 

The retrospective cohort study may suffer from incomplete data since the investigator is relying on already-collected data and the outcome is most often fatality (easier to verify).

From Rothman, et al., “Cohort studies are poorly suited to study the effect of exposures that are hypothesized to cause rare diseases with long induction periods” (pp.108).
1)  Healthy Worker Effect (HWE):  The bias following form only including the employed in the study – a confounding effect may result since the employed must meet a minimal criterion of health that the at large population does not.
2)  Exposure misclassification:  This is more problematic in case-control studies but it can happen.  One example is using job as a proxy for exposure where the nature of the job and tasks may change over time.
3)  Outcome misclassification (Information Bias):  The disease may be incorrectly classified. 
4)  (Differential) Loss-to-follow-up:  Particularly problematic if associated w/ either exposure or outcome.  A f/up rate <60% likely to yield flimsy study results. 
5)  (MINOR) Selection Bias:  This isn’t terribly problematic at study onset since, in theory, the full source population has been enumerated and is under study.  Nevertheless, if tracing of subjects is related to exposure or outcome, then selection bias may result & mimic that often seen in case-control studies. 
A cohort study yields person-time, thus the Incidence rate is a common measure.  The most common measure of association, however, is the relative risk / rate ratio (RR).
Also known as Incidence case-control; cases obtained via sampling of the source population. 
Attempts to estimate results that a cohort study would have produced by a) including as many of the cases as possible generated by the population at risk, b) sampling from among the non-diseased individuals in the population at risk, c) comparing the frequency (odds) of exposure between the groups. 

Effectively two types:  the nested case-control (cohort-based) that is conducted in the context of a cohort study and, as such, requires the enumeration of the source population and its experience over the risk period.  The second type is your garden-variety case-control study – a community-based case control study.  In this type, the cases are defined first, the source population ascertained, then the controls sampled.
1)  Case-control study is especially useful and appropriate when the disease is RARE (and has a long induction time).
2)  When the community (study population) has a relatively HIGH prevalence of exposure.
3)  Also good when the occupational cohort is difficult to enumerate (e.g. farmers). 

1)  Case-control study is inefficient for rare exposures. 
2)  Selection of controls is of paramount importance, to wit:  “Controls should be selected from the same population – the source population – that gives rise to the study cases.”             

“Although case control studies do present more opportunities for bias and mistaken inference than cohort studies, these opportunities come as a result of the relative ease with which a case-control study can be mounted” (Rothman, et al., pp.112). 

1)  Recall Bias:  Since data is often collected directly from the cases & controls and the cases are obviously aware of their disease status, they may recall differently their exposure. 
2)  Interviewer Bias:  Bias may result if interviewers are poorly trained, not blinded, or o/wise interact w/ cases & controls differently. 
3)  Information Bias:  Primarily a concern since data collected (for exposure) after disease diagnosis – exposure misclassification.
4)  Selection Bias:  Results when cases and controls are not selected from the same (or similar) source population.
(Exposure) Odds Ratio:  Estimates the Rate Ratio (RR) when the case-control is properly conducted.  If the controls in the case-control study are still controls at study end, then the case-control odds ratio estimates the risk ratio obtained from a cohort design. 
Collection of data at one point in time; also known as a one-time survey.  Subjects selected without regard for exposure or disease status.  Good for assessing prevalence at a specific point in time.
Intended for assessing point or period prevalence.
Good for investigating risk factors for nonfatal, progressive diseases w/ no clear point of onset.
Exposure assessed at the same time as disease status.
Not designed for assessing disease incidence. 

Cannot arrive at any semblance of causality – temporality not considered in this type of study. 
1)  Length-biased sampling:  cases included over-represent cases w/ long duration and under-represent cases w/ short duration of illness. 

Prevalence; prevalence odds; prevalence odds ratio (POR). 

No comments:

Post a Comment