Academic journal article Environmental Health Perspectives

Biased Exposure-Health Effect Estimates from Selection in Cohort Studies: Are Environmental Studies at Particular Risk?

Academic journal article Environmental Health Perspectives

Biased Exposure-Health Effect Estimates from Selection in Cohort Studies: Are Environmental Studies at Particular Risk?

Article excerpt

Introduction

Although randomized controlled trials (RCTs) are often conducted in a highly selected set of participants, exposure in such studies is unrelated to the selection process because exposures come after the randomization of selected participants. Therefore, it is reasonable to believe that this initial selection process does not induce biased exposure-health effect associations (i.e., associations that are different from the true total causal effect of the exposure on the outcome, in the absence of chance associations, in the source population that was sampled to obtain the study sample), although the findings of RCTs may lack generalizability (i.e., that the causal effect in the sampled population would not be the same as the causal effect in a different population). In observational studies, the initial selection process may result in biased exposure-health effect associations. Past or current exposure status may influence selection into the study or into substudies, either because exposure or correlates of the exposure are related to prespecified eligibility criteria or because they influence participant availability or willingness to participate. If both the exposure and the outcome or their correlates (including past exposure and outcome) are related to participation, a study's exposure-health effect associations may not reflect the true total causal effect in the source population that was sampled to obtain the study sample, either because of selection bias (known as collider stratification bias in causal theory) (Hernan et al. 2004) or because the selection process is equivalent to conditioning on an intermediate between the exposure of interest and the outcome (Schisterman et al. 2009).

The impact of collider stratification bias is well recognized in the setting of case-control studies. By definition, participation is related to the outcome; if recruitment into a case-control study is related to the exposure of interest as well, the observed exposure-health effect association may not reflect the true causal effect in the sampled population (Hernan et al. 2004; Wacholder et al. 1992). Similarly, exposure-health effect associations in prospective cohort studies likewise may not reflect the true causal effect in the whole cohort (and by extension, the population from which the cohort was drawn) if both the exposure and outcome are related to loss to follow-up (Hernan et al. 2004). Perhaps less well appreciated is how the process of cohort formation can induce a similar problem. Structurally, this problem is the same as loss to follow-up. If the exposure and outcome, or their correlates, influence a person's initial eligibility or participation, the resulting exposure-health effect association may not reflect the causal effect in the source population.

The impact of conditioning on an intermediate is also well recognized in the epidemiologic literature. Termed "overadjustment" by some, in simulation, the resultant bias can be substantial (Rothman et al. 2008; Schisterman et al. 2009). However, as with collider stratification bias, it is less well recognized that the cohort formation process may induce this bias in specific situations.

Many studies of environmental toxicant exposures are likely susceptible to bias of the exposure-health effect estimates attributable to the study or substudy selection process--including both collider stratification bias and bias due to conditioning on an intermediate--for two reasons: a) Environmental exposures are often related to socioeconomic status (SES), which is known to predict participation (de Graaf et al. 2000; Howe et al. 2013; Mein et al. 2012; Weuve 2013), and b) exposure to an environmental toxicant at one time point is often reasonably correlated with exposure at other time points, and prior exposures (and their consequences) may be related to participation. Another, perhaps simpler way to think of this relies on the fact that environmental exposures and their correlates (e. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.