Cohort Studies
Three types of cohort studies
Retrospective, prospective, ambispective
Measures of association that can be estimated from a case-control study.
Relative measures of effect
True/false -- Participants are sampled with respect to their exposure status.
False -- everyone is sampled at a point or period of time.
Definition of a confounder
1) Risk factor for the outcome among the unexposed
2) Associated with exposure in the source population
3) Does not mediate the exposure-outcome relationship
Measures of association that can be estimated
Absolute and relative measures of risk and rates
Controls in a case-control study should be sampled independently of
Exposure
Cross-sectional studies have trouble determining
Temporality
Exception -- if questions are posed to define temporality
Definition of a mediator
An intermediate variable along the causal path between the exposure and outcome relationship
The statistical model that should be used to estimate the association between an exposure and binary outcome in a cohort study
Dist=binomial, link=log (relative) or identity (absolute)
Problems with convergence:
Dist=Poisson, link=log, repeated subject=id/type=unstr
Also the logistic regression IS sometimes OK to use in a cohort study with a binary outcome (rare outcomes) as a reasonable approximation of the RR- just not always. This is worth discussing in response to your prompt (Why shouldn’t a logistic regression be used?)
True/false -- case-control studies are just cohort studies done in reverse
False -- case-control studies are more efficient forms of cohort studies
Two types of cross-sectional studies
Descriptive and analytical
Effect measure modifier
Sources of bias in cohort studies
Information bias (non-differential classification, differential misclassification, holding covariates fixed over time)
Confounding (time-varying confounding)
Sampling hospital controls can introduce this
Selection bias - namely, Berkson's bias.
Outcome=binary
Model=logistic regression
Estimate=?
Prevalence odds ratio
Methodological and analytical strategies to control for confounding
Randomization (with sufficient sample size), stratification, regression adjustment, matching, standardization
True/false -- non-differential misclassification of the exposure always biases the effect estimate towards the null.
False
Types of control sampling in nested case-control studies.
Density sampling (IRR -- no rare disease assumption needed), cumulative sampling (OR -- need rare disease assumption if interpreting as RR, otherwise "odds" interpretation remains), case-cohort sampling (RR --no rare disease assumption needed)
Limitations: lead-time bias, limited temporality, other traditional biases (selection, information, confounding)
Confounding or effect measure modification?
Crude OR = 2.0
Strata 1 OR= 1.2
Strata 2 OR = 2.8
Effect measure modification