Dear class,
Thank you for attending lecture on Thursday. There were a couple points that I wanted to clarify:
First, we didn’t get to the bidirectional sampling slides, but I did mention an example of smoke from forest fire in relation to acute cardiac events as a type of question that could be examined with a control window sampled after the outcome. While it is true that a heart attack would not influence whether a forest fire occurred (or air pollution levels, etc.), Dr. Chan and I realized this may still have measurement error since individuals may be less likely to go outside, or more likely to use an air purifier or mask, after a heart attack. This demonstrates how tricky it is to find research questions where the outcome does not influence future exposure, and you must assume the outcome does not influence future exposure if you sample control windows after the outcome.
Second, attached are DAGs are to help explain why you A) need to stratify by the matching factor in a match case-control study when the matching factor is a confounder OR associated only with exposure; and B) do not have to stratify by the matching factor in a matched case-control study when the matching factor is only associated with the disease. The key is that if the matching factor is associated with the exposure (either because it’s a confounder or if it’s only associated with exposure), then the exposure distribution in the matched controls will not be representative of the exposure distribution in the population that gave rise to the cases. This may cause bias that can be addressed by stratifying by the matching factor. If the matching factor is only associated with the disease, than there is no “backdoor path” between the exposure and disease introduced by matching on the factor and your unadjusted analyses remain valid. Note, your adjusted results are also valid, though perhaps less precise. However, because you never know for sure if something is a confounder, you should stratify (adjust) for the matching factors in your analysis to avoid potential bias.
Please email me (cc Dr. Chan) if you have any questions regarding the matching and/or case-crossover lecture material.
Best,
Erin L. Van Blarigan, ScD
Associate Professor
Department of Epidemiology and Biostatistics
University of California, San Francisco