Causal Inference
Causal Diagrams for Disease Latency Bias Mahyar Etminan* Mahyar Etminan Ramin Rezaeianzadeh Mohammad Ali Mansournia Mahyar Etminan
Objective: Disease latency bias (DLB) can affect epidemiologic studies that examine the causal effects of exposures (eg, a drug) with a wide range of chronic diseases. Using causal directed acyclic graphs (cDAGs), we demonstrate four scenarios where disease latency can introduce bias into causal epidemiologic studies.
Methods: Epidemiologic studies have shown that benzodiazepines can increase the risk of dementia. Some of these studies could have been affected by DLB as the prodromal signs of dementia could have preceded drug use. We show 4 different cDAGs related to this question with variables A (benzodiazepine use), Y (diagnosed dementia), Y* (early dementia symptoms), U (unmeasured confounder), C (censored subjects) and M (mediator between Y* and A).
Figure 1. Biasing path through an unmeasured confounder. U represents the unmeasured confounder ‘insomnia’. DLB may be introduced when U is a common cause of Y* (early signs of dementia) and use of a benzodiazepine. A biasing path can be introduced through the path: Y←*Y←U→A.
Figure 2. Biasing path through reverse causality. Early signs of cognitive deficit can lead to benzodiazepine use years prior to diagnosis of dementia. A biasing path is created through Y*, acting as an unmeasured confounder: Y←Y*→A
Figure 3. Biasing path through selection bias. C is a collider on the path A→C←Y*→Y. Subjects experiencing early symptoms of dementia (Y*) and adverse events from benzodiazepine use are censored and analysis is done only in those who stay in the study (represented by a boxed C=0).
Figure 4. Biasing path through a mediator. The effect of Y* is mediated on A via M (number of physician visits). Early signs of dementia prompts patients to have more physician visits, and more likely to receive a benzodiazepine. Adjustment for M in these situations can close the biasing path.
Summary: DLB is an underrated bias that might methods such as probabilistic sensitivity analysis to better address it.
