Abstract
Background: We consider the analysis of nested, matched, case-control studies that have multiple biomarker measurements per individual. We propose a simple approach for estimating the marginal relationship between a biomarker measured at a single time point and the risk of an event. We know of no other standard software package that can perform such analyses while explicitly accounting for the matching. Results: We propose an application of conditional logistic regression (CLR) that can include all measurements and uses a robust variance estimator. We compare our approach to other methods such as performing CLR with only the first measurement, CLR with an average of all measurements, and Generalized Estimating Equations. In simulations, our approach is significantly more powerful than CLR with one measurement or an average of all measurements, and has similar to power to GEE but correctly accounts for the matching. We then apply our approach to the CLUE cohort to show that an increased level of the immune marker sCD27 is associated with non‐Hodgkin lymphoma (NHL) and, by evaluating the strength of the association as a function of time until diagnosis, that the an increased level is likely an effect of the disease as opposed to a cause of the disease. The approach can be implemented by the R function clogitRV available at https://github.com/sampsonj74/clogitRV.Conclusion: We offered an approach and software for analyzing matched case-control studies with multiple measurements. We demonstrated that these methods are accurate, precise, and statistically powerful.
Causal diagrams and the logic of matched case-control studies
