(242) Reducing Bias from Chronic Exposure in Case-Crossover and Case-Time-Control Studies: A Simulation Study

Publication Number: 916

Background: Case-crossover studies have been proposed as a design suitable for brief exposure that causes a transient change in risk of an acute-onset disease. In pharmacoepidemiology, drug use is often chronic or successive, which leads to bias in case-crossover studies due to within-subject dependency.

Objectives: In our simulation study, we show that bias due to within-subject exposure dependency associated with chronic exposure can be removed by the Mantel-Haenszel method or our recently proposed weighting method. We also show that when some patients are censored due to switching to another drug after chronic use of a previous drug, bias occurs which is similar to that due to exposure time trends.

Methods: In the simulation, we used case-crossover and case-time-control studies within a cohort of new drug users. Assume that every day, 60 patients enter the cohort after starting a drug of interest and continue it at least for 60 successive days. After the end of drug use, some patients are unexposed for at least 60 days. The other patients switch to another drug when the exposure period ends and are censored immediately. We varied the proportion who were censored from 0 to 30%. We assume that the rate of the outcome occurrence is 0.00005 per day during the unexposed post-exposure period and the rate is 0.0002 per day during the exposure period, so the rate ratio (RR) for the exposure is 4.0. We assume case-crossover and case-time-control studies are conducted to identify cases and control-subjects selected by the risk-set sampling with one case period and 1, 2, 3, 4, 5, 6, 10, 12, 15, 20, 30 or 60 control periods within the cohort during a time period of 200 days after the "steady state" is attained. Data was analyzed by the standard conditional logistic regression, Mantel-Haenszel method, and recently proposed weighting method.

Results: With no censoring, odds ratio (ORs) by standard conditional logistic regression (ORSCL) was unbiased when M=1 but increased with M and was about twice the true value when M=60. Those biases were removed by the Mantel-Haenszel or weighting method. When 30% of patients were censored, ORSCL was biased when M=1 (ORSCL=5.91) and further increased with M. This bias was removed by the Mantel-Haenszel or weighting method in conjunction with a case-time-control design.

Conclusions: The requirement for "brief exposure" may be relaxed in the case-crossover design, which may be reliably used for chronic or successive exposure using the Mantel-Haenszel or weighting method for analysis, in conjunction with a case-time-control design if the proportion of switchers is large, provided that the effect of a drug on the outcome occurrence is transient with no "carry-over" effect.