Bias Control In The Analysis Of Case–control Studies With Incidence Density Sampling

Abstract

Background: Previous simulation studies of the case–control study design using incidence density sampling, which required individual matching for time, showed biased estimates of association from conditional logistic regression (CLR) analysis; however, the reason for this is unknown. Separately, in the analysis of case–control studies using the exclusive sampling design, it has been shown that unconditional logistic regression (ULR) with adjustment for an individually matched binary factor can give unbiased estimates. The validity of this analytic approach in incidence density sampling needs evaluation. Methods: In extensive simulations using incidence density sampling, we evaluated various analytic methods: CLR with and without a bias-reduction method, ULR with adjustment for time in quintiles (and residual time within quintiles) and ULR with adjustment for matched sets and bias reduction. We re-analysed a case–control study of Haemophilus influenzae type B vaccine using these methods. Results: We found that the bias in the CLR analysis from previous studies was due to sparse data bias. It can be controlled by the bias-reduction method for CLR or by increasing the number of cases and/or controls. ULR with adjustment for time in quintiles usually gave results highly comparable to CLR, despite breaking the matches. Further adjustment for residual time trends was needed in the case of time-varying effects. ULR with adjustment for matched sets tended to perform poorly despite bias reduction. Conclusions: Studies using incidence density sampling may be analysed by either ULR with adjustment for time or CLR, possibly with bias reduction.