Application Of Inverse Probability Weighting In Medical Studies

Background:In the oncology clinical trials which aim to investigate whether the test treatment prolongs overall survival/progression free survival compared with control treatment, patients will be allowed to crossover from control group to test group for ethical reasons if the interim analysis shows a significant advantage on Disease Free Survival for patients randomly assigned to test group. However, such "selective crossover" will disturb the randomization and results in the informative censoring. Subsequent analysis using Cox model may get a biased estimator of Hazard Ratio. While in the observational studies with exposure that varies over time, standard approaches for adjustment for confounders are biased when there exist time dependent confounders that are also affected by previous exposures. A new class of model, Inverse Probability Weighting method, can be applied to those problems from the fields of RCT and longitudinal data.Objective:We will introduce the application of IPW for the cases of selective crossover and time dependent confounders and investigate the performance of IPW in different scenarios. Some suggestions will be provided when using IPW in the medical studies.Method:Based on BIG1-98trial, we generated data to simulate oncology trials and evaluated the performance of IPW when censoring time is independent/conditional dependent/dependent with survival time and with different crossover proportions, sample size and Hazard Ratio between groups. For the time dependent confounders, we will explain the advantage of IPW over other methods we often used by a simulated observation study. And with published data from OAI study, a multi-center prospective study focusing on osteoarthritis, we will introduce how to fit IPW model, using the estimation of the effect of physical activities on knee function performance among knee osteoarthritis patients.Main results&Conclusion:IPW method is able to correct the bias caused by selective crossover and the coverage rate of95%CI is around95%. For the simulated trial when censoring time conditional depends on survival time and sample size=1000, HR=0.8(exp(-0.2)), crossover proportion in the control group=0.2, the coverage rate of95%CI is95.4%, power46.3%and bias0.005(the true parameter=-0.2). While under the same conditions, the coverage rate of95%CI of Cox model is90.5%, power23.5%ans bias0.07(the true parameter=-0.2). The power of the method is mainly influenced by the true parameter of HR and the number of death events, while the crossover proportion has an impact on bias.For the problem of time dependent confounder, the results of simulation gives a demonstration why traditional methods don’t work and explains the general idea of IPW. By the effect of weighting, IPW method tries to create a pseudo-population following two important properties, first exposure is unconfounded by the time varying covariates, second the effect of the exposure on the outcome stays the same with that in the original population. Hence it follows that we can unbiasely estimate the effect of exposure by a standard analysis in the pseudo population. And for the case study, IPW model shows that the effect of very high level physical activities is significant higher than that of low level, with P=0.02and95%CI(0.06-0.67). Very high level of activities can slow down the decline of knee function performance among knee OA patients. Patients with knee OA should be encouraged to take more exercises in daily time to maintain their knee function.