All Times EDT
Keywords: latent subgroup, Bayes theorem, immune checkpoint inhibitor
Recently, immune checkpoint inhibitors of PD1/PDL1 have been actively developed as cancer treatment. Those whose expression level of the biomarker (BM) are over than a cutoff value, e.g. 50%, are regarded as BM-positive. But, a currently used subgroup classifier based on PD1/PDL1 expression level concerning activity of effector phase of T-cell may partly misclassify patient due to other immune system activity mainly or additionally related to cancer progression including priming phase and/or due to uncertainty of a cutoff value of BM. So there might latently exist a true classifier which can be applied to patient population who would receive PD1/PDL1 inhibitors.
Existing methods to estimate latent subgroup (Altstein and Li, Biometrics. 2013; 69:52-61) which use a framework likewise principal stratification in non-compliance setting with partially known principal strata can be applied to a situation in which true subgroup membership is known for those who are assigned to the one of two arms in clinical trial. However, latent subgroup membership is now supposed to be unknown for both treatment arms in situation of immune checkpoint inhibitors as mentioned above.
We propose a new statistical method to estimate treatment effect for latent subgroup which might be defined based on some completely unknown true classifier and individual patient's probability belong to the subgroup. Our methods utilize observed PD1/PDL1 classifier as prior information and update individual status based on Bayes theorem according to survival outcome data. In simulation study, we investigated identifiability of latent subgroup using proposed method. Our method showed identifiablility and gave unbiased estimate of latent subgroup treatment effect in scenarios that true classifier prognostic effect and treatment effect are in range of 0.2 to 0.67 and 0.1 to 0.7, respectively in Hazard Ratio scale. Also c-statistics of prediction showed 0.96 to 0.878 against 0.863 for PD1/PDL1 status only.