Bias reduction method for prior event rate ratio, with application to emergency department visit rates in patients with advanced cancer

TL;DR

This paper introduces a bias reduction method for the prior event rate ratio (PERR) that accounts for event dependence, demonstrated through simulations and applied to emergency visits in advanced cancer patients, enhancing PERR's robustness.

Contribution

The study proposes a conditional frailty method to reduce bias in PERR estimates when the assumption of event independence is violated, broadening its applicability.

Findings

Simulations showed bias correction with median bias around 5%.

Positive event dependence was identified in emergency visits among cancer patients.

The new method reduced bias compared to conventional PERR estimates.

Abstract

Objectives: Prior event rate ratio (PERR) is a promising approach to control confounding in observational and real-world evidence research. One of its assumptions is that occurrence of outcome events does not influence later event rate, or in other words, absence of 'event dependence'. This study proposes, evaluates and illustrates a bias reduction method when this assumption is violated. Study Design and Setting: We propose the conditional frailty method for implementation of PERR in the presence of event dependence and evaluate its performance by simulation. We demonstrate the use of the method with a study of emergency department visit rate and palliative care in patients with advanced cancer in Singapore. Results: Simulations showed that, in the presence of negative (positive) event dependence, the crude PERR estimate of treatment effect was biased towards (away from) the null value. The proposed method successfully reduced the bias, with median of absolute level of relative bias at about 5%. Dynamic random-intercept modelling revealed positive event dependence in emergency department visits among patients with advanced cancer. While conventional time-to-event regression analysis with covariate adjustment estimated higher rate of emergency department visits among palliative care recipients (HR=3.61, P<0.001), crude PERR estimate and the proposed PERR estimate were 1.45 (P=0.22) and 1.22 (P=0.57), respectively. Conclusions: The proposed bias reduction method mitigates the impact of violation of the PERR assumption of absence of event dependence. It allows broader application of the PERR approach.