64th ISI World Statistics Congress - Ottawa, Canada

64th ISI World Statistics Congress - Ottawa, Canada

Introducing ISENSE: An Index of Sensitivity to Non-exchangeability


Md Rashedul Hoque


  • Y
    Yi Qian
  • J
    J Antonio Aviña-Zubieta
  • M
    Mary A De Vera
  • L
    Lawrence McCandless
  • H
    Hui Xie


64th ISI World Statistics Congress - Ottawa, Canada

Format: CPS Abstract

Keywords: causal treatment effect, non-exchangeability

Session: CPS 12 - Statistical methodology IV

Monday 17 July 4 p.m. - 5:25 p.m. (Canada/Eastern)


The exchangeability of units between treatment groups is a key assumption for evaluating causal intervention effects in observational studies. The standard statistical methods assuming exchangeability can yield biased treatment effect estimates if the assumption does not hold. It is important to evaluate the robustness of study findings to the violation of the exchangeability assumption. Existing methods evaluate the sensitivity of treatment effect estimates to non-exchangeability due to unmeasured confounders only. In practice, non-exchangeability can occur for either unmeasured confounders or reverse causality. Also, those existing approaches require specifying the distribution and the number of unmeasured confounders.
We propose an index of sensitivity to non-exchangeability (ISENSE) to measure the impact of non-exchangeability on treatment effect estimates. Unlike many existing methods, ISENSE does not require imposing any assumptions regarding the distribution or number of unmeasured confounders, and it can handle both unmeasured confounders and reverse causality. ISENSE is a computationally inexpensive local sensitivity method based on a Taylor-series approximation to the non-exchangeability likelihood, evaluated at the parameter estimates under the exchangeability assumption. Two sub-models for the potential outcomes and treatment assignment contribute to the likelihood function for ISENSE construction. The ISENSE describes how a unit change in the non-exchangeability parameter from exchangeability displaces the maximum likelihood estimates (MLEs) of the parameters. One can interpret ISENSE intuitively through the unit-free ‘MinNE’ statistic values that capture the minimum non-exchangeability needed to cause important sensitivity.
We evaluate the performance of ISENSE using simulation studies and illustrate its use with an example using health administrative data from British Columbia, Canada. In the application, we assess the association between the adherence to hydroxychloroquine, a common medication for managing rheumatoid arthritis (RA) and systemic lupus erythematosus (SLE) patients, and physician billed-costs among newly diagnosed RA and SLE patients. In the retrospective population-based cohort study, we observe that for the treatment effect (hydroxychloroquine adherence) estimate, there is a potential sensitivity for modest (weak to moderate) departures from exchangeability. Consequently, the treatment effect estimates of hydroxychloroquine adherence could be sensitive to the violation of exchangeability either for unmeasured confounders or reverse causality. The application illustrates the usefulness of ISENSE to screen for potential sensitivity of treatment effect estimates to the assumption of exchangeability using observational data, and to quantify the reliability and strength of evidence for health policy decision makings.