Fractional imputation method for handling nonignorable missing exposure in survival analysis

Music City Center 

Nonignorable missing covariates frequently arise in survival analysis, leading to biased results when the missing mechanism is incorrectly assumed to be missing at random (MAR). Existing methods for addressing nonignorable missing covariates often rely on strong model identification assumptions, such as the use of instrumental variables, which are challenging to verify in practice. In this paper, we consider the setting where a one-dimensional covariate, referred to as the exposure, is subject to nonignorable missingness, whereas other covariates are fully observed. We propose a novel estimation procedure for parameters in the propensity score model for the exposure variable by assuming a Gaussian mixture form for the conditional density of the exposure given observed covariates among the subjects with a non-missing exposure. Fractional weights that depend on this conditional density as well as the parameters in the propensity score model can be constructed. We then conduct statistical inference for Cox regression parameters using the generated fractional weights. Our approach offers a flexible framework for handling nonignorable missing mechanisms and does not require the self-consistency assumption imposed by traditional multiple imputation methods. Monte Carlo simulations and real-world data applications demonstrate the efficacy of our proposed method, highlighting its potential to provide robust and reliable inferences in survival analysis settings with a nonignorable missing exposure.

Fractional imputation

Nonignorable missing

Survival analysis 

Lifetime Data Science Section