An index of local sensitivity to non‐ignorability for multivariate longitudinal mixed data with potential non‐random dropout

Abstract

Multivariate longitudinal data with mixed continuous and discrete responses with the possibility of non-ignorable missingness are often common in follow-up medical studies and their analysis needs to be developed. Standard methods of analysis based on the strong and the unverifiable assumption of missing at random (MAR) mechanism could be highly misleading. A way out of this problem is to start with methods that simultaneously allow modelling non-ignorable mechanism, which includes somehow troubling computations that are often time consuming, then we can use a sensitivity analysis, in which one estimates models under a range of assumptions about non-ignorability parameters to study the impact of these parameters on key inferences. A general index of sensitivity to non-ignorability (ISNI) to measure sensitivity of key inferences in a neighborhood of MAR model without fitting a complicated not MAR (NMAR) model for univariate generalized linear models and for models used for univariate longitudinal normal and non-Gaussian data with potentially NMAR dropout are well presented in the literature. In this paper we extend ISNI methodology to analyze multivariate longitudinal mixed data subject to non-ignorable dropout in which the non-ignorable dropout model could be dependent on the mixed responses. The approach is illustrated by analyzing a longitudinal data set in which the general substantive goal of the study is to better understand the relations between parental assessment of child's antisocial behavior and child's reading recognition skill.

Multivariate multilevel spline models for parallel growth processes: application to weight and mean arterial pressure in pregnancy

“…The MLMs and maximum likelihood procedure used in the SEM models to deal with missing data both assume that the data are missing at random, meaning that the missingness in the data can be explained by observed variables . If there is reason to suspect that the missing data mechanism is nonignorable, there are methods available to test the sensitivity of inferences to data being not missing at random for multivariate longitudinal data .…”

Advantages of Joint Modeling of Component HIV Risk Behaviors and Non-Response: Application to Randomized Trials in Cocaine-Dependent and Methamphetamine-Dependent Populations

“…No distinction was made between multiple active doses within a trial (e.g., 200 and 400 mg of modafinil) in coding the regressor for placebo vs. active. Intercept terms were excluded from these trivariate probit models to facilitate identification on latent scales (Mahabadi and Ganjali, 2010). …”

Second‐order local sensitivity to non‐ignorability in Bayesian inferences

“…This index has been applied in various likelihood-based settings. [8][9][10][11][12][13][14][15][16] Also in the Bayesian context, Zhang and Heitjan 17 and Xie 18 developed ISNI for the case of Bayesian modelling respectively in cross-sectional and longitudinal studies where they used the first-order derivative of the posterior mean instead of MLE in the frequentist approach.…”