Measuring the Impact of Nonignorable Missingness Using the R Package isni. (October 2018)

• Record Type: Journal Article
• Authors: Xie, Hui; Gao, Weihua; Xing, Baodong; Heitjan, Daniel F.; Hedeker, Donald; Yuan, Chengbo
• Is Part Of: Computer methods and programs in biomedicine. Volume 164(2018)
• Page Start: 207
• Page End: 220
• Publication Date: 2018-10
• Subjects: Analytical reliability -- Data quality -- Missing data -- Missing not at random -- Multivariate normal -- Selection model; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2018.06.014
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: This is the first software package dedicated to computing the index of local sensitivity to nonignorability (ISNI), which efficiently quantifies the sensitivity of model parameter estimates to nonignorable missingness in the outcome data. The computation of ISNI is straightforward and avoids the need to estimate complex nonignorable models. Our code implements the ISNI approach for the most commonly used regression models for continuous or discrete data, sampled with either independent or clustered outcomes. It allows for arbitrary incompleteness patterns caused by dropout or intermittent missed visits. Functions in our new R package, called isni, provide a consistent user interface and an automated analysis with sensible defaults. Results appear in simple, readily interpreted tables. Code is entirely in R and uses conventional modeling syntax. The package is available for download from cran. The isni package enables a systematic analysis that informs the quality, reliability and validity of empirical findings when some observations are missing. Abstract: Background and Objective : The popular assumption of ignorability simplifies analyses with incomplete data, but if it is not satisfied, results may be incorrect. Therefore it is necessary to assess the sensitivity of empirical findings to this assumption. We have created a user-friendly and freely available software program to conduct such analyses. Method : One can evaluate the dependence of inferences on the assumption of ignorability by measuring their sensitivity to its violation. One tool for such an analysis is the index of local sensitivity to nonignorability (ISNI), which evaluates the rate of change of parameter estimates to the assumed degree of nonignorability in the neighborhood of an ignorable model. Computation of ISNI avoids the need to estimate a nonignorable model or to posit a specific magnitude of nonignorability. Our newR package, namedisni, implements ISNI analysis for some common data structures and corresponding statistical models. Result : Theisni package computes ISNI in the generalized linear model for independent data, and in the marginal multivariate Gaussian model and the linear mixed model for longitudinal/clustered data. It allows for arbitrary patterns of missingness caused by dropout and/or intermittent missingness. Examples illustrate its use and features. Conclusions : TheR packageisni enables a systematic and efficient sensitivity analysis that informs evaluations of reliability and validity of empirical findings from incomplete data.