A Propensity-score-adjustment Method for Nonignorable Nonresponse

Semiparametric maximum likelihood estimation with data missing not at random

“…In the fully parametric approach model identification and model misspecification can be a problem and sensitivity analysis has been recommended (Scharfstein, Rotnitzky, & Robins, ; Rotnitzky et al, ; Verbeke et al, ; Tsiatis, ). Sverchkov () and Riddles, Kim, & Im () proposed another fully parametric approach using a different model specification based on (i) and (ii) . This new parametric model approach of Riddles, Kim, & Im () is practically more attractive because one can verify from the respondents, but it is still subject to model misspecification problems because it is a fully parametric approach.…”

“…Sverchkov () and Riddles, Kim, & Im () proposed another fully parametric approach using a different model specification based on (i) and (ii) . This new parametric model approach of Riddles, Kim, & Im () is practically more attractive because one can verify from the respondents, but it is still subject to model misspecification problems because it is a fully parametric approach. Chang & Kott () and Wang, Shao, & Kim () considered generalized method of moments estimation using a response mechanism only, but their method is generally less efficient than the likelihood‐based one.…”

A propensity score adjustment method for regression models with nonignorable missing covariates

“…Chang and Kott (2008) and Wang et al 25 (2014) proposed the Generalized Method of Moment (GMM) (Hansen, 1982) approach. Riddles (2013) and Riddles et al (2015) proposed a maximum likelihood method for estimating propensity scores to consider parameter estimation for a PS model with nonignorable missing data, treating the calibration condition as auxiliary information to improve inference using the theory of GMM. 30 For a detailed coverage of the PS method, see Kim and Shao (2013).…”

“…Although we have proposed an efficient and robust estimation method for re-90 gression models with nonignorable missing covariates, the response probabilities are unknown in general and need to be estimated consistently from the sample (Riddles et al, 2015). In what follows, we propose some alternative approaches.…”

“…We also compare our method with that of Riddles et al (2015) and The results are given in Table 4. In Table 4, the PS-LS estimators are com-270 puted using the corresponding estimated propensity scores; CQR 5 and CQR 9…”