ARC Laboratory Sharing: HF's comments

42 Modeling Achievement Trajectories When Attrition Is Informative (Present by Xue-Lan)

HF's comments

This study addressed the influence of ignoring nonignorable missingness in longitudinal studies and suggested using a discrete-time survival process to monitor dropout. Listwise deletion, mean imputation, and a dual-process approaches were applied to the NELS data and simulated data. Results indicated that estimates of intercepts were not influenced by approaches but slopes and variances were, particularly when a large propotion of data were missing and there is a strong departure from MAR. Findings highlighted the importance of dealing missing data cautiously and the incorporation of a survival model to a shared-parameter or random-effect-dependent missing data model.

Comments:

1) The findings did not surprise me where more data were missing not at random, MAR and imputation approaches yielded severe biased estimates on growth models. But why did these approaches lead to greater bias on variance, covariance, and slopes, but not on intercepts?

2) For the NELS data, it is not clear that the NMAR-model estimates were not significantly different from those yielded from the MAR or listwise deletion approaches.

3) It will be better to describe the recovery of the dual-process approach on simulated data.

4) Model fit was missing on table 6.

5) It will be better to use plausible values in the growth model or to use multilevel IRT approach to estimate growth.

6) The paper did not explain why regression coefficients were constrained as 0, 2, 4 for the linear slope and 0,4,16 for the quadratic slope.

Re: HF's comments

by QIU Xuelan - Wednesday, 7 November 2012, 12:16 AM

1. At the first time point, they were no missing values and the estimated intercept means did not differ very much between the models. Only the listwise deletion will overestimate the intercept since the weak students with low score drop out.

2. From table 2, it is clear that the dual-process NMAR model were not substantially different from those that from MAR. Therefore, it was assumed that missing data are MAR.

The listwise-delection approach will overestimate the intercept mean than the NMAR model. Moreover, when there were no risk factors, the slope are not different between the listwise delection and NMAR model. But, where there were risks, the slopes were overestimated with listwise-delection than NMAR model.

3. Exactly.

4. Thank you for your carefulness. I even did not notice it.

5. Yes. I'm thinking how to solve the problem within the framewok of IRT.

6. It was assumed that the intercept have equal varaiance 1 across three timepoint. Therefore, in figure 1 growth model, there are three 1 under intercept. However, the authors did not explain why the residual variance will be different for linear slope and quadratic slope.