Longitudinal Regression Conditioning on ContinuationPublic Deposited
Add to collection
You do not have access to any existing collections. You may create a new collection.
Downloadable ContentDownload PDF
MLADaza, Eric. Longitudinal Regression Conditioning On Continuation. Chapel Hill, NC: University of North Carolina at Chapel Hill Graduate School, 2015. https://doi.org/10.17615/mybs-gw34
APADaza, E. (2015). Longitudinal Regression Conditioning on Continuation. Chapel Hill, NC: University of North Carolina at Chapel Hill Graduate School. https://doi.org/10.17615/mybs-gw34
ChicagoDaza, Eric. 2015. Longitudinal Regression Conditioning On Continuation. Chapel Hill, NC: University of North Carolina at Chapel Hill Graduate School. https://doi.org/10.17615/mybs-gw34
- Last Modified
- March 19, 2019
- Affiliation: Gillings School of Global Public Health, Department of Biostatistics
- Individuals in a longitudinal study may have missing data for multiple reasons, including intermittent missed visits or permanent study drop out. Additionally, individuals may experience a truncating event, such as death, past which the outcomes of interest are no longer meaningful. Kurland and Heagerty (2005) developed a method to conduct regression conditioning on being alive (RCA), which constructs inverse-probability weights (IPWs) of the dropout probability among continuing individuals used to fit generalized estimating equations (GEE). RCA has since been extended to allow for intermittent missingness (IM) of outcomes (Shardell and Miller, 2008). We further extend these methods to simultaneously accommodate different mechanisms for dropout and IM, and call our method regression conditioning on continuation (RCC). RCC is illustrated using data from a recent study of mother-to-child transmission of HIV to draw inference on mean infant weights subject to truncation from infant death or HIV infection. Currently, there is no widely available software for conducting RCA. We present the xtrccipw command in Stata, which can estimate the dropout IPWs required by RCC if there is no IM. These IPWs estimated using xtrccipw are then used as weights in a GEE estimator using the glm command, completing the RCC method. In the absence of truncation, the xtrccipw and glm commands can also be used in a weighted GEE analysis. The xtrccipw command is demonstrated by analyzing two example datasets and the original Kurland and Heagerty (2005) data. A fundamental weakness of most non-sampling IPW methods is that the missing-data model is unknown and yet must be correctly specified to obtain consistent mean-outcome estimates. We extend the RCC approach to use augmented estimating equations (AEE) in what we call the augmented RCC (ARCC) method. In addition to the missing-data model specified by IPW-GEE methods, AEE approaches specify a model for the outcome joint probability. However, only one of these two models need be correct for the corresponding mean-outcome estimator to be consistent, making such techniques doubly robust to model misspecification. The empirical bias of the ARCC and RCC estimators are characterized and compared in a simulation study, and the ARCC method is applied to the mother-to-child HIV transmission study.
- Date of publication
- August 2015
- Resource type
- Rights statement
- In Copyright
- Cole, Stephen
- Herring, Amy
- Adair, Linda
- Hudgens, Michael
- Preisser, John
- Doctor of Public Health
- Degree granting institution
- University of North Carolina at Chapel Hill Graduate School
- Graduation year
- Place of publication
- Chapel Hill, NC
- There are no restrictions to this item.
This work has no parents.