Geng, Ming
(2015)
Marginal structural Cox proportional hazards model for data with measurement errors.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
In causal inference, interest often lies in estimating the joint effect of treatment on outcome at different time points. Marginal structural models are particularly useful for this purpose when a time-dependent confounder exists in the causal path between the treatment assigned in the previous time and the outcome. These models provide a consistent estimate when treatment is measured perfectly. In practice however, treatment may be subject to measurement error. Many studies have shown that measurement error in treatment can result in underestimating its effect. One approach proposed in the literature for correcting this problem is the marginal structural measurement-error model. It requires using a validation data set in which both the true treatment and the observed treatment are available to correct the bias. In this study, we developed a new method which combines the marginal structural Cox proportional hazards model, the regression calibration method, and the Bayesian method to account for measurement error in treatment without the need for a validation data set. Moreover, instead of fitting a traditional pooled logistic regression model, a weighted Cox proportional hazards model is implemented to reduce bias. The performance of our proposed method was assessed through the simulation study. Our simulation results show that the bias is reduced even with an approximate value of the parameter of the prior distribution. Our sensitivity analysis also shows that the estimated treatment effect is robust to the choice of the prior distribution. We applied our proposed method to estimate the effect of highly active antiretroviral therapy on the incidence of acquired immune deficiency syndrome or death among HIV-positive patients using a data set in which the observed treatment assignment was subject to misclassification. Public Health Significance: Measurement errors can happen in medical studies despite good intentions. In general, either a validation data set or the replicates of the observed predictor are needed to correct for bias in estimation. Our study provides a new method in causal inference for correcting bias caused by measurement errors when investigators only have the main data set in which the observed treatment is measured only once at each time point.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
28 January 2015 |
| Date Type: |
Publication |
| Defense Date: |
10 November 2014 |
| Approval Date: |
28 January 2015 |
| Submission Date: |
16 December 2014 |
| Access Restriction: |
3 year -- Restrict access to University of Pittsburgh for a period of 3 years. |
| Number of Pages: |
53 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
School of Public Health > Biostatistics |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
Bayesian, marginal structural Cox model; misclassification, time-dependent confounder, treatment causal effect. |
| Date Deposited: |
28 Jan 2015 16:09 |
| Last Modified: |
01 Jan 2018 06:15 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/23893 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[img]](http://d-scholarship.pitt.edu/style/images/fileicons/application_pdf.png)
![[feed]](/images/feed-icon-32x32.png){kind=icon}