Application of Marginal Structural Models (MSMs) to Irregular Data Settings

  • Edmore Chamapiwa

Student thesis: Phd

Abstract

The uptake of Marginal Structural Models (MSMs) in causal inference research has been increasing and continues to do so. The increase in use of this methodological approach is attributed to its ability to give unbiased treatment estimates when treatments are time varying and confounded by time varying covariates. However, MSMs have been widely used in clinical settings where individuals are observed and have their treatments reviewed at regular time intervals. In healthcare treatment settings, individuals can be observed and treated at irregular time intervals and this irregularity may well be related to previous treatments and it can also influence future treatment. Not much is known about performance of MSMs in irregular data settings. On the basis of the foregoing, this thesis sought to investigate application of MSMs to irregular data settings The overall aim of this thesis was to to investigate applicability and performance of Marginal Structural Models to estimate causal effects in health data settings where individual patients may be observed at different (or “irregular”) time intervals, and these intervals themselves may be related to treatment and confounder histories. A new class of Inverse Probability of Treatment Weighted (IPTW) estimators for irregular data was proposed to address the aim of the thesis. Four aspects of application of MSMs to irregular data settings were addressed. First, the new class of estimators was tested to irregular data for different clinical settings in a simulation study where the true effects were known or could be computed. Performance of the IPTW etimators for irregular data was compared with performance of competing stratification based regression methods. Second, the proposed estimators were assessed on their sensitivity to treatment model specification and applicability of machine learning approaches to model treatment model estimation in irregular data settings was also explored. Third, performance of IPTW estimators for irregular data was explored in three positivity settings: non-violation, near-violation and violation. Finally, applicability of the estimators to real life healthcare data was demonstrated through a ‘blinded’ simulation study – a new variant of a simulation study design that was proposed and used in this thesis. Results of parameter estimates from analyses of irregular data generated from different clinical settings showed that the proposed IPTW estimators for irregular data can give unbiased parameter estimates for MSMs. Performance of the estimators to different specifications of the treatment model showed that they are sensitive to model specification. The results from machine learning approaches to estimate the treatment model from simulated data showed that IPTW estimators for irregular data can be successfully used to estimate MSM parameters when the distribution of the treatment model is not known. Performance of this class of estimators was at least comparable to A-IPTW and TMLE in scenarios of near-violation and violation of positivity. Application of IPTW estimators for irregular data in a ‘blinded’ simulation study setting indicated that these estimators can be successfully used in a ‘real life’ healthcare data setting.
Date of Award1 Aug 2019
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorDavid Reeves (Supervisor), David Springate (Supervisor), Darren Ashcroft (Supervisor) & Evangelos (Evan) Kontopantelis (Supervisor)

Cite this

'