ESTIMATING THE COMPARATIVE EFFECTIVENESS OF LONGITUDINAL INTERVENTIONS

Purpose:

The aim of this paper is to illustrate the gains from using modern causal methods, such as targeted maximum likelihood estimation (TMLE), when estimating the comparative effectiveness of longitudinal interventions, using observational data.

Method:

Large, observational databases are increasingly used to answer questions of comparative effectiveness. The longitudinal structure of these datasets allows  the estimation of  the effects of interventions that change over time. Examples include the treatment of chronic diseases such as diabetes and hypertension, where decisions such as when to initiate a treatment or introduce a concomitant mediation are updated over time. For decision makers to compare the consequences of alternative longitudinal interventions, first, it is essential to carefully define the strategies of interest. Second, when estimating the effects of time-varying treatment from observational studies, appropriate statistical methods are required to address both baseline and time-dependent confounding.

As standard regression approaches cannot adjust for time-dependent confounding, inverse probability of treatment weighting (IPTW) is often used in such settings. A recently proposed semi-parametric method, longitudinal TMLE can improve on the properties of IPTW estimation, because it does not rely on the correct estimation of the treatment mechanism only, but also exploits information from outcome regressions, and it can be coupled with machine learning.

This talk aims to introduce longitudinal TMLE, and contrast it with IPTW, while addressing a high profile question in clinical decision making: what is the optimal timing and quantity of caloric intake for critically ill children?  We re-analyse a clinical trial to estimate the effect of alternative feeding regimes on the probability of being discharged from the Paediatric Intensive Care Unit (PICU) by a given day. While feeding is measured daily, this intervention was not randomly assigned in this study.

Result:

We find that before adjusting for confounding, patients who followed the  regime “Never
feed", were discharged earlier from the PICU than patients who followed the regime “Feed from day 3”, or “Feed when off mechanical ventilation”. When adjusting for baseline and time-varying confounders, both with IPTW and TMLE, most of this difference disappears, with TMLE reporting narrower confidence intervals than IPTW.

Conclusion:

We conclude that longitudinal TMLE offers a flexible estimation approach that merits wider application in comparative effectiveness research, to inform decision making.