Murthy Mittinty, Senior Lecturer in Biostatistics at the University of Adelaide, spoke at our April meeting on his work about targeted maximum likelihood estimation (TMLE) for causal inference in observational studies. This is the topic of his current research, following on from many years of work on causal inference. At the beginning of his talk, Murthy described the targeted maximum likelihood estimation, which is a semiparametric doubly‐ robust method that improves the chance of true parameter estimation by allowing for flexible estimation using (nonparametric) machine‐ learning methods, or super learner. He provided a step‐ by‐ step guided implementation of TMLE and illustrated it in a simulation scenario based on dental epidemiology.
To attain causal inference from observational studies, methods such as G-estimation, inverse probability treatment weighting, or targeted maximum likelihood estimation (TMLE) are preferred over traditional regression approaches, which are biased under misspecification of a parametric outcome model. He claimed that the assumptions such as positivity, consistency, exchangeability, and faithfulness needs to be made when using TMLE. Doubly robust methods, which require correct specification of either exposure or outcome model have been proposed as an improvement over simple IPTW methods. He demonstrated that the assumptions about correct model specification and positivity (ie, when a study participant has zero probability of receiving the treatment) were nearly violated when implementing TMLE to a realistic scenario based on dental epidemiology.
In conclusion Murthy demonstrated that his research provides a concise and reproducible educational introduction to TMLE for a binary outcome and exposure. The user should gain sufficient understanding of TMLE from this introductory tutorial to be able to apply the method in practice. It was an entertaining and engaging presentation through our first zoom meeting in the SA Branch.
Key reference: Schuler & Rose (2017). Targeted maximum likelihood estimation for causal inference in observational studies. AJE; 185(1): 65-73