Comparing the g-formula fitted with generalized linear models (GLMs) and g-formula fitted with machine learning methods (MLMs).

Kamath, Tapan (2022) Comparing the g-formula fitted with generalized linear models (GLMs) and g-formula fitted with machine learning methods (MLMs). Master's Research Project 1, Medical Pharmaceutical Sciences.

Preview	Text mMPS_2022_KamathT.pdf Download (1MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (144kB)

Abstract

Introduction: One challenge that many variables in a dynamic treatment regime face is confounding, especially time-varying confounding. The g-formula is a method used to study dynamic treatment regimes. It helps in estimating the effectiveness of a model by using the counterfactual theory of causation. Data and methods: To get insight into the problem of model specification in this study we fitted the g-formula with generalized linear models (GLMs) and machine learning methods (MLMs) and then compared them in terms of their predictions with the help of loss functions. Using a simulation study we try to determine the predictivity of each model from different classes. Results: The density plots revealed that the generalized linear model (GLM) and the lasso predictions are very close to each other and are similar to the validation dataset. The random forest plots were far off from the validation dataset. The generalized linear model (GLM) and lasso model loss values were very close to each other while the random forest loss values were much higher than the other models. Conclusion: Determining the predictivity is an important step towards determining the performance of the models in causal effect estimation. After looking at the density plots and the loss values, both the lasso model and the GLM model were recommended for future studies on causal effect estimation.

Item Type:	Thesis (Master's Research Project 1)
Supervisor name:	Hak, E. and Bijlsma, M.J.
Degree programme:	Medical Pharmaceutical Sciences
Thesis type:	Master's Research Project 1
Language:	English
Date Deposited:	27 Jun 2022 12:51
Last Modified:	27 Jun 2022 12:51
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/27417

Actions (login required)

View Item