Post-graduate theses
Current Record: 14 of 127
|
Identifier |
000457260 |
Title |
Bayesian causal feature selection from observational and limited experimental data |
Alternative Title |
Μπεϋζιανή αιτιακή επιλογή χαρακτηριστικών από παρατηρησιακά και περιορισμένα πειραματικά δεδομένα |
Author
|
Λελόβα, Κωνσταντίνα
|
Thesis advisor
|
Τριανταφύλλου, Σοφία
|
Reviewer
|
Τσαμαρδίνος, Ιωάννης
Καμαριανάκης, Ιωάννης
|
Abstract |
In medical research, the selection of variables that contribute to an optimal predictive model
or aid in uncovering associations between treatment, outcome, and pre-treatment variables
poses a paramount goal. However, one of the most crucial challenges faced by doctors is
the selection of treatments that will optimize individual patient outcomes. This objective
can be effectively addressed by framing it as the problem of feature selection for predicting
post-intervention outcomes using pre-intervention variables.
Experimental data from randomized controlled trials allow for unbiased estimation of the
probability of post-treatment outcomes. However, such data have limited sample sizes and
may be underpowered to accurately estimate conditional effects. Observational data contain
many more samples but in most realistic cases, the presence of confounding variables makes
it difficult to establish causal relationships. Thus, identifying a set of appropriate covariates
and adjusting for their influence to mitigate confounding bias is not always possible from
the observational data alone.
This thesis argues that the combination of experimental and observational data may help
to improve the prediction of the post-intervention outcome and lead to an unbiased conditional
treatment effect estimation.
We propose a Bayesian feature selection method for finding the Markov boundary from
the observational data and using the concepts of feature selection, Bayesian inference, and
Bayesian regression, we extend a recently proposed method that combines large observational
and limited experimental data to identify adjustment sets and improve the estimation
of causal effects for a target population. [40] This method was developed for multinomial
distributions with Dirichlet priors and closed-form solutions and we present its extension
for data sets with both binary and continuous explanatory variables when the outcome is
binary or ordinal. In healthcare settings, the ordinal data is of great importance as it allows
for the nuanced measurement of patient outcomes and a significant gap exists in effective
methods for predicting post-interventional outcomes in this case.
We test our method in a simulated data set under different conditions. Results indicate that
our method (a) demonstrates high performance in accurately identifying the correct Markov
boundary for both binary and ordinal cases, even when applied to small observational data
sets, (b) exhibits strong performance in identifying the optimal set Z that when included in
a model, yields the best prediction for the post-intervention outcome P(Y |do(X), Z). The experiments
were conducted using limited experimental and large observational data samples,
respectively. When dealing with ordinal data, it is essential to have a larger set of experimental
data compared to the binary case.
|
Language |
English |
Issue date |
2023-07-19 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Mathematics and Applied Mathematics--Post-graduate theses
|
|
Type of Work--Post-graduate theses
|
Permanent Link |
https://elocus.lib.uoc.gr//dlib/8/8/b/metadata-dlib-1689319668-892667-8183.tkl
|
Views |
473 |