Each method below is listed with the project in which it was applied.
Cognitive Modeling
Reinforcement learning models of the Iowa Gambling Task (delta-rule value updating, outcome-representation learning, prospect valence learning); sequential sampling models (the diffusion decision model, the racing diffusion model, the linear ballistic accumulator); joint reinforcement learning–sequential sampling models, including the joint likelihood of choice and response time and the linking function between learned value and drift rate; and neural diffusion decision models with factor-analytic structure. Applied in the Iowa Gambling Task joint model and the lasso factor-analytic neural diffusion decision model.
Bayesian Estimation
Model implementation in Stan: custom likelihood functions for the joint density of choice and response time, with the comparison models implemented in the same framework. Estimation and diagnostics: weakly informative and shrinkage (Bayesian lasso) priors, convergence diagnostics (R̂, effective sample size, divergent transitions), posterior predictive checking, and simulation-based parameter recovery. Applied in the two modeling projects above.
Machine Learning and Interpretability
Tree-based ensembles with cross-validated evaluation; grouped and stratified resampling for clustered data; class-imbalance handling by threshold adjustment, synthetic oversampling of mixed categorical–continuous features, and class-balanced bootstrapping; cluster bootstrap inference for model comparison; and Shapley additive explanation values computed out-of-fold, with dummy-column contributions grouped back to their parent variables. Applied in the dyadic fertility-desire prediction project.
Causal Inference
Conditional average treatment effect estimation with generalized random forests: Robinson orthogonalization with forest-estimated nuisance functions, cross-fitted honest estimation, adaptive-neighborhood weighting, and doubly robust AIPW summaries, with calibration testing of the average effect and of heterogeneity preceding any subgroup interpretation. Applied in the COVID-19 well-being project.
Software and Experimental Methods
Design and implementation of response-time-sensitive tasks in PsychoPy, including control of motor confounds across items, deployed in live data collection. Single-trial beta estimation and ROI-level summarization within an atlas parcellation on the analysis side of the joint neural–behavioral project; fMRI preprocessing in SPM and structural equation modeling in Mplus at coursework level. Programming in R, Stan, Python, and MATLAB, with formal training in computer science (B.S., Seoul National University).
Sample Code
Each repository contains a self-contained excerpt of the modeling approach used in the corresponding project, written against synthetic data. These are illustrative samples rather than full analysis pipelines; complete code and materials are available on request.
igt-rl-ssm Stan · R
A reinforcement learning sequential sampling model: value updating, the value-to-drift linking function, and a racing diffusion likelihood in Stan, with an R script that simulates data, fits the model, and runs a parameter recovery check.
View on GitHub →lasso-fa-nddm Stan · R
The joint behavioral–neural model: single-trial decision parameters as latent factors, region-level responses loading onto them, a Bayesian lasso prior on the loading matrix, and anchor-based identification.
View on GitHub →dyadic-ml-sample Python
A prediction pipeline for clustered data: couple-grouped stratified cross-validation, oversampling and threshold selection fit within the training fold, and out-of-fold Shapley attribution with dummy columns grouped back to their parent variables.
View on GitHub →causal-forest-sample R · grf
Heterogeneous treatment effect estimation from fitting through interpretation: causal forest fitting, the calibration test used before subgroup interpretation, AIPW average effects, variable importance, and best linear projection.
View on GitHub →