Precision and Robustness in Adaptive Testing: An IRT Analysis within an Online Convex Optimization Framework (Dissertation)

My dissertation addresses the finite-sample performance of Computerized Adaptive Testing (CAT) algorithms by incorporating the Online Convex Optimization (OCO) framework. On the precision front, I demonstrate that classical information-maximization methods constitute no-regret algorithms and construct anytime-valid confidence intervals using martingale concentration inequalities. On the robustness front, I adapt the Upper Confidence Bound algorithm from multi-armed bandit literature to design item selection strategies that account for the exploration-exploitation trade-off in early-stage estimation.

On the Estimation of Standard Errors for the Ability Estimate in Item Response Models

We propose a precise and practical method for calculating standard errors of ability parameters in dichotomous IRT models. Because the information function represents the asymptotic variance of the MLE, it may diverge from the actual variance in finite samples. We advocate a plugin estimator of empirical variance and develop an efficient algorithm to reduce computational complexity. I derive the asymptotic validity theorems and implement the procedures in R.

Family, Health, and the Market

In collaboration with Prof. Kuan-Ming Chen at NTU Economics, I analyze government administrative data to study wage inequality trends in Taiwan using the Abowd-Kramarz-Margolis (AKM) model, intra-household tax decision-making, and the relationship between pet ownership and fertility. This work involves linking and analyzing large-scale datasets from Taiwan’s Ministry of Finance and Ministry of Agriculture.

Evaluation of an Education Program Using Item Response Theory

Working with Prof. Yu-Chang Chen, I apply the two-parameter IRT model to assess student abilities on the PaGamO online learning platform (100K+ students, 1M+ item bank). I develop adaptive testing decision algorithms from a Bayesian decision theory perspective and investigate the consistency of Bayesian adaptive testing under the Rasch model.

Neural Correlates of Emotion as an Arbitrator to Reconcile the Conflicts within Dual Processing in the Context of Decision Making under Risk

We performed behavioral and fMRI experiments to probe the role of emotion when people face a dual-process induced cognitive conflict under a risky condition. I was engaged in the analysis of fMRI data and in developing an estimation method for the parameters in the utility function.

Analysis of Cognitive Structures Underlying Financial Behavior

This project grew out of an industry-academia collaboration with Alliance Bernstein. We investigated possible factors that caused biases in investment and finance. I applied exploratory factor analysis to uncover the latent structure/factors of financial biases.

Coalition without Trust: The Intra-Brain Connectivity and Inter-Brain Synchronization of Herd Behaviors in an Economic Bubble Game

This project explored humans’ decision processes in stock markets. We simulated a stock market and asked subjects to compete (recorded using hyperscanning fMRI). I analyzed the behavioral and fMRI data and used k-means clustering to identify different response patterns among participants.