Graduate Seminar

Location:  RI 209

Speaker: Isaac Kwarteng

PhD Student supervised by Yang Zhao

Title: Regression with Non-monotone Missing Covariates

 Abstract:

The question:
When covariates are observed in overlapping subsets, must we discard every incomplete record? This research develops a way to let each partial record contribute while keeping the assumptions visible.

Core idea:
Predict the hidden complete-data score from what is observed, then correct that prediction using inverse-probability weighting:
\[ m_i(\beta) + \frac{C_i}{\pi_i} \{ U_i(\beta) - m_i(\beta) \}. \] Pseudo-likelihood learns nuisance structure across arbitrary patterns.

What will be established:
Identification, robustness conditions, computation, sandwich variance, diagnostics, and finite-sample performance against complete-case, MI, IPW, and AIPW benchmarks.

Take-home message:
Partial records are not empty records: use them deliberately, diagnose overlap, and report the assumptions.