Research
Selected Statistical Work
My graduate work in statistics included independent methodological research along with applied projects in causal inference, statistical learning, and uncertainty quantification. The selections below focus on the methodological questions, assumptions, and limitations that shaped each analysis.
Technical detail is included where it clarifies what a project was designed to examine and what its results can reasonably support.
RandomForestSpecCheck: A permutation-based random forest diagnostic for linear mixed models
Linear mixed models are used for correlated data such as repeated measurements on a subject, students within classrooms, or patients within clinics. They model that correlation directly, and in doing so they introduce assumptions about the fixed effects, the random effects, and the residual covariance structure. A model can be wrong in any of those places. The usual diagnostics address one failure at a time and require the analyst to guess in advance what may have gone wrong.
RandomForestSpecCheck begins from a different observation. If the mean structure is correctly specified, the conditional residuals should carry no predictable information about the covariates. The procedure fits a random forest from those residuals onto covariates the analyst names and scores it by its out-of-bag coefficient of determination, so that fitting residual noise earns it nothing. The null distribution is generated by permuting residuals within clusters, which severs the relationship between residuals and predictors while preserving the clustered design of the data. The analyst supplies the variables to examine. The forest is left to search for the functional form.
A model is flagged only when the observed statistic exceeds both the upper percentile of the permutation null and a practical threshold. The second condition matters in large samples, where an effect too small to be substantively useful can still be statistically detectable.
ranger and dplyrThe behavior of the procedure was mapped by simulation across a range of cluster counts, cluster sizes, and noise levels, and it was then applied to data from the Framingham Heart Study.
The same simulations established where the diagnostic is not informative, and that is as much a part of the result as the power curves. It detects conditional mean structure only, so misspecified variance structure does not register. The within-cluster permutation preserves cluster membership, which leaves a wholly omitted cluster-level effect invisible to the test. The forest can only find structure in the variables it is given, and power against subtle departures depends on the number of clusters, the observations within each one, and the noise level, so a null result in a small or noisy study is weak evidence that a model is adequate. The procedure is best understood as a screening instrument rather than a verdict on a model.
The work was completed as sole-authored independent research and written for submission to a statistics methodology journal. It has not been peer reviewed.
Causal analysis of food insecurity and type 2 diabetes
The target was the average treatment effect of experiencing food insecurity on the subsequent development of type 2 diabetes, estimated from three pooled cycles of the National Health and Nutrition Examination Survey spanning 2013 to 2018. Estimation used augmented inverse probability weighting, with g-computation as a comparison. The estimated effect was close to zero and not statistically significant across the specifications we considered.
The methodological content of the project was more consequential than the estimate. NHANES is primarily cross-sectional, so exposure and outcome are usually observed at the same point in time and the temporal ordering a causal claim depends on is difficult to establish. Five causal assumptions were stated and audited before estimation, and one of the two planned questions was abandoned because it could not be identified from the available data.
A conformal prediction framework for multi-label movie genre classification
A multi-label classifier can be accurate on average and still give no indication of when a particular prediction is uncertain. This project paired a fully fine-tuned DistilBERT model with conformal prediction so that each of 2,386 films received a set of possible genres rather than a single thresholded prediction, then examined what the resulting coverage guarantee cost. Fifteen configurations were compared across score functions, fine-tuning depth, and calibration scheme. The selected model reached 90.2% empirical coverage, and the average prediction set contained 8.62 of the 20 possible genres. The guarantee held. The sets were often too large to be useful, which is the more instructive half of the result.
Fine-tuning BERT models for recipe classification
Twenty-two fine-tuned configurations of BERT Base, BERT Large, and RoBERTa were compared on classifying more than 20,000 recipes as vegetarian or not from description and ingredient text alone. The best held-out accuracy, 0.869, came from BERT Base rather than from either larger architecture. Across the twenty-two runs, validation accuracy was almost unrelated to held-out accuracy, with a correlation of 0.09. That is a caution about selecting a model on a single validation split, and it would change how I designed the comparison if I ran it again.
Current project-based work is described under Statistical Services. My academic background and teaching history are on the About page.