Her paper was recently published in Statistics in Medicine and is also available on arXiv. Congratulations to Hillary, who recently started a position in the Department of Data Sciences at DFCI!
rats to Hillary, who recently started a position in the Department of Data Sciences at the DFCI.

Her approach offers an efficient and scalable way to model survival times in the presence of many potential predictors, automatically selecting important variables and identifying whether their effects are fixed or random—without requiring prior knowledge.

To address this, former Department of Biostatistics at UNC-Chapel Hill student Hillary Heiling developed an approximation to the popular Cox model for survival data. This approximation allowed her to adapt tools from prior work on modeling correlated binary and count outcomes to the survival setting

However, it is not known in advance which variables' effects vary across sites, which are fixed, or which variables are even associated with survival at all. Existing random effects models for survival data are also not scalable beyond a handful of predictors to answer these questions.

Random effects models help account for this variability across sites, enabling more accurate assessment of whether a treatment truly affects survival, rather than confounding results with site-level differences. This leads to more reliable and generalizable conclusions.

In biomedical research, data are often collected across multiple hospitals or sites. These groups may differ in ways can that influence the relationship between a treatment and patient outcomes—such as survival time.

Yusha is an assistant professor in the Bios dept at @uncpublichealth.bsky.social and the @unclineberger.bsky.social, where she serves as a member of the Biostatistics Core. She recently joined our department and we are so lucky to have her!

sph.unc.edu/adv_profile/...

Thank Mike!!

Thanks Tim!!

Semi-supervised analysis is almost always what people really want, especially when they aren’t playacting for study section or tenure

Unsupervised: we don’t know anything about it

Supervised: we know everything about it

Neither is common in serious research. Thanks for fighting the good fight!

Thanks Mike!

It’s not enough to just publish markers or signatures, but also provide a robust and feasible means to measure such markers on the a clinic, as well as clearly delineate the clinical properties of such markers, such as thresholds, their treatment implications, and their generalizability