DONDENA Seminar - Subharup Guha

You may follow the seminar at the following link.

“Whose Population Are We Describing? Generalizable Comparisons from Pooled Observational Studies” 

SPEAKER: Subharup Guha (Dartmouth College)

ABSTRACT: Researchers frequently combine several observational studies in order to compare groups defined by exposure or by disease subtype. The samples are rarely representative of the population that the conclusions are meant to describe. Weighting is the standard remedy, and every weighting scheme implicitly defines a pseudo-population that the resulting estimate actually describes. Inverse probability weights and overlap weights construct pseudo-populations in which the groups are equally prevalent, and such a population may bear little resemblance to the one of policy interest. The target population can instead be specified by the researcher rather than determined by the estimator. Known population proportions from an external source, such as a census or a national registry, are supplied as inputs. Among all weightings consistent with those proportions, the procedure selects the one that retains the most information, as measured by effective sample size, and familiar estimators arise as special cases of the construction. The same weights support a wide range of estimands, including group-specific means, dispersions, quantiles, and correlations among several outcomes. The approach is illustrated using seven academic medical centers whose patient composition differs sharply from national prevalence and from one another. Relative to existing methods, the procedure recovers roughly sixty percent more usable information from the same patients while describing a population that matches the national composition. The same structure arises well outside biomedicine, in pooled household surveys, in multi-site trials whose external validity is in question, in regional administrative registers of uneven coverage, and in local historical records used to characterize a national population.

BIO: Subharup Guha is Professor of Biomedical Data Science at the Geisel School of Medicine at Dartmouth. His research develops methods for combining evidence across studies whose samples do not represent the population of interest. The work concerns causal comparison under selection, together with the question of which population a reported estimate actually describes. Applications range across cancer epidemiology, health services research, health disparities, and genomics. He previously held faculty positions at the University of Florida and the University of Missouri, and he serves as an associate editor for Statistics in Medicine.