Most of the Department of Health Care Policy’s research involves existing data from previous studies, rather than clinical trials. This type of research requires specialized techniques that HCP’s statisticians bring to each project while also performing their own research to advance the field of statistics.

One challenge of working with healthcare datasets is missing data; subjects don’t answer a question, for example, or there’s a technical error. Without these data, the study results may not be valid. Addressing missing data is the primary research topic of HCP faculty member Yulei He, MS, PhD. He is bringing this expertise to CanCORS (Cancer Care Outcomes Research and Surveillance), the National Cancer Institute’s multisite study of quality and patterns of cancer care and their impact on patient outcomes. As the study involves 12,000 patients and 700–800 variables, there is ample opportunity for missing data. For example, in the cancer registry, some of the key treatment variables are underreported. He developed models to “impute”—or correct—these misreported variables using information from medical records data.

HCP faculty member Mary Beth Landrum, PhD, applies techniques that aim to infer causes—rather than just associations—from observational data. For example, a quality-of-care study she is working on compares outcomes for cancer patients served by VA hospitals and by Medicare; these settings serve very different populations, making it difficult to draw valid comparisons. Using a technique called propensity scores, Landrum can find comparable treatment groups within each population, making it possible to compare outcomes. She’s also working on a study with Nancy L. Keating, MD, MPH, looking at which cancer treatments are effective; in this study she is using both propensity score methods and another technique for making causal inferences—instrumental variables (IV)—to determine which practices are effective.

One of the difficulties in measuring quality of care is that researchers generally can’t run clinical trials, randomly assigning patients with the same condition to hospitals and analyzing the outcomes. Researchers must instead use statistics on data such as hospital records and patient evaluations to profile hospitals. HCP faculty member Sharon-Lise Normand, PhD, brings an expertise in profiling to research projects at HCP and to her position as director of Mass-DAC, the data-coordinating center that collects, analyzes, and reports on the quality of care for cardiac patients in Massachusetts' hospitals.

Another ongoing national study with leadership from HCP researchers is the Consumer Assessments of Healthcare Providers and Systems (CAHPS) survey. HCP faculty member Alan M. Zaslavsky, PhD—who has served on several advisory panels on the U.S. census—brings his expertise in survey analysis to this project  He developed analysis procedures that are incorporated in software used to prepare health plan and hospital quality reports throughout the country. Results of an annual CAHPS survey of about 700,000 Medicare beneficiaries, analyzed at HCP, are included each year in the handbook beneficiaries use to choose their health plans, and the survey’s assessments drive plans’ quality-improvement activities.

Zaslavsky has an ongoing interest in correlating survey data with administrative records when survey data alone can’t provide all the answers. In the census, data for administrative sources such as tax records can be used to estimate the number and characteristics of households missed by the census.  By using multiple data sources, statisticians can create more complete data.

Faculty member A. James O’Malley, PhD, is researching in a relatively new area for statisticians: social network analysis. Although this analysis has historically been the domain of sociologists, statisticians are now developing new statistical methods to address a broader range of questions. One study O’Malley is working on looks at the effect of behavior-based health traits on a friendship network. Are people with similar body mass index more likely to form ties? Do health traits spread throughout a population by way of friendship ties? Who is central, and potentially most influential, in these networks?

This type of analysis has already shown that people with similar body mass index tend to form ties, that smokers tend to form ties with other smokers, and that nonsmokers tend to dissolve ties with smokers. Information of this kind might enable health care providers to take advantage of these networks to influence the health of many. And, because it is a relatively new application of statistics, there is potential for O’Malley to make contributions to the methodology.

The variety of challenges at HCP keeps the statisticians stimulated. “It’s the fun of being a statistician,” Zaslavksy said. “This is a great place to work.”