Uncovering surprises in home health services' effects post-hospitalization. Patients seen at home by a nurse, occupational therapist, or physical therapist tend to readmit to hospital at a high rate, but this might be expected based on the higher-than-average severity of their conditions. With this accounted for (statistically controlled) one would expect a reversal of the readmission effect. Not true for any of the seven US hospital populations that colleagues and I examined. A conference presentation and a journal submission, provisionally accepted for publication by The American Statistician, explored the reasons why. Our article went on to specify the conditions under which such a reversal could occur in the context of the common analytic technique of logistic regression.
Explaining the incidence of surgical errors. A researcher working with a national nursing organization forged the opportunity to collect a wealth of data on the occurrence of wrong-site surgeries (e.g., removal of the wrong kidney) as well as on staffing, staff education, and climate in several hundred US hospitals. She sought to assemble these many types of objective and subjective data into a model that could help predict incidence of these catastrophic errors. Because such errors are rare given the vast number of surgeries performed, standard linear regression would have been ineffectual. We helped her devise more general linear models -- using logistic and Poisson regression -- that were more sensitive to distinctions between the rare and the "extremely rare." We also advised on scale development (combining survey responses into scale scores for greatest validity and reliability) and on model specification (maximizing the model's stability and sensitivity by choosing the most promising variables and the best number to include).
Making medical office procedures more efficient. An internal medicine practice kept detailed records on telephone contacts with patients but lacked the ability to analyze the data in depth. We used analysis of variance, time series modeling, and data visualizations to identify peaks and lulls at specific times of year, week, and day. We also helped the clinicians evaluate the effect of changes in office practices over time. Our report helped the office to plan its level of staff availability and thus achieve greater efficiency in its contacts with patients. We also gave them tools to plan schedules so that the staff members most able to handle specific types of questions (on prescriptions, referrals, etc.) would be able to match cyclical demands.
Understanding bullying among nurses and its implications for staff turnover. For a Ph.D. dissertation, a nurse-researcher had surveyed 500 nurses to learn about ways in which they might have been mistreated by fellow nurses; how these harmful behaviors might have related to one another; and how, combined with demographic factors, they might explain some nurses' intention to leave the profession. We guided her through power transformations, to make the variables more amenable to modeling, and factor analysis, to sort two dozen behaviors into four underlying types (Punishment, Belittling, Exclusion, and Undermining). We used these factors in multiple regression and analysis of covariance (ANCOVA) models to explain the outcome of intention to leave. Along the way we checked diagnostic statistics and graphs to see whether any relationships hinged on gender, ethnicity, or nursing experience or nursing specialty. We advised and supported her as she underwent a successful Ph.D. defense.
The research led to publication of three articles in peer-reviewed nursing research journals, the latter article exploring the most valid and efficient ways of gauging the extent to which bullying affects a workplace. We found that our new instrument comprised of just four survey questions was superior to a popular one that required 22. In response to this paper we have fielded many researchers' requests for help in replicating our approach.