Uncovering psychiatric test information with graphical techniques of Exploratory Data Analysis

This article illustrates how Exploratory Data Analysis (EDA) can complement conventional statistical methods in evaluating psychiatric tests. Using one recent EDA computer program, we evaluated the ability of repeated psychiatric screening tests (the General Health Questionnaire [GHQ]) to predict medical and psychiatric service use in a Health Maintenance Organization (HMO), the Harvard Community Health Plan (HCHP). Using a stratified random sample of 244 new HCHP enrollees and viewing three-dimensional graphs of their data from multiple perspectives, we found two subpopulations: low GHQ scorers, for whom the tests did not predict service use; and high scorers, for whom they did. Surprisingly, improving scores forecast increased use and chronically high scores predicted diminished use. Using another stratified random sample of 213 new HCHP enrollees, and with scatterplot matrices from another interactive computer program, we found that high and unchanging GHQ scores forecast HMO dropout. We examine possible interpretations--for example, that chronically distressed patients may become immobilized, diminish service use, and ultimately leave the HMO. We also explain how EDA methods may help uncover elusive results in other data (e.g., mental health outcomes).