Title: Letter to editor: Incident depression increases medical utilization in Medicaid patients with hypertension
Abstract: We read with interest the Original Research article by Breunig et al. [1] and as investigators experienced in using Medicaid datasets for clinical research, we would like to comment on the methodology used by the authors. We are primarily concerned about the difference-in-difference (DID) analytic methods employed in this study and about some of the conclusions drawn. While interrupted time series analyses employing differencing techniques have increasingly been used in health care to demonstrate the effects of policy or pharmacotherapy changes over time, their applicability to demonstrating clinical/ patient changes in this cohort is somewhat problematic. Calculating difference scores in a trend analysis is usually based on having a series of observations over time prior to and after the policy or therapy intervention, with some statisticians recommending at least 25–30 observed time periods over which to calculate the difference scores to establish a trend [2,3]. In this investigation, medical utilization was aggregated for 12 months before and after the first diagnosis of depression among cases, and was compared to the medical utilization in a propensitymatched control group over the same time period. Thus, the research design appears to be a simple pre–post comparison with control group. The preferred method for analyzing pre–post design data is the general linear model framework using the year 1 utilization data prior to the development of depression variable as one main independent covariate, and the depression/control condition variable as the second main covariate predicting year 2 utilization or the difference between year 1 and year 2 utilization as the dependent variable, in order to eliminate systematic bias, regression to the mean and reduce error variance [4–6]. It is not clear from the authors’ description whether year 1 utilization was controlled for in their regression analyses. Moreover, the DID regression approach is heavily dependent on incorporating various assumptions regarding known and unknown related conditions or variables to maximize its internal validity. Implementation of these assumptions requires additional computational adjustments and sensitivity analyses to ensure proper inference or generalizability. In most clinical research designed to inform clinical practice, these covariates would be added to the general linear model regression equation in a transparent manner, according to the hypothesized relationships among variables, and the effect of each related covariate is demonstrated/presented, not only to validate the clinical/conceptual rationale employed, but also to allow clinicians to compare the case-mix of the research cohort employed to that of their own caseloads, to which the final results will, hopefully, be generalized. However, in this DID study, the conceptual rationale for the inclusion and definition of these covariates is not provided (e.g., the Charleson Comorbidity Index, or the comorbid conditions included because they were not in the