1018 — Checking for Seasonal Changes in Patient Outcomes
Christiansen CL (Center for Health Quality Outcomes and Economic Research, Bedford, MA) , Miller DR
(Center for Health Quality Outcomes and Economic Research, Bedford, MA), Fincke BG
(Center for Health Quality Outcomes and Economic Research, Bedford, MA), Tseng CL
(Center for Health Care Knowledge Management, DVA New Jersey Health Care System), Rajan M
(Center for Health Care Knowledge Management, DVA New Jersey Health Care System), Pogach L
(Center for Health Care Knowledge Management, DVA New Jersey Health Care System)
Previous work shows seasonal differences in patient outcomes, e.g., glucose control in diabetics, vitamin D insufficiency, and alcohol use. Our objective is to describe methodology to test for seasonal effects and interpret regression coefficients related to seasons using sophisticated methodology in the literature but not widely adopted.
We use a cohort of Veterans Health Administration (VHA) clinic users with diabetes and 5 years of A1C values for the application and description of the methodology. Investigation of potential monthly fluctuations began with plots of average A1C values across time (months). Next, A1C values were regressed on patient characteristics, time, and season in a linear model with individuals and VHA stations as random effects. Seasonal variation was tested using two trigonometric terms: sine and cosine, using one year as the period for the observed fluctuation cycle. Their coefficient values were used to estimate the amplitude and phase shift in the standard cosine function and will be detailed in the presentation.
Graphs of monthly A1Cs showed a clear seasonal pattern throughout the 5 years. In the regression model, the effects for the sine/cosine variables were statistically significant (p<.0001). The amplitude of the curve, representing the maximum deviation due to months/seasons in a year from the regression line, was 0.05, a clinically meaningful fluctuation. Additionally, estimates of phase shifts led to identification of February as the maximum of the cosine curve and August as the minimum.
Graphs help identify seasonal patterns in patient outcomes. The addition of two variables in regression models was used to determine statistically significant seasonal effects, to interpret the magnitude of the difference, and to help pinpoint when the maximum and minimum values occur. In the example of A1C values in diabetic veterans we found a fluctuation of plus/minus 0.05 between late-winter and late-summer when the maximum and minimum occurred.
Checking for seasonality using methodology currently in the literature should be used more widely in health services research. Excluding significant seasonal effects could give false results, particularly if outcomes are compared in time periods where the maximum and minimum values occur.