Time Series Forecasts of Poison Center Call Volume

Abstract

We tested the hypothesis that time series analysis can provide accurate predictions of future poison center telephone call volume by a prospective stochastic time series modeling of calls to a university-based regional poison center. All callers evaluated and managed during two sequential years had the time and date of the call recorded in a computer database. Time series variables were formed for poison center calls per hour. Prediction models were developed from the 1992 data and included four types: raw observations, moving average, means with moving average smoothing, and autoregressive integrated moving average. Forecasts from each model were tested against observations from the first 26 weeks of 1993. Each model's adequacy was tested on residuals by autocorrelation functions, integrated periodograms, linear regression, and differences among the variances. A total of 44,584 calls were received in 1992 and 24,781 in the first half of 1993. Large periodic variations in call volume with time of day were found (p < 0.001). The model based on arithmetic means of each hour of the week with three-point moving average smoothing yielded the most accurate forecasts and explained 58.5% of the variation observed in the 1993 test series (p < 0.001). Time series analysis can provide powerful, accurate short range forecasts of future poison center telephone call volume. Simpler, less expensive models performed best in this study.