The Reliability and Accuracy of Time Series Model Identification

Abstract

The most widely employed procedure for interrupted time series analysis consists of a two-step procedure: (1) determining the ARIMA model by examining the pattern of autocorrelations and partial autocorrelations; and (2) employing a general linear model solution after the effect of dependency has been removed. In order to determine the reliability and accuracy of model identification, 12 extensively trained subjects were each asked to identify 32 different computer generated time series. Six commonly occurring models were employed with different levels of dependency (high, medium, or low) and different numbers of data points (N=40 and N=100). The overall accuracy, 28%, was affected by the number of data points, the type of model, and the degree of dependency.