Model identification effects on interrupted time-series analysis

Monte Carlo simulation was conducted to evaluate 12 procedures under varying conditions of dependence and nonnormality. Models examined include white noise, random walks, AR(1), AR(2), and MA(1). A minimum of 1000 replications was applied in each condition. All models were evaluated in null conditions to evaluate Type I error rates. Empirical error rates were tested for significance using McNemar's test for correlated proportions. Only procedures showing adequate control of Type I errors should be considered for power analysis. The most powerful procedure can be determined by minimum Type II error rate among procedures with comparable Type I error control. Sample sizes range from small (N = 10) to moderate (N = 50).