Bootstrap inference on a nonlinear time series model of advertising effects

This paper deals with the analysis of a nonlinear time series model of the effects of advertising. Given the nonlinear nature of the process it is not possible to rely on the asymptotic inference. Furthermore, we can not provide an (asymptotic) pivotal statistic. Our solution is the application of bootstrap techniques. In particular, we find that the double bootstrap procedure provides good results. In this case, the choice of model-based time series resampling, sieve bootstrap or moving-blocks (circular blocks) bootstrap seems to have negligible effects on the confidence intervals of the parameters.