Within the context of threshold regressions, we show that asymptotically-valid likelihood-ratio-based confidence intervals for threshold parameters perform poorly in finite samples when the threshold effect is large. A large threshold effect leads to a poor approximation of the profile likelihood in finite samples such that the conventional approach to constructing confidence intervals excludes the true threshold parameter value too often, resulting in low coverage rates. We propose a conservative modification to the standard likelihood-ratio-based confidence interval that has coverage rates at least as high as the nominal level, while still being informative in the sense of including relatively few observations of the threshold variable. An application to thresholds for US industrial production growth at a disaggregated level shows the empirical relevance of applying the proposed approach.
Bibliographical notePublisher Copyright:
© 2018 Walter de Gruyter GmbH, Berlin/Boston 2018.
- confidence interval
- finite-sample inference
- inverted likelihood ratio
- threshold regression