面板数据Fisher型平稳性检验结果矛盾的问题咨询

Absolutely, this scenario is not only possible but actually quite common when working with Fisher-type panel unit root tests. Let’s walk through why this happens and how to make a confident decision.

Why Does This Conflict Occur? #

The four Fisher statistics (inverse chi-square P, inverse normal Z, inverse logit L*, modified inverse chi-square Pm) all combine p-values from individual cross-sectional unit root tests (like ADF), but they use different weighting schemes and distribution assumptions, which leads them to react differently to patterns in your data:

• The Z statistic (inverse normal) averages transformed p-values (converted to standard normal quantiles). It’s more sensitive to the central tendency of p-values, so if a small subset of cross-sections has very large p-values (indicating non-stationarity), this can dilute the overall signal and prevent Z from reaching significance.
• P, L*, and Pm place more weight on small p-values (which signal stationarity in individual cross-sections). For example, P is calculated as -2Σln(p_i)—small p-values translate to large ln(p_i) values, so they dominate the sum. This means even if a few cross-sections are non-stationary, the majority of stationary units can push these statistics to reject the null.

How to Decide Which Result to Trust? #

Follow these steps to resolve the conflict:

1. Dig into individual cross-section results
   • Pull up the p-values from each cross-sectional unit root test. Identify which cross-sections have large p-values (failing to reject the null of non-stationarity). Ask: Are these cross-sections outliers? Do they have shorter time series? Is there a structural break in their data that you didn’t account for?
2. Evaluate test robustness
   • The Pm (modified inverse chi-square) statistic is generally more robust to heteroskedasticity and cross-sectional dependence than the others. If your panel has these features (common in real-world data), prioritize Pm as your primary reference.
   • If you’re working with an unbalanced panel, the Z statistic tends to perform worse than P, L*, and Pm because it’s less tolerant of unequal time-series lengths across cross-sections.
3. Cross-validate with other tests
   • Run alternative panel unit root tests like LLC, IPS, or Pesaran’s CADF (which accounts for cross-sectional dependence explicitly). If most tests reject the null of panel non-stationarity, that reinforces the conclusion from P, L*, and Pm.
   • If the conflicting results persist, consider splitting your panel: test the stationary subset and non-stationary subset separately, or investigate whether the non-stationary cross-sections can be explained by economic theory (e.g., a sector with persistent shocks).

Final Takeaway #

Conflicting Fisher statistics aren’t a red flag—they’re just telling you that different combination methods are picking up different signals in your data. Focus on understanding why Z isn’t rejecting the null (usually driven by a small set of non-stationary cross-sections) and use robustness checks to confirm your conclusion.

内容的提问来源于stack exchange，提问作者Eunhyuk Elf