Zivot-Andrews Structural Break Test

The ADF test assumes no structural break in the data-generating process. If a break exists, the ADF tends to under-reject the unit-root null (low power). The Zivot-Andrews test adds a structural-break dummy and searches endogenously for the best break date, using its own — more negative — critical values to account for this selection.

Model A allows a one-time level shift, Model B allows a change in the trend slope, and Model C allows both. Choose based on economic reasoning: if the series is expected to shift permanently in level (e.g., after a policy change), use A; if only the growth rate changes, use B; if both are plausible, use C. Reporting all three and discussing the results is also common practice.

If the null is rejected, the series is treated as stationary around the estimated break point. Incorporate the break dummy into subsequent regression models (ARDL, VAR, etc.) and re-test for cointegration if needed, using break-robust cointegration procedures.

The original test is for a single time series. For panel data, panel-specific unit root tests that allow for cross-sectional dependence and structural breaks (e.g., Carrion-i-Silvestre et al.) are more appropriate than applying ZA equation by equation.

The number of augmentation lags k is typically chosen by information criteria (AIC or BIC) or by the Ng-Perron sequential testing procedure. A common default is to allow up to 12 lags for monthly data or 4 for quarterly data, then select the model minimising the criterion.