Testing for Structural Breaks in Small Samples
IMF Working Papers, March 1, 2008
Source details
- Canonical URL
- Testing for Structural Breaks in Small Samples
Other formats
Bibliographic details
- Authors: Sergei Antoshin, Andrew Berg, Marcos R Souto
- Published: March 1, 2008
- Series: IMF Working Papers
Summary
- Bai and Perron (2006) show their approach for testing for multiple structural breaks in time series works well in large samples but exhibits substantial deviations in both the size and power of their tests in smaller samples.
- The authors propose modifying Bai and Perron’s methodology for small samples by using Monte Carlo simulations to determine sample-specific critical values each time the test is run.
- Simulations are used to produce practical suggestions on handling serial correlation, model misspecification, and the use of alternative test statistics for sequential testing.
- For most types of data generating processes in samples with as low as 50 observations, the proposed modifications perform substantially better.
Key findings
- Bai and Perron (2006) methodology: performs well in large samples but shows substantial deviations in size and power in small samples.
- Proposed modification: determine sample-specific critical values via Monte Carlo simulations each time the test is applied.
- Performance claim: proposed modifications perform substantially better for most data generating processes in samples with as low as 50 observations.
Methodological approach
- Use Monte Carlo simulations to obtain sample-specific critical values under each execution of the test.
- Address issues raised by small samples through simulation-driven calibration rather than relying on asymptotic critical values.
- Consider alternative test statistics for sequential testing as part of the modified procedure.
Practical recommendations
- Implement Monte Carlo-derived, sample-specific critical values whenever applying multiple structural-break tests in small samples.
- Adjust testing procedures to account explicitly for serial correlation.
- Guard against model misspecification by using simulation evidence to guide test choice and calibration.
- Consider alternative test statistics for sequential testing informed by small-sample simulation results.
Implication for empirical work
- Researchers working with time series samples of modest size (including samples with as low as 50 observations) should prefer simulation-calibrated critical values over standard asymptotic critical values to improve size and power properties.
- Attention to serial correlation and model specification is crucial in small-sample structural-break testing; simulation-based adjustments can materially improve inference.
Source: Testing for Structural Breaks in Small Samples — Sergei Antoshin, Andrew Berg, Marcos R Souto; March 1, 2008. IMF Working Paper No. 2008/075.
Content in this bundle
- _wp0875 — Size Tests—Weiner DGP with No Breaks