Questions: Chow Test and Detection of Structural Breaks

Question 1

An econometrician estimates a single regression pooling pre- and post-recession data, unaware that the true coefficients shifted after the recession. Compared to the true split-sample estimates, the pooled estimates will be:

• A. Unbiased but less precise, since pooling only reduces degrees of freedom
• B. Biased, because they average over two different regimes and represent neither accurately
• C. More efficient, since using all data always reduces variance
• D. Identical to the split-sample estimates, since OLS minimizes overall residuals in either case

Answer: (B)

Ignoring a structural break biases the coefficient estimates because the pooled regression constrains the slope and intercept to be the same across periods when they actually differ. The fitted coefficients will be a weighted average of the two regime coefficients — accurately describing neither period. Out-of-sample forecasts based on these estimates will be systematically wrong if the forecast period is in a different regime than the estimation sample.

Question 2

A researcher tests for a structural break by computing a Chow-like F-statistic at every possible break date in their sample and reports the largest one as significant. Why is this approach problematic?

• A. The F-distribution is not valid for time series data regardless of the search procedure
• B. Searching across dates inflates the false-positive rate because finding the maximum over many tests exploits random variation, so standard F critical values are too small
• C. The procedure is invalid because the break date must be chosen after looking at the residuals
• D. Splitting a sample into two periods always violates the OLS assumption of homoskedasticity

Answer: (B)

This is the data-snooping problem. If you test enough possible break dates, you will almost certainly find one where the split looks significant by chance — even in stable data. The standard F critical values assume you chose the break date independently of the data. The Quandt-Andrews test addresses this by computing the statistic at every candidate date and taking the maximum, then comparing it against specially derived critical values that account for the search over break dates.

Question 3

The Chow test is a fundamentally new testing procedure, distinct from the F-test for joint significance.

• T. True
• F. False

Answer: False

The Chow test is the F-test logic applied to a specific restriction: that regression coefficients are the same across two subperiods. The restricted model pools all data (imposing coefficient equality); the unrestricted model estimates separate regressions for each period. The F-statistic measures whether the reduction in RSS from splitting justifies the extra parameters. Understanding this connection makes the Chow test easier to remember and apply — it's the same framework, just a different restriction.

Question 4

The classic Chow test requires the researcher to specify the break date before looking at the data.

• T. True
• F. False

Answer: True

This is a genuine limitation of the classic Chow test. The test is valid only when the break date is chosen independently of the data — for example, because it corresponds to a known policy change, financial crisis, or institutional shift. When the break date is unknown and chosen by searching the data, the standard critical values are no longer valid because the test statistic is the maximum of many correlated tests. The Quandt-Andrews test was designed precisely for the unknown-break-date case.

Question 5

What is the null hypothesis of the Chow test, and what does rejecting it tell you about your regression model?

Model answer: The null hypothesis is that the regression coefficients (intercept and all slopes) are identical across the two subperiods — i.e., there is no structural break. Rejecting the null means at least one coefficient differs significantly between periods, indicating that the relationship between variables changed at the proposed break date. This implies the pooled model is misspecified and that separate regressions for each period provide a better description of the data.

Rejection of the Chow test null does not tell you *which* coefficient changed or by how much — only that the pooled restriction is rejected. Follow-up analysis can examine whether the intercept, one slope, or all coefficients changed. The practical implication is that forecasts, policy simulations, or causal inferences based on the pooled model are unreliable if they extrapolate across the break.