Structural Break Testing

Definition

Structural break testing refers to procedures for detecting and estimating parameter instability in econometric models — changes in means, trends, slopes, or variances at one or more unknown dates within the sample. The central challenge is that ignoring genuine breaks biases inference toward spurious persistence (unit roots, long memory, integrated GARCH (IGARCH)), while falsely imposing breaks inflates the risk of spurious rejections.

Key Ideas

A stationary series around a broken trend can fool standard Dickey-Fuller (DF) tests into non-rejection of the unit root null (Perron 1989, the "structural break as mimicking unit root" result).
The break date is unknown and should be treated as endogenous (Zivot-Andrews 1992); treating it as known risks data snooping on the Nelson-Plosser data.
With multiple breaks, conditioning on incorrectly located earlier breaks distorts inference for later ones; global SSR minimization over all partitions is preferred (Bai-Perron 1998).
Structural change in the regressors themselves (not just the coefficients) invalidates standard F-test size; the fixed-regressor bootstrap controls size in this case (Hansen 2000).
The same logic applies to variance: unaccounted generalized autoregressive conditional heteroskedasticity (GARCH) regime changes cause estimated $\alpha+\beta \to 1$ , a spurious finding of IGARCH (Hillebrand 2004).

How It Works

Classical Tests for Coefficient Constancy (Chow 1984 §10)

The following tests all address $H_0: \beta_t = \beta$ (constant) vs. a changing-coefficient alternative. They predate the Zivot-Andrews / Bai-Perron literature and remain standard diagnostics.

Chow (1960) F-test — requires a known break date $T_B$ . Estimate separate regressions on $[1, T_B]$ and $[T_B+1, T]$ ; the ratio of the improvement in sum of squared residuals (SSR) to the restricted SSR, scaled by degrees of freedom, is $F$ -distributed under $H_0$ . Critical limitation: the test is not valid when $T_B$ is chosen by inspecting the data.

Cumulative Sum (CUSUM) and CUSUM² (Brown, Durbin, Evans 1975) — based on recursive residuals $w_t = (y_t - x_t'\hat{\beta}_{t-1})/(1 + x_t'(X_{t-1}'X_{t-1})^{-1}x_t)^{1/2}$ . The CUSUM statistic $W_r = \sum_{t=k+1}^r w_t / \hat{\sigma}$ is compared against straight-line boundaries $\pm a\sqrt{T-k}(1 + 2(r-k)/(T-k))$ . CUSUM² uses cumulated squared recursive residuals and detects variance change. Both have asymptotically known distributions under $H_0$ .

Quandt (1960) maximum likelihood-ratio test (LRT) — maximizes the Chow F-statistic over all interior break dates $\lambda = T_B/T \in [\lambda_0, 1-\lambda_0]$ . Precursor to the Zivot-Andrews statistic; critical values are not standard $F$ and must be tabulated for each $(\lambda_0, k)$ .

Nyblom (1983) LMP test — locally most powerful test of $H_0: V = 0$ against the random-walk time-varying-parameter (TVP) alternative $\beta_t = \beta_{t-1} + \eta_t$ . The test statistic is the score evaluated at the null: $L = T^{-2}\sum_{t=1}^T S_t' (X'X)^{-1} S_t / \hat{\sigma}^2, \qquad S_t = \sum_{s=1}^t x_s(y_s - x_s'\hat{\beta})$ Asymptotically distributed as a functional of Brownian bridges; critical values tabulated.

Pagan (1980) score test — also tests $H_0: V = 0$ ; similar score-based approach. Detects any form of coefficient variation, not just the random-walk alternative.

Perron (1989) — Single Known Break

Perron (1989) demonstrates that standard DF tests have zero asymptotic power against trend-stationary (TS) alternatives with a single known structural break. Three models are considered, distinguished by the form of the break in the deterministic component; each has a distinct null, alternative, and limiting distribution.

Break date and fraction. Let $T_B$ be the break date, $\lambda = T_B/T \in (0,1)$ the break fraction, and define: $DU_t = \mathbf{1}(t > T_B)$ (post-break level dummy), $DT^*_t = t - T_B$ for $t > T_B$ , else 0 (post-break trend), $D(T_B)_t = \mathbf{1}(t = T_B + 1)$ (pulse dummy at the break).

Model A — Crash (intercept shift).

Null: $y_t = \mu + d D(T_B)_t + y_{t-1} + e_t$ (unit root with level pulse)
Alternative: $y_t = \mu_1 + \beta t + (\mu_2 - \mu_1) DU_t + e_t$ (TS with permanent level shift)
Test regression (additive-outlier (AO) form): $y_t = \hat\mu + \hat\beta t + \hat\delta DU_t + \hat\alpha y_{t-1} + \sum_{j=1}^k c_j \Delta y_{t-j} + \hat{e}_t$

Model B — Changing growth (slope shift).

Null: $y_t = \mu_1 + y_{t-1} + (\mu_2 - \mu_1) DU_t + e_t$ (random walk with drift change)
Alternative: $y_t = \mu + \beta_1 t + (\beta_2 - \beta_1) DT^*_t + e_t$ (TS with slope break)
Test regression (innovational-outlier (IO) form): $y_t = \hat\mu + \hat\beta_1 t + \hat\gamma DT^*_t + \hat\alpha y_{t-1} + \sum_{j=1}^k c_j \Delta y_{t-j} + \hat{e}_t$

Model C — Both. Both an intercept and slope shift are allowed; the test regression includes $DU_t$ , $t$ , and $DT^*_t$ .

Theorem 1 (zero power under the alternative). Under the trend-stationary alternative, the probability limit of $\hat\alpha$ from the test regression satisfies: for Model A (crash), $\bar\alpha \to [(\mu_1-\mu_2)^2 A] / [(\mu_1-\mu_2)^2 A + \sigma^2_e] < 1$ — power is positive but $\bar\alpha \not\to 0$ ; for Model B (changing growth), $\bar\alpha \to 1$ exactly — the test is asymptotically inconsistent against slope breaks.

Theorem 2 (limiting distribution under the null). Under the unit root null (with $\sigma^2 = \sigma^2_e + \sigma^2_u$ , where $\sigma^2_e$ is the regression error variance and $\sigma^2$ is the long-run variance):

$T(\hat\alpha^i - 1) \Rightarrow \frac{H_i}{K_i}, \qquad t_{\hat\alpha^i} \Rightarrow \frac{(\sigma/\sigma_e) H_i}{(g_i K_i)^{1/2}}$

where $H_i$ , $K_i$ , $g_i$ are functionals of a standard Brownian motion $W(r)$ that depend on $\lambda$ and differ across Models A, B, C. Because the break variables $(DU_t, DT^*_t)$ are nuisance regressors under the null, the limiting distributions are non-standard and shift with $\lambda$ .

Critical values. The 5% critical value for the $t$ -statistic is approximately: Model A $\approx -3.76$ ; Model B $\approx -3.96$ ; Model C $\approx -4.24$ (tabulated for each $\lambda \in \{0.1,\ldots,0.9\}$ in Tables IV–VI). The standard DF value (with constant and trend) is $-3.41$ ; ignoring the break makes rejection too easy.

AO versus IO. The Additive Outlier (AO) model assumes the break in the mean is instantaneous — used for Model A (1929 crash). The Innovational Outlier (IO) model allows the break to propagate gradually through the AR polynomial — used for Model B (1973 oil shock). The two forms require different test regressions and, for the AO model, the two-step procedure of first estimating the break-corrected series and then testing on residuals.

Empirical results. Applied to 13 Nelson-Plosser macroeconomic series: 11 reject the unit root (real gross national product (GNP), nominal GNP, industrial production, employment, wages at 1%; real per capita GNP, GNP deflator, money stock, common stock prices, real wages at 2.5%); non-rejections: consumer prices, velocity, interest rate. Quarterly real GNP (not in NP dataset): $\hat{t} = -3.98$ , $\hat\alpha = 0.86$ , lag $k=10$ .

Exogeneity postulate. The break dates 1929 and 1973 are treated as exogenous, known constants. This delivers sharper critical values than searching over $\lambda$ , but raises data-snooping concerns addressed by Zivot-Andrews (1992); see the Zivot-Andrews (1992) subsection below.

Erratum. Perron-Vogelsang (1993) corrected errors in the AO-model asymptotic distributions. The IO-model results and all empirical conclusions of the paper are unaffected.

See Perron (1989).

Zivot-Andrews (1992) — Endogenous Break Date

Treats $T_B$ as unknown; searches over $\lambda \in \Lambda$ (a trimmed interior of $(0,1)$ ):

$\inf_{\lambda \in \Lambda} t_\alpha(\lambda)$ This minimizes the t-statistic over all candidate break dates, giving greatest weight to the trend-break hypothesis. Critical values are tabulated for this statistic (stricter than Perron's known-break values). Zivot-Andrews also show the break is frequently identified one period too early, a mis-timing issue.

Bai-Perron (1998) — Multiple Breaks at Unknown Dates

Model:

$y_t = x_t' \beta + z_t' \delta_j + u_t$

for regime $j = 1,\ldots,m+1$ , with $T_j$ unknown.

Estimation: minimize SSR over all (m+1)-partitions with minimum segment length h; dynamic programming achieves this efficiently.

sup F(k; q): tests stability against k fixed known breaks; generalizes Andrews (1993) sup F.

UDmax F(M, q): $\max_{k=1,\ldots,M} \sup_{\Lambda_k} F_T(\lambda_1,\ldots,\lambda_k; q)$ with equal weights $a_k = 1$ ; null is stability, alternative is unknown $k \leq M$ breaks.

WDmax F(M, q): $\max_{k=1,\ldots,M} \tfrac{c(q,\alpha,1)}{c(q,\alpha,k)} \cdot \sup F_T(k; q)$ ; weights equalize marginal p-values across k so that power does not decline as the true number of breaks grows. Critical values tabulated for $M=5$ , $\alpha=0.05$ .

Bai (1999) sequential LR test: null l breaks vs. l+1; statistic is the normalized drop in optimal SSR from l to l+1 breaks; limiting distribution has known closed-form density over the search interval, yielding analytic critical values.

Hansen (2000) — Systems with Non-Stationary Regressors

Extends structural change testing to systems where regressors $x_{ni}$ may themselves be subject to structural change or non-stationarity. Key insight: standard SupF/ExpF/AveF tests are not invariant to structural change in the regressors — a significant test could reflect instability in $x$ rather than in $\beta$ .

Fixed-regressor bootstrap: treats all regressors (including lagged dependents) as if fixed; generates bootstrap null distributions that are valid even under arbitrary structural change in $x$ and under heteroscedastic errors. Achieves near-correct size in small samples, unlike asymptotic critical values.

Bayesian Random Shift Models — McCulloch-Tsay (1993)

Classical break tests require specifying the number of break points. A fully Bayesian alternative treats each observation's shift indicator as a latent variable, eliminating the model-selection problem.

RLAR (Random Level-shift AR): The mean level evolves as $\mu_t = \mu_{t-1} + \delta_t\beta_t$ , where $\delta_t \overset{iid}{\sim}\mathrm{Bernoulli}(\varepsilon)$ and $\beta_t \sim \mathcal{N}(0, \tau^2)$ is the jump magnitude. The autoregressive (AR) process $y_t = \mu_t + x_t$ runs around this stochastically shifting mean.

RVAR (Random Variance-shift AR): The innovation variance shifts multiplicatively — $\sigma_t^2 = \sigma_0^2 \cdot G_t$ where $G_t$ cumulates the multiplicative jumps $\beta_j^2$ at times when $\delta_j = 1$ .

Gibbs sampler. All full conditional posteriors are conjugate and closed-form: AR coefficients ( $\mathcal{N}$ ), innovation variance ( $\chi^{-2}$ ), jump sizes ( $\mathcal{N}$ or $\chi^{-2}$ ), shift probability $\varepsilon$ ( $\mathrm{Beta}$ ). The binary indicators $\delta_t$ are sampled from their posterior odds ratios. Drawing 3 indicators jointly rather than one at a time improves convergence.

Probit extension: Replace constant $\varepsilon$ with $\varepsilon_t = \Phi(w_t'\gamma)$ where $w_t$ are exogenous predictors (Albert-Chib 1993 data augmentation for $\gamma$ ). Enables prediction of future shifts: given a known upcoming event, the model returns $P(\text{shift at }t \mid w_t)$ .

Contrast with Markov-switching VAR (MS-VAR): Markov-switching models (Hamilton 1989) allow persistent states — once a regime change occurs, the system stays in the new state until the next switch. RLAR/RVAR assume each period's shift is independent and identically distributed (i.i.d.) Bernoulli, making them better suited to isolated, infrequent jumps rather than persistent regime shifts.

Gasoline application: Monthly US price changes (1978–1991, $n=158$ ). RVAR posterior shift probability $\hat\varepsilon \approx 0.11$ ; detected shifts align with OPEC events. RVAR-Probit: $P(\text{shift} \mid \text{world event}) \approx 0.73$ vs. $P(\text{shift} \mid \text{no event}) \approx 0.12$ . See McCulloch-Tsay (1993).

Bayesian Changepoint in Markov-Switching Model — Kim and Nelson (1999b)

When the parameter instability occurs within a regime-switching model — i.e., the hyperparameters of the Markov-switching process themselves shift — classical sup-F tests are non-standard because the transition probability $q_{00}$ is a nuisance parameter absent under the null. Kim and Nelson (1999b) resolve this by treating the changepoint as an absorbing Markov state $D_t$ :

$\Pr[D_t=0|D_{t-1}=0] = q_{00}, \qquad \Pr[D_t=1|D_{t-1}=1] = 1$

so that once the economy transitions to the post-break regime, it stays there permanently. Conditional on $D_t$ , the MS parameters shift: $\mu^*_{S_t,t} = \mu_{S_t} + \mu_{S_t,\text{post}}D_t$ and $\sigma^2_t = (1-D_t)\sigma^2_0 + D_t\sigma^2_1$ . The Bayesian framework eliminates the nuisance-parameter problem: $q_{00}$ appears in the joint posterior and is integrated out via the Gibbs sampler. The changepoint posterior $p(\tau|\text{data})$ is a byproduct of sampling the $D_T$ sequence.

Marginal likelihoods (Chib 1995/1998 reduced-run decomposition) compare four models: no break, break in shift parameters only, break in variance only, break in both. Applied to U.S. real gross domestic product (GDP) 1953:II–1997:I, the break in shift parameters model wins ( $\ln m = -247.01$ ), locating the break at 1984:Q1 and identifying a narrowing boom–recession gap as the dominant source of the Great Moderation.

See Great Moderation, Markov-Switching VAR, Kim-Nelson (1999b).

Why It Matters

Perron's lesson: macroeconomic persistence tests are compromised whenever the deterministic component is misspecified; unit root evidence may simply reflect un-modelled shifts in mean or trend growth.
Multiple breaks: a single-break specification when two breaks are present can itself generate a spurious unit root; the Bai-Perron framework guards against this.
Variance analogue: Hillebrand (2004) shows the same spurious-persistence mechanism operates in GARCH — regime changes masquerade as IGARCH, motivating structural break tests before interpreting volatility persistence.
Forecasting: Pesaran-Timmermann (2004) show that including some pre-break data can reduce forecast mean squared error (MSE), even when the true AR coefficients shift at the break — using only post-break data is not always optimal.

Open Questions

Power against smooth parameter evolution (gradual structural change) remains limited for all discrete-break procedures; Smooth transition autoregressive (STAR)-type tests (Lin-Teräsvirta) are complementary.
Distinguishing structural breaks from genuine long memory in finite samples is unresolved; see Long Memory and Fractional Integration.
Optimal choice of upper bound M on the number of breaks is somewhat arbitrary; information criteria for breaks (Bayesian Information Criterion (BIC) analogues) are an alternative.
Cointegrated VAR extensions (Hansen-Johansen 1999, Johansen et al. 2000) extend the methods to systems with long-run relations; likelihood ratio tests for m vs. m+k changes have been developed.