Testing for Unit Roots in the Presence of Structural Breaks

Imagine you’re tracking India’s GDP per capita over the last 50 years. You’d see a long period of steady, modest growth, and then, right around 1991, the line seems to “jump” and take off at a much steeper angle. If you were trying to model this, you’d face a critical question: Is this data “wandering” aimlessly (what economists call non-stationary), or did something fundamental just change its path? If you run a standard statistical test, it might get confused by that big 1991 jump and tell you the whole series is unpredictable. This is one of the biggest challenges in time series econometrics, and it’s a problem that can lead to completely wrong conclusions.

This “jump” is what we call a structural break-a sudden, permanent change in the underlying structure of the data. Ignoring it can fool our most common statistical tools. Today, we’re exploring how to handle this, focusing on the groundbreaking solution proposed by economist Pierre Perron in 1989.

What is a unit root and why do we care? 🔗

Before we dive into the solution, let’s quickly refresh our memory on the problem. In econometrics, we are obsessed with whether a time series is stationary or non-stationary.

• A stationary series is one that, while it may fluctuate, tends to return to a constant long-term mean. Think of it like a thermostat: the temperature may go up or down, but it’s always pulled back toward the setting. Shocks are temporary.
• A non-stationary series has no such anchor. It follows what’s called a “stochastic trend,” or more famously, a “random walk.” If it gets hit by a shock, it never returns to its old path. The shock is permanent.

A common cause of non-stationarity is having a unit root. The “unit root hypothesis” is the null hypothesis (the default assumption) in many tests that a series is, in fact, non-stationary.

Why does this matter? Because running a regression analysis with non-stationary data is extremely dangerous. You can find a high “R-squared” and significant t-statistics between two variables that have absolutely nothing to do with each other. This is called spurious regression. For example, you might find a statistically “perfect” relationship between mango production in India and the number of tech startups in Canada, simply because both series were trending upward over time. To avoid this, we must first test our variables for unit roots using tools like the Augmented Dickey-Fuller (ADF) test.

The big problem: When a ‘jump’ looks like a ‘wander’ 🔗

Here’s the catch. Standard unit root tests, like the ADF, are built on the assumption that the series’ “structure” (its mean and trend) is stable over the entire sample period. But what about our 1991 GDP example? Or the 2008 global financial crisis? Or the 2020 pandemic?

These events are structural breaks. When an ADF test looks at a series that is stationary *before* the break and stationary *after* the break (but at a new level or on a new trend), it doesn’t see two stable periods. It sees one giant, chaotic series and misinterprets the single, large jump as evidence of a permanent, random shock. It concludes the series must have a unit root.

In technical terms, the presence of an ignored structural break causes the ADF test to lose power. It becomes heavily biased toward the non-rejection of its null hypothesis. In plain English: The test tells you the series is non-stationary (has a unit root) even when it’s actually stationary around a broken trend. This is a massive problem, as it could lead a researcher to difference their data unnecessarily, or worse, abandon their model altogether.

Enter Pierre Perron (1989): Accounting for the break 🔗

In his 1989 paper, “The Great Crash, the Oil Price Shock, and the Unit Root Hypothesis,” Pierre Perron challenged the growing consensus that most macroeconomic variables had unit roots. He argued that this conclusion was largely an illusion created by failing to account for a few massive, world-changing events.

His solution was elegant and intuitive: If you know a big event happened, don’t ignore it-model it directly. Instead of letting the structural break contaminate the test, you should explicitly include it in the test regression. By “soaking up” the effect of the break, the test can then properly examine the behavior of the series *before* and *after* the event.

This became known as the Perron (1989) test. It’s an exogenous break test, which is a key distinction. “Exogenous” means the researcher must specify the break date (which we’ll call `TB`) *in advance*. You can’t just let the data “pick” its own biggest break; you must have an external, theory-based reason, such as a major policy change, a war, or a financial crisis. This is different from later “endogenous” break tests (like the Zivot-Andrews test) which are designed to *find* an unknown break date for you.

How the Perron test works: Using dummy variables 🔗

The “magic” of the Perron test comes from adding simple dummy variables to the standard unit root test equation. A dummy variable is just a variable that takes a value of 0 or 1 to indicate the absence or presence of some effect. Perron’s original paper proposed three main models to capture different *types* of breaks. Let’s look at the dummy variables required, as mentioned in the topic summary.

The ‘level dummy’ for a shift in intercept 🔗

This is used to model a “crash” or a sudden, permanent shift in the *level* of the series. The series’ growth rate (slope) doesn’t change, but its baseline “intercept” does.

• The break it models: A one-time jump up or down.
• The real-world example: Think of the 1991 Indian economic reforms. The *potential level* of India’s GDP arguably shifted to a new, permanently higher path almost overnight, even if the underlying growth *rate* took time to change.
• The dummy variable (Level Shift): This dummy, often labeled `DU`, takes the value 0 for all periods *before* the break date `TB` and the value 1 for all periods *after* the break.

 DU = { 0 if t <= TB { 1 if t > TB 

The ‘trend dummy’ for a change in growth rate 🔗

This is used to model a change in the *slope* of the trend line. The series might not have “jumped,” but its long-term growth rate has fundamentally changed.

• The break it models: A change in the steepness of the trend.
• The real-world example: The 1973 oil crisis. Perron argued this event didn’t just cause a one-time drop in GDP; it permanently slowed the *rate of growth* for many Western economies for years to come.
• The dummy variable (Trend Shift): This dummy, often labeled `DT`, takes the value 0 for all periods *before* the break and the value `t – TB` (time minus the break period) for all periods *after* the break. This clever design lets it capture a new, separate trend line that starts at 1, 2, 3… right after the break.

 DT = { 0 if t <= TB { t - TB if t > TB 

The ‘pulse dummy’ for a one-time spike 🔗

This variable is slightly different. It’s designed to capture a single, massive spike (or dip) that happens in one period and then vanishes. This is often called an “additive outlier.”

• The break it models: A one-period anomaly.
• The real-world example: A single quarter of disastrously low agricultural output due to a severe drought, after which production returns to its normal trend. Or a stock market “flash crash” that lasts one day.
• The dummy variable (Pulse): This dummy, often labeled `D(TB)`, takes the value 1 *only* for the single period *immediately following* the break (`t = TB + 1`) and 0 for all other periods.

 D(TB) = { 1 if t = TB + 1 { 0 otherwise 

Perron’s three main test models (A, B, and C) combine these dummies to test for a level shift, a trend shift, or both happening at the same time.

Building the augmented test regression 🔗

So, how do we use these? We don’t just run a new test. We take the familiar Augmented Dickey-Fuller (ADF) test regression and simply… add the dummies.

For example, a test for a simple level shift (like Model A) would look something like a standard ADF regression, but with the `DU` (level dummy) variable added as an extra regressor. This allows the regression to estimate the size of that jump.

The “Augmented” part of the ADF test is, of course, still critical. After we run this new regression (with our dummy variables included), we must examine the residuals (the errors). If those residuals show signs of autocorrelation (meaning today’s error is correlated with yesterday’s error), we must “clean” the regression. We do this by adding lagged difference terms of the dependent variable (e.g., `Δy(t-1)`, `Δy(t-2)`, etc.) as additional regressors until the autocorrelation is gone. This ensures our final test statistic is reliable.

Interpreting the results (and a crucial warning) 🔗

Once you have your final, augmented regression, you look at the t-statistic for the `y(t-1)` variable (the one testing for the unit root). But-and this is the most important part-you cannot use the standard Dickey-Fuller critical values.

Think about it: we’ve fundamentally changed the test by adding these deterministic dummy variables. The statistical distribution of the t-statistic is no longer the same. Perron (1989) had to solve this problem by running massive simulations to generate an entirely new set of critical values specifically for his tests.

So, the final step is to compare your calculated t-statistic to the Perron-specific critical values, which depend on the model you used (A, B, or C) and the location of the break. If your t-statistic is *more negative* than the Perron critical value, you can finally reject the null hypothesis of a unit root.

In doing so, you’ve successfully demonstrated that the series is not a “random walk.” It is, in fact, a trend-stationary process that was merely interrupted by a structural break. You’ve seen through the illusion and found the true, stable process underneath.

What do you think? When you look at economic data for your own country or industry, what major events (like a policy change, a new technology, or a crisis) come to mind? How do you think ignoring those events might change a company’s or a government’s economic forecast?

References

1. https://www.aptech.com/blog/unit-root-tests-with-structural-breaks/
2. https://www.ssc.wisc.edu/~bhansen/718/Perron1989.pdf
3. https://www.bu.edu/econ/files/2019/01/structural-change-oxford.pdf