A spurious unit root from omitting the intercept
Fitting an autoregressive model without an intercept to data far from zero produces an apparent unit root. The effect is shown with Lake Huron's water levels, together with the small-sample bias that remains when the model is correctly specified.
Fitting an autoregressive model without an intercept to a series that fluctuates around a large value almost always produces an estimate consistent with a unit root. In most such cases the conclusion is wrong.
The first-order autoregressive model is the simplest time-series model there is, and it can be misapplied in a way that appears reasonable. This note shows the mistake on a classic dataset, Lake Huron's water levels, explains why it happens, and then describes a second, subtler bias that remains even when the model is right. A simulation reproduces both effects.
The relevant literature is as follows.
- Kendall (1954) and Marriott and Pope (1954) derived the small-sample bias of the least-squares autoregressive estimate, approximately when the mean is estimated.
- Dickey and Fuller (1979) tabulated the distribution of the estimator under a unit root, the basis of the tests now used to decide whether a series is a random walk, with separate tables for models with and without a constant.
- Brockwell and Davis (2002) use the Lake Huron series (annual levels, 1875-1972) as a running example of a stationary autoregressive process; it ships with R as
LakeHuron.
The model and its estimator
In its simplest form, the AR(1) model without a constant says
and the least-squares estimate of is a ratio:
For a stationary series with mean zero, , and the precision of is governed by the Fisher information , so its variance is approximately , as in least-squares estimation of the AR(1) coefficient in three regimes.
Example: Lake Huron
In an exercise in my mathematical statistics course, the test data were the annual water levels of Lake Huron from 1875 to 1972, a classic series that ships with R. The levels are around 579 feet above sea level and vary by only a foot or two from year to year.
Applying the estimator above directly to the levels gave , whatever the sample size. Taken at face value that is a unit root: a random walk with no tendency to return to any level.
Source of the apparent unit root
The result is an artefact. The model without a constant assumes the series fluctuates around zero. Lake Huron fluctuates around 579. In the ratio estimator, both sums are dominated by that large mean, roughly in each, and the small fluctuations that carry the actual dynamics barely register. The ratio is driven towards one by the level, not by persistence.
The same artefact appears with any series measured far from zero. A room controlled by a thermostat at 21 °C returns to that temperature every time a window is opened and closed; it is plainly not a random walk. Yet an AR(1) without a constant, fitted to its hourly readings, gives an estimate very close to 1, because the level of 21 dominates fluctuations of a fraction of a degree.
Adding an intercept, or equivalently centring the data on its mean first, removes the artefact. Both approaches gave : a persistent but clearly stationary series. Lake levels have memory, but they do return to their average.
A second-order model fitted better still, with coefficients of about 1.02 and −0.24, an improvement confirmed by the Akaike information criterion.
Small-sample bias
Even when the model is specified correctly, the least-squares estimate of is biased towards zero in short series, a result going back to Kendall and to Marriott and Pope in the 1950s. A Monte Carlo experiment with a true value of , 20,000 simulated series and a known mean of zero gave average estimates of about 0.65 with five observations and 0.77 with fifty. When the mean has to be estimated as well, the bias is considerably larger; Kendall's approximation puts it at about . The bias shrinks roughly in proportion to , but for the short series common in economics it is not negligible. With twenty annual observations and an estimated mean, Kendall's approximation puts the expected estimate near 0.63 when the true value is 0.8. A shock with a true half-life of about 3.1 years then appears to halve in about 1.5, and a government forecasting how long an economic downturn will weigh on revenue would plan for a recovery twice as fast as the data support.
The simulation below illustrates three results: without an intercept, a stationary series far from zero yields ; with an intercept the estimate centres near the truth, minus a small-sample bias of roughly ; and even with a known mean, short series underestimate persistence.
Each run simulates 2,000 stationary AR(1) series with around an adjustable level.
Suggested settings:
- Lake Huron. Press the preset: level 580, 98 years. Without an intercept the average estimate is 1.000, in every run, an apparent unit root. With an intercept it is about 0.765, consistent with Kendall's approximation.
- Lower level. Reduce the level to 0. The two histograms nearly merge: around zero, the intercept hardly matters. At a level of only 10, about six standard deviations of the series, the no-intercept estimate is already above 0.99. A large level is not required to produce the artefact.
- Shorter series. With the level at 0, set the length to 5 years. Even the correctly specified estimator now averages about 0.65 instead of 0.8. At 50 years it averages about 0.77. Short series underestimate persistence.
- Short series with an intercept. At level 580 and 10 years, the estimate with an intercept falls far below 0.8, much further than the known-mean case: estimating the mean uses information where it is most limited.
Effect of the missing intercept. Write with a stationary AR(1) around zero with variance . Without an intercept,
using and . When is much larger than , the fraction vanishes and , whatever the true . For Lake Huron, feet and the fluctuations have a variance of the order of 1 square foot, so the correction term is about . With an intercept, the level is subtracted before the ratio is formed, and only the small-sample bias remains.
Two points apply even to the simplest time-series model:
- Model the level. A missing intercept can turn a stationary series into an apparent random walk. Many statistical packages include the mean by default; hand-written estimators often do not.
- Expect small-sample bias. With few observations, persistence is underestimated, and inference based on large-sample formulas is optimistic.
Both matter well beyond this example. Many conclusions about whether economic series are persistent or mean-reverting rest on these estimates.
References
- M. G. Kendall, "Note on bias in the estimation of autocorrelation", Biometrika, 41(3/4), 403-404, 1954.
- F. H. C. Marriott and J. A. Pope, "Bias in the estimation of autocorrelations", Biometrika, 41(3/4), 390-402, 1954.
- D. A. Dickey and W. A. Fuller, "Distribution of the estimators for autoregressive time series with a unit root", Journal of the American Statistical Association, 74(366), 427-431, 1979.
- P. J. Brockwell and R. A. Davis, Introduction to Time Series and Forecasting, 2nd ed., Springer, 2002.