Factor-Augmented Panel Unit-Root Inference

Why Estimate Factors Before Testing for a Unit Root?

A panel can share persistent movements because its units are exposed to common economic, institutional, or environmental shocks. If those movements are omitted, a unit-root test may attribute shared persistence to each unit’s own dynamics. Cross-sectionally augmented unit-root tests address dependence using cross-sectional averages. A factor-augmented approach instead estimates latent common movements and incorporates them into the test. It introduces an important additional issue: the factors are estimated, not observed.

The inferential target must be stated precisely. A test may ask whether the idiosyncratic component of each series has a unit root after accounting for common factors. That is not automatically the same as testing whether the observed series is stationary: the common factor itself may be nonstationary. Nor does rejecting a panel-wide null imply that every unit is stationary.

Definition: A factor representation for a panel series is $$ y_{it}=d_{it}+\lambda_i'F_t+u_{it}, $$ where $y_{it}$ is the observed outcome for unit $i$ at time $t$, $d_{it}$ is a specified deterministic component, $F_t$ is a vector of $r$ unobserved common factors, $\lambda_i$ is the unit’s factor-loading vector, and $u_{it}$ is its idiosyncratic component. A unit-root null for the idiosyncratic component is $H_0:\rho_i=1$ in $u_{it}=\rho_i u_{i,t-1}+v_{it}$, against a stationary alternative such as $H_a:|\rho_i|<1$.

This decomposition clarifies the interpretation. If $F_t$ contains a stochastic trend, $u_{it}$ could be stationary while $y_{it}$ remains nonstationary. A factor-augmented idiosyncratic test therefore asks whether the unit-specific component has a stochastic trend, conditional on the common component being represented adequately. Conclusions about $y_{it}$ itself require considering both components.

Including Estimated Factors in the Test

In practice, the researcher estimates the factor space from the panel, commonly using principal components, and uses the estimated factors $\hat F_t$ as controls. Factor estimates are only identified up to a rotation: replacing $F_t$ by a nonsingular transformation and adjusting the loadings leaves the common component unchanged. The relevant object is the factor space and its span, not a unique labeling of the individual factors.

One factor-augmented ADF specification uses both factor levels and changes as controls:

\Delta y_{it} = a_i+b_i t+\phi_i y_{i,t-1} +\eta_i'\hat F_{t-1} +\kappa_i'\Delta\hat F_t +\sum_{j=1}^{p_i}c_{ij}\Delta y_{i,t-j} +\varepsilon_{it}.

Here $p_i$ is the chosen lag order. The factor terms allow common movements to enter the conditional dynamics; the lagged differences help accommodate serial correlation in the regression disturbance. Under a suitable factor representation and specification, the coefficient $\phi_i$ is used to assess idiosyncratic persistence. In particular, the unit-root null is $\phi_i=0$, with a stationary alternative $\phi_i<0$. This regression is a framework, not a universal recipe: the deterministic terms, factor treatment, and limiting distribution must match the particular test and assumptions.

An alternative implementation estimates the common component and forms residuals $\hat u_{it}=y_{it}-\hat d_{it}-\hat\lambda_i'\hat F_t$, then applies a unit-root regression to those residuals. These two approaches share the aim of separating common and idiosyncratic dynamics, but their finite-sample behavior and critical values need not coincide. A researcher should use a test’s stated construction and calibration rather than treating the alternatives as interchangeable.

Worked Example: Interpreting a Factor-Augmented Regression

Suppose a hypothetical annual panel contains 30 regions. A pre-specified factor procedure estimates two common factors. For one region, the fitted regression has $\hat\phi_i=-0.18$, with the factor level and change controls and one lag of $\Delta y_{it}$. The standard error reported by the regression software is 0.07.

The raw ratio is $$ \frac{\hat\phi_i}{SE(\hat\phi_i)} = \frac{-0.18}{0.07} = -2.571. $$ This is the computed test statistic from the reported coefficient and standard error. Its negative sign is consistent with mean reversion under the specified alternative.

It is not sufficient to compare $-2.571$ to an ordinary standard normal or single-series ADF cutoff. The test uses estimated factors, and the unit-root statistic has a nonstandard distribution whose calibration depends on the factor and deterministic-term treatment. Suppose a suitable factor-augmented bootstrap gives a 5% left-tail critical value of $-2.90$ and a bootstrap $p$-value of 0.08. Since $-2.571>-2.90$ and $0.08>0.05$, this example does not reject the idiosyncratic unit-root null at the 5% level. The coefficient estimate alone does not establish stationarity.

For the calculation to have the intended interpretation, the selected factor space must capture the relevant common dependence, the lag specification must adequately address serial correlation, and the bootstrap calibration must represent the null and the factor-estimation procedure. These are substantive requirements, not consequences of a negative coefficient.

Factor-Number Selection and Calibration

The number $r$ of factors is usually unknown. Too few factors leave common dependence in the residuals and can distort test size. Too many can absorb unit-specific variation, reduce power, and make factor estimates unstable. Information criteria, eigenvalue-based rules, and scree plots are common selection tools, but they are not interchangeable and none guarantees that the selected number is correct in a finite sample.

Selection should be based on the factor structure and a declared criterion, not chosen simply because a particular count produces a preferred unit-root conclusion. It is also useful to examine a small, defensible range of factor counts. If the substantive conclusion changes sharply with $r$, that sensitivity is evidence that the inference depends on factor specification.

Worked Example: Comparing Candidate Factor Counts

Suppose a hypothetical panel has 40 units and 80 periods. A specified information criterion is computed for three candidate factor counts; smaller values are preferred. The results are:

Candidate factors $r$	Criterion value
1	12.4
2	8.1
3	9.0

Among these candidates, $r=2$ minimizes the criterion because $8.1<9.0<12.4$. This is a selection according to the stated criterion, not proof that the true number of factors is two.

Now suppose the panel unit-root procedure, calibrated separately at each count, gives $p$-values of 0.03, 0.08, and 0.12 for $r=1,2,3$, respectively. The criterion-selected specification does not reject at the 5% level, while the one-factor specification does. Reporting only the $r=1$ result would conceal a material modeling sensitivity. A careful report identifies $r=2$ as the primary criterion-based choice and presents the other results as sensitivity checks, without treating them as equally justified primary specifications.

The test distribution also matters. Even if factors were known, a unit-root statistic generally does not follow the usual $t$ distribution under the null. Estimating factors adds generated-regressor uncertainty and can change finite-sample behavior further. Thus, standard errors conditional on $\hat F_t$, followed by ordinary $t$ critical values, do not in general provide valid unconditional inference.

A design-appropriate bootstrap can help calibrate finite-sample inference. For example, a null-imposed bootstrap may generate panels under the unit-root null while preserving the relevant dependence structure, then re-estimate the factors and recompute the test statistic in each replication. Re-estimating factors is important: holding $\hat F_t$ fixed represents them as known and omits a source of variation. If uncertainty about factor-number selection is part of the inferential target, the selection step should also be repeated in each replication. Whether a particular bootstrap is valid depends on its assumptions and construction; repeating steps alone does not guarantee validity.

Worked Example: Accounting for Factor-Estimation Uncertainty

Consider a hypothetical test of the panel null using 500 bootstrap replications. In each replication, the null is imposed, two factors are re-estimated, and the panel statistic is recalculated. The observed statistic is $-2.40$, and 34 of the 500 bootstrap statistics are at least as far into the left tail.

Using the common finite-sample adjustment, $$ \hat p = \frac{34+1}{500+1} = \frac{35}{501} \approx 0.0699. $$ Thus the estimated bootstrap $p$-value is 0.0699. At $\alpha=0.05$, this procedure does not reject the null. The conclusion is not that the null has a 6.99% probability of being true; the $p$-value measures how often a statistic at least this extreme occurs under the bootstrap null procedure.

Suppose an otherwise similar calculation that holds the original factor estimates fixed gives a $p$-value of 0.041. That calculation would reject at 5%, but it conditions on the estimated factors and does not account for their re-estimation. The contrast illustrates why treating estimated factors as observed can matter. It does not prove that the re-estimating bootstrap is correct: the bootstrap must also reproduce the relevant dependence and obey the assumptions supporting its calibration.

A Short Justification for Re-Estimating the Factors

The reason to repeat factor estimation in a bootstrap follows from the statistic’s dependence on the estimated factor space. Write the statistic abstractly as $T(y,\hat F)$, where $y$ denotes the observed panel. If factors were known, uncertainty would be represented by variation in $T(y,F)$ while $F$ stayed fixed. But in the actual procedure, $\hat F$ is itself a function of the sample, say $\hat F=g(y)$. Consequently, $$ T(y,\hat F)=T\bigl(y,g(y)\bigr). $$

Across repeated samples, both the panel $y$ and the factor estimate $g(y)$ vary. A bootstrap that holds $\hat F$ fixed reproduces variation in $T(y,\hat F)$ conditional on that one estimated factor path; it omits variation induced by changing $g(y)$. Re-estimating factors in each bootstrap sample instead approximates the sampling variation of the full procedure $T(y,g(y))$, subject to the bootstrap’s validity conditions. This argument explains the uncertainty being addressed, not a universal guarantee about bootstrap accuracy.

What to Report and Common Pitfalls

Factor-augmented inference is most useful when the researcher’s target and factor treatment are transparent. Report the unit-root null, whether it concerns observed outcomes or idiosyncratic components, the deterministic terms, the factor estimator and selected count, the lag choices, and the method used to obtain critical values or $p$-values. Include sensitivity to plausible factor counts when that choice is consequential. For panel-wide tests, state whether the alternative allows some units to be stationary or requires a stronger common alternative; aggregation does not erase heterogeneity in the unit-level processes.

Pitfall: treating estimated factors as known. Using an ordinary $t$ reference distribution because a regression includes $\hat F_t$ ignores the nonstandard unit-root null and factor-estimation uncertainty. Use a calibration justified for the specified factor-augmented test.
Pitfall: assuming factor removal proves stationarity. A stationary estimated idiosyncratic component does not establish that the observed series is stationary if common factors can contain stochastic trends.
Pitfall: choosing the factor count by the desired conclusion. A count selected because it yields rejection is outcome-driven. State a factor-selection rule and show whether reasonable alternatives change the inference.
Pitfall: claiming every unit is stationary after a panel rejection. A panel statistic’s alternative depends on its construction. Interpret the result using that alternative, not as a blanket statement about all units.

Key takeaway: Factor augmentation makes the unit-root question conditional on an estimated common-component structure. Valid inference requires a clearly defined target, a defensible factor count, and a test calibration that reflects both the unit-root null and the fact that factors were estimated.

Check Your Understanding

Answer each question using the distinction between common and idiosyncratic dynamics.

In the factor representation, what component does an idiosyncratic unit-root test target?
Why can a stationary idiosyncratic component coexist with a nonstationary observed series?
A factor-augmented test reports an ordinary regression $t$-statistic. Why is that statistic not automatically compared with a standard normal cutoff?
An information criterion selects two factors, but the test rejects only when one factor is used. What should be reported and why?
What source of uncertainty is missed when a bootstrap holds estimated factors fixed in every replication?