Why Persistence Changes the Comparison
Difference GMM and system GMM use different equations and moment conditions to estimate dynamic relationships in panels. Their comparison becomes especially important when the outcome is persistent: a high current outcome tends to be followed by a high next-period outcome. In that setting, the instruments used for the differenced equation can be weak, while the additional instruments used by system GMM may improve precision only if their stronger assumptions are credible.
The practical question is not simply which estimator reports the smaller standard error. A useful comparison asks three separate questions: are the instruments relevant for the endogenous regressors, are the moment conditions valid under the data-generating process, and are the resulting estimates stable across defensible specifications? Persistence affects the first question, but it does not settle the other two.
The usual difference-GMM moment conditions require that the time-varying errors have no serial correlation that would invalidate the selected lagged levels, and that past outcomes are appropriately uncorrelated with current errors. In particular, for suitable dates \(s\geq 2\), the differenced-equation moments take the form \(E[y_{i,t-s}\Delta\varepsilon_{it}]=0\). The differencing removes \(\alpha_i\), but it also makes the lagged outcome in the differenced equation, \(\Delta y_{i,t-1}\), potentially difficult to predict using old levels when the series is highly persistent.
System GMM adds moments for the level equation. A typical instrument is \(\Delta y_{i,t-1}\), with a restriction such as \(E[\Delta y_{i,t-1}(\alpha_i+\varepsilon_{it})]=0\). This restriction does not follow merely from the absence of serial correlation in \(\varepsilon_{it}\). It also requires a suitable initial-condition or mean-stationarity restriction: changes in the outcome must be uncorrelated with the unit effect. System GMM therefore trades additional identifying information for additional assumptions.
Instrument Strength in a Persistent Process
Instrument validity and instrument strength are distinct. Validity says that an instrument is uncorrelated with the equation error, so its moment condition is correct. Strength concerns whether the instrument predicts the endogenous regressor well enough to identify its coefficient precisely. A valid instrument can be weak; a strong association does not, by itself, establish validity.
A stationary AR(1) illustration makes the role of persistence concrete. Temporarily consider deviations from a unit-specific mean following \(y_{it}=\rho y_{i,t-1}+\varepsilon_{it}\), with \(|\rho|<1\), serially uncorrelated errors of variance \(\sigma^2\), and a stationary process. Then \(\operatorname{Var}(y_{it})=\sigma^2/(1-\rho^2)\). For the differenced equation, \(y_{i,t-2}\) is a standard instrument for \(\Delta y_{i,t-1}\). Its correlation with that regressor is:
The correlation is negative in this centered stationary illustration, but instrument relevance depends on its magnitude, not its sign. As \(\rho\) approaches one, the magnitude approaches zero. Intuitively, adjacent outcome levels become very similar, so their difference contains relatively little predictable information from a lagged level. This is the weak-instrument concern for difference GMM under high persistence.
The corresponding level-equation instrument has a related limitation in this stylized stationary process. The covariance between \(\Delta y_{i,t-1}\) and \(y_{i,t-1}\) is \(\sigma^2/(1+\rho)\), and their correlation is \(+\sqrt{(1-\rho)/2}\). Thus, as persistence approaches one, the standardized relationship between the lagged difference and the level regressor also becomes small in this illustration. System GMM can improve performance in many finite samples by adding level-equation information, but its added moments are not a guarantee of strong instruments for every highly persistent process.
Worked Example: How Persistence Changes a Correlation
Suppose a hypothetical stationary process has \(\rho=0.50\) in one setting and \(\rho=0.95\) in another. Use the correlation formula above to compare the relevance of \(y_{i,t-2}\) for \(\Delta y_{i,t-1}\). The calculation describes a stylized process, not observed empirical data.
For \(\rho=0.50\), the correlation is \(-\sqrt{(1-0.50)/2}=-\sqrt{0.25}=-0.500\). Its absolute magnitude is \(0.500\).
For \(\rho=0.95\), the correlation is \(-\sqrt{(1-0.95)/2}=-\sqrt{0.025}\approx-0.1581\). Its absolute magnitude is about \(0.1581\), substantially smaller.
This calculation shows why difference GMM may lose precision as persistence rises: the lagged-level instrument becomes less correlated with the differenced regressor on a standardized scale. It does not establish that any particular empirical instrument is invalid, nor does it provide a universal threshold for deciding whether an instrument is sufficiently strong. In an applied panel, relevance depends on the sample, time span, instrument set, and other regressors as well as on \(\rho\).
What System GMM Adds—and What It Assumes
System GMM stacks the differenced and level equations and uses moments from both. The difference-equation moments retain their usual requirements. The level-equation moments add restrictions linking changes in the outcome to the unit effect and the level-equation error. A common rationale is that the process is mean-stationary around its unit-specific mean: changes in the outcome are uncorrelated with persistent unit characteristics. The exact admissible lag timing depends on the model and the assumed error structure.
This additional restriction can be consequential. If units with systematically high \(\alpha_i\) also tend to have systematically rising or falling outcomes, then \(\Delta y_{i,t-1}\) may be correlated with \(\alpha_i\). In that case, the level-equation moment fails, even if the difference-equation moments remain valid. The additional moments can then shift the system-GMM estimate away from the parameter of interest rather than simply making it more precise.
Tests of overidentifying restrictions and difference-in-Hansen tests can help assess compatibility of moments with the data, but they cannot prove the assumptions true. A test may have low power, particularly in finite samples, and a non-rejection should not be reported as proof that every added level moment is valid. Conversely, a rejection calls for investigation of the specification and moment set; it does not identify the precise source of failure on its own.
Worked Example: Separating Precision From Evidence on Validity
Suppose a hypothetical study estimates the same persistence parameter with a carefully restricted instrument set. Difference GMM gives \(\hat{\rho}=0.91\) with reported asymptotic \(SE=0.12\); system GMM gives \(\hat{\rho}=0.84\) with reported asymptotic \(SE=0.05\). Assume the reported standard errors are suitable for the stated large-sample Wald calculation. Compare the estimates and their approximate 95% intervals.
For difference GMM, the approximate interval is \(0.91\pm1.96(0.12)\). Since \(1.96(0.12)=0.2352\), the interval is \([0.6748,1.1452]\), or approximately \([0.675,1.145]\).
For system GMM, the approximate interval is \(0.84\pm1.96(0.05)\). Since \(1.96(0.05)=0.0980\), the interval is \([0.742,0.938]\). The system-GMM interval is narrower, so that estimator is more precise under these reported standard errors. Both point estimates indicate substantial persistence, but their numerical difference is not itself evidence that one is correct.
The intervals are approximate confidence intervals under the relevant asymptotic approximation; they are not probabilities that the fixed parameter lies inside the particular intervals. More importantly, the smaller system-GMM standard error does not establish that its additional level moments are valid. A defensible conclusion also requires examining the economic and institutional plausibility of the level restriction, the instrument specification, diagnostics, and sensitivity to reasonable changes in the moment set.
A Structured Comparison
Estimator choice should be guided by the assumptions that are credible for the application, not by a universal ranking. The following sequence keeps the main issues separate.
Look at the estimated dynamics and the behavior of the outcome over time. High persistence raises concern that lagged levels predict differenced outcomes weakly; it does not by itself prove that difference GMM is unusable or that system GMM is preferable.
For difference GMM, justify the error and initial-condition assumptions supporting the selected lagged-level moments. For system GMM, separately justify the added restriction connecting lagged changes to the unit effect and level-equation error.
Check whether the chosen instruments carry meaningful predictive information, and use a parsimonious set appropriate to the panel. A large instrument count can weaken diagnostics and overfit endogenous variables, so adding every available lag is not a remedy for weak relevance.
Report estimates, standard errors, diagnostics, and sensitivity to defensible lag choices. A large change when the system moments are added is a reason to scrutinize the extra restrictions, not a reason to select the estimate with the smaller standard error automatically.
Worked Example: Interpreting a System-GMM Sensitivity Check
Consider a hypothetical application with a highly persistent outcome. A restricted difference-GMM specification estimates \(\rho=0.90\). Adding the level-equation moments gives a system-GMM estimate of \(\rho=0.78\), with a smaller reported standard error. Suppose a difference-in-Hansen test for the added moment subset has \(p=0.03\), while the difference-equation diagnostics do not indicate an obvious failure of their maintained assumptions.
At a prespecified \(\alpha=0.05\), the test result is evidence against the joint validity of the tested added moment subset under the test's assumptions, because \(0.03<0.05\). It does not prove which particular moment is invalid, and it does not prove the difference-GMM specification is correct. The 0.12 change in the point estimate, from \(0.90\) to \(0.78\), also signals sensitivity to the added restrictions.
A careful report would explain why the level moments might or might not be credible in this application, disclose the test and the chosen instrument set, and investigate defensible alternatives. The smaller system-GMM standard error is not enough to override the diagnostic evidence. Nor should the test result be treated mechanically as a complete decision rule: its power and the plausibility of the underlying restrictions still matter.
Common Mistake: Treating High Persistence as an Automatic Verdict
A common mistake is to reason, “The outcome is highly persistent, so difference GMM has weak instruments; therefore system GMM must be the right estimator.” The first part identifies a genuine risk, but the conclusion does not follow. System GMM adds moment conditions whose validity depends on restrictions beyond those needed for the differenced equation. In addition, standardized instrument relevance for the level equation can also be limited in highly persistent processes.
The correction is to treat persistence as a reason to investigate strength and precision, not as a substitute for assessing assumptions. Difference GMM may be less precise but rest on more credible moments in a given application. System GMM may be advantageous when its additional restrictions are plausible and its instruments informative. Either estimator can perform poorly if the instrument set is unsuitable, the moments are invalid, or the available panel is uninformative.
Check Your Understanding
Answer each question by distinguishing instrument relevance from moment validity.
- Why does the magnitude of the correlation between \(y_{i,t-2}\) and \(\Delta y_{i,t-1}\) decline as \(\rho\) approaches one in the stationary AR(1) illustration?
- What additional restriction, beyond the usual differenced-equation moments, is needed to use lagged differences as instruments in the level equation?
- Can a smaller system-GMM standard error establish that its additional moments are valid? Explain.
- In the stylized correlation formula, calculate the absolute correlation for \(\rho=0.80\).
- If a difference-in-Hansen test rejects the added system moments, what does that result indicate, and what does it not establish?