Persistence And Weak Internal Instruments

Why Persistence Matters for Internal Instruments

Dynamic panel models use past outcomes to explain current outcomes. When outcomes are persistent, their levels can be highly informative about future levels but much less informative about changes. This distinction matters when estimating a dynamic model after differencing: a lagged level may be a valid candidate instrument for a differenced lagged outcome, yet still predict that outcome poorly.

The issue is not simply whether an instrument is statistically associated with a regressor in principle. Instrument strength affects how much information the data provide about the coefficient. Weak relevance can make an instrumental-variable estimate highly variable and can increase its finite-sample bias. These concerns are especially important in short panels, where the available time history is limited.

Consider the dynamic panel model \(y_{it}=\rho y_{i,t-1}+\eta_i+v_{it}\), where \(y_{it}\) is the outcome for unit \(i\) at time \(t\), \(\eta_i\) is a time-invariant unit effect, and \(v_{it}\) is a transitory innovation. Differencing removes \(\eta_i\), leaving \(\Delta y_{it}=\rho\Delta y_{i,t-1}+\Delta v_{it}\), where \(\Delta y_{it}=y_{it}-y_{i,t-1}\). A common internal instrument candidate for \(\Delta y_{i,t-1}\) is the earlier level \(y_{i,t-2}\).

The focus here is the instrument’s relevance: how strongly \(y_{i,t-2}\) predicts \(\Delta y_{i,t-1}\). Whether this instrument is valid depends on assumptions about the innovations and initial conditions. Those assumptions must be considered separately; a strong first stage cannot repair an invalid instrument.

A Simple Persistence Calculation

To isolate the relevance problem, suppose the outcome after removing its unit-specific mean follows a stationary AR(1) process: \(y_{it}=\rho y_{i,t-1}+v_{it}\), with \(0<\rho<1\). Assume the innovations have mean zero, variance \(\sigma_v^2\), are serially uncorrelated, and are independent of the earlier outcome. Under stationarity, the variance of \(y_{it}\) is \(\sigma_y^2=\sigma_v^2/(1-\rho^2)\). The differenced dynamic equation has \(\Delta y_{i,t-1}\) as its regressor, and \(y_{i,t-2}\) as the candidate instrument.

The first-stage slope in a regression of \(\Delta y_{i,t-1}\) on \(y_{i,t-2}\), with an intercept, is the covariance divided by the variance of the instrument. The covariance is:

Thus the first-stage slope equals \(-(1-\rho)\). As \(\rho\) approaches one, this slope approaches zero: a one-unit difference in the lagged level predicts an increasingly small difference in the outcome. The negative sign is expected in this particular first stage: conditional on a higher earlier level, the subsequent change tends to be lower, because the process tends to move back toward its mean.

The slope alone depends on the scale of the instrument, so it is also useful to calculate the correlation. Under the same assumptions, \(\operatorname{Var}(\Delta y_{i,t-1})=2(1-\rho)\sigma_y^2\). Therefore:

The magnitude of this correlation shrinks toward zero as persistence increases. In a simple first stage with one instrument, its squared correlation is the population \(R^2\). The formula therefore makes the loss of predictive power visible on a standardized scale as well. It describes the stationary model for \(\rho<1\); at \(\rho=1\), the stationary variance formula no longer applies, so the limiting comparison should not be treated as a stationary model at exactly one.

From Weak Relevance to Imprecise Estimates

In a simple just-identified instrumental-variable regression, the estimated slope can be written as the sample covariance between the instrument and the outcome divided by the sample covariance between the instrument and the regressor. If the latter covariance is small, small changes in the observed sample covariances can produce large changes in the ratio. This is the basic reason that weak relevance makes the estimate unstable.

The same point appears in the large-sample variance expression. For a centered instrument \(Z\), regressor \(X\), and regression disturbance \(u\), the just-identified estimator has an approximate variance proportional to:

where \(N\) is the number of independent units contributing observations under the simplified setup. If the association \(E[ZX]\) is small, its square in the denominator is smaller still, and the variance can become large. In panel applications the exact variance depends on the dependence structure and estimation procedure, but the weak-denominator intuition remains important.

Weak instruments also create finite-sample bias concerns. In standard linear instrumental-variable settings, estimates with weak instruments often move toward the corresponding non-instrumental-variable estimate; this is a tendency, not a guarantee about every sample or every model. In the differenced dynamic equation, ordinary least squares is itself problematic because \(\Delta y_{i,t-1}\) contains \(v_{i,t-1}\), while \(\Delta v_{it}=v_{it}-v_{i,t-1}\) contains that same innovation with the opposite sign. Under serially uncorrelated innovations, their covariance includes \(-\sigma_v^2\). Thus bias toward ordinary least squares is not necessarily bias toward the true parameter.

Weak relevance also complicates conventional inference. A large standard error may produce a wide confidence interval, but the concern can be more fundamental: when the first stage is very weak, approximations used to justify familiar large-sample tests and intervals may be unreliable. A narrow-looking interval from a procedure that assumes strong identification should not be taken as reassurance without examining the first stage and the method of inference.

What the Calculation Does—and Does Not—Show

The stationary AR(1) calculation gives a clear mechanism: as \(\rho\) rises toward one, adjacent outcome levels become similar, and changes become relatively difficult to predict from a lagged level. The lagged level may still have substantial variation, but relevance depends on how that variation aligns with the differenced regressor, not on the instrument’s variance alone.

The calculation is an illustration, not a universal finite-sample formula for every dynamic panel. Results can depend on initial conditions, time-series length, controls, additional regressors, and how the instrument is scaled or combined with other instruments. In particular, a first-stage slope can look numerically small simply because of measurement units. Standardized association and partial predictive power help distinguish a scaling issue from genuinely weak relevance.

Nor does the calculation establish validity. An instrument that strongly predicts \(\Delta y_{i,t-1}\) may still be correlated with the equation’s disturbance, while a valid instrument may predict the regressor so weakly that it provides little usable information. Dynamic panel analysis therefore has two separate questions: whether the assumptions justify the instrument, and whether the data show that it is informative enough for stable estimation.

Common Mistakes

• Confusing persistence with instrument strength. A highly persistent outcome has strongly related levels over time, but that does not mean a lagged level strongly predicts a change. The relevant first stage is for the differenced regressor.
• Treating validity and relevance as the same condition. Exogeneity does not guarantee predictive power, and predictive power does not guarantee exogeneity. Both issues require attention.
• Interpreting a weak first stage as proof of invalidity. Weak prediction is evidence about relevance, not a direct test of whether the instrument is correlated with the structural disturbance.
• Assuming more observations automatically fix weak identification. More independent units can improve information, but when the first-stage relationship is very small, precision can improve slowly and finite-sample problems can remain consequential.
• Reading an IV point estimate without its first stage. A plausible-looking estimate can be unstable if the instrument barely predicts the regressor. Report and assess the first-stage evidence alongside the structural estimate.