Tracing Stock-Market Wealth into Local Payrolls, with Clearer Causal Assumptions

From prediction to a causal claim

Economic variables influence one another over time. Panel vector autoregressions, or PVARs, model these relationships across multiple units, such as counties, using the past values of several variables. But a model that predicts payroll growth from earlier wealth changes does not, by itself, establish that wealth caused the growth.

In an arXiv preprint, Raimondo Pala of the University of Rome Tor Vergata adapts an external-instrument approach from structural time-series models to PVARs. An instrument supplies variation intended to isolate the economic shock of interest from other forces affecting the outcome. The paper combines a causal interpretation of the immediate response with procedures for estimating subsequent responses and their uncertainty.

Whose effect is being measured?

The central theoretical contribution is an average effect called the μ-LATE. In the basic formulation, it gives more weight to units whose policy variable responds more strongly to the instrument. It is therefore not necessarily the average effect across all counties or households.

That interpretation depends on substantial assumptions. Both the outcome’s response to the policy variable and the policy variable’s response to the instrument must be linear. Conditional on the model’s controls, the instrument must be independent of potential outcomes and response coefficients, and it must affect the outcome only through the policy variable. It must also move that variable in the same direction across units and have a nonzero average effect on it.

The precise weights depend on the design. For an instrument combining a common market movement with county-specific wealth exposure, including state-by-quarter effects changes which counties drive the estimate. In the application, the paper describes greater weight on populous counties whose wealth exposure differs most from their state’s average.

Interpreting the later response requires additional restrictions: the policy shock must be serially uncorrelated and recoverable from current and past data, while the instrument must be uncorrelated with other shocks at every date. These are conditions for the causal reading, not conclusions guaranteed by fitting the model.

Avoiding misleading precision

Pala addresses two estimation problems. First, removing county-specific averages from a dynamic panel creates a bias that can distort responses at longer horizons. A half-panel jackknife correction combines estimates from the full time span and its two halves to remove the leading bias term.

Second, instruments can be weak: they may provide too little independent movement in the variable being studied. Conventional confidence intervals can then be unreliable. The paper constructs Anderson–Rubin confidence sets and, for multiple instruments, conditional likelihood ratio sets, with asymptotic coverage results under its stated assumptions.

These procedures can honestly report very little information. In the weakest simulated design, more than half of the Anderson–Rubin sets were unbounded. That is not evidence of an infinitely large effect; it signals that the data cannot reliably constrain its size.

The simulations also show why counties cannot simply be treated as independent observations. Common shocks move many counties together. Clustering uncertainty by county alone performed poorly, whereas accounting for both county and quarter dependence generally produced coverage closer to the intended level. The multiple-instrument simulations found that conditional likelihood ratio tests recovered some power lost by Anderson–Rubin tests using all instruments, but did not outperform using the relevant instrument alone.

What the county application finds

The application revisits an earlier study using 2,897 US counties over 92 quarters between 1992 and 2015. Its instrument combines county stock wealth measured eight quarters earlier with the current aggregate stock-market return. The baseline model includes eight lags, county effects, state-by-quarter effects, controls from the earlier study and population weights.

For a wealth gain equal to 1% of annual county labor income, the baseline estimated response of nontradable payroll is 2.83 basis points after four quarters, with a 95% Anderson–Rubin set from 1.55 to 4.70 basis points. The estimated increase builds within the first two years and then levels off over the reported horizon. Estimates from the earlier study’s local projections lie within the PVAR confidence sets at every horizon.

The longer-run path is less settled. Models with more lags produce continuing increases and wider uncertainty sets at later horizons. The supplement also shows that spending spillovers between counties can contaminate the estimated direct effect. Separating direct and spillover effects requires additional structure, including a known exposure mapping and constant effects for the interpretation established there.

The preprint thus offers a conditional framework for causal interpretation and uncertainty—not an assumption-free demonstration that stock-market gains permanently raise local payrolls.