Choosing economic controls twice to trace a shock

What happens to prices or production after an unexpected interest-rate change? Economists often answer with local projections: separate regressions estimating the relationship between a shock today and an outcome at each future horizon. But those estimates depend on what else the researcher accounts for. A short control list can miss relevant information; a large one can overwhelm the available observations.

In the third version of an arXiv preprint, Jooyoung Cha proposes a way to select controls from a large collection of economic variables and their past values. The method addresses estimation and uncertainty, not the separate problem of establishing that an observed shock has a causal interpretation.

Many small contributions, not just a few large ones

High-dimensional regression methods often exploit sparsity: the idea that relatively few predictors capture the important information. Exact sparsity means that the remaining coefficients are zero. Yet macroeconomic information may be spread across many variables, each making a modest contribution.

Cha allows many nonzero coefficients, provided their magnitudes satisfy structured restrictions. One example is polynomial decay: after ranking coefficients from largest to smallest, their sizes fall according to a power law with an exponent greater than one. This remains a form of approximate sparsity, not permission for an arbitrary collection of equally influential predictors.

The distinction matters because the theoretical error rates are expressed through that decay exponent rather than a specified number of important variables. The method relaxes exact sparsity; it does not eliminate assumptions about how concentrated predictive information must be.

Why select controls twice?

The procedure first predicts the observed shock from candidate controls. Separately, it predicts the future outcome from those controls, repeating this second selection for each horizon. It then combines both selected sets and estimates the outcome’s response to the shock using ordinary least squares.

The two selections address different risks. A variable may help explain predictable contamination in the shock measure even if it looks unimportant for predicting the outcome. Another may explain outcome variation without contributing much to the shock. Selecting only for one equation could miss information needed by the other.

Selection uses an orthogonal greedy algorithm. It adds the variable most strongly associated, after scale adjustment, with the current unexplained variation, then refits before choosing another. A high-dimensional version of the Akaike information criterion balances fit against model size along this sequence.

Controls required by the identifying strategy stay in every regression rather than competing for selection. The final uncertainty calculation uses heteroskedasticity- and autocorrelation-consistent standard errors, accommodating changing error variability and dependence over time.

Three simulation tests, three different lessons

Cha evaluates the method in a calibrated dynamic factor model containing 207 observed macroeconomic series. Each design uses 200 retained observations, 12 response horizons and 1,000 simulation repetitions. Because the model’s true response is known, the experiments can measure estimation error and how often nominal 95% confidence intervals contain it.

In the first design, the observed shock contains a predictable component built from lagged variables with decaying coefficients. The proposed method produces coverage closest to 95% across most horizons. Conventional small-control methods miss the true response too often at the first horizon. Debiased LASSO gives shorter intervals but also undercovers at early horizons; the proposed procedure trades wider intervals for more reliable coverage.

The second design leaves the shock clean but adds persistent, predictable variation to the outcome. The proposed method delivers low point-estimation errors and relatively short intervals. However, coverage falls a few percentage points below 95%. Some alternatives obtain better coverage with wider intervals. Better point estimates therefore do not establish uniformly better inference.

The third design makes shock contamination arise from common latent factors. Methods that summarize the panel through estimated factors gain an advantage. The proposed selection method does not dominate: factor-augmented local projections have better first-horizon coverage, while the factor-augmented vector autoregression generally has lower subsequent point errors. Selection and factor methods emerge as complements, suited to different information structures.

Feasible does not mean assumption-free

A monetary-policy application uses 122 FRED-MD series. Its specification has 1,654 potential regressors excluding the intercept, compared with 572 usable observations at the first horizon. The proposed estimates broadly align with established high-dimensional alternatives, illustrating feasibility rather than revealing a known true response.

The theoretical guarantees require stationary series, controlled temporal dependence, sufficiently high finite moments, well-behaved predictor relationships and restrictions on dimensional growth. They apply separately at a fixed horizon, not automatically to a confidence band covering the entire response path. The theoretical search limit also depends on the unknown coefficient-decay exponent; adaptation using one exponent-independent limit is not established.

Finally, selection cannot repair invalid shock identification or solve general noisy-proxy problems requiring instrumental variables. Selected controls adjust for nuisance variation; their appearance is not evidence of a distinct economic transmission channel.