Diagnostic Checks Can Cut Variance Without Changing the Estimand

Balance tests, pre-trend checks, and instrument-validity checks are usually treated like roadblocks before the real analysis begins. This paper argues they should become part of the estimate itself. The authors propose residualizing a baseline estimator against the vector of diagnostic check statistics, which strips away the part of sampling noise that those checks explain. In their framework, that single move can do three things at once: reduce distortions from selective reporting after looking at checks, lower variance without changing the estimand when the original model is correctly specified, and minimize worst-case bias under bounded local misspecification within the class of linear adjustments. They apply the method to the randomized controlled trial in Kaur et al. (2024) and find that, even though all balance checks pass comfortably, residualization makes the baseline point estimate larger and its standard error smaller. In that application, the gain is equivalent to about a 10% increase in sample size. The core message is simple: the information in a diagnostic check does not have to sit on the sidelines; it can help make estimates sharper and more stable.

A balance test can do more than say yes or no. It can also help make the final estimate steadier. That is the surprise here. In one trial from Kaur et al. (2024), all balance checks passed comfortably. Even so, the residualized estimate grew in size and its standard error fell. The gain was about a 10% increase in sample size. For anyone who has ever watched a result wobble after a few small choices, that is a big deal. It turns a check from a gate into a tool.

When a check becomes part of the estimate

The core idea is simple. Start with a baseline estimator, which is the main number used to measure an effect. Then line up that estimate with the vector of diagnostic check statistics. Those checks can cover covariate balance in randomized trials, pre-trends in event studies, and instrument validity in instrumental variables designs. The method removes the part of sampling noise that the checks explain. That leaves a residualized estimator. It keeps the same target when the original model is right. It also cuts variance. And it limits the worst bias under bounded local misspecification, which means small local errors in the model.

How residualization works in plain English

Think of the check statistics as side notes on the main estimate. If those side notes predict part of the wobble, the method subtracts that predicted piece away. That step is called residualizing. In plain terms, it means taking out the part of the estimate that the checks already explain. This matters because the remaining noise is less tied to those checks. The result can then be reported once, instead of being nudged after every look at the diagnostics. The paper frames that as a free lunch. The same move can improve honesty and precision at once.

Why this changes the way applied work can be read

This shifts the role of a check. A covariate balance test no longer acts only as a warning light. A pre-trend test no longer serves only as a stop sign. An instrument validity check no longer stays outside the estimate. The same statistics can help pick apart signal from noise. That is useful even when the checks look clean. The Kaur et al. application makes that clear. All balance checks passed, yet the residualized point estimate still moved up and the uncertainty moved down. The message is not that every check is bad. The message is that a passed check still contains information worth using.

What to watch next in applied economics

The next test is where the checks are not so polite. Event studies with shaky pre-trends would show whether the same residualization still helps when the diagnostics are noisy. RCTs with mild covariate imbalance could test the same idea against standard control adjustments. IV designs could probe whether validity checks add useful signal beyond the baseline estimate. The paper does not claim a cure for every misspecification. It does show a more modest and more practical prize. A diagnostic check can help build the estimate itself. That makes the check part of the answer, not just part of the audit trail.