Checking regression models when observations come in pairs

What the tests ask

The preprint studies undirected dyadic data: observations attached to unordered pairs, so the pair A–B is the same as B–A. Its central question is whether a chosen regression function correctly describes the average outcome given the observed explanatory variables. The method covers straight-line models and smooth nonlinear alternatives.

That is a question about model shape, not causation. Rejecting a straight-line specification means it misses systematic variation in the conditional average. It does not automatically invalidate the coefficients as a description of the best linear approximation.

The motivation is practical. In a selected review of 88 empirical economics papers using dyadic or other relational data, the authors identified linear specifications in 55. That review includes settings beyond the undirected framework developed here; it is not a representative census of all economics research.

How the method works

The researchers fit a regression by least squares and examine its residuals—the differences between observed and fitted outcomes. If the proposed conditional average is correct, residuals should average to zero across regions defined by the explanatory variables.

Two statistics summarize departures. The Kolmogorov–Smirnov test looks for the largest absolute departure across evaluation points. The Cramér–von Mises test averages squared departures over the observed explanatory-variable values. The calculations adjust for the fact that the regression parameters were estimated from the same data.

The difficult part is deciding how large a departure could arise from sampling variation. Pairs sharing a participant may be dependent, so treating every pair as independent can miscalibrate that decision.

The authors develop two bootstrap procedures, which generate artificial fluctuations using random weights. Their corrected first-order procedure removes a double-counting problem: ordinary participant-level weighting counts each pair’s own covariance contribution twice, once through each endpoint. This correction matters particularly when pairs are independent.

A second procedure handles cases where individual participants’ leading contributions vanish but interactions between participants still matter. Here, the limiting fluctuations need not follow a bell-shaped Gaussian distribution. Each pair receives the product of its two participants’ random weights. Additional terms account for parameter estimation and nonlinear curvature. Using the same weights throughout preserves dependence between contributions, including possible cancellation.

Their recommended envelope test uses the larger rejection threshold from the two procedures. Under the stated assumptions, its false-rejection probability approaches the chosen significance level in two Gaussian cases. In the other covered cases, its asymptotic false-rejection probability is no greater than that level, potentially making it conservative.

What the experiments found

The simulations examined independent pairs, weaker and stronger participant-level dependence, and a non-Gaussian case involving interactions, estimation effects and curvature. Each size setting used 20,000 simulated datasets and 999 bootstrap draws.

No individual bootstrap procedure worked reliably across every setting. The uncorrected method rejected too rarely with independent pairs. The augmented second-order method, used alone, rejected too often in the nondegenerate Gaussian designs. Omitting estimation and curvature corrections also produced substantial distortions in the non-Gaussian experiment.

The envelope offered protection across these designs, but finite-sample errors remained, especially with strong participant-level disturbances. It also detected the departures from the proposed regression considered in the power experiments; these results do not establish equal sensitivity to every possible misspecification.

Housing: distance is not the whole description

The application used 414 housing transactions from 2012–2013 at 259 recorded locations in Xindian District, New Taipei City. Each location became a participant, represented by its median transaction price per unit of floor area. The 33,411 pairs were compared using absolute differences in log prices.

Both envelope tests rejected models based only on geographic distance—even one allowing a separate average for every distance band. Neither rejected separate straight lines for pairs with matching versus mixed metro-access classifications at the 5% level: the reported p-values were 0.1315 and 0.2488.

Non-rejection does not prove those lines are correct. Conclusions about their precise shape were sensitive to the metro-distance cutoff. Nor does the analysis establish that metro access causes price differences.

Limits and status

The theory assumes that relabeling participants leaves the joint distribution unchanged and that collections of pairs with no participants in common are independent. It also requires smoothness, moment and identification conditions. The guarantees concern large samples under covered models, not universal finite-sample reliability.

This briefing concerns version 3 of an arXiv preprint. The supplied evidence does not establish peer-reviewed publication; the theoretical claims and empirical findings remain subject to scrutiny.