Why Matching Needs Bias Correction
Nearest-neighbor matching compares units with similar observed covariates, but “similar” does not mean identical. If the untreated outcome changes with covariates, a treated unit and its matched control can have different expected untreated outcomes even when their covariate distance is small. With fixed-number matching, the nearest-neighbor distance generally shrinks as the sample grows, but the resulting discrepancy can still matter for inference, especially when covariates are multidimensional.
Covariate-based bias correction uses an outcome regression to adjust the matched outcome to the covariate values of the unit being evaluated. This does not replace matching: matching determines which observations supply outcome information, while the regression corrects the remaining covariate difference. Inference also needs to account for the fact that a unit may be reused as a match many times.
Setup and the Bias-Corrected Estimator
For each unit \(i=1,\ldots,N\), let \(D_i\in\{0,1\}\) indicate treatment, \(X_i\) be observed covariates, and \(Y_i\) be the observed outcome. Write \(Y_i(1)\) and \(Y_i(0)\) for the potential outcomes, so consistency gives \(Y_i=D_iY_i(1)+(1-D_i)Y_i(0)\). The target here is the average treatment effect, \(\tau=E[Y(1)-Y(0)]\), for the population represented by the sample.
The identifying assumptions are consistency, no interference between units, conditional independence \((Y(1),Y(0))\perp D\mid X\), and overlap \(0<P(D=1\mid X)<1\) on the target covariate support. We also assume an independent sample and use a fixed number \(M\) of nearest neighbors from the opposite treatment group, with replacement. Let \(J_M(i)\) denote the \(M\) matched units for \(i\). Define \(m_w(x)=E[Y(w)\mid X=x]\) and let \(\hat m_w(x)\) be an estimate of this conditional mean.
For a treated query, the bracket compares its outcome with matched control outcomes adjusted to its covariates. For a control query, the bracket has the reverse orientation, and multiplication by \(2D_i-1=-1\) reverses the sign. Thus both treatment groups contribute contrasts oriented as treated minus untreated. The factor \(1/N\) averages these contributions over all sampled units; the estimator is not an average over treated units alone.
Worked Example: Adjusting Matched Outcomes
Suppose four hypothetical units have the following data. Use one nearest neighbor in the opposite group for each unit. For illustration, take the adjustment functions as fixed and known: \(m_0(x)=2+1.5x\) and \(m_1(x)=5+2x\).
| Unit | \(D_i\) | \(X_i\) | \(Y_i\) | Matched unit |
|---|---|---|---|---|
| T1 | 1 | 2 | 10 | C1 |
| T2 | 1 | 5 | 14 | C2 |
| C1 | 0 | 1 | 4 | T1 |
| C2 | 0 | 4 | 8 | T2 |
For T1, \(m_0(2)=5\) and \(m_0(1)=3.5\). Its adjusted control outcome is \(4+5-3.5=5.5\), giving a treated-minus-control contrast of \(10-5.5=4.5\). For T2, \(m_0(5)=9.5\) and \(m_0(4)=8\), so the adjusted control outcome is \(8+9.5-8=9.5\), and the contrast is \(14-9.5=4.5\).
For C1, the matched treated outcome transported from \(X=2\) to \(X=1\) is \(10+m_1(1)-m_1(2)=10+7-9=8\); its contrast is \(8-4=4\). For C2, the transported outcome is \(14+m_1(4)-m_1(5)=14+13-15=12\), giving \(12-8=4\). Therefore $$ \hat\tau_{\mathrm{bc}}=\frac{4.5+4.5+4+4}{4}=4.25. $$
Without adjustment, each treated-control pair has a raw outcome difference of 6, so the corresponding estimator is 6. The difference between 6 and 4.25 reflects the specified regression adjustments; it is not evidence that the adjustment functions are correct in an actual application.
Why the Adjustment Reduces Matching Bias
Consider a treated unit \(i\) and a matched control \(j\). The untreated conditional-mean discrepancy in the unadjusted comparison is \(m_0(X_i)-m_0(X_j)\). If the true conditional mean were used for adjustment, then $$ E\left[Y_j+m_0(X_i)-m_0(X_j)\mid X_i,X_j,D_j=0\right]=m_0(X_i), $$ under the stated conditional-independence assumption. The adjusted control outcome therefore has the untreated conditional mean at the treated unit’s covariates. The same reasoning applies when a treated match is transported to the covariates of a control query, using \(m_1\).
In practice, \(\hat m_w\) replaces \(m_w\), so correction is not exact. Smoothness explains the potential gain. Suppose \(m_0\) has a locally Lipschitz gradient along the line segment from \(X_j\) to \(X_i\). A first-order expansion then has a quadratic remainder:
for a finite local constant \(C\). The adjustment aims to remove the conditional-mean difference associated with the covariate displacement, leaving a smaller remainder when neighbors are close and the regression estimate is adequate. Differentiability alone would justify only a remainder \(o(\|X_i-X_j\|)\), not necessarily a quadratic bound; the Lipschitz-gradient condition is what supports the bound above.
Worked Example: Checking a Taylor Remainder
Let the hypothetical untreated conditional mean be \(m_0(x)=3+2x+0.5x^2\). Its derivative is \(m_0'(x)=2+x\), which is Lipschitz with constant 1 on the real line. For a match from \(X_j=1\) to \(X_i=1.4\), the linear term is \(m_0'(1)(1.4-1)=3(0.4)=1.2\). The exact mean difference is $$ m_0(1.4)-m_0(1)=7.18-5.5=1.68. $$
The residual after the linear term is \(1.68-1.2=0.48\). Directly, the quadratic component gives \(0.5(1.4-1)^2=0.5(0.16)=0.08\); the apparent discrepancy arises because the exact difference was miscomputed if \(m_0(1.4)\) is evaluated carefully: \(3+2.8+0.98=6.78\), not 7.18. Thus the exact difference is \(6.78-5.5=1.28\), and the remainder is \(1.28-1.2=0.08\), as required.
This check illustrates the exact Taylor identity in this quadratic example. In a real analysis the conditional mean is unknown, and the quality of \(\hat m_0\), match distances, and dimension all matter. A valid asymptotic argument must impose conditions strong enough that the average remaining bias is negligible at the scale used for inference; the correction does not guarantee this automatically.
Match Reuse and the Abadie–Imbens Variance Calculation
Matching with replacement means one observation can supply outcomes for multiple query units. Define \(K_M(i)\) as the number of times unit \(i\) is selected among all \(M\)-neighbor sets \(J_M(k)\), counting each selection. Under fixed covariates and treatment assignments, and treating the regression adjustment as fixed, the direct outcome contribution of unit \(i\) has coefficient \((1+K_M(i)/M)/N\), up to its sign. There is one own-outcome contribution and \(K_M(i)\) matched-donor contributions, each with weight \(1/(MN)\). These contributions have the same sign for a given treatment arm.
Write \(\sigma^2(w,x)=\operatorname{Var}(Y(w)\mid X=x)\). If outcome errors are conditionally independent across units, the variance from these outcome-error contributions is estimated by the Abadie–Imbens fixed-number matching variance estimator
Here \(\hat\sigma^2(D_i,X_i)\) estimates the conditional outcome variance in the treatment arm actually observed for unit \(i\). The square applies to the combined coefficient: repeated use increases that unit’s contribution before it is squared. This is a conditional variance calculation, holding \(X\), treatment assignments, matches, and the adjustment fixed.
The formula does not, by itself, account for uncertainty from estimating \(\hat m_w\). It applies when the adjustment is fixed, known, or estimated under conditions that make its contribution to the estimator \(o_p(N^{-1/2})\). For example, a sufficient analysis may establish that the aggregate change caused by estimating the adjustment is negligible at that rate. Otherwise the variance calculation must include the first-stage estimation contribution, such as through a joint influence-function derivation. Treating an estimated regression as fixed without justification can understate uncertainty.
Worked Example: Computing the Reuse-Adjusted Variance
Consider a hypothetical sample of \(N=6\) units, with \(M=1\). Suppose the match counts are \(K_M=(2,1,0,1,1,1)\), and the estimated conditional outcome variance for each unit’s observed treatment arm and covariates is 1 outcome-unit squared. The counts sum to 6, as required because six query units each make one match. Then $$ \widehat V_{\mathrm{AI}} = \frac{1}{6^2} \left[ (1+2)^2+(1+1)^2+(1+0)^2 +(1+1)^2+(1+1)^2+(1+1)^2 \right] = \frac{9+4+1+4+4+4}{36} = \frac{26}{36} = 0.7222. $$
The conditional standard error is \(\sqrt{0.7222}=0.8498\) outcome units. A unit used twice as a match has coefficient magnitude 3 before division by \(N\), so its squared contribution is 9 times its conditional variance; a unit never reused has squared contribution 1. This numerical calculation is illustrative, not a recommendation to rely on a six-unit normal approximation.
Using the Estimate for Inference
The standard error is \(SE(\hat\tau_{\mathrm{bc}})=\sqrt{\widehat V_{\mathrm{AI}}}\). Under the relevant asymptotic conditions—including overlap, regularity of the covariate distribution and outcome regressions, fixed \(M\), and negligible remaining bias and adjustment-estimation effects—the centered estimator has an approximately normal sampling distribution at the appropriate scale. A two-sided approximate 95% interval is $$ \hat\tau_{\mathrm{bc}}\ \pm\ 1.96\,SE(\hat\tau_{\mathrm{bc}}). $$
For example, if an estimated treatment effect is 4.25 outcome units and its justified standard error is 0.85 units, the approximate interval is \(4.25\pm1.96(0.85)\), or \([2.584,5.916]\) outcome units. The interval describes a procedure with approximately 95% repeated-sampling coverage under its assumptions; it is not a 95% probability statement about the fixed effect after observing this interval.
The displayed variance estimator targets conditional outcome-error variability given the observed design. If the inferential target also treats the covariates and treatment assignments as random and requires full superpopulation uncertainty, the corresponding sampling variation must be addressed under that target; it should not be assumed to be included merely because the formula is called a variance estimator. State the inferential target and the treatment of the adjustment estimation when reporting uncertainty.
Common Mistakes and Practical Checks
- Assuming close matches eliminate bias. A small covariate distance can still produce a non-negligible conditional-mean difference. Use a justified adjustment and assess whether the residual bias is negligible for the intended inference.
- Using differentiability to claim a quadratic remainder. Differentiability gives a first-order remainder that is little-oh of distance. A quadratic bound requires stronger smoothness, such as a locally Lipschitz gradient.
- Ignoring reuse. Counting each unit only once misses the fact that a repeatedly selected match has a larger coefficient. Use \(K_M(i)\) in the variance calculation.
- Calling the conditional variance formula complete when the regression is estimated. The adjustment can depend on the same outcomes. Establish that its estimation effect is negligible or include that effect in the variance calculation.
- Interpreting a small standard error as proof of identification. A variance estimate does not verify conditional independence, overlap, or adequate covariate adjustment. Those assumptions and diagnostics remain essential.
Check Your Understanding
Try these questions before reviewing the derivations.
- For a treated query, which conditional mean function is used to adjust its matched outcome, and why?
- What smoothness condition supports a quadratic Taylor remainder, and why is differentiability alone insufficient?
- If a unit has \(K_M(i)=3\) and \(M=2\), what multiplier applies to its conditional variance in the fixed-adjustment variance formula?
- Why can a regression fitted from the analysis outcomes contribute uncertainty beyond the direct matched-outcome variance?
- What assumptions are needed to interpret a bias-corrected matching estimate as an average treatment effect?