Why Corrected Estimates Need Corrected Inference
A local polynomial regression discontinuity (RD) estimate uses observations near a treatment cutoff to estimate the outcome jump there. Because the fit is local, its estimate can have smoothing bias: even if the treatment effect is identified at the cutoff, the fitted curves may not recover the limiting values exactly. Bias correction adjusts the estimate, but the adjustment itself is estimated from data and therefore uncertain. A confidence interval must account for both sources of sampling variability.
This tutorial concerns a sharp RD design. Let \(R_i\) be unit \(i\)'s running variable, \(c\) the cutoff, \(D_i=1(R_i\ge c)\) the treatment indicator, and \(Y_i\) the outcome. The target is the local treatment effect at the cutoff, the difference between the right- and left-hand limits of the conditional mean outcome. For a causal interpretation, the conditional mean functions of both potential outcomes must be continuous at \(c\), treatment must follow the sharp cutoff rule, and units must not precisely manipulate their running-variable values in a way that invalidates the local comparison. The running-variable distribution must also provide observations on both sides near the cutoff.
Inference additionally relies on regularity conditions: observations are independent (or the variance estimator accounts for their dependence), conditional outcome variances are finite, the running-variable density is positive near \(c\), and the conditional mean functions are sufficiently smooth on each side. The local polynomial approximation and bandwidths must satisfy conditions under which the remaining bias is negligible relative to the standard error after correction.
How the Robust Standard Error Arises
The central point is that the corrected estimate is a difference of two estimates obtained from the same observations. Those estimates are generally correlated. The variance identity for a difference gives
Dropping either the bias-estimate variance or the covariance term does not, in general, produce the variance of the corrected estimator. In practice, the local polynomial fits and their bias correction are linear in the outcomes conditional on the running-variable values. The corrected estimate can therefore be written as a weighted sum \(\hat\tau_{bc}=\sum_i w_i^{bc}Y_i\). With independent observations and heteroskedasticity allowed, a corresponding robust variance estimate has the form \(\hat V_{bc}=\sum_i (w_i^{bc})^2\hat\varepsilon_i^2\), using residual estimates appropriate to the local fits. This combined-weight calculation incorporates the correction and its covariance with the original estimate.
A short derivation shows why the covariance is essential. Write \(A=\hat\tau_p\) and \(B=\hat B_p\). Then
The minus sign in \(A-B\) creates the negative covariance term. If \(A\) and \(B\) move together across repeated samples, their shared variation can partly cancel in the difference; if they are negatively correlated, the corrected estimate may be more variable. Thus the robust standard error is not automatically larger than the uncorrected one.
Confidence Intervals and Tests
Under the stated smoothness, sampling, support, and bandwidth conditions, the corrected estimator is asymptotically normal when centered at the target effect and scaled by its robust standard error. In particular, the bias remaining after correction must be small relative to that standard error. For a two-sided confidence level \(1-\alpha\), an approximate confidence interval is
Here \(z_{1-\alpha/2}\) is the corresponding standard normal quantile; for a 95% interval it is approximately 1.96. A test of \(H_0:\tau=0\) uses \(z=\hat\tau_{bc}/\widehat{SE}_{bc}\), with a two-sided p-value calculated from the standard normal distribution. These are large-sample approximations, not exact finite-sample guarantees.
Worked Example: Incorporating Covariance in the Standard Error
Suppose a hypothetical local linear RD analysis estimates a cutoff jump of 5.2 outcome points, with an estimated leading bias of 0.4 points. The estimated variances are \(0.10^2=0.0100\) for the original jump and \(0.05^2=0.0025\) for the bias estimate, and their estimated covariance is 0.0030 points squared. Find the corrected estimate and a 95% robust interval.
The bias-corrected estimate is \(5.2-0.4=4.8\) outcome points. The variance of the difference is
Therefore, \(\widehat{SE}_{bc}=\sqrt{0.0065}=0.08062\) points. Using 1.96 for a 95% interval gives a margin of \(1.96(0.08062)=0.1580\) points and endpoints \(4.8-0.1580=4.642\) and \(4.8+0.1580=4.958\). The interval is approximately 4.642 to 4.958 outcome points.
The covariance is mathematically possible because its magnitude, 0.0030, is less than \(\sqrt{0.0100(0.0025)}=0.0050\), the maximum allowed by the Cauchy–Schwarz inequality. The interval quantifies uncertainty under the RD and approximation assumptions; it does not mean there is a 95% probability that the fixed treatment effect lies in this particular interval.
Worked Example: Testing a Zero Cutoff Effect
In a hypothetical study, suppose the corrected effect is 2.4 points and the robust standard error is 0.8 points. Test \(H_0:\tau=0\) against \(H_a:\tau\ne0\) at \(\alpha=0.05\), assuming the large-sample normal approximation is appropriate.
The test statistic is \(z=2.4/0.8=3.00\). The two-sided p-value is \(2P(Z\ge3.00)\approx0.00270\), where \(Z\) has a standard normal distribution. Since 0.00270 is less than 0.05, reject \(H_0\) at the 5% level. The data provide evidence of a nonzero cutoff effect under the model and design assumptions.
For context, the 95% interval is \(2.4\pm1.96(0.8)=2.4\pm1.568\), or 0.832 to 3.968 points. It excludes zero, in agreement with the two-sided test. Neither the p-value nor the interval establishes that the assumptions hold.
Bandwidths, Diagnostics, and Failure Modes
Bias estimation uses a higher-order fit to approximate the leading bias of the primary fit. For example, when \(p=1\), the main fit is local linear and a common bias-estimation fit uses \(q=2\). The bandwidth \(h\) determines which observations contribute to the main estimate; \(b\) determines the observations used for bias estimation. Both fits must have adequate local support on each side. The ratio of \(b\) to \(h\) and the bandwidths themselves matter for the approximation and variance; the standard error must be computed for the actual corrected estimator, not borrowed from a fit with different weights.
A useful review of an analysis includes the following checks:
- Confirm that there are observations sufficiently close to the cutoff on both sides, and that the local fits are not effectively determined by only a few observations.
- Inspect the fitted behavior near the cutoff and residual variation for extreme observations or poor local approximation. A smooth-looking plot is informative but cannot verify the identifying assumptions.
- Check that the polynomial degree, kernel, and bandwidth choices are reported and that the bias-correction fit has enough support to estimate the leading bias.
- Examine whether the corrected estimate and interval are sensitive to reasonable bandwidth choices. Large changes can indicate weak local information or sensitivity to the approximation.
- Use a variance estimator that matches the sampling structure. If observations are dependent within groups, an independent-observation variance estimate may be inappropriate.
Robust bias correction addresses a particular problem: smoothing bias in the local polynomial estimate and uncertainty introduced by estimating that bias. It does not repair a design with invalid cutoff assignment, precise manipulation, inadequate support, an incorrect outcome measurement, or a poor smoothness approximation. Nor does the word “robust” mean that the result is insensitive to every modeling choice.
Worked Example: Comparing Bandwidth Results Carefully
Suppose a hypothetical analysis reports these robust bias-corrected estimates and standard errors for an outcome measured in days:
| Main bandwidth \(h\) | Corrected estimate | Robust standard error |
|---|---|---|
| 0.8 | 1.5 days | 0.7 days |
| 1.0 | 1.2 days | 0.6 days |
| 1.2 | 0.9 days | 0.7 days |
For \(h=1.0\), the approximate 95% interval is \(1.2\pm1.96(0.6)=1.2\pm1.176\), or 0.024 to 2.376 days. For \(h=0.8\), it is \(1.5\pm1.372\), or 0.128 to 2.872 days. For \(h=1.2\), it is \(0.9\pm1.372\), or −0.472 to 2.272 days.
The point estimates change as the local window changes, while all three intervals are fairly wide and overlap. This is a reason to report the sensitivity and assess local support, not a license to select the bandwidth with the preferred estimate or significance result. Each interval is an approximation conditional on the bandwidth-specific fit and assumptions.
How This Approach Compares With Alternatives
| Approach | What it does | Main caution |
|---|---|---|
| Uncorrected local polynomial interval | Uses \(\hat\tau_p\) and its standard error. | Can have poor coverage when smoothing bias is appreciable relative to sampling uncertainty. |
| Bias-corrected estimate with original standard error | Subtracts \(\hat B_p\) but uses the uncertainty calculation for \(\hat\tau_p\) alone. | Does not generally account for variance of the correction or its covariance with the estimate. |
| Robust bias-corrected inference | Uses \(\hat\tau_{bc}\) and a standard error for the combined corrected estimator. | Still depends on smoothness, local support, and suitable bandwidth and variance calculations. |
| Undersmoothing | Uses a smaller bandwidth so smoothing bias is negligible relative to standard error, without the same bias correction. | May substantially increase variance and depends on choosing a bandwidth that achieves the required bias control. |
The second row is a particularly common mistake. Subtracting an estimated bias changes the estimator; it does not leave the estimator’s sampling variance unchanged. Robust bias-corrected inference is designed to pair the corrected point estimate with uncertainty calculated for that corrected estimate.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- If \(\operatorname{Var}(\hat\tau_p)=0.040\), \(\operatorname{Var}(\hat B_p)=0.009\), and \(\operatorname{Cov}(\hat\tau_p,\hat B_p)=0.006\), what are the corrected variance and standard error?
- A corrected estimate is 3.0 units with robust standard error 1.0 unit. Give an approximate 95% confidence interval.
- Why is it generally incorrect to subtract an estimated bias and then use only the original estimate’s standard error?
- Name two conditions or diagnostics that matter even when robust bias correction is used.
Solutions. (1) The variance is \(0.040+0.009-2(0.006)=0.037\) units squared, so the standard error is \(\sqrt{0.037}=0.1924\) units. (2) The interval is \(3.0\pm1.96(1.0)\), or 1.04 to 4.96 units. (3) The bias estimate is random and correlated with the original estimate, so the variance of their difference includes both the bias-estimate variance and the covariance term. (4) Examples include adequate observations on both sides, sufficient smoothness and a controlled remainder bias, a variance estimate appropriate for dependence, and sensitivity checks across reasonable bandwidth choices.