Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
77,629 characters · 20 sections · 72 citation commands
Local Composite Quantile Regression for Regression Discontinuity
\doublespace
JEL Classification: C18, C21.
Keywords: Regression Discontinuity, Treatment Effect, Local Composite Quantile Regression.
\doublespace
Over the past few decades, regression discontinuity (RD) has become a popular quasi-experimental method to identify the local average treatment effect. In its simplest form, the sharp RD design, a unit $i$ receives treatment if and only if an underlying variable $X_i$ attains a pre-specified cutoff. Under some smoothness assumptions, units in a small neighborhood of the cutoff share similar characteristics so that the difference between the outcomes above and below the cutoff can be interpreted as the local average treatment effect. For recent discussions of the literature and background references, see, e.g., cattaneo2019practical (cattaneo2019practical, cattaneo2020practical).
Central to a large portion of the empirical and theoretical work on RD is the use of local linear regression (LLR, see fan1996local), while the goal of this paper is to introduce another nonparametric smoother to the estimation and inference in sharp as well as fuzzy RD designs. Although LLR is the best linear smoother, it is possible to use an alternative estimator in RD if we consider the larger class of nonlinear smoothers. One such example is the local composite quantile regression (LCQR) method in kai2010local, who show that LCQR can have some efficiency gain against LLR when data are non-normal. The literature related to LCQR is growing. Several recent examples include kai2011new, zhao2014efficient, li2016local, and huanglin2020, although none of these studies focuses on RD. To use LCQR, the researcher chooses a finite number of quantile positions, say $5$, and uses a local polynomial to estimate the quantile (the intercept) at each of the 5 quantile positions. Averaging the 5 quantile estimates gives an estimate for the conditional mean, whose values above and below the pre-specified cutoff are the key components in sharp and fuzzy RD designs.
Our paper makes the following contributions to the large and growing literature on RD. First, we introduce the LCQR method to sharp and fuzzy RD designs. Numerical evidence for the efficiency gain of using LCQR instead of LLR in RD is also provided. Second, similar to the robust $t$-test in calonico2014robust for conducting inference on causal effects, we propose a t-test that adjusts both the bias and the standard error of the LCQR estimator, and the resulting confidence intervals are shown to have good coverage probabilities. Third, we further discuss a new bandwidth selector that is based on an adjusted mean squared error (MSE). As a byproduct of our research, we also develop an R package rdcqr that implements the LCQR method in this paper. The rdcqr package can be downloaded from \url{https://github.com/xhuang20/rdcqr}.
The rest of the paper unfolds as follows. Section 2 sets up the notation for RD designs and introduces the LCQR method. Section 3 presents the bias correction and standard error adjustment for inference on the boundary. Section 4 briefly discusses several extensions, such as allowing for covariates and kink RD designs. Section 5 presents a simulation experiment. An empirical illustration of LCQR for RD is provided in Section 6 using the data from lee2008randomized. Section 7 concludes. The Online Supplement contains all technical details and proofs, as well as additional figures and tables.
We introduce the notation for RD and the application of LCQR in RD in this section.
Consider the triplet $(Y_i,X_i,T_i)$, $i=1,\cdots,n$ in a standard RD setup that nests both the sharp and fuzzy RD designs. $Y_i$ is the observed outcome for individual $i$, $X_i$ is the variable that determines the assignment of treatment, and $T_i$ equals $1$ if individual $i$ receives treatment and $0$ otherwise. For individual $i$ to receive treatment, a threshold value for $X_i$ is set. Without loss of generality, we assume the threshold is $0$, and the treatment assignment rule becomes that individual $i$ is assigned to the treatment group if $X_i \geq 0$. Throughout the paper, the signs $+$ and $-$ will be used to denote data or quantities associated with $X_i \geq 0$ and $X_i < 0$, respectively.
We use a general nonparametric model to describe the relationship between $Y_i$ and $X_i$:
where $m_{Y_+}, m_{Y_-}, \sigma_{\epsilon_{Y_+}}, \sigma_{\epsilon_{Y_-}}, \epsilon_{Y_+}$ and $\epsilon_{Y_-}$ are the conditional mean functions, conditional standard deviation functions, and error terms with unit variance, respectively.
In a sharp RD design, whether individual $i$ receives treatment is determined by whether $X_i \geq 0$, so $T_i=I\{X_i \geq 0\}$. One is interested in measuring the average treatment effect at the threshold $0$:
where
A t-statistic for testing a hypothesized treatment effect $\tau_0$ reads:
where, under the i.i.d. assumption, we have
In a fuzzy RD design, the treatment assignment rule remains the same, but $T_i$ is not necessarily equal to $1$ when $X_i \geq 0$ or $0$ when $X_i < 0$. It is customary to use a nonparametric function to describe the probability of receiving treatment:
The treatment effect measurement becomes
where $m_{Y_+}(0)$, $m_{Y_-}(0)$ are as defined in ((ref)) above, and similarly,
The corresponding t-statistic is given by
where
and $\text{Var}(\hat{\tau}_{\text{fuzzy}})$ depends on the variances and covariances of $\hat{m}_{Y_+}(0)$, $\hat{m}_{Y_-}(0)$, $\hat{m}_{T_+}(0)$ and $\hat{m}_{T_-}(0)$, which we provide in the Online Supplement (see (A.16) in Section S.2) to conserve space.
To conduct the t-test, we thus need nonparametric estimates for all quantities in (ref) at the boundary point $0$.
LLR is the leading method to estimate quantities in (ref). LCQR is introduced in kai2010local as an alternative to the LLR method due to its potential efficiency gains with non-normal errors. The probable advantage under non-normality is the key reason we propose to use LCQR in RD as normality can be easily violated in practice.
The same LCQR method can be applied to (ref). Because of this similarity, we describe the LCQR method based on (ref). For $k=1,\cdots,q$, let $\tau_k = k/(q+1)$ be the equally spaced $q$ quantile positions and $\rho_{\tau_k}(r) = \tau_k r -r I(r < 0)$ be the $q$ check loss functions in quantile regression. The loss function for LCQR at the point $x$ is defined as
where $a_k$ is the k-th quantile estimand, $b$ is a slope restricted to be the same across quantiles, $h_{Y+}$ is the bandwidth, and $K$ is a kernel function. Here we use the linear approximation to the conditional mean function in (ref). Let $F_{Y|X}$ be the cumulative distribution function of $Y$ given $X$ so that $a_k = F_{Y|X = x}^{-1}(\tau_k)$ is the k-th quantile, and $b$ is $m_{Y_+}^{(1)}(x)$, the first derivative of the conditional mean function at $x$. Using a single slope $b$ allows us to combine information across quantiles, since a specification allowing for $b_k$ in (ref) would be equivalent to estimating each quantile separately. Our point of interest in (ref) is $x = 0$.
The parameter $q$ plays an important role here. Once $q$ is chosen, there will be $q$ quantiles corresponding to the $q$ quantile positions $\tau_k$. Since we use equally spaced quantile positions, the median will always be selected as the middle quantile if $q$ is odd. For example, if $q=3$, the quantile positions are $(0.25,0.5,0.75)$. The goal is to spread the quantile positions on the interval $(0,1)$ so that LCQR can combine information in multiple quantiles. The value of $q$ is decided by the researcher. Results in kai2010local and our discussion later on show that a small number such as 5, 7 or 9 may be adequate for many common non-normal errors. It can be viewed as a hyperparameter and can be tuned.
Minimizing (ref) w.r.t. $a_k$ and $b$ yields $(\hat{a}_1,\cdots,\hat{a}_q,\hat{b})$. The LCQR estimator for $m_{Y+}(x)$ is defined as
and $\hat{b}$ is the estimator for $m_{Y_+}^{(1)}(x)$, the first derivative of $m_{Y_+}(x)$. In a similar fashion, $\hat{m}_{Y_-}(x)$, $\hat{m}_{T_+}(x)$, and $\hat{m}_{T_-}(x)$ can be obtained at $x = 0$. The average in (ref) combines information across different quantile estimators. The hope is to possibly improve the relative efficiency of individual nonparametric quantile estimator to the nonparametric mean estimator based on LLR.
kai2010local discuss the asymptotic bias and variance of the LCQR estimator at an interior point $x$. The asymptotic properties at a boundary point $x$ are discussed in kai2010local_supp. Given the results in kai2010local_supp, one can immediately implement the t-test for the sharp RD in (ref) with bias correction if needed. To implement the t-test for the fuzzy RD in (ref), additional covariance expressions for (ref) need to be estimated and we provide these results in Lemma 3 in the Online Supplement. The asymptotic results for the bias and variance of $\hat{m}_{Y_+}(x)$, $\hat{m}_{Y_-}(x)$, $\hat{m}_{T_+}(x)$, and $\hat{m}_{T_-}(x)$ follow those in kai2010local_supp, and are given in Lemma 2 in the Online Supplement.
We make the following assumptions. Let $c$ be a positive constant of supp($K$).
Following (ref), we define, for $j=0$, $1$, $2$, $\cdots$,
(ref) are used for Taylor series expansions in the proof. The bias expressions require only second-order differentiability of the conditional mean function. Higher-order differentiability permits the study of small order terms in the expansion. The support of the kernel is assumed to be $(-\infty,\infty)$ in (ref). If the kernel has a bounded support such as $[-1,1]$, we would have $\mu_{+,j}(c) = \int_{-c}^{1} u^j K(u) du$. Similar changes can be made to the $\nu$ variables. The differentiability in (ref) is also used for Taylor expansions in the proof, and we require the positivity of the density at the boundary as it appears in the denominator of the bias expression. The symmetric error distribution assumption in (ref) helps to remove an extra term in the bias. When the error distribution is asymmetric, this extra term can be removed during bias correction. In (ref), $h_{Y_+} \rightarrow 0$ and $n_+h_{Y_+} \rightarrow \infty$ help establish the consistency of $\hat{m}_{Y_+}$. The assumption $n_+h_{Y_+}^7 \rightarrow 0$ is used to establish the asymptotic normality of the bias-corrected and s.e.-adjusted $t$-statistic. (ref) translates to $h \rightarrow 0$, $nh \rightarrow \infty$, and $nh^7 \rightarrow 0$ as $n \rightarrow \infty$ if we use a single bandwidth $h$ for the data above and below the cutoff, and $n$ is the total number of observations. Similar assumptions can be found in calonico2014robust for LLR.
Next we provide two theorems on the LCQR estimator for sharp and fuzzy treatment effects defined in (ref) and (ref), respectively. Let $\mathbf{X}$ be the $\sigma$-field generated by all $X_i$.
Proof of (ref) is a straightforward application of Theorem 2.1 in kai2010local_supp. As long as a symmetric kernel is used, the value of $a_{Y_+}(c)$ remains the same if kernel moments on the other side of the cutoff are used. Generally, we have $b_{Y_+}(c) \neq b_{Y_-}(c)$ unless error distributions on both sides of the cutoff are assumed to be the same.
When the bandwidths above and below the cutoff are assumed to be the same, (ref) reduces to the following corollary.
The asymptotic results for the fuzzy RD are similarly provided in the theorem below.
From (ref) and using a single bandwidth $h$, we see a lot of similarities between LCQR and LLR: both biases have order $O(h^2)$, and both variances have order $O((nh)^{-1})$. In fact, the asymptotic bias expression is identical for the two methods. The variance from LCQR, however, is smaller than that from LLR under many non-normal error distributions, for which we provide numerical results in the next subsection.
In addition, in the special case of $q = 1$, we can show that
where $\tau_{kk'}=\tau_{k}\wedge\tau_{k'}-\tau_{k}\tau_{k'}$, $k$, $k'=1$, $2$, ..., $q$, $c_k=F_{\epsilon_{Y_+}}^{-1}(\tau_k)$, and $F_{\epsilon_{Y_+}}$, $f_{\epsilon_{Y_+}}$ are the c.d.f. and p.d.f. of the error distribution, respectively. This result similarly holds for $b_{Y_-}(c)$. However, (ref) does not hold for $q \geq 2$ in general.
We provide some numerical evidence in support of using LCQR in RD. To facilitate the presentation of the results, we consider (ref) with a single bandwidth and replace the conditional density $f_{X_+}(0)$ with the unconditional density $f_{X}(0)$ so that the product $n_+ f_{X_+}(0)$ becomes $nf_{X}(0)$. Similarly, $n_- f_{X_-}(0)$ is also replaced by $nf_{X}(0)$.
Under standard assumptions, the MSE for the LLR estimator of $\tau_{sharp}$ is given by (see, e.g., imbens2012optimal)
where
Let $a(c) = a_{Y_+}(c)= a_{Y_-}(c)$. By assuming errors on both sides of the cutoff have the same distribution and letting $b_{Y}(c) = b_{Y_+}(c) = b_{Y_-}(c)$, we can further simplify the variance expression in (ref) to have
where the two MSEs differ by a factor $\frac{b_Y(c)}{b(c)}$ for the variance term. Minimizing these two MSEs gives optimal bandwidths. Substituting the bandwidths into the MSEs leads to the optimal values, $\text{MSE}^{\text{opt}}_{\text{LLR}}$ and $\text{MSE}^{\text{opt}}_{\text{LCQR}}$. Similar to kai2010local, we define the asymptotic relative efficiency (ARE) of LCQR with respect to LLR on the boundary as
When $q = 1$, the ratio in ((ref)) simplifies to $\left[\frac{1}{q^2}\sum_{k=1}^{q}\sum_{k'=1}^{q} \frac{\tau_{kk'}}{f_{\epsilon_{Y_+}}(c_k)f_{\epsilon_{Y_+}}(c_k')}\right]^{-4/5}.$ When $q \geq 2$, this simplification does not hold. Yet we can evaluate the ratio in (ref) for many common distributions. Using the triangular kernel as an example, we calculate the ratio for the five error distributions listed in (ref). Using the Epanechnikov kernel gives similar results.
The ARE for boundary points in (ref) largely follows the same pattern for interior points as shown in Table 1 of kai2010local. When errors are normally distributed (Distribution 1), increasing $q$ will quickly bring ARE close to one, reflecting a very small efficiency loss of LCQR with respect to LLR. For non-normal errors (Distributions 2 - 5), there can be large efficiency gains. The column with $q=1$ in (ref) is identical for both interior and boundary points, a result of using a symmetric kernel function. When $q \geq 2$, the numbers differ from those in Table 1 of kai2010local. Overall, (ref) provides the numerical evidence in support of using LCQR in RD when data are non-normal, i.e., the corresponding ARE mostly exceeds 1.
The goal of inference is related to but different from that of estimation. While one still needs a precise $\hat{\tau}_{\text{sharp}}$ or $\hat{\tau}_{\text{fuzzy}}$, much of the effort is spent on making $t$-statistics in (ref) behave like a standard normal random variable. Common approaches include reducing the impact of bias by under-smoothing, directly adjusting the bias, combining bias correction and standard error (s.e.) adjustment, etc. This section studies the bias-corrected and s.e.-adjusted $t$-statistics based on LCQR for conducting inference in sharp and fuzzy RD designs.
From (ref), the bias-corrected estimator is given by
where $\widehat{\text{Bias}}(\hat{\tau}_{\text{sharp}}) = \widehat{\text{Bias}}(\hat{m}_{Y_+}) - \widehat{\text{Bias}}(\hat{m}_{Y_-})$ with
and $\hat{m}_{Y_+}^{(2)}$ and $\hat{m}_{Y_-}^{(2)}$ are the estimators for the second derivatives $m_{Y_+}^{(2)}$ and $m_{Y_-}^{(2)}$, respectively, which can be computed using LCQR with a second-order polynomial.
Simply replacing $\hat{\tau}_{\text{sharp}}$ in (ref) with $\hat{\tau}_{\text{sharp}}^{\text{bc}} $ may still lead to undercoverage of the resulting confidence intervals. calonico2014robust observed that the key to fix this problem is to take into consideration the additional variability introduced by bias correction. Similar idea is applied to inference based on LCQR, i.e., instead of using $\text{Var}(\hat{\tau}_{\text{sharp}})$ in the denominator of (ref), we use $\text{Var}(\hat{\tau}_{\text{sharp}} - \widehat{\text{Bias}}(\hat{\tau}_{\text{sharp}}))$. The additional variability due to bias correction increases the adjusted variance, and the s.e./variance adjustment naturally improves the coverage of the resulting confidence intervals.
By the i.i.d. assumption, we have
where the expressions for $\text{Var}(\hat{m}_{Y_+})$ and $\text{Var}(\hat{m}_{Y_-})$ are available in kai2010local_supp, kai2010local and Lemma 2. We contribute to the literature by deriving the explicit forms of the variances of bias and the covariances in (ref).
Following (ref), the confidence interval for the nominal $95\%$ coverage becomes $\hat{\tau}_{\text{sharp}} - \widehat{\text{Bias}}(\hat{\tau}_{\text{sharp}}) \pm 1.96 \sqrt{\text{Var}(\hat{\tau}_{\text{sharp}} - \widehat{\text{Bias}}(\hat{\tau}_{\text{sharp}}))}$, which incorporates the variability introduced by bias correction and will thus improve the coverage probability.
\setcounter{theorem}{0}
Similar to the sharp case above, we also propose the adjusted $t$-statistic for the fuzzy RD design. Following feir2016weak, we use a null-restricted t-statistic to help eliminate the size distortion due to possibly weak identification in the fuzzy RD. Use (ref) to rewrite the the null $H_0: \tau_{\text{fuzzy}} = \tau_0$ as
Define $\tilde{\tau}_{\text{fuzzy}} = (\hat{m}_{Y_+} - \tau_0 \hat{m}_{T_+}) - (\hat{m}_{Y_-} - \tau_0 \hat{m}_{T_-})$. The bias-corrected $\tilde{\tau}_{\text{fuzzy}}$ is given by
where
We note that $\widehat{\text{Bias}}(\hat{m}_{Y_+})$ and $\widehat{\text{Bias}}(\hat{m}_{Y_-})$ are defined in (ref), while $\widehat{\text{Bias}}(\hat{m}_{T_+})$ and $\widehat{\text{Bias}}(\hat{m}_{T_-})$ are defined similarly by replacing $Y$ with $T$. All the bias terms can be estimated using the result in Lemma 2 in the Online Supplement. To illustrate the components in the adjusted variance, it is helpful to consider first the data above the cutoff (the expressions for the data below the cutoff result by replacing $+$ with $-$):
To operationalize the variance adjustment process, we derive all the required covariance terms in the Online Supplement.
\setcounter{theorem}{3}
In this subsection, we propose a method to revise the MSE-optimal bandwidth by taking into consideration both bias correction and s.e. adjustment. It is well-known in the nonparametric literature that the MSE-optimal bandwidth, when used in inference, often induces undercoverage of conventional confidence intervals. The root cause is that the MSE-optimal bandwidth, $h_{\text{MSE}}=O(n^{-1/5})$, leads to $n h_{\text{MSE}}^5 \rightarrow \text{a constant}$, while we need $nh^5 \rightarrow 0$ in order to ignore bias and use undersmoothing without bias correction. Hence using $h_{\text{MSE}}$ in (ref) could lead to poor coverage of confidence intervals. Using a bias-corrected t-statistic without adjusting the variance, $\frac{\hat{\tau}_{\text{sharp}} - \widehat{\text{Bias}}(\hat{\tau}_{\text{sharp}})-\tau_0}{\sqrt{\text{Var}(\hat{\tau}_{\text{sharp}})}}$, does not solve this problem; see the discussion in Section 2 of calonico2014robust.
It is helpful to revisit the bandwidth selection process in an MSE-optimal setting in order to find a solution. We use a single bandwidth $h$ for illustration. Consider the usual MSE when estimating a conditional mean function $m(\cdot)$ by $\hat{m}(\cdot)$:
It is clear that the MSE-optimal bandwidth aims to balance two terms, $\text{Bias}^2$ and $\text{Variance}$. After correcting the bias for the numerator of the t-statistic, an optimal bandwidth should balance $\text{Bias}^2(\hat{m}-\widehat{\text{Bias}}(\hat{m}))$, not $\text{Bias}^2(\hat{m})$, with the variance term. Given $\text{Bias}(\hat{m}) = O(h^2)$, we need to expand the bias expression up to $O(h^3)$ in order to compute $\text{Bias}(\hat{m}-\widehat{\text{Bias}}(\hat{m}))$. In addition, after incorporating variance adjustment due to bias correction, the adjusted MSE takes the following form
where $\text{adj. Var}(\hat{m})$ can take the form in (ref) in a sharp RD design. The following theorem presents the bandwidth result associated with the adjusted MSE in a sharp RD design, assuming we work with data above the cutoff.
The result for the fuzzy design case can be obtained similarly. (ref) assumes that two separate bandwidths are used for the data above and below the cutoff. Similar result holds for $h_{Y_-}$. On each side of the cutoff, a single bandwidth is used in both estimation and bias correction. (ref) provides the expression of this bandwidth.
\setcounter{theorem}{3}
In this section, we extend the discussion to several related topics that are of either practical or theoretical importance. Within the framework of LLR, these topics are studied in calonico2018jasa,calonico2019covariates,calonico2020optimal, calonico2020optimalbandwidth. Within the framework of LCQR, we develop the fixed-$n$ approximations for estimation and inference in both sharp and fuzzy RD, while we also provide a brief discussion on several other topics.
Results in previous sections are based on asymptotic approximations as $n \rightarrow \infty$, while calonico2018jasa point out that fixed-$n$ approximations and the associated Studentization can help local polynomial regression retain the automatic boundary carpentry property in coverage errors. We therefore study the fixed-$n$ approximations of the asymptotic results introduced in early sections for LCQR. Essentially, we re-derive (ref) by keeping $n$ and $h$ fixed. Details of the fixed-$n$ version of the t-statistics in (ref) are given in Propositions 1 and 2 and the associated proofs in Section S.3 of the Online Supplement.
As an example, consider the fixed-$n$ variance of the estimator $\hat{m}_{Y_+}$ in Lemma 5:
where $e_q$ is a vector of ones, $S_{nY_+}$ and $\Sigma_{nY_+}$ are defined in Section S.1 Equation (A.7). Compared to the first-order asymptotic variance in Lemma 2, it is clear that the fixed-$n$ expression does not use any of the kernel moments such as $\mu_{+,1}$. The $o_p$ term at the end is a result of approximating the non-differentiable objective function of LCQR with a differentiable, quadratic function; see Proof of Lemma 5 in Section S.3 for details.
Our simulation presented later on suggests that fixed-$n$ approximations improve the coverage of LCQR-based confidence intervals when the sample size is small, and work particularly well under heteroskedasticity. In Table (ref), confidence intervals based on fixed-$n$ approximations are found comparable to asymptotic confidence intervals, though fixed-$n$ results seem more conservative and give slightly higher coverage. When the sample size is decreased, fixed-$n$ approximations give better coverage (see Table 6 in Section S.4.4 of the Online Supplement). The rdcqr package offers both asymptotic and fixed-$n$ approximations in the computation of bias, s.e., and adjusted s.e.
Although we discuss the bandwidth that results from minimizing the adjusted MSE of the bias-corrected LCQR estimator, other bandwidth choices that satisfy (ref) are also allowed. One possibility is the coverage error (CE)-optimal bandwidth suggested by calonico2020optimalbandwidth, who propose to derive the optimal bandwidth by minimizing the coverage error of confidence intervals for the treatment effect. calonico2020optimal further show that this bandwidth choice can make bias-corrected confidence intervals attain the minimax bound on coverage errors. It will thus be interesting to derive such a CE-optimal bandwidth for the LCQR-based confidence interval, which, however, would require a considerable amount of work beyond the scope of this paper. We leave it for future research. Our simulation evidence presented in Section 5, however, shows that the bandwidth in (ref) provides good coverage probabilities along with an accurate estimation of the treatment effect.
Many RD applications come with covariates in addition to the treatment assignment variable $X$. Let $Z_{\text{cv}}$ denote the additional covariates. Depending on the nature of the data, adding covariates could help reduce bias and make treatment estimates more precise. One can simply add these covariates to LLR to adjust the treatment estimate; see imbens2008regression (imbens2008regression) for an example. More recently, calonico2019covariates gives an MSE expansion for the covariate-adjusted RD estimator, and provides the corresponding bandwidth selection, bias correction and s.e. adjustment results.
Using a single bandwidth $h$, similar to Equations (1) and (2) in calonico2019covariates, we can write the treatment equation without and with covariates as
where all estimates are LCQR estimates. (ref) are derived based on (ref) while (ref) follows from the following LCQR objective function,
Since (ref) leads to a nonlinear problem with no closed-form solution, we cannot use partitioned regression to directly express $\tilde{\tau}$ as a linear function of other parameters in the equation and study its asymptotics. Instead, one needs to retool the asymptotic methods in kai2010local_supp, kai2010local to incorporate the covariates based on (ref). As a result, all asymptotic results in previous sections need to be revised to reflect the presence of covariates. Developing the rigorous asymptotic results for (ref), similar to calonico2019covariates, will be a useful addition to make LCQR more appealing to applied researchers, and require future work.
For interested readers who want to try LCQR with covariates in RD, we offer an ad hoc approach. The companion rdcqr package offers a function to estimate $\tilde{\tau}$ in (ref), though no bias-correction or s.e. is currently provided. We suggest to proceed by using the bias and adjusted s.e. of $\hat{\tau}$ in (ref) for the $t$-statistic for $\tilde{\tau}$. This ad hoc approach is not completely unwarranted. imbens2008regression wrote: “If the conditional distribution of $Z$ given $X$ is continuous at $x=c$, then including $Z$ in the regression will have little effect on the expected value of the estimator for $\tau$ ...” and $Z$ refers to covariates. Hence we conjecture that, although $\tilde{\tau}$ and $\hat{\tau}$ are different, they could be numerically close to each other, and their variance differences could also be relatively small in practice. As a result, using the bias and adjusted s.e. of $\hat{\tau}$ for $\tilde{\tau}$ in the t-test probably will not hurt much. Indeed, our simulation in Section S.4.6 of the Online Supplement shows that this ad hoc approach works well for the data generating process (DGP) used in calonico2019covariates. When $\tilde{\tau}$ and $\hat{\tau}$ differ to a large extent, one cannot use the ad hoc approach.
While the interest of sharp and fuzzy RD designs lies in the levels (i.e. conditional mean functions) at the cutoff, kink RD designs focus on the derivatives of regression functions; see e.g. card2015inference. kai2010local_supp, kai2010local show that LCQR can also have efficiency gains when used for estimating derivatives. Therefore, it is natural to extend the LCQR method to kink RD designs. In Table 7 of the Online Supplement, we report the simulation outcome that shows LCQR could outperform the local polynomial regression for estimating derivatives in a sharp kink design when data are non-normal. We, however, do not further explore kink RD designs in this paper, since they would involve higher-order terms in Taylor series expansions, and are more challenging in practical implementations.
In this section, we conduct a Monte Carlo study to investigate the finite sample properties of the LCQR method. We focus on the sharp RD design and separate the Monte Carlo study into two parts, one for estimation (Section (ref)), and the other for inference (Section (ref)).
The sharp designs calibrated to lee2008randomized (lee2008randomized, Lee) and ludwig2007does (ludwig2007does, LM) are used in the DGP:
where the conditional means are given by
and
Hence the treatment effects are $0.04$ and $-3.45$ in ((ref)) and ((ref)), respectively.
We use the same five error distributions in kai2010local to simulate $\epsilon_i$. These five error distributions are listed in Table (ref) above, which lead to five DGPs: for DGP 1, $\epsilon_i \sim N(0,1)$; ...; for DGP 5, $\epsilon_i \sim \text{mixture normal }0.95N(0, 1) + 0.05N(0,10^2)$.
In addition, the homoskedastic and heteroskedastic specifications in kai2010local are used for simulating the standard deviation:
We set $n = 500$ with 5000 replications. The data are i.i.d. draws in all replications. The triangular kernel is used in all estimations.
Since we also study the coverage probability of confidence intervals, we compare the LCQR results with the robust confidence interval of calonico2014robust. For ease of exposition, we first summarize in Table (ref) the types of the estimators reported later on.
We set $q = 7$ for all LCQR estimators. $\hat{\tau}_{\text{1bw}}^{\text{robust,bc}}$ is obtained by using the R package rdrobust. We use the option bwselect = mserd in estimation and the option bwselect = cerrd in inference so that $\hat{\tau}_{\text{1bw}}^{\text{robust,bc}}$ is MSE-optimal and CE-optimal in (ref) and (ref), respectively. The local linear estimator, $\hat{\tau}_{\text{1bw}}^{\text{llr}}$, is obtained by applying the main bandwidth used in $\hat{\tau}_{\text{1bw}}^{\text{robust,bc}}$.
To calculate the bandwidth in (ref) and (ref) for LCQR, we use the rule-of-thumb bandwidth selector described in Section 4.2 of fan1996local to compute quantities such as $C_2$ and $C_3$ in (ref). The bandwidth in (ref) or (ref) is then used to perform the LCQR estimation, bias correction, and s.e. adjustment. Unlike the LLR estimator, the LCQR estimator has no closed-form expression, and is obtained from the iterative MM algorithm.
In this subsection, we compare LCQR with LLR for the treatment effect estimation.
Without bias correction, (ref) suggests that the two LCQR estimators, $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ and $\hat{\tau}_{\text{2bw}}^{\text{cqr}}$, are less accurate than LLR, though the numerical difference is small. This result probably is not a surprise since we use the bandwidth based on the adjusted MSE, which is not MSE-optimal for $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ and $\hat{\tau}_{\text{2bw}}^{\text{cqr}}$. Since the employed bandwidth balances the bias of the bias-corrected estimator and adjusted variance in (ref), it will be interesting to investigate its bias-correction performance. (ref) presents the bias of the bias-corrected LCQR and LLR estimators for estimating the treatment effect in (ref) with homoskedasticity. It shows that LCQR produces accurate estimates after bias-correction. The improvement can be large, depending on the error distribution; similar results for other models can be found in Figure 3 in the Online Supplement. It suggests that bias correction will help center the confidence intervals.
Table (ref) further presents the standard errors of the studied estimators to facilitate comparison. The s.e. of $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ is consistently smaller than that of $\hat{\tau}_{\text{1bw}}^{\text{llr}}$, indicating the efficiency gain of LCQR against LLR. Although DGP 1 has normal errors, $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ and $\hat{\tau}_{\text{1bw}}^{\text{llr}}$ do not use the same bandwidth, so the s.e. of $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ also appears smaller than that of $\hat{\tau}_{\text{1bw}}^{\text{llr}}$. As the DGP moves away from normal errors, the LCQR estimator achieves various levels of efficiency gains compared to the LLR estimator. For example, the s.e. of $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ in DGP 5 of Table (ref) Panel B is $0.137$, compared to $0.255$ for $\hat{\tau}_{\text{1bw}}^{\text{llr}}$, so the LCQR/LLR standard error ratio is close to 50% in this example.
While reading Table (ref), it is important to bear in mind that the s.e.s of estimators with a superscript “bc" are not suitable to assess the efficiency of the estimators. They are computed by incorporating the additional variability due to bias correction. However, they shed light on the length of confidence intervals. Table (ref) shows that the adjusted s.e. of $\hat{\tau}_{\text{1bw}}^{\text{cqr,bc}}$ could be smaller or larger than that of $\hat{\tau}_{\text{1bw}}^{\text{robust,bc}}$, so the simulation results for comparing LCQR with LLR appear to be mixed after bias correction.
In addition, using two bandwidths above and below the cutoff also gives mixed results in terms of bias correction for the LCQR estimator, though it is clear that the two-bandwidth approach leads to a small decrease in s.e. for LCQR.
This subsection studies the coverage probability of the confidence intervals based on the LCQR estimator, using the adjusted s.e. in (ref) and the adj. MSE-based bandwidth selector. The results are summarized in Table (ref), where the nominal coverage probability is set to $95\%$.
Without bias correction or s.e. adjustment, confidence intervals based on the LCQR estimators, $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ and $\hat{\tau}_{\text{2bw}}^{\text{cqr}}$, are found to have poor coverage probabilities in Table (ref), which highlights an important difference between estimation and inference: despite the excellent finite sample properties of LCQR in Table (ref), one has to perform both bias correction and s.e. adjustment in order to produce the desired coverage of confidence intervals. Furthermore, the coverage probability of LCQR-based confidence intervals is lower than that of LLR-based confidence intervals. For example, it is $92.2\%$ under DGP 1 in Panel A of Table (ref), compared to $93.4\%$ of the LLR-based confidence interval. Although $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ tends to have a smaller bias and a smaller s.e. than $\hat{\tau}_{\text{1bw}}^{\text{llr}}$, it appears that the bias in $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ is still not small enough to center the estimator close to the true value. A smaller s.e. will further contribute to the decrease in coverage probabilities. This explains the poor coverage for confidence intervals based on $\hat{\tau}_{\text{1bw}}^{\text{cqr}}$ and $\hat{\tau}_{\text{2bw}}^{\text{cqr}}$.
With bias correction and s.e. adjustment, the proposed bandwidth selectors in (ref) and (ref) lead to good coverage across the five DGPs in Table (ref). It is important to recall that the presented LCQR estimators use the bandwidths in (ref) and (ref), which are not designed to optimize the coverage probability of confidence intervals, while the confidence intervals for $\hat{\tau}_{\text{1bw}}^{\text{robust,bc}}$ in Table (ref) are optimized to minimize coverage errors. In this respect, it is reasonable to conclude that LCQR with bandwidths in (ref) and (ref) offers very competitive results.
In the last row of each panel in (ref), we also report the coverage probability based on fixed-$n$ approximations. The benefit of using fixed-$n$ approximations becomes clearer in Table 6 of the Online Supplement, where the sample size decreases to 300 and fixed-$n$ approximations often give the best coverage.
In this section, we use the data from lee2008randomized to illustrate the practical usage of LCQR. To facilitate comparison, the findings based on LLR are also presented.
We revisit a classic example from lee2008randomized with $6,558$ observations depicted in Figure (ref)(a); see also imbens2012optimal. The horizontal running variable is the democratic vote share (margin of victory) in a previous election, while the vertical outcome variable is the democratic vote share in the election afterwards. Consistent with lee2008randomized and several follow-up studies, Figure (ref)(a) indicates that there is a positive impact of incumbency on re-election, i.e., a visible jump occurs at the threshold zero.
Figure (ref)(b) and (c) present the estimated impact of incumbency on re-election by LCQR and LLR methods, respectively. We consider a sequence of bandwidth values ranging from 0.05 to 1 with the step size 0.025: 0.05, 0.075, 0.1, ..., 1. This bandwidth sequence thus nests many common choices such as the MSE-optimal bandwidth of imbens2012optimal, which is about 0.3 for the studied data.
The comparison of Figure (ref)(b) and (c) shows that LCQR and LLR yield similar point estimates over a wide range of bandwidths for the studied lee2008randomized application. As the bandwidth increases, Figure (ref)(a) indicates that there are data points staying further away from the fitted regression line. These data points affect LCQR and LLR estimators in a different manner, leading to slightly disparate point estimates. Nevertheless, all the point estimates depicted in Figure (ref)(b) and (c) are significantly positive, since the (vertical) zero value is excluded from the shaded regions generated by $\pm$ 2 standard errors.
Most importantly, Figure (ref)(d) highlights that the standard error of the LCQR estimator is substantially smaller than that of LLR. The standard error ratio is mostly around 70% $\sim$ 80%, which can also be viewed by comparing the shaded regions in Figure (ref)(b) and (c). Moreover, this standard error ratio does not change much as the bandwidth varies. Thus, Figure (ref)(d) indicates that a confidence interval for the impact of incumbency on re-election based on the LCQR estimator could be considerably tighter than that by LLR.
Consider, for example, 0.3 as the adopted bandwidth. The LCQR estimate $\pm$ 1.96 standard error leads to the 95% confidence interval (0.068, 0.090), while the conventional 95% confidence interval by LLR is (0.065, 0.096). If bias correction is further accounted for at the adopted bandwidth 0.3, then the bias-corrected 95% confidence interval based on LCQR is (0.048, 0.090). This interval is comparable to the bias-corrected 95% confidence interval using the calonico2014robust approach, which is (0.046, 0.089). These empirical findings therefore lend credibility to our proposed LCQR approach.
In this paper, we study the application of LCQR in kai2010local to the estimation and inference in RD. We present numerical evidence for the efficiency gain of using LCQR in RD estimation, and also propose a bias-corrected and s.e.-adjusted t-statistic to improve the coverage of confidence intervals. Simulation results show good performance of the proposed method under several non-normal error distributions.
The current work can be extended in several directions. For instance, throughout the paper, we focus on the local linear composite quantile regression, while the general case, local $p$-th order polynomial composite quantile regression, can be similarly adopted. kai2010local_supp establish the asymptotic theory for the LCQR estimator in the general case. Following kai2010local_supp, one could extend our results on the bias-corrected and s.e.-adjusted t-statistic to allow for higher-order polynomials. In addition, it will be naturally appealing to formally explore LCQR for kink RD designs as well as RD designs using covariates. It will also be interesting to revisit some of the existing applications in RD with the proposed method as data may deviate from normality. Finally, on the computation side, instead of using the same bandwidth for estimation, bias correction and s.e. adjustment, one can refine the bandwidth selection process, which may also improve the estimation and coverage probability of LCQR. We leave these topics for future research.
We thank the Editor, the Associate Editor, and two anonymous referees for their comments that substantially improved the paper. The authors gratefully acknowledge the financial support of the Education Economics Center and the computational support of the Office of Research at Kennesaw State University, and have no conflicts of interest to disclose.
\spacing{1.45}
\setcounter{page}{1} \spacing{1.42}