EconBase
← Back to paper

Regression discontinuity design with right-censored survival data

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.

80,947 characters · 13 sections · 56 citation commands

Rendered from LaTeX for readability, not typeset faithfully. Citation keys are highlighted; maths is left as source; figures, tables and equation environments are summarised rather than reproduced; unrecognised commands are greyed out so nothing is silently dropped. Email addresses are removed.

Regression discontinuity design with right-censored survival data

abstract{ In this paper the regression discontinuity design is adapted to the survival analysis setting with right-censored data, studied in an intensity based counting process framework. In particular, a local polynomial regression version of the Aalen additive hazards estimator is introduced as an estimator of the difference between two covariate dependent cumulative hazard rate functions. Large-sample theory for this estimator is developed, including confidence intervals that take into account the uncertainty associated with bias correction. As is standard in the causality literature, the models and the theory are embedded in the potential outcomes framework. Two general results concerning potential outcomes and the multiplicative hazards model for survival data are presented.}

Introduction

The {{regression discontinuity design}} is a widely used technique for causal inference in economics, political science, and sociology (see van2008regression, imbens2008regression, lee2010regression, and cattaneo2022regression for recent reviews). In the {{regression discontinuity design}}, possible confounders are {`}controlled for{'} by exploiting that units are assigned to treatment based on whether their value of an observed covariate is above or below some known cut-off, the idea being that units with values of this observed covariate just above the known cut-off are similar to the subjects with values just below the cut-off.

For situations where the observed outcomes are noncensored random variables, the inference theory, building on that of local polynomial regression, is well developed (see e.g., fan1996local, hahn2001identification, and calonico2014robust,calonico2014supplement, and the reviews above). In this paper we extend the {{regression discontinuity design}} to the survival analysis setting. In particular, we study the {{regression discontinuity design}} applied to right-censored survival data in an intensity based counting process framework (see ABGK93, and ABG08,aalen2010history). The survival analysis models we study are all instances of the broad class of multiplicative intensity models, meaning that the modelling and inference revolves around the hazard rate function.

The setup is as follows: On a probability space $(\Omega,\mathscr{H},{\rm Pr})$, let $Z \in \mathbb{R}$ be an observed covariate; $X\in \{0,1\}$ a treatment indicator; $U$ a vector of unobserved confounders; $\widetilde{T}^0 \geq 0$ the potential outcome for the non-treated, and $\widetilde{T}^1 \geq 0$ the potential outcome under treatment (for potential outcomes theory, see, for example, holland1986statistics, morgan2015counterfactuals, imbens2015causal, and imbens2020potential\footnote{imbens2020potential is an excellent and very readable article comparing the potential outcomes theory to the directed acyclic graph theory associated with Judea Pearl. See pearl2009causality for the canonical treatise on DAGs, and pearl2018why or pearl2016causal for more accessible accounts.}). Since $\widetilde{T}^0$ and $\widetilde{T}^1$ take values on the positive half of the real line, we refer to these two random variables as {\it potential lifetimes}. That we are in a regression discontinuity setting, means that treatment is determined by the value of an observed covariate. We take $X = I\{Z \geq z_0\}$ for some known cut-off $z_0$, where $I\{A\}$ is the indicator function of the event $A$.\footnote{What we describe here is the {\it sharp} regression discontinuity design. In contrast, a {\it fuzzy} regression discontinuity design occurs when the probability of receiving treatment does not jump from zero to one at the cut-off, rather $z \mapsto {\rm Pr}(X = 1\mid Z = z)$ has a point of discontinuity at the cut-off. This paper deals only with the sharp regression discontinuity design.} The two potential lifetimes $\widetilde{T}^0$ and $\widetilde{T}^1$ are assumed to stem from distributions with hazard rate functions $\alpha_0(t,Z,U)$ and $\alpha_1(t,Z,U)$, respectively, meaning that

equation[equation omitted — 209 chars of source]

In analogy with the average treatment effect (the {{\sc ate}}) and conditional or local average treatment effects ({{\sc cate}}s or {{\sc late}}s) (see, e.g., imbens2015causal), the causal estimands of the present paper are all defined in terms $\alpha_0(t,Z,U)$ and $\alpha_1(t,Z,U)$ when the confounder $U$ is averaged out. Specifically, with

equation[equation omitted — 109 chars of source]

the main estimand in this paper is

equation[equation omitted — 70 chars of source]

evaluated in the cut-off $z = z_0$. The estimands $\theta(t,z_0)$ and $\Theta(t,z_0)$ are, we contend, natural survival analysis counterparts of the classical estimand in the standard {{regression discontinuity design}}, namely the average treatment effect at the cut-off, see, for example, Eq. (2.1) in imbens2008regression.

The paper proceed as follows. In Section (ref) we provide a general presentation of the potential outcomes framework as it applies to right-censored survival data studied in a counting process framework. In Section (ref), we narrow in on the {{regression discontinuity design}} and present a lemma that is key to making the {{regression discontinuity design}} feasible in the present setting. Section (ref) contains a discussion of various causal estimands, and assumptions under which they may be identified. In Section (ref) we lay out the assumptions made about the model generating the data, and present a special case of the covariate-localised Aalen additive hazards estimator. Section (ref) contains large-sample theory for the general version of the estimator, with results on variance estimation and bias correction in Sections (ref) and (ref), respectively.

Potential outcomes and right-censored data

Let $(\widetilde{T}^0,\widetilde{T}^1,X,Z,U,C)$ be random variables on the probability space $(\Omega,\mathscr{H},{\rm Pr})$. Here, $\widetilde{T}^0$ and $\widetilde{T}^1$ are the potential lifetimes corresponding to whether a subject is not treated or treated, respectively; $X \in \{0,1\}$ is an indicator of treatment; $Z$ is a vector of observed covariates; $U$ is a vector of unobserved possible confounders;\footnote{A confounder is a covariate that is associated with the outcome {\it and} with the treatment variable of interest. In the multiplicative intensity models studied here, this means that a covariate, to be a confounder, must have an effect on the hazard rate {\it and} be correlated with the treatment variable. If a covariate only satisfy one of these two conditions, then it is not a confounder.} and $C \geq 0$ is a censoring variable. We use $g$ as an index for untreated ($g = 0$) and treated ($g = 1$), and write, for example, $\widetilde{T}^g$ to indicate one of the two potential lifetimes. The units under study are observed over the time interval $[0,\tau]$, where $\tau < \infty$. Assume that, conditionally on $(Z,U)$, the potential lifetimes stem from distributions with hazard rate functions $\alpha_{0}(t,Z,U)$ and $\alpha_{1}(t,Z,U)$.\footnote{Here we model the unconfoundedness assumption $(\widetilde{T}^0,\widetilde{T}^1)\perp \!\!\! \perp X \mid (Z,U)$ directly. See yadlowsky2018bounds. In the {{regression discontinuity design}} this assumption is trivially satisfied imbens2008regression.}

What distinguishes the potential outcomes theory when right-censoring is present from when there is no censoring, is that with no censoring only one of the two potential lifetimes is observed for the same unit, while for right-censored data {\it at most} one of the potential lifetimes is actually observed for the same unit. This means that when there is no censoring we observe

equation[equation omitted — 81 chars of source]

while, when the data are right-censored we only observe $\widetilde{T}$ if it is smaller than the censoring time $C$, that is

equation[equation omitted — 72 chars of source]

where $T^0 = \min(\widetilde{T}^0,C)$ and $T^1 = \min(\widetilde{T}^1,C)$ are the possibly right-censored potential lifetimes. Throughout the paper, we assume that the two potential lifetimes are conditionally independent given the covariates, which we express by

equation[equation omitted — 105 chars of source]

and we work under the assumption of {\it random right-censoring} (see ABGK93), which means that $(\widetilde{T}^0,\widetilde{T}^1)$ is independent of the censoring time $C$ given $(X,Z,U)$, with symbols

equation[equation omitted — 112 chars of source]

Two potential lifetimes in turn leads to two indicators of noncensoring,

equation[equation omitted — 122 chars of source]

and to two potential counting and potential at-risk processes

equation[equation omitted — 136 chars of source]

Let $\mathcal{X} = \sigma(X,Z,U)$, define the filtrations $\mathscr{E}_{t}^g = \sigma(\{N^g(s),Y^g(s)\}_{s\leq t})$ for $g = 0,1$, and set

equation[equation omitted — 103 chars of source]

It is assumed that $X$, $Z$, and $U$ are realised at time zero, meaning that $\mathscr{G}_0^g = \mathcal{X}$. With respect to $\mathscr{G}_t^0$ and $\mathscr{G}_t^1$ we have, using the assumptions in (ref) and (ref) (see ABGK93), two {`}potential{'} local square integrable martingales $M^0$ and $M^1$, respectively, given by

equation[equation omitted — 133 chars of source]

This ends our description of our basic modelling assumptions. The reader familiar with the counting process approach to, and martingale methods in, survival analysis will see that the above is nothing more than what one gets the when taking the potential outcomes framework seriously and applying it to the standard intensity based counting process approach to survival analysis. In fact, the presentation so far is just a slight notational and semantic reformulation of the theory presented in Chapter III.2 of ABGK93. It is important to notice that the modelling undertaken up to this point has taken place in the two potential worlds, so to speak, culminating in the {`}potential world{'} martingales of (ref). We now proceed to the consequences of this model for quantities of this world, that is, the observed quantities. Denote the observed counting process

equation[equation omitted — 69 chars of source]

and the observed at-risk process

equation[equation omitted — 70 chars of source]

Let $\mathcal{E}_{t} = \sigma(\{N(s),Y(s)\}_{s\leq t})$ be the filtration generated by the observable counting and at-risk processes, and let $\mathcal{X}^{\rm obs} = \sigma(X,Z)$ be the {$\sigma$-algebra} generated by the observed covariates. Define the filtrations

equation[equation omitted — 169 chars of source]

where, as above, $\mathcal{G}_{0} = \mathcal{X}$ and $\mathcal{F}_0 = \mathcal{X}^{\rm obs}$. The next lemma says that the intensity process of the observed counting process $N$ with respect to $\mathcal{G}_t$ takes the form it it intuitively should take.

lemmaAssume that (ref) and (ref) hold. The $\mathcal{G}_{t}$-intensity of the observed counting process $N = XN^1 + (1 -X)N^0$ is $Y(t) \{X \alpha_1(t,Z,U) + (1 - X) \alpha_0(t,Z,U) \}$. In particular, \begin{equation} I_{X = g}M^g(t) = I_{X = g}\{ N^g(t) - \int_0^t Y^g(s)\alpha_{g}(s,Z,U)\,{\rm d} s \}, \end{equation} is a locally square integrable martingale with respect to $\mathcal{G}_t$.
proofDefine $\mathscr{G}_t = \mathscr{G}_t^0 \vee \mathscr{G}_t^1$. It follows directly from the theory on independent right-censoring ABGK93 that the $\mathscr{G}_t$-intensity of $N^g(t)$ is also $Y^g(t)\alpha_g(t,Z,U)$. Since $I_{X=g}$ is $\mathscr{G}_{t}$ measurable for all $t$, the $\mathscr{G}_t$-intensity of $I_{X=g}N^g(t)$ is $I_{X=g}Y^g(t)\alpha_g(t,Z,U)$ (this is a consequence of Theorem II.3.1 in ABGK93, considering $I_{X=g}$ as a constant process in $t$). Since $\mathcal{G}_{t}\subset \mathscr{G}_t$, the innovation theorem (see aalen1978nonparametric or ABGK93) entails that the $\mathcal{G}_t$-intensity process, $\lambda_g^{\mathcal{G}}(t)$ say, of $I_{X=g}N^g$ is \begin{equation} \lambda_g^{\mathcal{G}}(t) = {\rm E}\,\{ I_{X=g}Y^g(t) \alpha_g(t,Z,U) \mid \mathcal{G}_{t-}\}. \notag \end{equation} Since $I_{X=g}Y^g(t) = I_{X=g}Y(t)$, in terms of the observed at-risk process in (ref); $I_{X = g}$ is $\mathcal{G}_t$-measurable for all $t$, and $Y(t)$ is predictable with respect to $\mathcal{G}_{t}$, \begin{equation} \begin{split} \lambda_g^{\mathcal{G}}(t) & = {\rm E}\,\{ I_{X=g}Y^g(t) \alpha_g(t,Z,U) \mid \mathcal{G}_{t-}\}\\ & = {\rm E}\,\{ I_{X=g}Y(t) \alpha_g(t,Z,U) \mid \mathcal{G}_{t-}\} = I_{X=g}Y(t) \alpha_g(t,Z,U). \end{split} \notag \end{equation} This entails that $I_{X=g} M^g(t) = I_{X = g}\{N^g(t) - \int_0^t \lambda_g^{\mathcal{G}}(s)\,{\rm d} s\}$ is a $\mathcal{G}_t$ martingale, and, because it is a counting process martingale, it is locally square integrable (see ABGK93). Since $X$ is $\mathcal{G}_0$-measurable ($\mathcal{G}_0 = \mathcal{X}$) and does not depend on time, $XM^1 + (1 - X)M^0$ is also a locally square integrable martingale.

An important, though rather intuitive, thing to note in the preceding lemma is that it does {\it not} say that $M^g$ is a $\mathcal{G}_t$-martingale. Instead, the lemma says that $I_{X = g}M^g$ is a $\mathcal{G}_t$-martingale. This is intuitive because $\mathcal{G}_t$ is generated by the observables $N$ and $Y$ in (ref) and (ref), respectively, and if $X=0$, for example, then the path of $N^1$ is certainly not observable.

In the next section we specialise the potential outcomes model for right-censored survival data to the regression discontinuity setting.

The regression discontinuity setting

We retain the definitions and assumptions from the previous section, with the following two exceptions: To conform with the {{regression discontinuity design}}, we require $Z \in \mathbb{R}$ and set $X = I\{Z \geq z_0\}$ for some known cut-off $z_0$. Henceforth, we often refer to $Z$ as the {\it forcing variable}. Notice also that $\sigma(X)$ is in included in $\sigma(Z)$, with the consequence that $\mathcal{X} = \sigma(Z,U)$ and $\mathcal{X}^{\rm obs} = \sigma(Z)$, and thus $\mathcal{G}_t = \mathcal{E}_t \vee \sigma(Z,U)$ and $\mathcal{F}_t = \mathcal{E}_t \vee \sigma(Z)$ (compare with (ref)).

In order to make what follows clear, we now take a short detour via the standard regression discontinuity design, as presented, for example, in imbens2008regression or calonico2014robust. Let $y^0$ and $y^1$ be two real valued potential outcomes,\footnote{We use lowercase letters to distinguish these two potential outcomes from the potential at-risk processes introduced in Section (ref).} $Z \in \mathbb{R}$ the forcing variable, and $X = I\{Z \geq z_0\}$ the treatment indicator. The observed outcome is $y = Xy^1 + (1 - X)y^0$, and the estimand of interest is $\tau_{\rm srd} = {\rm E}\,( y^1 - y^0 \mid Z = z_0)$, called the average treatment effect at the cut-off (or threshold). To estimate $\tau_{\rm srd}$, the limits $\lim_{z \uparrow z_0}{\rm E}\,( y \mid Z = z)$ and $\lim_{z \downarrow z_0}{\rm E}\,( y \mid Z = z)$ are estimated using local polynomial regressions to the left and to the right of the cut-off, respectively. Since the roles of the hazard rate functions $\alpha_0(t,Z,U)$ and $\alpha_1(t,Z,U)$ in multiplicative intensity models are analogoues to those of the conditional expectations ${\rm E}\,(y^0 \mid Z)$ and ${\rm E}\,(y^1 \mid Z)$ in the standard {{regression discontinuity design}}, we would like to work with hazard rate functions that only depend on the forcing variable (and time). More to the point, if ${\rm E}\,\{\alpha_g(t,Z,U) \mid \sigma(Z)\}$ was the hazard rate function of $I_{X = g} N^g$ with respect to the filtration of observables, then the standard local polyonimial regression theory would be straighforward to mimic. This is not quite the case, but nearly, in a sense made precise by the next lemma.

lemmaLet $T,Z,U,N,Y$, $\mathcal{E}_t, \mathcal{F}_t$, and $\mathcal{G}_t$ be as defined in Section (ref). Assume that $Z$ has density $f_Z(z)$ with support $[z_1,z_2]$, and that $f_Z(z)$ is bounded above and below on $[z_1,z_2]$, and that $z \mapsto {\rm Pr}(T \geq \tau \mid Z = z)$ is bounded below on $[z_1,z_2]$. For each $t \in [0,\tau]$, define \begin{equation} \mathcal{Z}_t = \sigma(Y(t))\vee \sigma(Z), \notag \end{equation} and let $\xi_t$ be a non-negative $\mathcal{G}_t$-adapted and almost surely left-continuous process, with ${\rm E}\,\sup_{t \leq \tau}\lvert\xi_t\lvert^2\, < \infty$. Then the left-continuous modifications of the processes \begin{equation} {\rm E}\,(\xi_t \mid \mathcal{F}_{t-})\quadand\quad I_{T \geq t}{\rm E}\,(\xi_t \mid \mathcal{Z}_t) + I_{T < t}{\rm E}\,(\xi_t \mid \mathcal{F}_{t-}), \notag \end{equation} are indistinguishable.
proofThe proof of this lemma is in Appendix (ref).

Our application of this lemma is when $\xi_t$ is one of the hazard rate functions $\alpha_{0}(t,Z,U)$ or $\alpha_{1}(t,Z,U)$. The importance of this lemma derives from the fact that $Y(t){\rm E}\,(\xi_t \mid \mathcal{F}_{t-})$ is, at first sight, a complicated function as it may depend on the forcing variable $Z$ as well as the paths of $N(t)$ and $Y(t)$. But an implication of Lemma (ref), since $Y(t)$ is left-continuous, is that

equation[equation omitted — 162 chars of source]

and since $\mathcal{Z}_t = \{\{T \geq t\},\{T < t\},\emptyset,\Omega\}\vee \sigma(Z)$, the conditional expectation $Y(t) {\rm E}\,(\xi_t \mid \mathcal{Z}_t)$ is a function of the forcing variable only (see, for example, Lemma 1.13, p. 7 and the discussion on p. 106 in kallenberg2002foundations). This motivates defining the functions,\footnote{ That ${\rm E}\,(\xi_t \mid \mathcal{Z}_t) = {\rm E}\,\{I_{T \geq t}\xi_t \mid \sigma(Z)\}/{\rm Pr}\{T \geq t\mid \sigma(Z)\} + {\rm E}\,\{I_{T < t}\xi_t \mid \sigma(Z)\}/{\rm Pr}\{T < t\mid \sigma(Z)\}$ almost surely, can be proved following the steps in Ex. 34.4 of billingsley1995probability.}

equation[equation omitted — 177 chars of source]

Notice that if $\alpha_g(t,Z,U) \leq g(Z,U)$ for some $g$ such that ${\rm E}\,g(Z,U) < \infty$, then we can pass the derivative under the intergal sign in $\partial/\partial t\,{\rm E}\,\{S_g(t,Z,U) \mid \sigma(Z)\} = {\rm E}\,\{\partial/\partial tS_g(t,Z,U) \mid \sigma(Z)\}$,\footnote{The dominated convergence theorem extends to conditional expectations, see, e.g. cohen2015stochastic} and consequently we have the relation $I_{X = g}\bar{\alpha}_g(t,z) = - I_{X = g}(\partial \bar{S}_g(t,z)/\partial t)/ \bar{S}_g(t,z)$. We summarise the above in a lemma that is used repeatedly in the remainder of the paper.

lemmaFor $g = 0,1$ define \begin{equation} \bar{M}^g(t) = I_{X = g} \{ N^g(t) - \int_0^t Y^g(s)\bar{\alpha}_g(s,Z)\,{\rm d} s\}. \notag \end{equation} Then $\bar{M}^g(t)$ are local square integrable martingales with respect to the filtration $\mathcal{F}_t$.
proofFrom Lemma (ref) we have that $I_{X = g}Y^g(t) \alpha_g(t,Z,U)$ is the $\mathcal{G}_t$-intensity of $I_{X = g}N^g(t)$. By the innovation theorem (references above), the $\mathcal{F}_t$-intensity $\lambda^{\mathcal{F}}(t)$ of $I_{X = g}N^g(t)$ is $I_{X = g}Y^g(t) {\rm E}\,\{\alpha_g(t,Z,U)\mid \mathcal{F}_{t-}\}$. Since $I_{X=g}Y^g(t) = I_{X=g}Y(t)$, Lemma (ref) gives that $\lambda^{\mathcal{F}}(t)$ and $I_{X=g}Y^g(t) \bar{\alpha}_g(t,Z)$ are indistinguishable. Thus, $I_{X=g}\int_0^tY^g(s) \bar{\alpha}_g(s,Z)\,{\rm d} s$ is the compensator of the counting process $I_{X = g}N^g(t)$, and $\bar{M}^g(t)$ are locally square integrable martingales ABGK93.

Estimands

If the difference $\alpha_{1}(t,z,u) - \alpha_{0}(t,z,u)$ is functionally independent of $u$, meaning that $\alpha_{1}(t,Z,Y) - \alpha_{0}(t,Z,U)$ is a $\sigma(Z)$-measurable random variable, then

equation[equation omitted — 92 chars of source]

where $\bar{\alpha}_g(t,z)$ for $g = 0,1$ are defined in (ref) and $\theta(t,z)$ is defined in (ref). The functional independence assumption just introduced is crucial for whether $\theta(t,z_0)$ is identifiable or not. We state the functional independence assumption here for easy reference, note, however, that it is not assumed throughout.

assumptionThe difference $\alpha_1(t,Z,U) - \alpha_0(t,Z,U)$ is functionally independent of the confounder $U$.

If this assumption is dropped, we are only able to estimate the average treatment effect at the cut-off {\it among those at-risk}, a quantity we denote $\theta_{\rm risk}(t,z_0)$, it is

equation[equation omitted — 139 chars of source]

with cumulative $\Theta_{\rm risk}(t,z) = \int_0^t \theta_{\rm risk}(s,z)\,{\rm d} s$.\footnote{By the definition in (ref) we mean that $\theta_{\rm risk}(t,Z)$ is the $\sigma(Z)$-measurable function such that $ Y(t){\rm E}\,\{\alpha_{1}(t,Z,U) - \alpha_{0}(t,Z,U)\mid \mathcal{Z}_t\} = Y(t)\theta_{\rm risk}(t,Z)$ almost surely.} The assumption of the difference $\alpha_1(t,Z,U) - \alpha_0(t,Z,U)$ being functionally independent of the confounder is crucial for whether we are estimating $\theta(t,z_0)$ or the at-risk version $\theta_{\rm risk}(t,z_0)$, but is otherwise immaterial to the theory developed in the subsequent sections.

The data, the assumptions, and the estimator

In this section, we first elaborate on the assumptions made about what are in statistics and econometrics jargon, respectively, called the true model or the data generating process. Subsequently, in Section (ref) we introduce a special case of our estimator, and provide some theory for this special case. The general large-sample theory is deferred to Section (ref).

Data and assumptions

Let $(\widetilde{T}_i^0,\widetilde{T}_i^1,Z_i,U_i,C_i),\, i = 1,\ldots,n$ be independent replicates of $(\widetilde{T}^0,\widetilde{T}^1,Z,U,C)$, where this latter is as described in Sections (ref) and (ref). In particular, the forcing variable $Z \in \mathbb{R}$ and $X = I\{Z \geq z_0\}$ for a known cut-off $z_0$. This entails that $(N_1,Y_1),\ldots,(N_n,Y_n)$ are independent replicates of $(N,Y)$ with $N(t) = XN^1(t) + (1 - X)N^0(t)$ and $Y(t) = XY^1(t) + (1 - X)Y^0(t)$ as defined in Section (ref). As above, we assume that the processes $N_i$ and $Y_i$ are observed over the finite time interval $[0,\tau]$. The filtrations $\mathcal{G}_t$ and $\mathcal{F}_t$ are now $\mathcal{G}_t = \mathcal{E}_t \vee \sigma(Z_i,U_i,\,i = 1,\ldots,n)$ and $\mathcal{F}_t = \mathcal{E}_t \vee \sigma(Z_i,\,i = 1,\ldots,n)$, with $\mathcal{E}_t^g = \sigma(\{(N_i(s),Y_i(s)\}_{s \leq t},\,i = 1,\ldots,n)$. For $g = 0,1$, denote $M_1^g,\ldots,M_n^g$ the independent replicates of the martingale $M^g$ in (ref). These are orthogonal local square integrable martingales with respect to the filtration $\mathscr{G}_t^g$. Similarly, $I_{X_1 = g}M_1^{g}(t),\ldots,I_{X_n = g}M_n^{g}(t)$ are orthogonal local square integrable martingales with respect to the filtration $\mathcal{G}_t$ (see Lemma (ref)), and $I_{X_1 = g}\bar{M}_1^{g}(t),\ldots,I_{X_n = g}\bar{M}_n^{g}(t)$ are orthogonal local square integrable martingales with respect to the filtration $\mathcal{F}_t$ (see Lemma (ref)).

The survival functions associated with the hazard rates $\alpha_{0}(t,Z,U)$ and $\alpha_{1}(t,Z,U)$ are denoted $S_{g}(t,Z,U) = \exp\{- \int_0^t \alpha_{g}(s,Z,U)\,{\rm d} s\}$ for $g = 0,1$, and we set

equation[equation omitted — 129 chars of source]

Let $y_g(t,z)$ to be the conditional expectations of the at-risk process $Y^g(t)$ given $Z = z$. Using the assumption in (ref), these functions are

equation[equation omitted — 136 chars of source]

where $H(t)$ is the distribution function of the censoring variable $C$. Without further mention, the following is assumed throughout the paper

assumptionThe potential lifetimes and the censoring variable are independent given the covariates, that is $\widetilde{T}^0\perp \!\!\! \perp \widetilde{T}^1 \perp \!\!\! \perp C \mid (Z,U)$.
assumptionConditionally on $Z = z$ and $U=u$, the potential lifetimes stem from distributions with hazard rate functions $\alpha_0(t,z,u)$ and $\alpha_1(t,z,u)$ are continuous in $t$ for all $(z,u)$.
assumptionThe censoring time $C$ stems from a distribution with a continuous distribution function $H(t)$ that is such that $H(\tau) < 1$.
assumptionFor some $\kappa_0 > 0$ the following hold on an interval $(z_0 - \kappa_0,z_0 + \kappa_0)$ around the cut-off $z_0$: \begin{itemize}\itemsep-0.2em • The density $f_Z(z)$ is continuous and bounded away from zero; • The conditional hazards $\bar{\alpha}_0(t,z)$ and $\bar{\alpha}_1(t,z)$, defined in (ref), are $S$ times continuously differentiable in $z$ for all $t \in [0,\tau]$. \end{itemize}
assumptionFor some $\kappa > 0$, the kernel function $k \colon [0,\kappa] \to \mathbb{R}$ is bounded and nonnegative, zero outside its support, and positive and continuous on $(0,\kappa)$.

Assumption (ref) gives the martingale representation in (ref), and is the key to Lemma (ref), and thereby also to Lemma (ref). Assumption (ref) is standard in survival analysis, and is equivalent to requiring absolute continuity of the survival functions. The assumption also entails that the conditional hazards $\bar{\alpha}_0(t,z)$ and $\bar{\alpha}_1(t,z)$ are continuous in $t$ for all $z$. Assumption (ref) is needed to ensure that $y_g(t,z)$ is bounded below (a fact that is, for example, used in the proof of Theorem (ref)). Assumption (ref) allows for the estimation of the parameters of interest using a regression discontinuity design. For a discussion of the analogue of this assumption in the standard regression discontinuity design, see hahn2001identification, and imbens2008regression. In the next section a special case of our main estimator is presented.

A covariate-localised Aalen estimator

The estimator we propose for $\Theta(t,z_0)$ is a weighted version of the Aalen additive hazards estimator (Aalen80,Aalen89,Aalen93, ABG08, and ABGK93).\footnote{Throughout this section we assume, for notational convenience, that Assumption (ref) holds. If this assumption does not hold, then all the theory of this section translates directly to the estimation of $\Theta_{\rm risk}(t,z_0)$.} The idea is to fit local polynomial regression models in the vicinity of the cut-off $z_0$. Thus, {`}local{'} here refers to an interval on the real line on which the forcing variable $Z$ takes its values, and does {\it not} refer to the time axis. Specifically, the conditional expectations $\bar{\alpha}_{0}(t,z)$ and $\bar{\alpha}_{1}(t,z)$, as defined in (ref), are approximated on intervals to the left and to the right of $z_0$ by $p$th order local polynomials. Since, for $g = 0,1$, the conditional expectation $\bar{\alpha}_{g}(t,z)$ may be approximated by

equation[equation omitted — 206 chars of source]

where $\bar{\alpha}_{g}^{(\nu)}(t,z)$ is the $\nu$th derivative of $\bar{\alpha}_{g}(t,z)$ with respect to $z$, the local polynomial regression version of the Aalen additive hazards estimator that we introduce, is an estimator of $\int_0^t \bar{\alpha}_{g}^{(\nu)}(s,z)/\nu!\,{\rm d} s$ for $\nu = 0,1,\ldots,p$. We start by presenting the local linear estimator, and then move on to general results for $p$th order local polynomial estimators in Section (ref).

For a bandwidth $h > 0$ and a kernel function $K_h(u) = K(u/h)/h$ with $K(u) = k(-u)I\{u < 0\} + k(u)I\{u \geq 0\}$,\footnote{In principle, different kernels could be used on either side of the cut-off. For simplicity, we employ the same kernel on both sides as this does not affect the theory.} where $k$ is a function satisfying Assumption (ref), the local linear estimator $(\widehat{B}_{g,1},\widehat{B}_{g,1}^{(1)})$ is, for $g = 0,1$, given by

equation[equation omitted — 250 chars of source]

with

equation[equation omitted — 189 chars of source]

This is seen to be a version the Aalen additive hazards estimator estimator with kernel weights on the forcing variable. An estimator for $\Theta(t,z_0)$ is then given by

equation[equation omitted — 111 chars of source]

For $t$ fixed, this estimator is simply the difference between the intercepts of two local linear regression to the left and to the right of the cut-off. In other words, it is the analogue of the most common estimator of the average treatment effect at the cut-off in the standard regression discontinuity design, see, for example, the estimator $\widehat{\tau}_{\rm srd}$ in imbens2008regression.

Under the assumptions of Lemma (ref) below, the bias of the estimator $\widehat{\Theta}(t,h)$ can be described by

equation[equation omitted — 171 chars of source]

with

equation[equation omitted — 158 chars of source]

Here $\bar{\alpha}_{g}^{(2)}(s,z),\,g = 0,1$ are the unknown second derivates of the conditional expectations defined in (ref); while $e_{1,0} = (1,0)^{{\rm t}}$, and $\Gamma_1$, and $\vartheta_{1,2}$ are quantities that only depend on the chosen kernel, and need not be estimated from the data. They are

equation[equation omitted — 224 chars of source]

Here and elsewhere in the paper, the notation is borrowed from calonico2014robust,calonico2014supplement. If the bandwidth minimising the mean squared error is chosen, namely $h_n = cn^{-1/5}$ for some constant $c > 0$, then, conditionally on the filtration $\mathcal{F}_t$ (here Lemma (ref) is invoked), we have process convergence in the space of c{\`a}dl{\`a}g functions on $[0,\tau]$,

equation[equation omitted — 217 chars of source]

as $n \to \infty$, where $\bar{M}_{0,1}$ and $\bar{M}_{1,1}$ are independent bivariate Gaussian martingales with variation processes

equation[equation omitted — 227 chars of source]

The general version of this result is the content of Corollary (ref). A consistent estimator of $\langle \bar{M}_{g,1},\bar{M}_{g,1}\rangle_t$ is introduced in Section (ref).

To avoid or to get rid of the bias term that appears on the right hand side of (ref), two approaches are discussed in this paper. First, one can choose a bandwidth $h_n$ such that $nh_n^{5}\to 0$, meaning that the bandwidth must be {`}smaller{'} than the mean squared error optimal one. Second, the bias term may be removed by subtracting off a consistent estimate of the bias term. The standard approch is the following (see, e.g., fan1996local): Suppose that ${\rm Bias}_n(t,b_n)$ is consistent for $\int_0^t {\rm bias}(s)\,{\rm d} s$ (uniformly in $t$) as $n \to \infty$ and the so-called pilot bandwidth $b_n \to 0$, then (see Corollary (ref))

equation[equation omitted — 185 chars of source]

as $n \to \infty$, provided, among other things, that $h_n/b_n \to 0$. This latter condition ensures that the variability of the bias correction estimation disappears, meaning that $\bar{M}_{0,1}$ and $ \bar{M}_{1,1}$ are the Gaussian martingales from (ref) (the variation process does not change). As pointed out in the influential paper calonico2014robust, $h_n/b_n$ is never zero in finite samples, and therefore, the variability associated with the bias correction ought to be accounted for in the limiting distribution. In the present paper, results of this type are presented in Section (ref).

General theory

In this section we first consider the general $p$th order local polynomial regression estimator of the $\nu$th derivative function $\int_0^t \bar{\alpha}_g^{(\nu)}(s,z_0)\,{\rm d} s$, and derive a representation for the bias of this estimator. Next, in Section (ref), we present two central limit theorems for this estimator. Throughout this section, Assumption (ref) (the functional independence assumption) is, for notational convenience, assumed to hold. If this assumption does not hold, all subsequent results are true with $\theta(t,z_0)$ replaced by $\theta_{\rm risk}(t,z_0)$ (see the discussion in Section (ref)). Moreover, $g = 0,1$ is the index used to indicate the potential outcomes, or statistics depending on these, for non-treated and treated, respectively. Since the theory we develop is the same on both sides of the cut-off, a result concerning a quantity with subscript $g$ means that it applies for both $g = 0$ and $g = 1$. Most of the proofs of the claims made in the present section are deferred to the appendices.

The estimator and its bias

For an integer $p \geq 1$, let $r_p(x) = (1,x,x^2,\ldots,x^p)^{{\rm t}}$, and define the estimator $\widehat{B}_{g,p}$ by

equation[equation omitted — 183 chars of source]

with

equation[equation omitted — 133 chars of source]

and

equation[equation omitted — 108 chars of source]

If $J_{n,h}(t) = 0$, we take ${\rm d} \widehat{B}_{g,p}(t,h) = 0$ for $g = 0,1$. Let $H_p(h) = {\rm diag}(1,h^{-1},\ldots,h^{-p})$. Using that $H_p(h)r_p(z) = r_p(z/h)$ and $r_p(z) = H_{p}(h)^{-1}r_p(z/h)$, the estimator above can be expressed as

equation[equation omitted — 189 chars of source]

with

equation[equation omitted — 185 chars of source]

and, since $G_{g,p,n}(t,h)$ is positive definite if and only if $\Gamma_{g,p,n}(t,h)$ is positive definite, $J_{n,h}(t) = I\{\text{$\Gamma_{0,p,n}(t,h)$ and $\Gamma_{1,p,n}(t,h)$ are positive definite}\}$. Let $e_{p,\nu}$ be the $(p+1)$-dimensional column vector with its $(\nu+1)$th element equal to $1$, and all other elements equal to zero, for example, $e_{1,0} = (1,0)^{{\rm t}}$, $e_{2,1} = (0,1,0)^{{\rm t}}$, $e_{3,2} = (0,0,1,0)^{{\rm t}}$, and so on. Denote $\bar{\alpha}_{g}^{(\nu)}(t,z)$ the $\nu$th partial derivative of $\bar{\alpha}_{g}(t,z)$ with respect to $z$, so $\bar{\alpha}_{g}^{(0)}(t,z) = \bar{\alpha}_{g}(t,z)$. With this notation,

equation[equation omitted — 153 chars of source]

are estimators for $\int_0^t \bar{\alpha}_0^{(\nu)}(s,z_0)\,{\rm d} s$ and $\int_0^t \bar{\alpha}_1^{(\nu)}(s,z_0)\,{\rm d} s$, respectively. An estimator for the $\nu$th derivative of $\Theta(t,z) = \int_0^t \theta(s,z)\,{\rm d} s$ with respect to $z$ and evaluated in $z_0$ (see (ref)), based on a $p$th order polynomial regression estimator, is then

equation[equation omitted — 144 chars of source]

Before we state the bias lemma, we also need the following quantities,

equation[equation omitted — 161 chars of source]

and

equation[equation omitted — 82 chars of source]

The matrices $\Gamma_p$ and $\Psi_p$ are of dimension $(p+1)\times(p+1)$, while $\vartheta_{p,q}$ are $(p+1)$-dimensional column vectors. It will be assumed throughout that the kernel $K$ is chosen so that $\Gamma_p$, for all relevant $p$, is positive definite. This entails that the $\Gamma_{p}^{-1}$ is also positive definite, and that $\lVert\Gamma_{p}\rVert$ and $\lVert\Gamma_{p}^{-1}\rVert$ are finite, with $\lVert\cdot\rVert$ here denoting the matrix norm. The $\mathcal{F}_t$-compensator of $I_{X= g}N^g(t)$, namely $I_{X=g}\int_0^tY^g(s)\bar{\alpha}_g(s,Z)\,{\rm d} s$, is absolutely continuous. This entails that the $\mathcal{F}_t$-compensator of $\widehat{B}_{g,p}^{(\nu)}(t,h)$ is also absolutely continuous, the derivative of this compensator therefore exists, and we denote these derivatives ${\rm E}\,({\rm d} \widehat{B}_{g,p}^{(\nu)}(t,h) \mid \mathcal{F}_{t-})/{\rm d} t$, and ${\rm E}\,({\rm d} \widehat{A}_{g,p}^{(\nu)}(t,h) \mid \mathcal{F}_{t-})/{\rm d} t = e_{p,\nu}^{{\rm t}}\nu!{\rm E}\,({\rm d} \widehat{B}_{g,p}^{(\nu)}(t,h) \mid \mathcal{F}_{t-})/{\rm d} t$. Introduce the column vectors of partial derivatives

equation[equation omitted — 184 chars of source]

and the vector valued functions

equation[equation omitted — 239 chars of source]

We can now state the bias lemma.

lemma{\sc (Bias)} The condtions of Lemma (ref) hold, and Assumption (ref)(b) holds with $S \geq p+2$. Then \begin{equation} {\rm E}\,({\rm d}\widehat{B}_{g,p}(t,h)\mid \mathcal{F}_{t-})/{\rm d} t = J_{n,h}(t)\beta_{g,p}^{(\nu)}(t,z_0) + H_{p}(h)h^{p+1}\mathfrak{b}_{g,p}(t) + H_p(h)O_p(h^{p+2}), \notag \end{equation} uniformly in $t \in [0,\tau]$, as $nh \to \infty$ and $h \to 0$.
proofThe proof is in Appendix (ref).
remarkThis lemma is a relative of Theorem 3.1 in fan1996local. The difference between $p - \nu$ odd and $p - \nu$ even discussed in relation to that theorem, does not apply here, however. This is becase the kernel $k$ is not symmetric to the left nor to the right of the cut-off.

Limiting normality

The central limit theorems coming up concern the sequences

equation[equation omitted — 145 chars of source]

properly normalised. Since $\widehat{\Theta}_{p}^{(\nu)}(t,h) = \widehat{A}_{1,p}^{(\nu)}(t,h) - \widehat{A}_{0,p}^{(\nu)}(t,h)$, and the theory for both sides of the cut-off are the same, we consider only $\widehat{A}_{g,p}^{(\nu)}(t,h) - \int_0^t \bar{\alpha}_g^{(\nu)}(s,z_0)\, {\rm d} s$ for a generic $g = 0,1$. These sequences are

equation[equation omitted — 224 chars of source]

and can be decomposed in different ways depending on which filtration we analyse it with respect to. From Lemma (ref) we have that $I_{X=g}M_i^g(t) = I_{X=g}\{N_i^g(t) - \int_0^tY_i^g(s)\alpha_g(s,Z_i,U_i)\,{\rm d} s\}$ are martingales with respect to the filtration $\mathcal{G}_t = \mathcal{E}_t \vee \mathcal{X}$, and from Lemma (ref) that $\bar{M}_i^g(t) = I_{X_i = g}\{N_{i}^g(t) - \int_0^t Y_i^g(s) \bar{\alpha}(s,Z_i)\,{\rm d} s\}$ are martingales with respect to the filtration $\mathcal{F}_{t} = \mathcal{E}_t \vee \mathcal{X}^{\rm obs}$ of observables. Recall that $\mathcal{X}^{\rm obs} \subset \mathcal{X}$, and that the confounders are not measurable with respect to $\mathcal{F}_t$ (see Sections (ref) and (ref) for the definitions). This, in turn, leads to two decompositions of the sequence in (ref), and, consequently, to two different central limit theorems for this sequence. Write,

equation[equation omitted — 281 chars of source]

where ${\rm Bias}_{g,p,n}(t,h)$ is, using Lemma (ref),

equation[equation omitted — 310 chars of source]

The first term on the right in (ref) can be written

equation[equation omitted — 152 chars of source]

where $\bar{M}_{g,p,n}(t,h)$ is a martingale with respect to the filtration $\mathcal{F}_t$, namely

equation[equation omitted — 272 chars of source]

where $\bar{M}_i^g$ are the $\mathcal{F}_t$-martingales of Lemma (ref).

theorem{\sc ($\mathcal{F}_t$-clt)} The condtions of Lemma (ref) hold. As $nh \to \infty$ and $h \to 0$ \begin{equation} (nh)^{1/2} (\bar{M}_{0,p,n}(\cdot,h),\bar{M}_{1,p,n}(\cdot,h)) \Rightarrow (\bar{M}_{0,p},\bar{M}_{1,p}), \notag \end{equation} where $\bar{M}_{0,p}$ and $\bar{M}_{1,p}$ are orthogonal mean zero Gaussian martingales with variation processes \begin{equation} \langle \bar{M}_{g,p},\bar{M}_{g,p}\rangle_{t} = \frac{1}{f_Z(z_0)} \int_0^t \frac{\bar{\alpha}_{g}(s,z_0)}{y_{g}(s,z_0)}\,{\rm d} s \, \Gamma_p^{-1}\Psi_p\Gamma_p^{-1}. \end{equation}
proofNote first that $nh\langle \bar{M}_{0,p,n}(\cdot,h),\bar{M}_{1,p,n}(\cdot,h)\rangle_t = 0$ for all $t$ and $n$, and is therefore zero in the limit. Assume, without loss of generality, that $z_0 = 0$. Define $\widetilde{\Gamma}_{1,p}(t) = y_{1}(t,z_0)f_Z(z_0)\Gamma_{p}$ and $\widetilde{\Gamma}_{0,p}(t) = y_{0}(t,z_0)f_Z(z_0)H_p(-1)\Gamma_{p}H_p(-1)$, and set $D_{g,p,n}(t,h) = J_{n,h}(t)\Gamma_{g,p,n}(t,h)^{-1} - \widetilde{\Gamma}_{g,p}(t)^{-1}$ for $g = 0,1$. Then \begin{equation} \bar{M}_{g,p,n}(t,h) = \xi_{g,p,n}(t,h) + r_{g,p,n}(t,h), \notag \end{equation} say, where, \begin{equation} \xi_{g,p,n}(t,h) = \int_0^t \widetilde{\Gamma}_{g,p}(s)^{-1}n^{-1}\sum_{i=1}^nI_{X_i=g}K_h(Z_i)r_p(Z_i/h)\,{\rm d} \bar{M}_i^g(s), \end{equation} and \begin{equation} r_{g,p,n}(t,h) = \int_0^t D_{g,p,n}(s,h)n^{-1}\sum_{i=1}^nI_{X_i=g}K_h(Z_i)r_p(Z_i/h)\,{\rm d} \bar{M}_i^g(s). \notag \end{equation} For any $\nu = 0,1,\ldots,p$, the predictable quadratic variation of $e_{p,\nu}^{{\rm t}} r_{g,p,n}(t,h)$ with respect to $\mathcal{F}_t$ is, using Lemma (ref) in the appendix, \begin{equation} \begin{split} &e_{p,\nu}^{{\rm t}}\langle r_{g,p,n}(\cdot,h),r_{g,p,n}(\cdot,h)\rangle_{t}e_{p,\nu}\\ & = \frac{1}{(nh)^2}\sum_{i=1}^nI_{X_i = g}K(Z_i/h)^2 \int_0^t e_{p,\nu}^{{\rm t}}D_{g,p,n}(s,h)r_p(Z_i/h)r_p(Z_i/h)^{{\rm t}}D_{g,p,n}(s,h)e_{p,\nu}\\ & \qquad \qquad \qquad \qquad \times Y_i^g(s)\bar{\alpha}_g(s,Z_i)\,{\rm d} s\\ & \leq \frac{1}{(nh)^2}\sum_{i=1}^nI_{X_i = g}K(Z_i/h)^2 \int_0^t \lVertD_{g,p,n}(s,h)\rVert^2\lVertr_p(Z_i/h)\rVert^2\,Y_i^g(s)\bar{\alpha}_g(s,Z_i)\,{\rm d} s, \end{split} \notag \end{equation} Since $K(z/h)\,\lVertr_p(z/h)\rVert^2 = K(z/h)\sum_{\nu=0}^p (z/h)^{2\nu} \leq K(z/h)\sum_{\nu=0}^p \kappa^{2\nu}$ and $K(u)$ is bounded (see Assumption (ref)), \begin{equation} \begin{split} & e_{p,\nu}^{{\rm t}}\langle r_{g,p,n}(\cdot,h),r_{g,p,n}(\cdot,h)\rangle_{t}e_{p,\nu} \\ & \qquad \quad \lesssim \sup_{t \in [0,\tau]}\lVertD_{g,p,n}(t,h)\rVert^2 \frac{1}{(nh)^2} \sum_{i=1}^n K(Z_i/h) \int_0^{\tau} Y_i^g(s)\bar{\alpha}_g(s,Z_i)\,{\rm d} s. \end{split} \notag \end{equation} From Lemma (ref)(i)--(iii) combined with Lemma (ref) in the appendix, we have that $\sup_{t\in[0,\tau]}\lVertD_{g,p,n}(t,h)\rVert = o_p(1)$ as $nh \to \infty$ and $h \to 0$. For $h \kappa \leq \kappa_0$ (see Assumption (ref)), $\bar{\alpha}_g(t,z)$ is bounded, thus $(nh)^{-1} \sum_{i=1}^n K(Z_i/h) \int_0^{\tau} Y_i^g(s)\bar{\alpha}_g(s,Z_i)\,{\rm d} s\lesssim (nh)^{-1} \sum_{i=1}^n K(Z_i/h) = O_p((nh)^{-1/2})$, from which $e_{p,\nu}^{{\rm t}}\langle r_{g,n,p}(\cdot,h),r_{g,n,p}(\cdot,h)\rangle_{\tau}e_{p,\nu}=o_p((nh)^{-3/2})$. Therefore, for $\nu = 0,\ldots,p$, $\sup_{t\in[0,\tau]}\lvert(nh)^{1/2}e_{p,\nu}^{{\rm t}}r_{g,p,n}(t,h)\lvert\, = o_p(1)$, by Lenglart{'}s inequality jacod2003limit. The predictable quadratic variation of $(nh)^{1/2}\xi_{g,p,n}$ with respect to $\mathcal{F}_{t}$ is \begin{equation} \langle \xi_{g,p,n}(\cdot,h),\xi_{g,p,n}(\cdot,h)\rangle_t = \frac{1}{n}\int_0^t \widetilde{\Gamma}_{g,p}(s)^{-1} \bar{\Psi}_{g,p,n}(s,h) \widetilde{\Gamma}_{g,p}(s)^{-1}\,{\rm d} s, \notag \end{equation} where $\bar{\Psi}_{g,p,n}(t,h) = n^{-1}\sum_{i=1}^n I_{X_i = g} K_h(Z_i)^2 r_{p}(Z_i/h)r_{p}(Z_i/h)^{{\rm t}} Y_i^g(t)\bar{\alpha}_g(t,Z_i)$. By Lemma (ref)(viii) in the appendix $h\bar{\Psi}_{1,p,n}(t,h) \to_p y_1(t,z_0)\bar{\alpha}_1(t,z_0)f_{Z}(z_0)\Psi_p$ and $h\bar{\Psi}_{0,p,n}(t,h) \to_p y_0(t,z_0)\bar{\alpha}_0(t,z_0)f_{Z}(z_0)H_p(-1)\Psi_pH_p(-1)$ uniformly in $t$ as $nh \to \infty$ and $h \to 0$. We conclude that \begin{equation} \begin{split} & \frac{1}{nh}\int_0^t\widetilde{\Gamma}_{g,p}(s)^{-1}h\bar{\Psi}_{g,p,n}(s,h)\widetilde{\Gamma}_{g,p}(s)^{-1} \,{\rm d} s \\ & \qquad\qquad = \frac{1}{nh}\big\{\frac{1}{f_Z(z_0)} \int_0^t \frac{\bar{\alpha}_{g}(s,z_0)}{y_{g}(s,z_0)}\,{\rm d} s \,\Gamma_{p}^{-1}\Psi_{p}\Gamma_{p}^{-1} + o_p(1)\big\}, \end{split} \end{equation} as $nh \to \infty$ and $h \to 0$, for all $t$. Next, we show that $(nh)^{1/2}\xi_{g,p,n}$ satisfies a Lindeberg condition. Write $H_{g,p,i}^n(t,h) = (nh)^{-1/2}\widetilde{\Gamma}_{g,p}(s)^{-1}I_{X_i = g}K(Z_i/h)r_p(Z_i/h)$ so that $\xi_{g,p,n}(t,h) = \sum_{i=1}^n\int_0^t H_{g,p,i}^n(s,h)\,{\rm d} \bar{M}_i^g(s)$. Note that for each $i$ and $t$ and $\nu$, \begin{equation} \lverte_{p,\nu}^{{\rm t}}H_{g,p,i}^n(t,h)\lvert^2\, \leq (nh)^{-1/2}\,\lVert\widetilde{\Gamma}_{g,p}(t)^{-1}\rVert^2 \lVertK(Z_i/h)r_p(Z_i/h)\rVert^2, \notag \end{equation} by Lemma (ref) in the appendix. Now, $\lVertK(Z_i/h)r_p(Z_i/h)\rVert\, \leq K_{\max}(\sum_{\nu = 0 }^{p}\kappa^{2\nu})^{1/2}$, where $K_{\max} = \sup_{u}K(u)$ is bounded by Assumption (ref). Furthermore, $\lVert\widetilde{\Gamma}_{g,p}(s)^{-1} \rVert\, = \lverty_g(s,z_0)f_Z(z_0)\lvert^{-1}\,\lVert\Gamma_{p}^{-1}\rVert$, by Assumption (ref) we have that $y_g(t,z_0) \geq y_g(\tau,z_0) = \{1 - H(\tau)\}\bar{S}_g(\tau,z_0) > 0$, and by Assumption (ref)(a) that $f_{Z}(z_0) > 0$. Therefore, \begin{equation} \sup_{s\in [0,\tau]} \lVert\widetilde{\Gamma}_{g,p}(s)^{-1} \rVert \leq \lverty_g(\tau,z_0)f_Z(z_0)\lvert^{-1}\,\lVert\Gamma_{p}^{-1}\rVert \eqqcolon C. \notag \end{equation} This shows that \begin{equation} \max_{\nu = 0,\ldots,p}\sup_{t \in [0,\tau]} \lverte_{p,\nu}^{{\rm t}}H_{g,p,i}^n(t,h)\lvert\, \leq \frac{C K_{\max}(\sum_{\nu = 0 }^{p}\kappa^{2\nu})^{1/2}}{(nh)^{1/2}}, \notag \end{equation} and, consequently, for all $\nu = 0,\ldots,p$, we have that for any $\varepsilon > 0$, \begin{equation} \int_0^t \sum_{i=1}^n \,\lverte_{p,\nu}^{{\rm t}}H_{g,p,i}^n(s,h)\lvert^2\, I\{\lverte_{p,\nu}^{{\rm t}}H_{g,p,i}^n(s,h)\lvert\, \geq \varepsilon\}Y_i^g(s)\bar{\alpha}_g(s,Z_i)\,{\rm d} s \overset{p}\to 0 , \notag \end{equation} for all $t \in [0,\tau]$ as $nh \to \infty$. That is, the Lindeberg condition holds. By the Rebolledo type central limit theorem in andersen1982cox, this entails that $(nh)^{1/2}\xi_{g,p,n} \Rightarrow \bar{M}_{g,p}$ as $nh \to \infty$ and $h \to 0$. Since \begin{equation} (nh)^{1/2}\bar{M}_{g,p,n} = (nh)^{1/2}\xi_{g,p,n} + (nh)^{1/2}r_{g,p,n}, \notag \end{equation} and $(nh)^{1/2}r_{g,p,n} \to_p 0$ uniformly in $t$, Lemma VI.3.31 in jacod2003limit yields $(nh)^{1/2}\bar{M}_{g,p,n} \Rightarrow \bar{M}_{g,p}$ as $nh \to \infty$ and $h \to 0$. Because $\bar{M}_{0,p,n}$ and $\bar{M}_{1,p,n}$ are orthogonal, joint convergence follows.

From the above theorem it is seen that the predictable quadratic variation of $\bar{M}_{g,p,n}$ is $\langle \bar{M}_{p,n}(\cdot,h),\bar{M}_{p,n}(\cdot,h)\rangle_t = O_p((nh)^{-1})$, so, in particular

equation[equation omitted — 151 chars of source]

Moreover, from Lemma (ref) the bias is seen to be

equation[equation omitted — 150 chars of source]

This shows that the mean squared error optimal bandwidth is $h = c n^{1/(2p+3)}$ for some constant $c > 0$. Combining the bias lemma with Theorem (ref), we get the following corollary.

corollaryThe condtions of Lemma (ref) hold. As $nh \to \infty$ and $h \to 0$, if $nh^{2p + 3} \to c $ for a constant $c \geq 0$, then \begin{equation} (nh^{2\nu + 1})^{1/2}\{\widehat{A}_{g,p}^{(\nu)}(\cdot,h) - \int_0^{\cdot}\bar{\alpha}_g^{(\nu)}(s,z_0)\,{\rm d} s \} \Rightarrow e_{p,\nu}^{{\rm t}}\nu!(\bar{M}_{g,p} + c^{1/2}\mathfrak{B}_{g,p}), \notag \end{equation} for $g = 0,1$, with $\mathfrak{B}_{g,p}$ as defined in (ref).
proofBy Lemma (ref) and the decomposition in (ref), \begin{equation} \begin{split} & \widehat{A}_{g,p}^{(\nu)}(t,h) - \int_0^t \bar{\alpha}_g^{(\nu)}(s,z_0)\,{\rm d} s = e_{p,\nu}^{{\rm t}}\nu! h^{-\nu} \bar{M}_{g,p,n}(t,h) + e_{p,\nu}^{{\rm t}}\nu! h^{p+1 - \nu} \mathfrak{b}_{g,p}(t)\\ & \qquad\qquad\qquad + \int_0^t \{J_{n,h}(s) - 1\}\bar{\alpha}_g(s,z_0)\,{\rm d} s + O_p(h^{p + 2 - \nu}). \end{split} \notag \end{equation} Upon multiplying $ \widehat{A}_{g,p}^{(\nu)}(t,h) - \int_0^t \bar{\alpha}_g^{(\nu)}(s,z_0)\,{\rm d} s$ by $(nh^{2\nu + 1})^{1/2}$, the first term on the right is $e_{p,\nu}^{{\rm t}}\nu! (nh)^{1/2}\bar{M}_{g,p,n}(t,h)$. By Theorem (ref) and the Cram{\'e}r--Wold device, $e_{p,\nu}^{{\rm t}}\nu! (nh)^{1/2}\bar{M}_{g,p,n}(\cdot,h)\Rightarrow e_{p,\nu}^{{\rm t}}\nu! \bar{M}_{g,p}$ as $nh \to \infty$ and $h \to 0$, where $\bar{M}_{g,p}$ is the Gaussian martingale of said theorem. For the second term on the right, i.e., the bias term, $e_{p,\nu}^{{\rm t}}\nu! (nh^{2p+3})^{1/2}\mathfrak{b}_{g,p}(t)\to e_{p,\nu}^{{\rm t}}\nu! c^{1/2}\mathfrak{b}_{g,p}(t)$ as $nh^{2p+3} \to c \geq 0$, uniformly in $t$ since $\mathfrak{b}_{g,p}(t)$ is bounded, see Assumption (ref)(b). For the third term on the right, we can, by Lemma (ref)(iii) in the appendix, find $n_0$ and $h_0$ such that $\sup_{t \in [0,\tau]}\lvertJ_{n,h}(t) - 1\lvert\, = 0$ a.s., for all $n \geq n_0$ and $h \leq h_0$. Thus, for all $n \geq n_0$ and $h \leq h_0$, $(nh^{2\nu + 1})^{1/2}\sup_{t\in [0,\tau]}\lvert\int_0^t \{J_{n,h}(s) - 1\}\bar{\alpha}_g(s,z_0)\,{\rm d} s\lvert \, \leq (nh^{2\nu + 1})^{1/2}\sup_{t\in [0,\tau]}\{J_{n,h}(s) - 1\}\int_0^{\tau}\bar{\alpha}_g(s,z_0)\,{\rm d} s = 0$ a.s., using that $\bar{\alpha}_g(t,z_0)\geq 0$. Finally, $(nh^{2\nu + 1})^{1/2}O_p(h^{p+2 - \nu}) = O_p((nh^{2p+5})^{1/2}) \to 0$ as $nh^{2p+3} \to c$, and this convergence is uniform in $t$ by Lemma (ref). Since $\mathfrak{b}_{g,p}(t)$ is continuous in $t$, the claim follows from Proposition VI.3.17 and Lemma VI.3.31 in jacod2003limit.

The two results above, Theorem (ref) and Corollary (ref), pertain to the sequences of $\mathcal{F}_t$-martingales $H_p(h)\bar{M}_{g,p,n} = \widehat{B}_{g,p}(t,h) - \int_{0}^t {\rm E}\,\{{\rm d}\widehat{B}_{g,p}(s,h)\mid \mathcal{F}_{s-}\}$. Considering these martingales essentially means that, in the central limit theorem, we average out the confounders but condition on the forcing variable.\footnote{There is a parallel here to the difference between the observed information and the Fisher information in likelihood inference for regression models, where, in the latter, the covariates are averaged out.} We now turn to a central limit theorem relative to the filtration $\mathcal{G}_t = \mathcal{E}_t \vee \mathcal{X}$, that is, the filtration with respect to which the confounder is measurable. Because the $\mathcal{F}_t$-martingales $\bar{M}_{g,p,n}$ analysed above are {\it not} martingales with respect to $\mathcal{G}_t$, these results are slightly more involved. Consider the decomposition

equation[equation omitted — 222 chars of source]

where ${\rm Bias}_{g,p,n}(t,h)$ is as defined in (ref); $M_{g,p,n}(t,h)$ is the $\mathcal{G}_t$-martingale

equation[equation omitted — 180 chars of source]

see Lemma (ref), and $L_{g,p,n}(t,h)$ is

equation[equation omitted — 109 chars of source]

where $Q_{g,p,n}(s,h)$ is the average of i.i.d. random variables given by

equation[equation omitted — 156 chars of source]

with $\Delta_{g,i}(t)$ being the difference between true hazard and the conditional hazards defined in (ref), that is

equation[equation omitted — 120 chars of source]

It turns out that the $\mathcal{G}_t$-martingales $M_{g,p,n}$ have the same limiting variance as the $\mathcal{F}_t$-martingales $\bar{M}_{g,p,n}$, however, the sequence $L_{g,p,n}$ is of the same order as the martingales, and thus contribute to the asymptotic variance of the estimator sequence. One might therefore view the variance stemming from $L_{g,p,n}$ as extra variance induced by not averaging out the confounder $U$. For the next theorem, in addition to Assumptions (ref)--(ref), we impose a Lipschitz condition and a boundedness condition on the hazard rate functions of the potential lifetimes, as well as on the conditonal hazard rate functions. These assumptions are likely much stronger than necessary.

assumptionThere are constants $\ell_g$, $\bar{\ell}_g$, $\alpha_{g,\max}$, and $\bar{\alpha}_{g,\max}$ for $g = 0,1$ such that $\alpha_g(t,z,u) \leq \alpha_{g,\max}$, $\alpha_g(t,z) \leq \bar{\alpha}_{g,\max}$ for all $t,z,u$, and $\lvert\alpha_g(t,z,u) - \alpha_g(s,z,u)\lvert\, \leq \ell_g\, \lvertt - s \lvert$ for all $t,s \in [0,\tau]$ and all $z,u$, and $\lvert\bar{\alpha}_g(t,z) - \bar{\alpha}_g(s,z)\lvert\, \leq \bar{\ell}_g\, \lvertt - s \lvert$ for all $t,s \in [0,\tau]$ and all $z$.
theorem{\sc($\mathcal{G}$-clt)} Assumption (ref)(b) holds with $S \geq p+2$, Assumption (ref) is in force, and the distribution function $H(t)$ of $C$ has density $h(t)$ that is bounded on $[0,\tau]$. If $nh^{2p+3} \to c$ for some $c \geq 0$ as $nh \to \infty$ and $h \to 0$, then \begin{equation} (nh^{2\nu + 1})^{1/2}\{\widehat{A}_{g,p}^{(\nu)}(\cdot,h) - \int_0^{\cdot}\bar{\alpha}_g^{(\nu)}(s,z_0)\,{\rm d} s\} \Rightarrow e_{p,\nu}^{{\rm t}}\nu!(M_{g,p} + L_{g,p} + c^{1/2} \mathfrak{B}_{g,p}), \notag \end{equation} where $M_{g,p}$ and $L_{g,p}$ are $(p+1)$-dimensional mean zero Gaussian processes, with $(M_{0,p},L_{0,p})$ and $(M_{1,p},L_{1,p})$ independent, and finite-dimensional distributions characterised by \begin{equation} {\rm Var}\,M_{g,p}(t) = \frac{1}{f_{Z}(z_0)}\int_0^t\frac{\bar{\alpha}_{g}(s,z_0)}{y_{g}(s,z_0)}\,{\rm d} s\,\Gamma_{p}^{-1}\Psi_{p}\Gamma_p^{-1}, \notag \end{equation} and \begin{equation} {\rm Cov}\{M_{g,p}(s),L_{g,p}(t) \} = -\frac{1}{f_Z(z_0)}\int_0^t \int_0^{u \wedge s} \frac{c_g(x,u \vee s,z_0)}{y_g(x,z_0)y_{g}(u,z_0)}\, {\rm d} x\,\Gamma_{p}^{-1}\Psi_{p}\Gamma_p^{-1}, \notag \end{equation} and \begin{equation} {\rm Cov}\{L_{g,p}(s),L_{g,p}(t) \} = \frac{1}{f_Z(z_0)} \int_0^s \int_0^t \frac{y_g(u\vee v,z_0)c_g(u,v,z_0)}{y_g(u,z_0)y_g(v,z_0)}\,{\rm d} u \,{\rm d} v \,\Gamma_{p}^{-1}\Psi_{p}\Gamma_p^{-1}, \notag \end{equation} where $c_{g}(u,s,z) = {\rm E}\,(\Delta_{g}(u)\Delta_{g}(s)\mid Z = z,Y(u\vee s ) = 1)$; and the bias term $\mathfrak{B}_{g,p}$ is as defined in (ref).
proofThe proof is in Appendix (ref).
remarkThere is an interesting affinity between Theorem (ref) of the present paper and Theorem 2.1 in hjort1992inference. Hjort{'}s central limit theorem pertains to parametric hazard rate functions under model misspecification. In that theorem, an extra term corresponding to the $L_{g,p,n}$ appears. This term is due to the difference $\alpha(t,\theta_{\rm lf}) - \alpha_{0}(t)$ between the parametric model $\alpha(t,\theta)$, evaluated in the least false parameter value $\theta_{\rm lf}$, and the true hazard $\alpha_{0}(t)$, that is the hazard under which $N(t) - \int_0^t Y(s) \alpha(s)\,{\rm d} s$ is a martingale. In our case it is the difference $\alpha_{g}(t,Z_i,U_i) - \bar{\alpha}_{g}(t,Z_i)$ that cause the extra term to appear, but this difference is a close nonparametric relative of the difference studied by Hjort. The misspecification in Theorem (ref) occurs because the $\mathcal{F}$-hazard $\bar{\alpha}_{g}(t,Z_i)$ is a form of model misspecifiation when the analysis takes place with respect to the filtration $\mathcal{G}_t$.

For the results in the remainder of the paper, we use the $\mathcal{F}$-central limit theorem in Theorem (ref). There are two reasons for this. First, estimating the limiting variance of the martingale is more straightforward than estimating the variances and covariances appearing in Theorem (ref). Second, the common approach in the regression discontinuity literature is to condition on the forcing variable, and average out all other covariates (see the discussion in the Section (ref)).

Variance estimation

The probability limit of the process $nh\langle \bar{M}_{g,p,n}(\cdot,h),\bar{M}_{g,p,n}(\cdot,h)\rangle_t$ is $\langle \bar{M}_{g,p},\bar{M}_{g,p}\rangle_t$, for which an expression is given in (ref). Consistent estimators for the variance processes $e_{p,\nu}^{{\rm t}}H_p(h_n)\langle \bar{M}_{g,p},\bar{M}_{g,p}\rangle_tH_p(h_n)e_{p,\nu}$ can be developed along the lines of the standard variance estimator for the variance of the estimator of the cumulative regression coefficients in the Aalen additive hazards model (see, for example, hjort2021partly). One such estimator is

equation[equation omitted — 331 chars of source]
lemmaThe condtions of Lemma (ref) hold. As $nh \to \infty$ and $h \to 0$, \begin{equation} V_{g,p,n}(t,h) \overset{p}\to \langle \bar{M}_{g,p},\bar{M}_{g,p}\rangle_t \notag \end{equation} for each $t \in [0,\tau]$.
proofAssume that $z_0 = 0$. Using the martingale decomposition in Lemma (ref), we see that $V_{g,p,n}(t,h) = nh \langle \bar{M}_{g,p,n}(\cdot,h),\bar{M}_{g,p,n}(\cdot,h)\rangle_t + r_{g,p,n}(t,h)$. For $\nu = 0,\ldots,p$, the $\nu$th element of $r_{g,p,n}(t,h)$ is \begin{equation} e_{p,\nu}^{{\rm t}}r_{g,p,n}(t,h) = \frac{h}{n}\sum_{i=1}^n I_{X_i=g} K_h(Z_i)^2 \int_0^tJ_{n,h}(s) \{e_{p,\nu}^{{\rm t}}\Gamma_{g,p,n}(s,h)^{-1} r_{p}(Z_i/h) \}^{2}\,{\rm d} \bar{M}_i^g(s). \notag \end{equation} From Lemma (ref) we get the bound \begin{equation} \big(e_{p,\nu}^{{\rm t}}\Gamma_{g,p,n}(s,h)^{-1} r_{p}(Z_i/h) \big)^{2} \leq \lVert\Gamma_{g,p,n}(s,h)^{-1}\rVert^2\lVertr_{p}(Z_i/h)\rVert^2, \notag \end{equation} so the predictable quadratic variation of $e_{p,\nu}^{{\rm t}}r_{g,p,n}(t,h)$, relative to $\mathcal{F}_t$, is \begin{equation} \begin{split} & e_{p,\nu}^{{\rm t}}\langle r_{g,p,n}(\cdot,h),r_{g,p,n}(\cdot,h) \rangle_t e_{p,\nu}^{{\rm t}}\\ & \; = \frac{h^2}{n^2}\sum_{i=1}^n I_{X_i=g} K_h(Z_i)^4 \int_0^t J_{n,h}(s)\{e_{p,\nu}^{{\rm t}}\Gamma_{g,p,n}(s,h)^{-1} r_{p}(Z_i/h) \}^{4}Y_i^g(s)\bar{\alpha}_g(s,Z_i)\,{\rm d} s\\ & \; \leq \frac{1}{(nh)^2}\sum_{i=1}^n I_{X_i=g} K(Z_i/h)^4 \int_0^t \lVert\Gamma_{g,p,n}(s,h)^{-1}\rVert^4\lVertr_{p}(Z_i/h)\rVert^4Y_i^g(s)\bar{\alpha}_g(s,Z_i)\,{\rm d} s, \end{split} \notag \end{equation} By Lemma (ref)(i)--(iii) and Lemma (ref) in the appendix $\lVertJ_{n,h}(s)\Gamma_{1,p,n}(s,h)^{-1}\rVert$ converges in probability to $\lVerty_g(s,z_0)f_Z(z_0)\Gamma_{1}(s)^{-1}\rVert$ and $\lVertJ_{n,h}(s)\Gamma_{0,p,n}(s,h)^{-1}\rVert$ converges in probability to $\lVerty_g(s,z_0)f_Z(z_0)H_p(-1)\Gamma_{1}(s)^{-1}H_p(-1)\rVert$, uniformly in $s \in [0,\tau]$ as $nh \to \infty$ and $h \to 0$. Both limits are bounded by $\lverty_g(\tau,z_0)f_Z(z_0)\lvert^{-1}\,\lVert\Gamma_{p}^{-1}\rVert$ and $K(z/h) \lVertr_{p}(z/h)\rVert \leq K(z/h) \sum_{\nu = 0}^p\kappa^{2\nu}$, as argued in the proof of Theorem (ref). Therefore, \begin{equation} e_{p,\nu}^{{\rm t}}\langle r_{g,p,n}(\cdot,h),r_{g,p,n}(\cdot,h) \rangle_t e_{p,\nu} \lesssim (nh)^{-2}\sum_{i=1}^n K(Z_i/h)\int_0^t Y_i^g(s)\bar{\alpha}_g(s,Z_i)\,{\rm d} s. \notag \end{equation} For $h \kappa < \kappa_0$, the conditional hazard $\bar{\alpha}_g(t,z)$ is bounded and $\int_0^{\tau} Y_i(s)\bar{\alpha}_g(s,Z_i)\,{\rm d} s = O_p(1)$ for all $i$. Moreover, $(nh)^{-1}\sum_{i=1}^n K(Z_i/h)$ is a sum of i.i.d. random variables with finite variance (Assumption (ref)) converging in probability to $\int_{-\kappa}^{\kappa} K(u) f_Z(hu)\,{\rm d} u$ which is finite provided $h \kappa < \kappa_0$. This gives that $e_{p,\nu}^{{\rm t}}\langle r_{g,p,n}(\cdot,h),r_{g,p,n}(\cdot,h) \rangle_t e_{p,\nu} = O_p((nh)^{-1}) = o_p(1)$ as $nh \to \infty$. By Lenglart{'}s inequality we now conclude that $e_{p,\nu}^{{\rm t}} r_{g,p,n}(t,h) = o_p(1)$ for each $\nu$, and therefore $r_{g,p,n}(t,h) = o_p(1)$.

This means that a consistent estimator of the limiting variance of the sequence $(nh^{2\nu + 1})^{1/2}\{\widehat{A}_{g,p}^{(\nu)}(t,h) - \int_0^{t}\bar{\alpha}_g^{(\nu)}(s,z_0)\,{\rm d} s \}$ of Corollary (ref) is $h^{2\nu} (\nu!)^2 e_{p,\nu}^{{\rm t}} V_{g,p,n}(t,h)e_{p,\nu}$. In particular, with $\Phi(z)$ the standard normal cumulative distribution function,

equation[equation omitted — 144 chars of source]

are approximate $(1 - \alpha)100$ percent pointwise confidence intervals provided $nh^{2p+3} \to 0$. If $nh^{2p+3} \to c$ for $c>0$, these confidence intervals are not valid due to the bias term appearing in the limit in Corollary (ref). The topic of the next section is how this may be fixed.

Bias correction

In this section we follow the conventional bias correction approach in local polynomial regression (see, e.g., fan1996local) and study the estimators given by

equation[equation omitted — 199 chars of source]

where

equation[equation omitted — 254 chars of source]

Here $b$ is a so-called pilot bandwidth sequence, typically larger than $h$ (because it is harder to estimate the $(p+1)$th derivative than the $\nu$th derivative when $\nu \leq p$, as it is here). In (ref) the second term on the right is an estimator of the bias term appearing in the limiting distribution of Corollary (ref). Under certain conditions on the bandwidth sequences $h$ and $b$ made precise below, subtracting off a bias estimate removes the asymptotic bias term from the limiting distribution in Corollary (ref), even when $nh^{2p+3}$ tends to a positive constant. In particular, the mean squared error optimal bandwidth $h \propto n^{-1/(2p+3)}$ may be employed, and the limiting martingale processes are the same as those given in said corollary.

As pointed out by calonico2014robust, this large-sample approximation relies on the condition $h/b \to 0$ as $h,b \to 0$, which makes the variability of the bias correction estimate disappear. That is, provided $h/b \to 0$, the estimator $\widehat{\Theta}_{g,q}^{(p+1)}(t,b)$ in (ref) does not contribute to the limiting variance of $\widehat{\Theta}_{g,p,q}^{(\nu),{\rm bc}}(t,h,b)$. Since $h/b$ is never zero in finite samples, the idea of calonico2014robust is to remove this requirement, and instead let $h/b \to \rho > 0$ as $h,b\to 0$, and thereby get limiting distributions of the estimator sequence where the variability of the bias estimator is accounted for. These ideas are formalised in Corollary (ref) below.

By Lemma (ref) and using the decomposition in (ref) twice, we obtain

equation[equation omitted — 627 chars of source]

From Lemma (ref)(iii) in the appendix, this approximation holds with probability one when $n,h,b$ are so that $J_{n,h}(t) = 1$ and $J_{n,b}(t) = 1$ for all $t \in [0,\tau]$, thus ensuring that $(nh^{2\nu+1})^{1/2}\int_0^t \{J_{n,h}(s) - 1\}\bar{\alpha}_g^{(\nu)}(s,z_0)\,{\rm d} s\ + (nh)^{1/2}h^{p+1}e_{p,\nu}\nu! \kappa_p \int_0^t \{J_{n,b}(s) - 1\}\bar{\alpha}_g^{(p+1)}(s,z_0)\,{\rm d} s = 0$ for all $t$, almost surely.

corollaryThe condtions of Lemma (ref) hold. Assume that $\bar{\alpha}_g(t,z)$ for $g = 0,1$ are at least $q+2 \geq p + 3$ times continuously differentiable in $z$ for all $t$; that $n \min(h_n,b_n) \to \infty$, $\max(h_n,b_n) \to 0$, $h_n/b_n \to \rho \geq 0$ and $nh_n^{2p+3}\max\{h_n^2,b_n^{2(q - p)}\}$ tends to zero as $n \to \infty$, and that $h_n \leq b_n$ for all $n$. Then \begin{equation} (nh_n^{2\nu + 1})^{1/2}\{\widehat{\Theta}_{p,q}^{(\nu),{\rm bc}}(\cdot,h,b) - \Theta^{(\nu)}(\cdot,z_0)\} \Rightarrow e_{p,\nu}^{{\rm t}}\nu! (\bar{M}_{1,p}^{(\rho)} - \bar{M}_{0,p}^{(\rho)} ), \notag \end{equation} where $\bar{M}_{0,p}^{(\rho)}$ and $\bar{M}_{1,p}^{(\rho)}$ are orthogonal mean zero Gaussian martingales with variation processes \begin{equation} \langle \bar{M}_{g,p}^{(\rho)},\bar{M}_{g,p}^{(\rho)} \rangle_t = \langle \bar{M}_{g,p},\bar{M}_{g,p}\rangle_t + \rho^{2p+3} V_{g,p,q}(t) - 2\rho^{p+2} C_{g,p,q}(t), \notag \end{equation} with \begin{equation} V_{g,p,q}(t) = \Gamma_p^{-1}\vartheta_{p,p+1} e_{q,p+1}^{{\rm t}} \langle \bar{M}_{g,q},\bar{M}_{g,q} \rangle_t e_{q,p+1}\vartheta_{p,p+1}^{{\rm t}}\Gamma_{p}^{-1}, \notag \end{equation} and \begin{equation} C_{g,p,q}(t) = \langle \bar{M}_{g,p},\bar{M}_{g,q}\rangle_t e_{q,p+1} \vartheta_{p,p+1}^{{\rm t}}\Gamma_p^{-1}, \notag \end{equation} where $\langle \bar{M}_{g,r},\bar{M}_{g,r} \rangle_t$ for $r = p,q$ are as defined in (ref), while $\langle \bar{M}_{g,p},\bar{M}_{g,q}\rangle_t$ is the $(p+1)\times (q+1)$ matrix \begin{equation} \langle \bar{M}_{g,p},\bar{M}_{g,q}\rangle_t = \frac{1}{f_{Z}(z_0)} \int_0^1 \frac{\bar{\alpha}_g(s,z_0)}{y_g(s,z_0)}\,{\rm d} s \,\Gamma_{p}^{-1}\Psi_{p,q}(\rho) \Gamma_q^{-1}, \notag \end{equation} where $\Psi_{p,q}(\rho) = \int_0^{\infty}K(u)K(\rho u) r_{p}(u)r_q(\rho u)^{{\rm t}}\,{\rm d} u$.
proofIt suffices to look at one of the sides of the cut-off. Define the $\mathcal{F}$-martingales \begin{equation} \bar{M}_{g,p,q,n}(t,h_n,b_n) = \bar{M}_{g,p,n}(t,h_n) - (h_n/b_n)^{p+1}\Gamma_{p}^{-1}\vartheta_{p,p+1} e_{q,p+1}^{{\rm t}} \bar{M}_{g,q,n}(t,b_n), \notag \end{equation} so that $\bar{M}_{g,p,q,n}(t,h_n,b_n)$ is the martingale in the curly brackets in (ref). The predictable quadratic variation of $(nh_n)^{1/2}\bar{M}_{g,p,q,n}(t,h_n,b_n)$ is \begin{equation} \begin{split} & nh_n \langle \bar{M}_{g,p,q,n}(\cdot ,h_n,b_n),\bar{M}_{g,p,q,n}(\cdot ,h_n,b_n)\rangle_t = nh_n\langle \bar{M}_{g,p,n}(\cdot,h_n),\bar{M}_{g,p,n}(\cdot,h_n)\rangle_t\\ & \; + (nh_n)(h_n/b_n)^{2(p+1)} \Gamma_{p}^{-1}\vartheta_{p,p+1}e_{q,p+1}^{{\rm t}} \langle \bar{M}_{g,q,n}(\cdot,b_n),\bar{M}_{g,q,n}(\cdot,b_n)\rangle_t e_{q,p+1}\vartheta_{p,p+1}^{{\rm t}}\Gamma_{p}^{-1}\\ &\; - 2(nh_n) (h_n/b_n)^{p+1} \langle \bar{M}_{g,p,n}(\cdot,h_n),\bar{M}_{g,q,n}(\cdot,b_n)\rangle_t e_{q,p+1} \vartheta_{p,p+1}^{{\rm t}}\Gamma_p^{-1}. \end{split} \notag \end{equation} From Theorem (ref), we get that $nh_n\langle \bar{M}_{g,p,n}(\cdot,h_n),\bar{M}_{g,p,n}(\cdot,h_n)\rangle_t \to_p \langle \bar{M}_{g,p},\bar{M}_{g,p}\rangle_t$ with an expression for this limit given in (ref), and also that the second term converges in probability to $\rho^{2p + 3}V_{g,p,q}(t)$. The third variation process in the third term is \begin{equation} \begin{split} &\langle \bar{M}_{g,p,n}(\cdot,h_n),\bar{M}_{g,q,n}(\cdot,b_n)\rangle_t\\ & \; = \frac{1}{n b_n}\int_0^t J_{n,h_n}(s) \Gamma_{g,p,n}(s,h_n)^{-1}b_n\bar{\Psi}_{g,p,q,n}(s,h_n,b_n)J_{n,b_n}(s)\Gamma_{g,q,n}(s,b_n)^{-1}, \end{split} \end{equation} where \begin{equation} \bar{\Psi}_{g,p,q,n}(s,h,b) = \frac{1}{n}\sum_{i=1}^n I_{X_i = g} K_{h}(Z_i)K_{b}(Z_i)r_{p}(Z_i/h)r_{q}(Z_i/b) Y_i^{g}(s)\bar{\alpha}_g(s,Z_i). \end{equation} By Lemma (ref)(xii) $b_n\bar{\Psi}_{1,p,q,n}(s,h_n,b_n) \to_p y_1(s,z_0)\bar{\alpha}_1(s,z_0)f_Z(z_0) \Psi_{p,q}(\rho)$ and (xiii) $b_n\bar{\Psi}_{0,p,q,n}(s,h_n,b_n) \to_p y_0(s,z_0)\bar{\alpha}_0(s,z_0)f_Z(z_0) H_{p}(-1)\Psi_{p,q}(\rho)H_{q}(-1)$, uniformly in $s \in [0,\tau]$. Combining this with Lemma (ref)(i)--(iii) and Lemma (ref), we have that the integrand in (ref) converges in probability, uniformly in $s \in [0,\tau]$, to $f_Z(z_0)^{-1}\{\bar{\alpha}_g(s,z_0)/y_g(s,z_0)\} \Gamma_{p}^{-1}\Psi_{p,q}(\rho) \Gamma_q^{-1}$, and $nb_n \langle \bar{M}_{g,p,n}(\cdot,h_n),\bar{M}_{g,q,n}(\cdot,b_n)\rangle_t \to_p \langle \bar{M}_{g,p},\bar{M}_{g,q}\rangle_t$. It now follows from Theorem (ref) that $(nh_n)^{1/2}\bar{M}_{g,p,q,n}(\cdot ,h_n,b_n) \Rightarrow \bar{M}_{g,p}^{(\rho)}$ for $g = 0,1$. Consider now second and the third term in (ref). The second term is a nonrandom continuous function, and $(nh_n)^{1/2} h_n^{p+1}b^{q - p} = (nh_n^{2p + 3} b_n^{2(q - p)})^{1/2}$ tends to zero by assumption. The third and final term is $O_p((nh^{2p+3})^{1/2}(h + b^{q-p})) = O_p((nh^{2p+3}\max\{h^2,b^{2(q-p)}\} )^{1/2}) = o_p(1)$ by assumption. Again, Proposition VI.3.17 and the Cram{\'e}r--Slutsky like Lemma VI.3.31 in jacod2003limit yield the result.

Concluding remarks

The ideas underlying designs such as the {{regression discontinuity design}}, the difference in difference design, and the intstrumental variable design are not bound to any particular estimand, model, or type of data. The statistical theory for these designs, however, are much more developed for the type of estimands, models, and data often encountered in economics, than for estimands, models, and data typically encountered in other fields of application. This paper is an attempt at taking one of these designs from the estimation of a conditional average treatment effect ({\sc cate}) based on uncensored data, to the estimation of a {{\sc cate}}-like object, namely the difference of two cumulative hazards, based on right-censored survival data.

A few directions the results of the present paper can be extended in are: First, the estimator developed in this paper can be used as a building block in the estimation of $\theta(t,z_0)$ using kernel smoothing techniques similar to those introduced by ramlau1983smoothing (research in this direction is underway). Second, the results of this paper is limited to the sharp {{regression discontinuity design}}, and ought to be extended to the fuzzy {{regression discontinuity design}}. Third, the estimator of this paper can be used to test whether the parameter of interest in a Cox regression model with unobserved confounders equals zero or not, but it can not provide asymptotically unbiased estimates of this parameter. Whether the {{regression discontinuity design}} can be used to identify the parameter of interest in proportional hazards models under confounding, can be studied.