Estimating Individual Dose-Response Curves under Unobserved Confounders from Observational Data
Extracted source text · arXiv 2410.15706
% This is samplepaper.tex, a sample chapter demonstrating the
% LLNCS macro package for Springer Computer Science proceedings;
% Version 2.20 of 2017/10/04
%
\documentclass[runningheads]{llncs}
%
\usepackage{array}
\usepackage{booktabs}
\usepackage{graphicx}
\usepackage{amsmath}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{amsfonts}
\usepackage{subfigure}
\usepackage{multirow}
\usepackage{enumitem}
\usepackage{verbatim}
\usepackage{xcolor}
\usepackage{hyperref}
\usepackage{makecell}
\usepackage{cite}
\usepackage{rotating}
\usepackage{tabularx}
\usepackage{adjustbox}
\usepackage{wrapfig}
\usepackage{appendix}
% \usepackage[square,sort,comma,numbers]{natbib}
% Used for displaying a sample figure. If possible, figure files should
% be included in EPS format.
%
% If you use the hyperref package, please uncomment the following line
% to display URLs in blue roman font according to Springer's eBook style:
% \renewcommand\UrlFont{\color{blue}\rmfamily}
\begin{document}
%
\title{Estimating Individual Dose-Response Curves under Unobserved Confounders from Observational Data}
%
\titlerunning{Individual Dose-Response Estimation with Hidden Confounders}
% If the paper title is too long for the running head, you can set
% an abbreviated paper title here
%
\author{Shutong Chen\inst{1} \and Yang Li\inst{1}}
%
\authorrunning{S. Chen, Y. Li}
% First names are abbreviated in the running head.
% If there are more than two authors, 'et al.' is used.
%
\institute{Shenzhen Key Laboratory of Ubiquitous Data Enabling, Tsinghua Shenzhen International Graduate School, Tsinghua University}
%\institute{Tsinghua-Berkeley Shenzhen Institute, Tsinghua University, China \\
%\email{\{cst21, yangli\}@mails.tsinghua.edu.cn}}%\\
%\url{http://www.springer.com/gp/computer-science/lncs} \\
%\email{yangli@mails.tsinghua.edu.cn}}%\{abc,lncs\}@uni-heidelberg.de}}
%
\maketitle % typeset the header of the contribution
%
\begin{abstract}
Estimating an individual's potential response to continuously varied treatments is crucial for addressing causal questions across diverse domains, from healthcare to social sciences. However, existing methods are limited either to estimating causal effects of binary treatments, or scenarios where all confounding variables are measurable. In this work, we present ContiVAE, a novel framework for estimating causal effects of continuous treatments, measured by individual dose-response curves, considering the presence of unobserved confounders using observational data. Leveraging a variational auto-encoder with a Tilted Gaussian prior distribution, ContiVAE models the hidden confounders as latent variables, and is able to predict the potential outcome of any treatment level for each individual while effectively capture the heterogeneity among individuals. Experiments on semi-synthetic datasets show that ContiVAE outperforms existing methods by up to 62\%, demonstrating its robustness and flexibility. %its superior performance compared to existing methods.
Application on a real-world dataset illustrates its practical utility.
\keywords{Causal inference \and Variational autoencoder \and Dose-response Curve.}
\end{abstract}
%
%
%
\input{1_intro}
\input{2_related_work}
\input{3_formulation}
\input{4_method}
\input{5_simulation_experiment}
\input{6_volunteer}
\section{Conclusions}
Existing methods for estimating unbiased causal effects are restricted either to treatments with binary or discrete values, or scenarios where all confounding variables are measurable. In this work, we propose ContiVAE, a novel framework enabling the estimation of individual dose-response curves in the presence of unobserved confounders. Taking advantage of a variational auto-encoder with a Tilted Gaussian prior distribution as well as an adjusted objective function, ContiVAE demonstrates robustness and flexibility in estimating individual dose-response curves across diverse circumstances, and captures individual differences effectively. In the experiments, we observed consistently the significant improvement of ContiVAE over various benchmark methods. Application to volunteer data also highlights its broader impact in real-world. For the future work, we consider investigating the theory of doubly robust~\cite{kennedy2017nondoublyrobust, bang2005doubly}, %further improvement of ContiVAE could be explored, such as
and refining ContiVAE into a doubly robust estimator for dose-response curves.
%\begin{theorem}
%This is a sample theorem. The run-in heading is set in bold, while
%the following text appears in italics. Definitions, lemmas,
%propositions, and corollaries are styled the same way.
%\end{theorem}
%
% the environments 'definition', 'lemma', 'proposition', 'corollary',
% 'remark', and 'example' are defined in the LLNCS documentclass as well.
%
%\begin{proof}
%Proofs, examples, and remarks have the initial word in italics,
%while the following text appears in normal font.
%\end{proof}
\section{Acknowledgement}
This work is supported in part by the Natural Science Foundation of China (Grant 62371270) and Shenzhen Key Laboratory of Ubiquitous Data Enabling (No.ZDSYS20220527171406015).
%
% ---- Bibliography ----
%
% BibTeX users should specify bibliography style 'splncs04'.
% References will then be sorted and formatted in the correct style.
%
\bibliographystyle{splncs04}
\bibliography{main}
%
%\begin{thebibliography}{8}
%\bibitem{ref_article1}
%Author, F.: Article title. Journal \textbf{2}(5), 99--110 (2016)
%\bibitem{ref_lncs1}
%Author, F., Author, S.: Title of a proceedings paper. In: Editor,
%F., Editor, S. (eds.) CONFERENCE 2016, LNCS, vol. 9999, pp. 1--13.
%Springer, Heidelberg (2016). \doi{10.10007/1234567890}
%\bibitem{ref_book1}
%Author, F., Author, S., Author, T.: Book title. 2nd edn. Publisher,
%Location (1999)
%\bibitem{ref_proc1}
%Author, A.-B.: Contribution title. In: 9th International Proceedings
%on Proceedings, pp. 1--2. Publisher, Location (2010)
%\bibitem{ref_url1}
%LNCS Homepage, \url{http://www.springer.com/lncs}. Last accessed 4
%Oct 2017
%\end{thebibliography}
\input{00_appendix}
\end{document}
\section{Introduction}
% background: most of studies focus on binary/discrete treatments, little attention has been paid to the continuous settings. In practice, many causal inquiries involves continuous-valued treatments. For instance, in medical field, but estimating individual dose-response curve is important.the challenge is ...
% existing methods also heavily rely on the unconfoundedness assumption, the problem is more challenging when there are unobserved confounders.
% existing methods address the continuous problem by ..., limitation is... for unobserved confounders, approaches are..., none of them address the problem simultaneously.
% our research problem is ..., to address the challenges, we proposed to extend CEVAE to continuous treatments. our method are different in ..., extensive experiments show that...
% our contribution is...
%( the rest of paper is organized as)
\begin{wrapfigure}[14]{r}{0.3\textwidth}
\vspace{-1.2cm}
\centering
\includegraphics[width=1.0\linewidth]{figures/dosecurve1.pdf}
\caption{\label{fig: dosecurve} An example of individual dose-response curves. Different colors denote curves of patients with various covariates $\mathbf{X}$ (e.g., age or gender).}
\end{wrapfigure}
While most of the literature and methods in causal inference focus on binary or discrete treatments~\cite{guo2020survey, yao2021survey, tarnet, cevae17, rosenbaum1987ipw, imai2014covariateCBPS, cider2022}, there has been relatively limited exploration in the continuous treatment settings. However, in practice, many causal inquiries involve continuous-valued treatments. For instance, in the medical field, making binary or discrete decision (e.g., whether to adopt a medicine or the type of treatment) is often inadequate, it is also critical to consider continuous-valued decisions, such as the dosage of the medicine, as dosage significantly impacts its effects. Moreover, the effects may vary considerably across patients with different conditions. Therefore, it is important to estimate the individual causal effects of continuous treatments, typically represented as individual dose-response curves (see Figure~\ref{fig: dosecurve}).
%compared to binary treatment scenarios where only one counterfactual outcome for each patient needs to be estimated,
%compared to binary treatment scenarios where only one counterfactual outcome for each patient needs to be estimated,
The fundamental problem in estimating causal effects from observational data lies in estimating the \textit{counterfactuals}, i.e., predicting what would happen if a patient received a different treatment, as well as in mitigating the spurious effects caused by \textit{confounders} \textendash \ variables that affects both the treatment and the outcome \textendash\ through properly adjusting for them. The nature of continuous treatments exacerbates the challenges~\cite{scigan2020, schwab2020learningdrnet, fong2018covariateCBGPS}: (1) they involve an infinite number of unobserved counterfactuals, unlike binary treatment scenarios where only one needs to be estimated, and the process of adjusting for confounders becomes more complex and less intuitive; (2) estimating individual dose-response curves requires the counterfactual estimator to effectively capture the individual differences. Furthermore, confounders are not always available in observational data, as factors like social status and living conditions are difficult to quantify or may not be recorded, making them \textit{unobserved confounders}. The presence of unobserved confounders adds complexity to the problem, as not all the confounders can be adjusted for to obtain unbiased causal effects.
To estimate causal effects of continuous treatments, one category of approaches extend methods for discrete treatments, including Generalized Propensity Score (GPS)~\cite{imai2004causalgps} and DRNets~\cite{schwab2020learningdrnet}. %generalizing the propensity scores~\cite{imai2004causalgps} and extending multi-task neural networks by dividing the range of treatment values into sub-intervals~\cite{tarnet, schwab2020learningdrnet}.
Another adopts generative networks to predict counterfactual outcomes, such as SCIGAN~\cite{scigan2020}. % and GANITE~\cite{yoon2018ganite}.
However, such methods rely on the ignorability assumption, which requires the observation and adjustment of all confounders. %Regarding unobserved confounders, common practices often utilize Instrumental Variables (IV) or proxy variables, which are measurable indicators correlated to the hidden confounders~\cite{cider2022, montgomery2000measuringproxies}, despite requiring strong assumptions about the underlying causal model or the structure of the proxies and hidden confounders. Alternatively,
Causal Effects Variational Auto-Encoder (CEVAE)~\cite{cevae17} tackles the challenges of unobserved confounders by leveraging deep latent-variable models to learn them as latent variables and approximately recover causal effects, but its architecture is restricted to binary treatments.% of binary treatments,
%with weaker assumptions about the structure of both the proxies and the unobserved confounders. Although these methods individually address the challenges posed by continuous treatments or hidden confounders, few can effectively tackle both.
%We propose a novel framework, ContiVAE, that utilizes a variational auto-encoder (VAE)
%We extend CEVAE~\cite{cevae17} which is tailored for binary treatments to continuous settings, and propose a novel framework named ContiVAE. To tackle the distinctive challenges posed by continuous treatments, ContiVAE significantly diverges from the architecture of CEVAE. Specifically, while CEVAE adopts multi-task networks similar to TARNET~\cite{tarnet}, which are trained separately to estimate the potential outcomes for treated and control patients respectively, our method utilizes shared multi-layer perceptron (MLP) networks that directly incorporate $T$ as part of the inputs. This allows for modeling complex function between the outcome, treatments and the proxies. %Since the model attempts to reconstruct $T$ in the decoder, its influence will be retained in the hidden layers to ensure accurate reconstruction.
In this work, we aim to address the problem of estimating individual dose-response curves in the presence of unobserved confounders. We introduce ContiVAE, a novel framework that conceives unobserved confounders as latent variables that influence observed data generation. ContiVAE utilizes a variational auto-encoder (VAE) to simultaneously model the distribution of hidden confounders in latent space and their relationships with the observed data (i.e., the covariates $\mathbf{X}$, treatment $T$ and outcome $Y$). We will later show that under specific conditions, individual dose-response curves can be estimated by deriving the joint probability $p(\mathbf{X}, \mathbf{Z}, T, Y)$. %Specifically, ContiVAE's encoder utilizes neural networks to approximate the posterior of the latent variables, mapping observed data to a latent space distribution. Conversely, the decoder attempts to reconstruct the observed data based on samples from the latent space, parameterizing $p(\mathbf{X}, T, Y | \mathbf{Z})$ with neural networks.
To effectively capture individual heterogeneity in dose-response curves, we assume a Tilted Gaussian distribution as the prior latent distribution instead of the commonly used Normal distribution~\cite{kingma2013autonormalvae}. % The Tilted Gaussian distribution is created by exponentially tilting the Gaussian distribution by its norm~\cite{floto2022tilted}.
Unlike the Normal distribution, where the maximum probability density concentrates at a single point, in the Tilted Gaussian distribution, it lies on the surface of a sphere, resulting in an exponentially larger volume %than that of the normal distribution
as the latent dimension increases. Consequently, points in the latent space can be better separated, and individual differences can be represented more effectively. Additionally, latent-variable models for causal inference tend to overly prioritize the reconstruction of high-dimensional or correlated $\mathbf{X}$, leading to incorrect estimation~\cite{rissanen2021criticalcevaefault}. %Finally,
To mitigate this, we introduce a scaling parameter $\lambda$ for the reconstruction loss of $\mathbf{X}$, to reduce its contribution to the total loss.
%Without $\lambda$, the model tends to overly prioritize the reconstruction of $\mathbf{X}$ while neglecting the reconstruction of other elements, i.e., the outcome and treatment, resulting in suboptimal performance. This phenomenon arises in real-world datasets where $\mathbf{X}$ is highly dimensional, and is also observed in~\cite{rissanen2021criticalcevaefault} when high correlations exist within $\mathbf{X}$.
Our work draws inspiration from CEVAE~\cite{cevae17}, which investigates the application of deep latent-variable models in causal inference for binary treatments, but we make significant modifications to tackle the distinct challenges posed by continuous treatments and mitigate some practical issues such as its tendency to overemphasize covariate reconstruction.
Extensive experimental results show that ContiVAE consistently outperforms several baselines and state-of-the-art models by up to 62\% across various scenarios, demonstrating its robustness and flexibility. Additionally, we apply ContiVAE to a real-world dataset tracking volunteer activities, to estimate the effects of a volunteer's social relationships on her retention. This illustrates how ContiVAE can effectively yield valuable insights to answer real-life causal inquiries.
Our contribution can be summarized as:
\begin{itemize}
\item We introduce a novel framework, ContiVAE, for estimating individual dose-response curves, filling the gap in identifying causal effects of continuous treatments in the presence of unobserved confounders.
\item Extensive experimental results demonstrate that our method achieves state-of-the-art performance in estimating individual dose-response curves across datasets with varying sizes and complexities.
\item The application of the proposed method to a real-world volunteer dataset shed light on the causal relationship of a volunteer's social network position on retention.
\end{itemize}
\section{Related Work}
\label{sec: relatedwork}
%\subsection{Causal Inference for Discrete Treatments}
%Most of methods developed in the causal inference literature focus on estimating causal effects of discrete treatments, especially binary treatments. For methods that rely on the three commonly used assumptions (SUTVA, Unconfoundedness and Positivity) including Inverse Propensity Weighting (IPW)~\cite{rosenbaum1987ipw}\cite{austin2011introductionpropensity}, TARnet~\cite{tarnet}, etc., we refer readers to a survey paper by%~\citeauthor{yao2021survey}~\cite{yao2021survey}. To cope with time-varying confounders, a common solution is to utilize recurrent neural networks (RNN) to capture the historical information of observational data~\cite{cider2022} \cite{timeseriesdeconfounder}. %~\cite{cider2022} proposed to utilize recurrent neural networks (RNN) to capture the historical information of covariates and learn the balanced confounder representations for treated and control units by minimizing the similarity of their distributions in the representation space.~\cite{timeseriesdeconfounder} also uses RNN with multi-task multi-perceptrons to build a factor model overtime and infer latent variables that render treatment assignments conditionally independent.
%re-weighting methods (e.g., Inverse Propensity Weighting (IPW)), matching methods (e.g., BNR), tree-based methods (e.g., Bayesian Additive Regression Trees (BART)), and deep representation learning methods such as TARnet, a popular model that learns representations of confounders regularized by IPM (Integral probability metric) to enforce the similarity between the distributions of different treatment groups in the representation space, and train separate counterfactual predictors using the learned representations.
\subsection{Causal Inference for Continuous Treatments}
%Propensity score plays an essential roles in the field of causal inference. Defined as the conditional probability of receiving a treatment given covariates, propensity score is utilized to adjust for observed confounders and estimate treatment effects through matching~\cite{abadie2016matchingpsm, rosenbaum1985psm}, weighting~\cite{rosenbaum1987modelpsweighting, hirano2003efficientpsweighting}, regression~\cite{heckman1998matchingregression} or their combination.
Many techniques for estimating continuous treatment effects extend propensity score-based methods to continuous settings and estimate the generalized propensity score (GPS)~\cite{imai2004causalgps, fong2018covariateCBGPS}. %Instead of modeling the treatment, another category of methods directly models the outcome by regressing it on the treatment and covariates. However, the performance of these models is highly sensitive to the specification of the outcome model or the treatment model. To address this challenge, a non-parametric
Addtionally, doubly robust frameworks combine propensity score with outcome regression models, mitigating the impact of model misspecification and %for continuous treatments have also been proposed~\cite{kennedy2017nondoublyrobust} to mitigate the issue of model misspecification of the propensity score, by combining propensity score with the outcome model,
ensuring unbiased estimation of average treatment effects when either of the model is correct.
%Traditional methods for estimating casual effects of continuous treatments extend methods for binary or discrete methods, including Generalized Propensity Score that extends the concept of propensity score, and
Recently, neural networks have been increasingly adopted in the field of causal inference due to their non-parametric nature and capacity to model complex functions. TARNET~\cite{tarnet}, a classic framework designed for discrete treatments, is consisted of a shared backbone and multi-task layers trained separately to predict the potential outcome of each treatment value. DRNets~\cite{schwab2020learningdrnet} extends TARNET to continuous treatments by dividing the dosage range of continuous treatments into subintervals and adding a multi-task head for each dosage interval. %enables estimating individual dose-response curves for multiple treatments with continuous dosages. In addition to the shared backbone and multi-task layers for treatments, it divides the dosage range of each treatment into $E$ equally-sized intervals and adding multi-task heads for each dosage interval.
%Instead of extending methods for discrete treatments,
Alternatively, SCIGAN~\cite{scigan2020} utilizes Generative Adversarial Networks (GAN) to model counterfactual distribution. Specifically, it trains a counterfactual generator and a discriminator adversarially, until the generator can produce counterfactual outcomes close enough to the true distribution to fool the discriminator.
\subsection{Causal Inference with Unobserved Confounders}
Most existing methods for estimating causal effects rely on the ignorability assumption (no unobserved confounders), and adjust for all measurable confounders to derive the unbiased estimates of treatment effects. %However, when there are unobserved confounders, treatment effects become unidentifiable without additional assumptions.
While Covariate Balancing Propensity Score (CBPS) (and its continuous extension)~\cite{imai2014covariateCBPS, fong2018covariateCBGPS} offers certain robustness without this assumption, it is limited to linear propensity model.
%Covariate Balancing Propensity Score (CBPS)~\cite{imai2014covariateCBPS} improves the covaraite balancing without relying on the ignorability assumption, and increases robustness of propensity score estimation under mild model misspecification. %does not require the ignorability assumption. As an enhancement of propensity score, CBPS improves the robustness of propensity score estimation under mild misspecification of the propensity score model. Regardless of the presence of unobserved confounders, it improve the balancing of the covariates.
%CBGPS~\cite{fong2018covariateCBGPS} extends CBPS to continuous settings, but is constrained by the linear assumption of the propensity model. % and will fail if the true treatment model is non-linear.
More common practices utilize proxy variables~\cite{cider2022, montgomery2000measuringproxies} \textendash\ measurable variables correlated to the unobserved confounders. While treating the proxies as ordinary confounders is incorrect and can lead to bias, other methods for unbiased estimation~\cite{pearl2012measurementpearlproxies} require strong assumptions about the structure of both the proxies and the hidden confounders (e.g., they are both categorical), which are often unrealistic in practical settings. %proposed a method to identify the unbiased treatment effects by first inferring the joint distribution $p(\mathbf{X}, \mathbf{Z})$ between the proxies $\mathbf{X}$ and the unobserved confounders $\mathbf{Z}$. However, such method require strong assumptions about the structure of $\mathbf{X}$ and $\mathbf{Z}$ (e.g., they should be categorical), which is typically untestifiable in practice.
Alternatively, Causal Effects Variational Auto-Encoder (CEVAE)~\cite{cevae17} proposed to model the hidden confounders as latent variables and use a Variational Auto-Encoder (VAE) to simultaneously discover the hidden confounders and how they affect treatment and outcome. By recovering the joint probability $p(\mathbf{X}, \mathbf{Z}, T, Y)$ of the hidden confounders, treatment, outcome and the observed covariates from observational data, the individual treatment effects become identifiable. However, the architecture of CEVAE is restricted to binary treatments, and few methods can address the challenges of unobserved confounders and continuous treatments simultaneously. %However, CEVAE is designed for binary treatments, and none of the aforementioned methods dealing with hidden confounders (except for CBGPS) is applicable for continuous treatments.
%For causal inference with unobserved confounders, a lot of methods use proxy variables and treat them as ordinary confounders, which rely on strong assumptions on the exact nature of hidden confounders and will induce bias, while~\cite{cevae17} proposed a model named CEVAE based on variational autoencoder (VAE) to estimate the latent space summarizing the confounders and the causal effects, which make weaker assumptions about the data generating process and the structure of the unobserved confounders.
%Compared to binary treatments, causal inference for continuous treatments is more challenging and understudied. Due to the continuous nature of treatments, there are uncountably many potential outcomes for per unit, nevertheless only one is observed for each unit or even across all units~\cite{bahadori2022end}. A majority of methods are build upon balancing methods for multiple treatments or utilize the generalized propensity score (GPS)~\cite{imai2004causalgps, imbens2000roledoseresponse}.~\cite{scigan2020} proposed a novel idea of using a modified generative adversarial networks (GAN) named SCIGAN that train a generator for predicting counterfactual outcomes, confronting a discriminator that can distinguish the pair of factual treatment and outcome from generated ones. %Although SCIGAN is reported to outperform existing methods such as GPS and GANITE~\cite{yoon2018ganite} on synthetic datasets, the response functions of real-world datasets are often very complicated, more likely resulting in overfitting and mode collapse. Furthermore, training of methods based on GAN may also suffer from problems like unstable or vanished gradients and highly sensitivity to hyperparameters~\cite{goodfellow2020generative}.
\section{Problem Formulation}
We assume a causal model of Figure~\ref{fig: causal_graph}. Unlike in~\cite{cevae17} where $T$ is a binary treatment, here $T$ takes continuous values, e.g., the dosage of a medication. $Y$ is an outcome, e.g., the recovery rate, $\mathbf{Z}$ is the unobserved confounders, e.g., living conditions, and covariates $\mathbf{X}$ are noisy proxies of $\mathbf{Z}$, e.g., blood pressures. Note that conditioned on $\mathbf{Z}$, $\mathbf{X}$ and $T$ is independent, so the joint probability can be factorized by $p(\mathbf{X}, \mathbf{Z}, T, Y) = p(\mathbf{Z}) p(\mathbf{X}|\mathbf{Z}) p(T|\mathbf{Z}) p(Y|\mathbf{Z}, T)$.
\begin{wrapfigure}[18]{r}{0.3\textwidth}
\vspace{-1cm}
\begin{center}
\includegraphics[width=0.25\textwidth]{figures/causalgraph.pdf}
\end{center}
\caption{\label{fig: causal_graph}The causal graph model. $T$ is a treatment, $Y$ is an outcome, $\mathbf{Z}$ is the unobserved confounders, and $\mathbf{X}$ is measurable covariates affected by $\mathbf{Z}$.}
\end{wrapfigure}
Suppose we receive observations $(\mathbf{x}_i, t_i, y_i)$ for $N$ samples, where $(\mathbf{x}_i, t_i, y_i)$ represents realisations of random variables $(\mathbf{X}, T, Y)$. Here, $\mathbf{x}_i \in \mathbb{R}^{d_x}$ is the covariate vector, $t_i$ is the observed treatment, and $y_i \in \mathbb{R}$ is the observed outcome of the $i$th sample. Without loss of generality, we assume the continuous treatment $T$ falls within the [0, 1] interval. The distribution of $(\mathbf{x}_i, t_i, y_i)$ is affected by unobserved confounders $\mathbf{z}_i$. Following the potential outcome framework~\cite{guo2020survey, rubin1984bayesianlypotentialoutcome}, we denote the potential outcome, which is the outcome $Y$ if applied an intervention $T=t$, as $Y|\text{Do}(T=t)$ or simply $Y(t)$. Unobserved potential outcomes are referred to as counterfactuals.
Our goal is to estimate the individual dose-response curve as
\begin{equation}
\mu(\mathbf{x}, t) = \mathbb{E}[Y(t)|\mathbf{X}=\mathbf{x}],
\end{equation}
which is a direct extension of the individual treatment effect (ITE). ~\cite{cevae17} proved that with $p(\mathbf{X}, \mathbf{Z}, T, Y)$ recovered, the ITE of binary treatment $\mathbb{E}[Y(1)|\mathbf{X}=\mathbf{x}] - \mathbb{E}[Y(0)|\mathbf{X}=\mathbf{x}]$ is identifiable. Similarly, we prove in the Appendix~\ref{apdx: proof} that with the same premise, the individual dose-response curve $\mathbb{E}[Y(t)|\mathbf{X}=\mathbf{x}]$ is identifiable.
\section{Methodology}\label{sec: method}
We propose ContiVAE to estimate individual dose-response curve with the presence of hidden confounders by leveraging a variational auto-encoder (VAE) to learn the distribution of unobserved confounders as latent variables and approximately estimate $p(\mathbf{X}, \mathbf{Z}, T, Y)$. We first derive the variational lower bound of ContiVAE, then provide a detailed explanation of its neural network architecture and the design of the prior distribution. Finally, we briefly describe the training and inference process of ContiVAE.
%In this work, we propose a novel method, ContiVAE, which extends CEVAE~\cite{cevae17} to continuous treatment scenarios. %Although the theoretical derivation of CEVAE is not confined to binary treatments, its model architecture is specifically tailored for such scenarios. Given the critical role of neural network architecture in predicting counterfactuals and the inherent challenges in designing an appropriate network~\cite{tarnet}, the model structure designed for binary treatments can not be directly extended to continuous ones.
%To tackle the unique challenges posed by continuous treatments, we significantly alter the network architectures as well as the prior distribution of latent variables. To mitigate the issue of over-prioritizing the reconstruction of high-dimensional or self-correlated covariates $\mathbf{X}$, ContiVAE scales the reconstruction loss of $\mathbf{X}$ to reduce its impact. In this section, we first revisit the approximate recovery of $p(\mathbf{X}, \mathbf{Z}, T, Y)$ using variational auto-encoders and the derivation of the variational lower bound, as introduced in~\cite{cevae17}. Then, we introduce the design of ContiVAE in details.
% challenges are:
% 1. 为了学习Individual dose-response curve, 需要模型能够从每个sample给出curve上的一个点，不仅学习到整条curve的形状，还要能够学到每个sample的特征，因为不同sample的curve可能有很大的区别。
% 2. 对于binary treatment，每个sample的counterfactual，即需要预测的unobserved outcome 只有1个，但对于continuous treatment，每个sample的counterfactual有无穷多个，增加了训练的难度。
% 3. 需要在比较宽的神经网络中保持treatment标量的影响，对于binary treatment来说，可以采取tarnet的方式分为两个网络用multi-task方式训练，但是对于continuous treatment来说无法采用无穷多个网络，（但是这个好在VAE会重构t，因此网络会试图在中间保留住t的信息？）
\subsection{Deriving the Variational Lower Bound}
ContiVAE models the hidden confounders $\mathbf{Z}$ as latent variables, and utilizes VAE to estimate the relationships between $\mathbf{Z}$ and $(\mathbf{X}, T, Y)$. Specifically, we assume $\mathbf{Z}$ follows a prior distribution $p(\mathbf{Z})$. For each observed sample $(\mathbf{x}, t, y)$, the generative model takes a sample $\mathbf{z}$ to reproduce the observational data, maximizing
\begin{align}
\log p(\mathbf{x}, t, y) &= %\log \int_{\mathbf{z}} p(\mathbf{z})\, p(\mathbf{x}, t, y|\mathbf{z})\,d\mathbf{z}\nonumber\\
\log \int_{\mathbf{z}} p(\mathbf{x}, \mathbf{z}, t, y)\, d\mathbf{z} = \log \int_{\mathbf{z}} p(\mathbf{z})\, p(\mathbf{x}, t, y|\mathbf{z})\,d\mathbf{z}\nonumber\\
&= \log \int_{\mathbf{z}} p(\mathbf{z}) \, p(\mathbf{x}|\mathbf{z}) \, p(t|\mathbf{z})\, p(y|t, \mathbf{z}) \, d\mathbf{z},
\label{eq:goal}
\end{align}
where the posteriors $p(\mathbf{x}|\mathbf{z})$, $p(t|\mathbf{z})$, $p(y|t, \mathbf{z})$ are parameterized by neural networks with parameters $\mathbf{\theta}$. As the prior $p(\mathbf{Z})$ is unknown, we approximate the posterior $p(\mathbf{z}|\mathbf{x}, t, y)$ with a neural network $q_\phi(\mathbf{z}|\mathbf{x}, t, y)$ which takes $(\mathbf{x}, t, y)$ as inputs and serves as the encoder of the VAE. Since minimizing the KL-divergence of $q_\phi(\mathbf{z}| \mathbf{x}, t, y)$ and $p(\mathbf{z}| \mathbf{x}, t, y)$ is to minimize
%\centerline{
%\begin{minipage}{1.0\textwidth}
%{\footnotesize
\begin{align}
& \mathcal{D}\left[q_\phi(\mathbf{z}| \mathbf{x}, t, y) \parallel p(\mathbf{z}| \mathbf{x}, t, y) \right] \nonumber\\
&= \mathbb{E}_{\mathbf{z} \sim q_\phi(\mathbf{z}| \mathbf{x}, t, y)}\left[ \log q_\phi(\mathbf{z}| \mathbf{x}, t, y) - \log p(\mathbf{z}| \mathbf{x}, t, y) \right] \nonumber\\
&= \mathbb{E}_{\mathbf{z} \sim q_\phi(\mathbf{z}| \mathbf{x}, t, y)} \left[ \log q_\phi(\mathbf{z}| \mathbf{x}, t, y) - \log p(\mathbf{z}) - \log p(\mathbf{x}, t, y | \mathbf{z}) + \log p(\mathbf{x}, t, y)\right] ,
\end{align}
%}%
%\end{minipage}
%}
the variational lower bound (ELBO) of our VAE model can be written as
\begin{align}
\mathcal{L} = \sum_{i=1}^N \mathbb{E}_{\mathbf{z}_i \sim q_\phi(\mathbf{z}_i| \mathbf{x}_i, t_i, y_i)} \left[ \log p_\mathbf{\theta}(\mathbf{x}_i, t_i, y_i | \mathbf{z}_i)
+ \log p(\mathbf{z}_i) - \log q_\phi(\mathbf{z}_i| \mathbf{x}_i, t_i, y_i) \right]]
\end{align}
\subsection{The Network Architecture}
With the ELBO derived, the next step is to specify neural networks $p_{\theta_x}(\mathbf{x}|\mathbf{z})$, $p_{\theta_t}(t|\mathbf{z})$, $p_{\theta_y}(y | t, \mathbf{z})$ and $q_\phi(\mathbf{z}|\mathbf{x}, t, y)$. Additionally, for an unseen sample with covariates $\mathbf{x}$, we need two auxiliary networks $q(t|\mathbf{x})$ and $q(y|t, \mathbf{x})$ to predict its treatment value and outcome respectively.
%CEVAE assumes a standard Gaussian prior, i.e., $p(\mathbf{Z}) \sim \mathcal{N}(0, \mathbf{I})$. Its model architecture is inspired by TARNET~\cite{tarnet}. For instance, it models $p(y|t, \mathbf{z})$ as $ (1-t)\cdot p_{\theta_0}(y|0, \mathbf{z}) + t \cdot p_{\theta_1}(y|1, \mathbf{z})$, where $p_{\theta_0}(y|0, \mathbf{z})$ and $p_{\theta_1}(y|1, \mathbf{z})$ are Normal distributions (or Bernoulli depending on the dataset used), with distribution parameters estimated by separate neural networks. Other components that involve $t$ as an input also adopt similar multi-task structure. This design aims to prevent the diminishing influence of $t$ in high-dimensional hidden layers when directly inputted into a neural network. As for continuous treatment cases, DRnet~\cite{schwab2020learningdrnet} divides the treatment range into subintervals and employs multi-task heads for each subinterval, but this introduces additional hyperparameters that adds complexity to model selection. For our ContiVAE, we opt for single neural networks directly involve $t$ as part of the input, rather than adopting the multi-task structure similar to TARNET. Since ContiVAE is trained to reconstruct $t$ in the decoder, the influence of $t$ and its information will be preserved in previous layers to ensure its reconstruction at the end.
\begin{figure}[t!]
\centering
\subfigure[Decoder.]{
\includegraphics[width=.65\linewidth]{figures/decoder.pdf}
\label{fig: decoder}
}
\subfigure[Encoder.]{
\includegraphics[width=.8\linewidth]{figures/encoder.pdf}
\label{fig: encoder}
}
\caption{The model architecture of ContiVAE. Blue rectangles represent MLP networks with $h$ units per hidden layer, while green rectangles denote output layers. Circles represent probability distributions.}
\label{fig: model}
\vspace{-0.2cm}
\end{figure}
%The structure of the decoder of ContiVAE is shown in Figure~\ref{fig: decoder}. Specifically, we parameterize $p_{\theta_x}(\mathbf{x}|\mathbf{z})$ as a Gaussian distribution with its mean and standard deviation estimated by the two output layers upon a shared multi-layer perceptions network, i.e.,
The structure of ContiVAE is shown in Figure~\ref{fig: model}. Specifically, in the decoder(see Figure~\ref{fig: decoder}), we parameterize $p_{\theta_x}(\mathbf{x}|\mathbf{z})$ as a Gaussian distribution (or Bernoulli, etc., depending on the characteristics of each element in $\mathbf{X}$), with its mean and standard deviation estimated respectively by two output layers upon a shared multi-layer perceptions (MLP) network, taking $\mathbf{z}$ as the input:
\begin{equation}
p_{\theta_x}(\mathbf{x}| \mathbf{z}) = \mathcal{N}(\mathbf{\mu}_\mathbf{x}, \mathbf{\sigma}_\mathbf{x}^2), \, \mathbf{\mu}_\mathbf{x} = f_1(\mathbf{z}), \, \mathbf{\sigma}_\mathbf{x} = \text{Softplus}(f_2(\mathbf{z})),
\end{equation}
The $\text{Softplus}$ activation ensures the resulting $\mathbf{\sigma}_\mathbf{x}$ is positive, $f_1$ and $f_2$ are neural networks. Similarly, as $T$ is a continuous variable, we model $p_{\theta_t}(t|\mathbf{z})$ as Gaussian,
\begin{gather}
p_{\theta_t}(t|\mathbf{z}) = \mathcal{N}(\mu_t, \sigma_t^2), \, \mu_t = f_3(\mathbf{z}), \, \sigma_t = \text{Softplus}(f_4(\mathbf{z})).
\end{gather}
To estimate the outcome $y$ corresponding to a given treatment value, the neural networks for $p_{\theta_y}(y|t, \mathbf{z})$ take the concatenated vector of $t$ and $\mathbf{z}$ as input:
\begin{gather}
p_{\theta_y}(y|t, \mathbf{z}) = \mathcal{N}(\mu_{yt}, \sigma_{yt}^2), \, \mu_{yt} = f_5(t \parallel \mathbf{z}), \, \sigma_{yt} = \text{Softplus}\left(f_6(t \parallel \mathbf{z})\right).
\end{gather}
where operation $t \parallel \mathbf{z}$ represents concatenating $t$ with $\mathbf{z}$. During training, the input $t$ for $p_{\theta_y}(y|t, \mathbf{z})$ is the observed treatment of each sample, while during inference, it corresponds to a treatment value of interest.
The structure of the encoder is represented in Figure~\ref{fig: encoder}. Following~\cite{cevae17}, we use $q(t|\mathbf{x})$ and $q(y|t, \mathbf{x})$ to predict $t$ and $y$ for an unseen sample $\mathbf{x}$:
{\small
\begin{gather}
q(t|\mathbf{x}) = \mathcal{N}(\tilde{\mu_t}, \tilde{\sigma_t}^2), \, \tilde{\mu_t} = g_1(\mathbf{x}), \, \tilde{\sigma_t} = \text{Softplus}(g_2(\mathbf{x})).\\
q(y|t,\mathbf{x}) = \mathcal{N}(\tilde{\mu_{yt}}, \tilde{\sigma_{yt}}^2), \, \tilde{\mu_{yt}} = g_3 \circ g_1(\mathbf{x}, \tilde t), \, \tilde{\sigma_{yt}}= \text{Softplus}(g_4 \circ g_2(\mathbf{x}, \tilde t)).
\end{gather}
}%
Subsequently, $\mathbf{z}$ is encoded using the predicted results $\tilde t$ and $\tilde y$ along with $\mathbf{x}$:
{\small
\begin{equation}
q(\mathbf{z}|\mathbf{x}, t, y) = \mathcal{N}(\mu_\mathbf{z}, \sigma_\mathbf{z}^2), \, \mu_\mathbf{z} = g_5 \circ g_3(\mathbf{x}, \tilde t, \tilde y), \, \sigma_\mathbf{z} = \text{Softplus}(g_6 \circ g_4(\mathbf{x}, \tilde t, \tilde y)).
\end{equation}
}%
The architecture of ContiVAE diverges significantly from CEVAE, which is inspired by TARNET~\cite{tarnet} and leverages distinct neural networks to model the probability distributions depending on $T$ under different treatment state ($t=1/0$). For instance, it adopts two neural networks trained on separate data to infer $p(y|1, \mathbf{z})$ and $p(y|0, \mathbf{z})$ respectively, taking solely $\mathbf{z}$ as input. Our design design contrasts with DRNets~\cite{schwab2020learningdrnet} as well, which extends TARNET by partitioning the dosage range of continuous treatments into multiple subintervals and adopts a similar multi-task strategy for each subinterval. Their design aims to prevent the diminishing influence of $t$ in high-dimensional hidden layers when directly inputted into a neural network. However, this strategy is unsuitable and unnecessary in our case. Since ContiVAE is trained to reconstruct $t$ in the decoder, the influence of $t$ and its information will be preserved in previous layers to ensure its reconstruction at the end. Therefore, we adopts single MLP networks for all treatment values, directly incorporating $t$ as part of the input. This design reserves flexibility and avoids introducing hyperparameters and additional complexity to model selection like in DRNets.
\subsection{Prior Distribution}
\begin{wrapfigure}[16]{r}{.5\textwidth}
\vspace{-1.5cm}
\centering
\includegraphics[width=0.9\linewidth]{figures/titledgaussian.pdf}
\caption{\label{fig:tilted}An example~\cite{floto2022tilted} of the Tilted Gaussian distribution in 2D, $\tau = 3$. The region of maximum density lies on the surface with a radius of 3.}
\end{wrapfigure}
Now we examine the prior distribution of $\mathbf{Z}$. %While CEVAE assumes a standard Gaussian latent distribution, an in-depth investigation of CEVAE~\cite{rissanen2021criticalcevaefault} has revealed that it estimates causal effects incorrectly when the latent variable is misspecified. Moreover,
ContiVAE adopts a Tilted Gaussian prior distribution~\cite{floto2022tilted}, rather than a standard Gaussian prior commonly employed in various methods~\cite{rezende2014stochasticvae, kingma2013autonormalvae} and in \cite{cevae17}.
Given our objective of estimating individual dose-response curves, which can vary significantly among individuals, the model must not only recover the curves of each sample % based on a single observed point per sample
but also capture individual heterogeneity at the same time. The normal distribution concentrates its high probability density region around a point, which becomes relatively smaller as the latent dimension increases. Minimizing the KL-divergence with a normal prior distribution will force the latent variables of different samples into the same point, thereby hindering the effective representation of individual heterogeneity. %which may result in a failure to capture individual differences.
%To address these issues, instead of the normal distribution, we adopt the Tilted Gaussian distribution proposed in~\cite{floto2022tilted} as the prior distribution $p(\mathbf{Z})$.
Opting for the Tilted Gaussian distribution helps address these issues.
The tilted Gaussian, denoted as $\mathcal{N}_\tau(0, \mathbf{I})$, is created by exponentially tilting the Gaussian distribution by its norm. In this distribution, the maximum probability density region lies on the surface of a sphere whose radius is $\tau$, rather than a single point. As the latent dimension increases, the volume of this region is exponentially larger compared to that of standard Gaussian as the latent dimension increases. The probability density function $\rho_\tau(\mathbf{z})$ of $\mathcal{N}_\tau(0, \mathbf{I})$ is defined as
\begin{equation}
\rho_\tau(\mathbf{z}):= \frac{e^{\tau}\Vert\mathbf{z}\Vert}{\mathbf{Z}_\tau} \frac{e^{-\frac{1}{2}\Vert\mathbf{z}\Vert^2}}{\sqrt{2\pi} \mathbf{I}^{d_\mathbf{z}}}, \mathbf{Z}_\tau := \mathbb{E}_{\mathbf{z} \sim \mathcal{N}(0, \mathbf{I})}[e^{\tau \Vert \mathbf{z} \Vert}],
\end{equation}
where $d_\mathbf{z}$ is the dimension of $\mathbf{Z}$, and $\rho_0(\mathbf{z})$ is equal to the density function of $\mathcal{N}(0, \mathbf{I})$. An example of $\rho_\tau(\mathbf{z})$ in 2-D is visualized in Figure~\ref{fig:tilted}. More detailed information of the Tilted Gaussian distribution, readers are referred to~\cite{floto2022tilted}. With a tilted Gaussian prior, the KL-divergence term is calculated by
\begin{gather}
\mathcal{D}\left[q_\phi(\mathbf{z}| \mathbf{x}, t, y) \parallel p(\mathbf{z}) \right] = \frac{1}{2}(\Vert \mu_\mathbf{z} \Vert - \Vert \mu_\tau^* \Vert)^2,\\
\mu_\tau^* = \text{argmin}_\mu -\tau\sqrt{\frac{\pi}{2}}\mathbf{L}_{\frac{1}{2}}^{\frac{d_\mathbf{z}}{2}-1} \left(-\frac{\Vert \mu \Vert^2}{2} \right)+ \frac{\Vert \mu \Vert^2}{2},
\end{gather}
where $\mathbf{L}$ is the generalized Laguerre polynomial. Throughout the paper, we set $\tau$ to 3.
\subsection{Optimization Objective}
An investigation of CEVAE~\cite{rissanen2021criticalcevaefault} uncovered the problem of over-emphasizing the reconstruction of $\mathbf{X}$ when there exists high correlations in $\mathbf{X}$, which may lead to neglecting the hidden confounders and directly adjusting for covariates. We encountered the same problem when applying on real world datasets with very high-dimensional $\mathbf{X}$. To address this, we introduce a scaling parameter $\lambda \le 1.0$ to reduce the emphasis on the reconstruction loss of $\mathbf{X}$. The value of $\lambda$ will be selected through $k$-fold cross validation in practice. During training, we utilizes Stochastic Gradient Descent to minimize the loss function
\begin{align}
\mathcal{L} &= \sum_{i=1}^N - (\,\lambda \cdot \log p_{\theta_x}(\mathbf{x}_i|\mathbf{z}_i) + \log p_{\theta_t}(t_i|\mathbf{z}_i) + \log p_{\theta_y}(y_i|t_i, \mathbf{z}_i) \nonumber\\
&+ \log q(t_i|\mathbf{x}_i) + \log q(y_i|t_i, \mathbf{x}_i) \,) + \mathcal{D}\left[q_\phi(\mathbf{z}_i| \mathbf{x}_i, t_i, y_i) \parallel p(\mathbf{z}_i)\right].
\end{align}
During inference, to obtain individual dose-response curve of a given $\mathbf{x}$, we input it into the model and gather the expected output of $p(y|t, \mathbf{z})$ as the potential outcome at $t$.
\section{Experiments on Synthetic Datasets} \label{sec: simulation}
%It is difficult to evaluate models for estimating continuous treatment effects on real world datasets, due to the impossibility to obtain the counterfactual outcomes and thereby rendering groundtruths unaccessiable. Although there have been well-established benchmark datasets with groundtruths for binary treatment scenarios, no such datasets exist for continuous treatment settings. In this study, we leverage an evaluation method akin to that proposed in prior work~\cite{scigan2020} to investigate the performance of our model on semi-synthetic datasets.
Due to the difficulty in obtaining real world datasets for causal inference with groundtruths, we evaluate our method on semi-synthetic datasets, seeking answers to the following questions: (1) How does the performance of ContiVAE compare to various baseline methods and state-of-the-art models? (2) How do varing hyperparameter configurations, especially choices of $\lambda$ influence ContiVAE's performance? (3) How does ContiVAE's performance change with increasing selection bias in the observational data?
\subsection{Experimental Setup}
\subsubsection{Datasets.} We conduct our experiments on two real world datasets: News~\cite{schwab2020learningdrnet} and The Cancer Genomic Atlas (TCGA)~\cite{weinstein2013cancertcga}\footnote{We use the same version of datasets as in \href{https://github.com/d909b/drnet}{https://github.com/d909b/drnet}.}.
\paragraph{TCGA.} The TCGA dataset contains gene expression data of various types of tumors for 9,659 patients. Following ~\cite{scigan2020}, we select the 4,000 most variable genes for each patient as covariates $X$, and normalize them to the [0, 1] interval using min-max normalization. Each patient's covariate is further normalized to have norm 1. We consider the single treatment scenario and evaluate on 4 manually designed dose-response curves, as listed in Table~\ref{tab:tcga_curves}. Each dose-response curve $f^j(\mathbf{x}, t)$ is characterized by a set of parameters $\mathbf{v}_1^j$, $\mathbf{v}_2^j$, $\mathbf{v}_3^j$, which are randomly sampled from $\mathcal{N}(0, 1)$ and normalized to have norm 1.
We randomly assign intervention $t$ to each patient, where $t$ is sampled from distribution $\text{Beta}(\alpha, \beta^j)$. Here, $\beta^j = \frac{\alpha-1}{t^{j*}} + 2 -\alpha$, where $t^{j*}$ is the optimal dosage for response curve $f^j$, ensuring that the mode of the beta distribution equals to $t^{j*}$. $\alpha$ determine the skewness of the distribution, thus controlling the dosage selection bias. When $\alpha=1$, $\text{Beta}(\alpha, \beta^j)$ reduces to a uniform distribution between 0 and 1. After assigning $t_i$ to a patient with covariate $\mathbf{x}_i$, its observed outcome is generated by adding a Gaussian noise $\epsilon \sim \mathcal{N}(0,0.02)$ to $f^j(\mathbf{x}_i, t_i)$.
\begin{table}[t]
\begin{adjustbox}{center}
\begin{tabular}{ccc}
\Xhline{0.8pt}
No. & Dose-Response Curve & Optimal Dosage \\
\specialrule{0.7pt}{1pt}{1pt}
\multirow{2}*{1} & \scalebox{0.9}{\multirow{2}*{$ C\left((\mathbf{v}_1^1)^T \mathbf{x} + 12t\left(t- k\right)^2\right), \text{where } k = 0.75 \frac{(\mathbf{v}_2^1)^T \mathbf{x}}{(\mathbf{v}_3^1)^T
\mathbf{x}}$}} & \scalebox{0.9}{$\frac{k}{3}$ if $k\ge0.75$}\\
& & \scalebox{0.9}{1 if $k < 0.75$} \\
\specialrule{0.5pt}{1pt}{1pt}
2 & \scalebox{0.9}{$C\left( (\mathbf{v}_1^2)^T \mathbf{x} + \sin \left(\pi \left(\frac{\mathbf{v}_2^{2T}\mathbf{x}}{\mathbf{v}_3^{2T}\mathbf{x}}\right) t\right)\right)$} & \scalebox{0.9}{$t_2^* =\frac{(\mathbf{v}_3^{2})^T \mathbf{x}}{2 (\mathbf{v}_2^{2})^T \mathbf{x}}$}\\
\specialrule{0.5pt}{1pt}{1pt}
3 & \scalebox{0.9}{$ C \left( (\mathbf{v}_1^3) ^T \mathbf{x} + 12 (\mathbf{v}_2^3) ^T \mathbf{x}t - 12 (\mathbf{v}_3^3)^T \mathbf{x} t^2 \right)$} & \scalebox{0.9}{$t_3^*=\frac{(\mathbf{v}_2^3)^T \mathbf{x}}{ 2 (\mathbf{v}_3^3)^T \mathbf{x}}$ }\\
\specialrule{0.5pt}{1pt}{1pt}
\multirow{3}*{4} & \multirow{3}*{\scalebox{0.9}{$ C \left( \cos ( (2+ (\mathbf{v}_1^4)^T \mathbf{x} ) \pi t + (\mathbf{v}_2^4)^T \mathbf{x} \pi ) + (\mathbf{v}_3^4)^T \mathbf{x}\right) $}} & \scalebox{0.9}{$b$ if $b = \frac{2-\mathbf{v}_2^{4T} \mathbf{x}}{ 2 + \mathbf{v}_1^{4T}\mathbf{x}} \le 1$}\\
& & \scalebox{0.9}{0 if $\cos (\pi (\mathbf{v}_2^4) ^T \mathbf{x}) > \cos(\pi (2+ (\mathbf{v}_1^4) ^T\mathbf{x} + (\mathbf{v}_2^4)^T\mathbf{x}))$} \\
& & \scalebox{0.9}{1 else.}\\
\specialrule{0.8pt}{1pt}{1pt}
\end{tabular}
\end{adjustbox}
\caption{The 4 dose-response curves used to generate outcomes using covariates $\mathbf{x}$ from TCGA dataset, and their corresponding optimal dosage $t^*$. When $t = t^{j*}$, the individualized dose-response curve $f^j(\mathbf{x}, t)$ reaches its maximum value.}
\label{tab:tcga_curves}
\vspace{-0.3cm}
\end{table}
\begin{table}[t]
\begin{adjustbox}{center}
\begin{tabular}{ccc}
\Xhline{0.8pt}
No. & Dose-Response Curve & Optimal Dosage \\
\specialrule{0.7pt}{1pt}{1pt}
\multirow{2}*{1} & \scalebox{0.9}{ \multirow{2}*{$C\left( \mathbf{u}^T \mathbf{V}_1^1 \mathbf{x} + 12t\left(t- k\right)^2\right), \text{where } k = 0.75 \frac{ \mathbf{u}^T \mathbf{V}_2^1\mathbf{x}}{\mathbf{u}^T \mathbf{V}_3^1
\mathbf{x}}$}} & \scalebox{0.9}{$\frac{k}{3}$ if $k\ge0.75$}\\
& & \scalebox{0.9}{1 if $k < 0.75$} \\
\specialrule{0.5pt}{1pt}{1pt}
2 & \scalebox{0.9}{$ C\left( \mathbf{u}^T \mathbf{V}_1^2 \mathbf{x} + \sin \left(\pi \left(\frac{\mathbf{u} ^T \mathbf{V}_2^2 \mathbf{x}}{\mathbf{u}^T \mathbf{V}_3^2\mathbf{x}}\right) t\right)\right)$} & \scalebox{0.9}{$t_2^* =\frac{\mathbf{u}^T \mathbf{V}_3^2 \mathbf{x}}{2 \mathbf{u}^T \mathbf{V}_2^2 \mathbf{x}}$}\\
\specialrule{0.5pt}{1pt}{1pt}
3 & \scalebox{0.9}{$C \left( \mathbf{u}^T \mathbf{V}_1^3 \mathbf{x} + 12 \mathbf{u}^T \mathbf{V}_2^3 \mathbf{x}t - 12 \mathbf{u}^T \mathbf{V}_3^3 \mathbf{x} t^2 \right)$} & \scalebox{0.9}{$t_3^*=\frac{\mathbf{u}^T \mathbf{V}_2^3 \mathbf{x}}{ 2 \mathbf{u}^T \mathbf{V}_3^3 \mathbf{x}}$} \\
\specialrule{0.5pt}{1pt}{1pt}
\multirow{3}*{4} & \scalebox{0.9}{\multirow{3}*{$C \left( \cos ( (2+ \mathbf{u}^T \mathbf{V}_1^4 \mathbf{x} ) \pi t + \mathbf{u}^T \mathbf{V}_2^4 \mathbf{x} \pi ) + \mathbf{u}^T \mathbf{V}_3^4 \mathbf{x}\right) $}}& \scalebox{0.9}{$b$ if $b = \frac{2-\mathbf{u}^T \mathbf{V}_2^4 \mathbf{x}}{ 2 + \mathbf{u}^T \mathbf{V}_1^4 \mathbf{x}} \le 1$} \\
& & \scalebox{0.9}{0 if $\cos (\pi \mathbf{u}^T \mathbf{V}_2^4 \mathbf{x}) > \cos(\pi (2+ \mathbf{u}^T \mathbf{V}_1^4 \mathbf{x} + \mathbf{u}^T \mathbf{V}_2^4 \mathbf{x}))$} \\
& & \scalebox{0.9}{1 else.} \\
\specialrule{0.8pt}{1pt}{1pt}
\end{tabular}
\end{adjustbox}
\caption{The 4 dose-response curves used to generate outcomes using covariates $\mathbf{x}$ and hidden confounders $\mathbf{u}$ from News dataset, and their corresponding optimal dosage $t^*$. When $t = t^{j*}$, the individualized dose-response curve $f^j(\mathbf{x}, \mathbf{u}, t)$ reaches its maximum value.}
\label{tab:news_curves}
\vspace{-0.5cm}
\end{table}
\paragraph{News.} The News dataset comprises 100,000 news articles sampled from the NY Times corpus. Each news article is associated with a 2,876-dimensional bag-of-word vector, serving as the covariates $\mathbf{X}$, as well as a 50-dimensional vector which can be think of as the browsing information of the reader, considered as the hidden confounders $\mathbf{U}$. We adopt the same method as with the TCGA dataset to generate observed data, and also examine four dose-response curves. However, while the dose-response curves on the TCGA dataset are functions of $\mathbf{x}$ and $t$, we explicitly set the dose-response curves on News dataset to be dependent on $\mathbf{u}$ by modifying paramters $\mathbf{v}_1^j, \mathbf{v}_2^j, \mathbf{v}_3^j$ to $\mathbf{u}^T \mathbf{V}_1^j, \mathbf{u}^T \mathbf{V}_2^j, \mathbf{u}^T \mathbf{V}_3^j$, where $\mathbf{V}_1^j, \mathbf{V}_2^j, \mathbf{V}_3^j$ are matrices sampled from $\mathcal{N}^{50 \times 2876}$. The definition of the 4 dose-response curves are provided in Table~\ref{tab:news_curves}. Note that the hidden confounders $\mathbf{U}$ is not visible to models during training; they are solely used for data generation and obtaining groundtruths.
\subsubsection{Models.}
\begin{itemize}
\item \textbf{ContiVAE and ContiVAE-N.} In addition to our proposed ContiVAE, to investigate the performance enhancement of using Tilted-Gaussian distribution as the prior latent distribution, we experiment with a variant of ContiVAE that utilizes a normal latent distribution, referred to as ContiVAE-N.
\item \textbf{MLP.} As a baseline, we develop an multi-layer perceptron (MLP) network which takes covariates and dosages as input and predicts the outcomes.
\item \textbf{SCIGAN.} SCIGAN~\cite{scigan2020} uses GANs to estimate individualized dose-response curves for multiple continuous treatments. It trains a counterfactual generator and discriminator adversarially, then utilizes the trained generator to supervisely train an inference network, achieving the state-of-art performance. In this work, we focus on single treatment scenario, so we set the number of treatments in SCIGAN to 1 and use optimal hyperparameters given in~\cite{scigan2020}.
\item \textbf{GPS.} Generalized Propensity Score (GPS)~\cite{imai2004causalgps} extends Propensity Score Matching to continuous treatment settings. We implement GPS using the \verb|causal_curve| python package with a normal treatment model and 2nd degree polynomial outcome model.
\item \textbf{CBGPS.} Covariate Balancing Generalized Propensity Score (CBGPS)~\cite{fong2018covariateCBGPS} is an improvement of GPS that estimating the propensity score by optimizing the resulting covariate balancing. It increases robustness against model misspecification, and improves covariate balancing even in the presence of hidden confounders. %, without relying on the ignorability assumption.
We implement CBGPS using the \verb|CBPS| R package, with a linear treatment model and 2nd degree polynomial regression.
\end{itemize}
\subsubsection{Hyperparameters.} We select the best hyperparameters of MLP and our model for each dataset and dose-response curve through 5-fold cross validation, and use the best settings given in the original paper~\cite{scigan2020} for SCIGAN. %All models are compared with each other using their best settings. %as given in Table~\ref{}.
On News dataset, we train all models using Adam~\cite{kingma2014adam} optimizer at a learning rate of 0.0001 for 100 epochs, and on TCGA dataset with learning rate of 0.001 for 300 epochs.
\subsubsection{Metrics.}
For each dataset and dose-response curve, we compare the models by the root of Mean Integrated Square Error (MISE) and Dosage Policy Error (DPE). MISE, defined in Eq.~\ref{eq: mise}, measures the overall accuracy of the model in estimating potential outcomes across the entire dosage range: %how well the model estimates potential outcome over the entire dosage range.
\begin{equation}\label{eq: mise}
\text{MISE} = \frac{1}{N} \sum_{i=1}^{N} \int \left(y_i(t) - \hat y_i(t) \right)^2 dt
\end{equation}
DPE evaluates how accurately the model predicts the optimal dosage for each patient:
\begin{equation}\label{eq: dpe}
\text{DPE} = \frac{1}{N}\sum_{i=1}^{N} \left( y_i(t^*) -y_i(\hat t^*)\right)^2,
\end{equation}
where $t^*$ is the true optimal dosage, $\hat t^*$ is the predicted optimal dosage. %We use Romberg integration with 65 equally spaced samples of $y_i(t)$ and $\hat y_i(t)$ to compute $\sqrt{\text{MISE}}$, and Sequential Least Squares Programming (SLSQP) to determine the maxima of $\hat y_i(t)$ and compute $\sqrt{\text{DPE}}$.
All the results are obtained on a held-out test set comprising 20\% of the data, and averaged over 5 repeat runs.
% table with MISE and DPE together
\begin{table}[h!]
\centering
\begin{tabular}{c|ccccc}
\specialrule{1.3pt}{0pt}{0.5pt}
Method & & Curve 1 & Curve 2 & Curve 3 & Curve 4 \\
\specialrule{0.8pt}{0.5pt}{0.5pt}
\multirow{2}*{ContiVAE} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& \textbf{0.2593 $\pm$ 0.06}
& \textbf{0.1401 $\pm$ 0.01}
& $0.3718 \pm0.12$
& \textbf{0.3035$\pm$0.06} \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& \textbf{0.3204$\pm$0.07}
& $0.0043\pm0.002$
& $0.0084 \pm0.004$
& \textbf{0.0197 $\pm$ 0.003} \\
\multirow{2}*{ContiVAE-N} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& \textit{0.4471 $\pm$ 0.08}
& $0.2915 \pm 0.07$
& $0.6348 \pm 0.05$
& \textit{0.8165 $\pm$ 0.06} \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& \textit{0.2171 $\pm$ 0.60}
& $0.0060 \pm 0.01$
& $0.0143 \pm 0.02$
& \textit{0.0603 $\pm$ 0.04} \\
\specialrule{0.8pt}{0.5pt}{0.5pt}
\multirow{2}*{SCIGAN} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& 1.5720 $\pm$ 0.12
& $0.9367 \pm 0.80$
& $1.2003 \pm0.46$
& 3.7088 $\pm$ 1.07 \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& $1.7665\pm 2.34$
& $0.0108\pm 0.01$
& $0.0102 \pm 0.01$
& $1.9687 \pm1.00$ \\
\specialrule{0.8pt}{0.5pt}{0.5pt}
\multirow{2}*{MLP} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& $1.6674 \pm 0.66$
& $3.6899 \pm 0.18$
& $8.0742\pm 0.27$
& $7.5790\pm 3.07$ \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& 0.4805 $\pm$ 0.40
& $9.9710 \pm 0.06$
& $22.4423\pm 0.00$
& $13.9378 \pm 8.73$ \\
\specialrule{0.8pt}{0.5pt}{0.5pt}
\multirow{2}*{GPS} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& $3.6592 \pm 0.43$
& $0.3634 \pm 0.02$
& \textit{0.3716 $\pm$ 0.04}
& $8.1514 \pm 0.30$ \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& $7.3813 \pm 0.24$
& \textit{0.0010 $\pm$ 0.001}
& \textit{0.0026 $\pm$ 0.003}
& $0.1723 \pm 0.03$ \\
\specialrule{0.8pt}{0.5pt}{0.5pt}
\multirow{2}*{CBGPS} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& $2.7198 \pm 0.33 $
& \textit{0.2354 $\pm$ 0.01 }
& \textbf{0.2500 $\pm$ 0.03 }
& $6.8300 \pm 0.27 $ \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& $6.3100 \pm 3.22 $
& \textbf{0.0007 $\pm$ 0.00}
& \textbf{0.0023 $\pm$ 0.002}
& 0.1521 $\pm$ 0.07 \\
\specialrule{1.3pt}{0.5pt}{2pt}
\end{tabular}
\caption{Performances of all compared models for four curves on TCGA dataset. Metrics are reported as mean $\pm$ half-width of 0.95 confidence interval. The best results are highlighted in bold while the second best are italicized.}
\label{tab:tcga_results}
\vspace{-0.3cm}
\end{table}
\begin{table}[t!]
\centering
\begin{tabular}{c|ccccc}
\specialrule{1.3pt}{0pt}{0.5pt}
Method & & Curve 1 & Curve 2 & Curve 3 & Curve 4 \\
\specialrule{0.8pt}{0.5pt}{0.5pt}
\multirow{2}*{ContiVAE} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& \textbf{1.2357$\pm$ 0.09}
& \textbf{0.7534 $\pm$ 0.03}
& \textbf{0.0885 $\pm$0.004}
& \textbf{0.7849$\pm$0.02} \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& \textit{2.8568 $\pm$0.34}
& \textbf{0.1156 $\pm$0.01}
& \textit{0.0038$\pm$0.001}
& \textbf{0.2856$\pm$ 0.09} \\
\multirow{2}*{ContiVAE-N} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& $1.5663 \pm 0.19$
& $0.7868 \pm 0.02$
& $0.0958 \pm 0.01$
& $1.0697 \pm 0.05$ \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& $3.1903 \pm 0.57 $
& $0.1364 \pm 0.004$
& $0.0041 \pm 0.001$
& \textit{0.2993 $\pm$ 0.09} \\
\specialrule{0.8pt}{.5pt}{.5pt}
\multirow{2}*{SCIGAN} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& $2.8740 \pm 0.30$
& $1.6702 \pm 0.56$
& $0.1740 \pm 0.05$
& $3.4695 \pm 0.45$ \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& $3.5969 \pm 2.84$
& 0.3307 $\pm$ 0.28
& $0.0252 \pm 0.03$
& $0.8911 \pm0.56$ \\
\specialrule{0.8pt}{.5pt}{.5pt}
\multirow{2}*{MLP} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& \textit{1.2619 $\pm$ 0.04}
& \textit{0.7716 $\pm$ 0.01}
& $0.1244 \pm 0.003$
& \textit{0.8615 $\pm$ 0.01} \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& \textbf{2.7795 $\pm$ 0.40}
& \textit{0.1279 $\pm$ 0.005}
& $0.0275 \pm 0.002$
& 0.3631 $\pm$ 0.08 \\
\specialrule{0.8pt}{.5pt}{.5pt}
\multirow{2}*{GPS} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& $3.4418 \pm 1.15$
& $0.8472 \pm 0.04$
& \textit{0.0907 $\pm$ 0.001 }
& $6.1522 \pm 0.04$ \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& $8.3123 \pm 0.75$
& 0.2361 $\pm$ 0.03
& \textit{0.0038 $\pm$ 0.002}
& $8.6744 \pm 0.05$ \\
\specialrule{.8pt}{.5pt}{.5pt}
\multirow{2}*{CBGPS} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& $2.9787 \pm 0.22 $
& $0.8325 \pm 0.08 $
& $0.0921 \pm 0.001$
& $6.1483 \pm 0.03 $ \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& $7.7260 \pm 0.07 $
& $0.2567 \pm 0.31 $
& \textbf{0.0035 $\pm$ 0.001}
& $8.6988 \pm 0.02 $ \\
\specialrule{1.3pt}{.5pt}{2pt}
\end{tabular}
\caption{Performances of all compared models for four curves on News dataset.}
\label{tab:news_results}
\vspace{-0.6cm}
\end{table}
\subsection{Benchmark Comparison}
The experimental results of all compared models on the TCGA dataset and the News dataset are represented in Table~\ref{tab:tcga_results} and~\ref{tab:news_results} respectively. We found that ContiVAE consistently achieves the best or comparable performance across both datasets and all evaluated curves, with significantly lower $\sqrt{\text{MISE}}$ and $\sqrt{\text{DPE}}$ than other methods. This demonstrates ContiVAE's robustness and its ability to accurately estimate individual dose-response curves. Notably, ContiVAE with a Tilted Gaussian prior exhibits significant improvement over its variant with the Normal prior (ContiVAE-N), indicating that changing to Tilted Gaussian expands the high probability density region in latent space, enabling the model to better capture individual differences and handle complex curve shapes.
We also observed that parametric methods like GPS and CBGPS perform well then the underlying curve aligns with their assumptions (Curve 3). However, for complex curves (Curve 1 and 4), neural networks, including ContiVAE, significantly outperform them. %relies on the correct specification of their model parameters. On Curve 3, where the 2nd degree polynomial curve aligns with the settings of GPS and CBGPS, they generally outperform neural network-based methods. However, on more complex curves like Curve 1 and 4, %which are 3nd degree polynomial and cosine with dynamic phase shift,
%neural network-based methods significantly outperform GPS and CBGPS.
ContiVAE's strengths also lies in its flexibility, %exhibits greater flexibility and robustness,
achieving competitive results in both scenarios.
%\item SCIGAN also demonstrates a certain degree of robustness, yielding stable results across both datasets and all curves, but it falls short of ContiVAE in accurately fitting individual dose-response curves.
While MLP performs comparably to ContiVAE on the large News dataset (100,000), but its performance suffers significantly on the smaller TCGA dataset (9,659). This suggests that MLP requires more training data to converge while ContiVAE is more data-efficient. The results of MLP and ContiVAE on the News dataset with reduced sample size (see Appendix~\ref{apdx: mlpresults}) illustrate MLP's dependence on larger sample size.
\iffalse
\begin{sidewaystable}[ht]
\centering
\begin{tabular}{c|cccccccc}
\Xhline{1.3pt}
\multirow{2}*{Method} & \multicolumn{2}{c}{Curve 1} & \multicolumn{2}{c}{Curve 2} & \multicolumn{2}{c}{Curve 3} & \multicolumn{2}{c}{Curve 4} \\
& \small{$\sqrt{\text{MISE}}$} & \small{$\sqrt{\text{DPE}}$} & \small{$\sqrt{\text{MISE}}$} & \small{$\sqrt{\text{DPE}}$} & \small{$\sqrt{\text{MISE}}$} & \small{$\sqrt{\text{DPE}}$} & \small{$\sqrt{\text{MISE}}$} & \small{$\sqrt{\text{DPE}}$} \\
\Xhline{0.8pt}
Ours & \textbf{0.2593 $\pm$ 0.06} & \textbf{0.3204$\pm$0.07}
& \textbf{0.1401 $\pm$ 0.01} & $0.0043\pm0.002$
& $0.3718 \pm0.12$ & $0.0084 \pm0.004$
& \textbf{0.3035$\pm$0.06} & \textbf{0.0197 $\pm$ 0.003}\\
SCIGAN & $1.5720 \pm 0.12$ & $1.7665\pm 2.34$
& $0.9367 \pm 0.80$ & $0.0108\pm 0.01$
& $1.2003 \pm0.46$ & $0.0102 \pm 0.01$
& $3.7088 \pm 1.07$ & $1.9687 \pm1.00$\\
MLP & $1.6674 \pm 0.66$ & $0.4805 \pm 0.40$
& $3.6899 \pm 0.18$ & $9.9710 \pm 0.06$
& $8.0742\pm 0.27$ & $22.4423\pm 0.00$
& $7.5790\pm 3.07$ & $13.9378 \pm 8.73$\\
GPS**2 & $3.6592 \pm 0.43$ & $7.3813 \pm 0.24$
& $0.3634 \pm 0.02$ & $0.0010 \pm 0.001$
& \textbf{0.3716 $\pm$ 0.04} &\textbf{0.0026 $\pm$ 0.003}
& $8.1514 \pm 0.30$ & $0.1723 \pm 0.03$ \\
GPS**5 & $0.2840 \pm 0.01 $ & $0.6628 \pm 0.30$
& $0.2701 \pm 0.04 $ & $0.0009 \pm 0.001$
& $0.3576 \pm 0.01$ & $0.0017 \pm 0.001 $
& $1.7571 \pm 0.13$ & $0.1262 \pm 0.03$\\
\Xhline{1.3pt}
\end{tabular}
\caption{Performances of all compared models for four curves on TCGA dataset.}
\label{tab:my_label}
\bigskip\bigskip
\centering
\begin{tabular}{c|cccccccc}
\Xhline{1.3pt}
\multirow{2}*{Method} & \multicolumn{2}{c}{Curve 1} & \multicolumn{2}{c}{Curve 2} & \multicolumn{2}{c}{Curve 3} & \multicolumn{2}{c}{Curve 4} \\
& \small{$\sqrt{\text{MISE}}$} & \small{$\sqrt{\text{DPE}}$} & \small{$\sqrt{\text{MISE}}$} & \small{$\sqrt{\text{DPE}}$} & \small{$\sqrt{\text{MISE}}$} & \small{$\sqrt{\text{DPE}}$} & \small{$\sqrt{\text{MISE}}$} & \small{$\sqrt{\text{DPE}}$} \\
\Xhline{0.8pt}
Ours & \textbf{1.2357$\pm$ 0.09} & $2.8568 \pm0.34$
& \textbf{0.7534 $\pm$ 0.03} & \textbf{0.1156 $\pm$0.01}
& \textbf{0.0885 $\pm$0.004} & \textbf{0.0038$\pm$0.001}
& \textbf{0.7849$\pm$0.02} & \textbf{0.2856$\pm$ 0.09}\\
SCIGAN & $2.8740 \pm 0.30$ & $3.5969 \pm 2.84$
& $1.6702 \pm 0.56$ & 0.3307 $\pm$ 0.28
& $0.1740 \pm 0.05$ & $0.0252 \pm 0.03$
& $3.4695 \pm 0.45$ & $0.8911 \pm0.56$\\
MLP & $1.2619 \pm 0.04$ & \textbf{2.7795 $\pm$ 0.40}
& $0.7716 \pm 0.01$ & $0.1279 \pm 0.005$
& $0.1244 \pm 0.003$ & $0.0275 \pm 0.002$
& $0.8615 \pm 0.01$ & $0.3631 \pm 0.08$\\
GPS**2 & $3.4418 \pm 1.15$ & $8.3123 \pm 0.75$
& $0.8472 \pm 0.04$ & $0.2361 \pm 0.03$
& $0.0899 \pm 0.001$ & $0.0038 \pm 0.002$
& $6.1522 \pm 0.04$ & $8.6744 \pm 0.05$\\
GPS**5 & $3.5817 \pm 9.19 $ & $3.7998 \pm 0.08 $
& $0.8226 \pm 0.03 $ & $0.2096 \pm 0.03$
& $0.0907 \pm 0.001 $ & $0.0034 \pm 0.001$
& $1.7273 \pm $ 0.01& $0.7557 \pm 0.01$ \\
\Xhline{1.3pt}
\end{tabular}
\caption{Performances of all compared models for four curves on News dataset.}
\label{tab:my_label}
\end{sidewaystable}
\fi
%\subsection{Ablation Study}
\subsection{Parameter Sensitivities}
We evaluate ContiVAE's sensitivity to key parameters, especially the scaling parameter $\lambda$, using 5-fold cross validation on Curve 4 of the News dataset (Figure~\ref{fig: parameter}). %Specifically, we examine performance variations with respect to different values of the loss scaling parameter $\lambda$, the number of hidden units $h$ and the dimension of latent variable $d_\mathbf{z}$.
Notably, introducing $\lambda$ resolves the previously identified issue~\cite{rissanen2021criticalcevaefault} and decreasing $\lambda$ significantly enhances ContiVAE's performance. Increasing hidden units % in ContiVAE's hidden layers
generally improves performance, with $h=128$ being sufficient. %satisfactory results.
Lastly, while a high-dimensional latent space is crucial for real-world data complexity, excessively high dimensions can hinder correct estimation. In our case, $d_\mathbf{z}=20$ yield optimal results.%real datasets require a sufficiently large latent space to capture complex relations of $\mathbf{X}$ and $\mathbf{Z}$, a too high-dimensional latent variables may prevent correct estimation, thus the dimension of latent space requires careful selection in practice. In our cases, ContiVAE achieves optimal results with $d_\mathbf{z}=20$.
\begin{figure}[thbp]
\centering
\includegraphics[width=1.0\linewidth]{figures/parameter_all.pdf}
\caption{Performance of our model on Curve 4, News dataset with different settings of $\lambda$ in the loss function, the number of units in each hidden layer and the dimension of latent variable $Z$. The blue line and y-axis on the left represents $\sqrt{\text{MISE}}$, and the green line and y-axis on the right represents $\sqrt{\text{DPE}}$.}
\label{fig: parameter}
%\vspace{-1cm}
\end{figure}
\vspace{-1cm}
\subsection{Robustness to Selection Bias}
\begin{figure}[h!]
\vspace{-0.5cm}
\centering
\includegraphics[width=\linewidth]{figures/selection_all.pdf}
\caption{Performance of all models with increasing dosage selection bias $\alpha$ on Curve 1 of TCGA dataset and News dataset. When $\alpha = 1.0$, dosage assignment is a uniform distribution in [0, 1]. When $\alpha$ grows, the difference between $\beta$ and $\alpha$ increases, thus making the dosage distribution $\text{Beta}(\alpha, \beta)$ more skewed.}
\label{fig:selection_all}
\vspace{-0.3cm}
\end{figure}
To assess the robustness of each model to increasing levels of dosage selection bias, we evaluate their performance on Curve 1 of both datasets with varing $\alpha$ from 1.0 to 4.0 (Figure~\ref{fig:selection_all}. %which are controlled by the parameter $\alpha$ in our treatment distribution $\text{Beta}(\alpha, \beta)$. %When $\alpha=1.0$, it degenerates into a uniform distribution. As $\alpha$ grows, the difference between $\alpha$ and $\beta$ increases, resulting in an more skewed treatment distribution.
When there is no selection bias ($\alpha$=1), all neural network-based methods (ContiVAE, MLP and SCIGAN) yield similar results. However, ContiVAE distinguishes itself by maintaining consistent performance across all $\alpha$ values, significantly outperforming other models as the bias increased.
\iffalse
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=0.45\linewidth]{figures/selection_tcga0_mise.pdf}
\label{fig: selection_tcga_mise}
}
\hspace{-12pt}
\subfigure[]{
\includegraphics[width=0.45\linewidth]{figures/selection_tcga0_dpe.pdf}
}
\hspace{-12pt}
\subfigure[]{
\includegraphics[width=0.45\linewidth]{figures/selection_news0_mise.pdf}
\label{fig: selection_news_mise}
}
\hspace{-12pt}
\subfigure[]{
\includegraphics[width=0.45\linewidth]{figures/selection_news0_dpe.pdf}
\label{fig: selection_news_dpe}
}
\caption{(a) $\sqrt{\text{MISE}}$ and (b) $\sqrt{\text{DPE}}$ of all models with increasing dosage selection bias $\alpha$ on Curve 1, News dataset. When $\alpha = 1.0$, dosage assignment is a uniform distribution in [0, 1]. When $\alpha$ grows, the difference between $\beta$ and $\alpha$ increases, thus making the dosage distribution $\text{Beta}(\alpha, \beta)$ more skewed. }
\label{fig:enter-label}
\end{figure}
\fi
\section{Application on Volunteer Data}\label{sec:volunteer}
In this section, we apply ContiVAE to estimate the effects of a volunteer's social network position on retention. Here, the social network is formed as volunteers engage in various tasks together, with their position quantified by social network metrics like degree centralities. Volunteer retention indicates how actively a volunteer will continue engaging in voluntary activities in the future, and can be quantified by the number of tasks a volunteer undertakes in an upcoming time span. The retention of volunteers is a crucial issue faced by non-profit and community organizations~\cite{volunteer2003, al2015volunteerimpact, cuskelly2006volunteermanagement}. %the level of future participation, is a crucial area in sociology, economics and other fields. Previous research has identified a range of factors influencing volunteering decisions, including personal circumstances like social status and religious beliefs, as well as organizational aspects such as management style% and organizational experiences
%~\cite{volunteer2003}. %Recent years have witnessed an increasing number of online crowdsourcing systems utilized to organize volunteering efforts. In such crowdsourcing systems, volunteers are no longer organized by centralized, hierarchical management, but collaborate in a self-organized way via spontaneously initiating and participating in tasks.
%In online crowdsourcing systems, where volunteers collaborate in a decentralized manner rather than being centrally managed, their retention may be influenced by their relationships with one another.
%With most studies of organizational factors on volunteer retention focusing on centralized management in formal offline volunteer agencies~\cite{cuskelly2006volunteermanagement, al2015volunteerimpact, volunteer2003}, little attention has been paid to the structure of social network among volunteers.
Applying ContiVAE, we aim to answer to which extent, the different position a volunteer occupies in the social network would affect her future retention.
\subsection{Volunteer Data}
Our volunteer data is collected from Anti-Pandemic Pioneer, a mobile crowdsourcing platform launched during the COVID-19 epidemic to organize volunteer activities in Shenzhen, China. We collected 708,933 task participation logs from February 2020 to February 2022, involving 160,677 volunteers, each containing participation details such as timestamps, locations, task descriptions, volunteer IDs, and task IDs. Volunteer profiles, including age and task preferences, etc., are also collected as covariates $\mathbf{X}$.
We split the time span into two segments. In the first segment, we construct a network $\mathcal{G}(\mathcal{V}, \mathcal{E}, \mathcal{W})$ of volunteers based on their co-participation relationships, where $v_i$ represents a volunteer, the presence of an edge $e_{ij}$ indicates that $v_i$ and $v_j$ collaborated at least once, and the edge weight $w_{ij}$ denotes the frequency of co-participation. The treatment $T$ of a volunteer is one of the indicators including centrality and structural hole spanner metrics~\cite{burt}, % of the volunteer or their neighbors,
calculated from the constructed graph. For each volunteer, we quantify the observed outcome $Y$ by counting the number of her participated tasks in the second segment. %See Appendix~\ref{apdx: volunteer} for more details of the volunteer data.
%Using the constructed graph, we calculate a series of indicators to characterize the position of each volunteer in the graph as well as the structure of its surroundings. % as treatment variables.
%These indicators include centrality and structural hole spanner metrics defined in~\cite{burt} of a node and its neighboring nodes, as well as measurements of density of its ego-network (e.g., clustering coefficients).
\begin{figure}[t!]
%\vspace{-1cm}
\centering
\subfigure[]{
\includegraphics[width=0.3\linewidth]{figures/volunteer_degree.pdf}
}
\subfigure[]{
\includegraphics[width=0.3\linewidth]{figures/volunteer_hierarchy.pdf}
}
\subfigure[]{
\includegraphics[width=0.297\linewidth]{figures/volunteer_triangle.pdf}
}
\vspace{-0.3cm}
\caption{Averaged dose-response curves of the number of a volunteer's future participation with (a) the degree centrality, (b) the hierarchy, (c) the number of triangles in the ego-network.}
\label{fig:volunteer_result}
\vspace{-0.3cm}
\end{figure}
\subsection{Results}
%Due to space constraints, we only present the averaged curves depicting how the number of future tasks varies with the degree centrality, hierarchy as well as the number of triangles in ego-networks over all volunteers in Figure~\ref{fig:volunteer_result}.
From the dose-response curves predicted by ContiVAE, we find that volunteers with higher degree centrality, hierarchy or number of triangles in ego-network exhibit a higher intention to remain engaged. We present the averaged curves depicting how the number of future tasks varies with these indicators across all volunteers in Figure~\ref{fig:volunteer_result}. %the degree centrality, hierarchy as well as the number of triangles in ego-networks over all volunteers in Figure~\ref{fig:volunteer_result}.
The degree centrality of a volunteer is the number of collaborators normalized by the total number of volunteers. Hierarchy, as defined by Burt~\cite{burt}, is an indicator to measure the quality of structural hole spanners \textendash \ individuals that bridge otherwise disconnected groups in a social network, such as scientists collaborating across disciplines to connect researchers from diverse fields. A higher hierarchy value indicates a volunteer's collaborations are more concentrated on a single collaborator. Finally, the number of triangles in a volunteer's ego-network captures the density of connections among their collaborators. As Figure~\ref{fig:volunteer_result} suggests, occupying central position through collaboration with numerous colleagues strongly encourages a volunteer to continue participation. In addition, fostering a strong tie with a single collaborator or building a tightly-knit group also promotes volunteer engagement.
\begin{subappendices}
\renewcommand{\thesection}{\Alph{section}}
\section{Identifiability of Individual Dose-Response Curves}\label{apdx: proof}
\begin{theorem}
If $p(\mathbf{X}, \mathbf{Z}, T, Y)$ is recovered, $\mathbb{E}[Y(t)|\mathbf{X} = \mathbf{x}]$ is identifiable from observational data under the causal model in Figure~\ref{fig: causal_graph}.
\end{theorem}
\begin{proof}
We prove that $p(Y|\mathbf{X}, \text{do}(t))$ is identifiable if $p(\mathbf{X}, \mathbf{Z}, T, Y)$ is recovered, then $\mathbb{E}[Y(t)|\mathbf{X}] = \int_y y \cdot p(y| \mathbf{X}, \text{do}(t))\, dy$ is identifiable.
\begin{align}
p(Y|\mathbf{X}, \text{do}(t)) & \overset{1}= \int_\mathbf{Z} p(y|\mathbf{X}, \text{do}(t), \mathbf{Z}) \,p(\mathbf{Z}|\mathbf{X}, \text{do}(t)) \, d\mathbf{Z}\\
& \overset{2}= \int_\mathbf{Z} p(y|\mathbf{X}, t, \mathbf{Z}) \, p(\mathbf{Z}|\mathbf{X}) \, d\mathbf{Z}\label{eq: terms}
\end{align}
$\overset{1}=$ is given by Pearl's backdoor adjustment formula~\cite{pearl2016causalprimer}, and $\overset{2}=$ is by applying do-operation on $T$ in the causal graph, which makes $Z$ independent of $T$ and remains the relationship of $Y$ and other variables. Once $p(\mathbf{X}, \mathbf{Z}, T, Y)$ is recovered, expression~\ref{eq: terms} can be calculated, thereby completing our proof.
\end{proof}
\section{Experimental Results of MLP and ContiVAE on News Dataset with Small Sample Size}\label{apdx: mlpresults}
We maintain the same data simulation parameters and the test set size of the News dataset, but reduce the size of training set (0.8 $\times$ 100,000 samples) to match that of the TCGA dataset (0.8 $\times$ 9,695 samples). We then perform experiments using ContiVAE and MLP, and the results are listed in Table~\ref{tab:appendixmlp}.
\begin{table}[h!]
\centering
\begin{tabular}{c|ccccc}
\specialrule{1.3pt}{0pt}{0.5pt}
Method & & Curve 1 & Curve 2 & Curve 3 & Curve 4 \\
\specialrule{0.8pt}{0.5pt}{0.5pt}
\multirow{2}*{ContiVAE} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& \textbf{2.5800 $\pm$ 1.48}
& \textbf{0.8024 $\pm$ 0.01}
& \textbf{0.1005 $\pm$ 0.004}
& \textbf{1.5718 $\pm$ 0.73} \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& \textbf{4.2986 $\pm$ 3.74}
& \textbf{0.1367 $\pm$ 0.02}
& \textbf{0.0049 $\pm$ 0.001}
& \textbf{0.3196 $\pm$ 0.12} \\
\specialrule{0.8pt}{0.5pt}{0.5pt}
\multirow{2}*{MLP} & \scalebox{0.8}{$\sqrt{\text{MISE}}$}
& $3.4804 \pm 0.17$
& $3.9918 \pm 0.15$
& $0.2826 \pm 0.01$
& $7.9010\pm 1.04$ \\
& \scalebox{0.8}{$\sqrt{\text{DPE}}$}
& $7.8602 \pm 0.01$
& $10.3197 \pm 0.01$
& $0.4665 \pm 0.09$
& $8.5948\pm 0.15$ \\
\specialrule{1.3pt}{0pt}{2pt}
\end{tabular}
\label{tab:appendixmlp}
\caption{Performance of ContiVAE and MLP on the News dataset with reduced training sample size.}
\end{table}
\end{subappendices}
\begin{thebibliography}{10}
\providecommand{\url}[1]{\texttt{#1}}
\providecommand{\urlprefix}{URL }
\providecommand{\doi}[1]{https://doi.org/#1}
\bibitem{al2015volunteerimpact}
Al~Mutawa, O.: Impact of Volunteer Management Practice on Volunteer Motivation and Satisfaction to Enhance Volunteer Retention. Ph.D. thesis, Brunel University London (2015)
\bibitem{bang2005doubly}
Bang, H., Robins, J.M.: Doubly robust estimation in missing data and causal inference models. Biometrics \textbf{61}(4), 962--973 (2005)
\bibitem{scigan2020}
Bica, I., Jordon, J., van~der Schaar, M.: Estimating the effects of continuous-valued interventions using generative adversarial networks. In: Proc. of NeurIPS. pp. 16434--16445 (2020)
\bibitem{burt}
Burt, R.S.: Structural holes: The Social Structure of Competition. Harvard University Press, Cambridge, Mass (1992)
\bibitem{cuskelly2006volunteermanagement}
Cuskelly, G., Taylor, T., Hoye, R., Darcy, S.: Volunteer management practices and volunteer retention: A human resource management approach. Sport Management Review \textbf{9}(2), 141--163 (2006)
\bibitem{floto2022tilted}
Floto, G., Kremer, S., Nica, M.: The tilted variational autoencoder: Improving out-of-distribution detection. In: Proc. of ICLR (2022)
\bibitem{fong2018covariateCBGPS}
Fong, C., Hazlett, C., Imai, K.: Covariate balancing propensity score for a continuous treatment: Application to the efficacy of political advertisements. The Annals of Applied Statistics \textbf{12}(1), 156--177 (2018)
\bibitem{guo2020survey}
Guo, R., Cheng, L., Li, J., Hahn, P.R., Liu, H.: A survey of learning causality with data: Problems and methods. ACM Computing Surveys (CSUR) \textbf{53}(4), 1--37 (2020)
\bibitem{imai2014covariateCBPS}
Imai, K., Ratkovic, M.: Covariate balancing propensity score. Journal of the Royal Statistical Society Series B: Statistical Methodology \textbf{76}(1), 243--263 (2014)
\bibitem{imai2004causalgps}
Imai, K., Van~Dyk, D.A.: Causal inference with general treatment regimes: Generalizing the propensity score. Journal of the American Statistical Association \textbf{99}(467), 854--866 (2004)
\bibitem{kennedy2017nondoublyrobust}
Kennedy, E.H., Ma, Z., McHugh, M.D., Small, D.S.: Non-parametric methods for doubly robust estimation of continuous treatment effects. Journal of the Royal Statistical Society Series B: Statistical Methodology \textbf{79}(4), 1229--1245 (2017)
\bibitem{kingma2014adam}
Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)
\bibitem{kingma2013autonormalvae}
Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)
\bibitem{volunteer2003}
Locke, M., Ellis, A., Smith, J.D.: Hold on to what you’ve got: the volunteer retention literature. Voluntary Action \textbf{5}(3), 81--99 (2003)
\bibitem{cevae17}
Louizos, C., Shalit, U., Mooij, J.M., Sontag, D., Zemel, R.S., Welling, M.: Causal effect inference with deep latent-variable models. In: Proc. of NeurIPS (2017)
\bibitem{cider2022}
Ma, J., Dong, Y., Huang, Z., Mietchen, D., Li, J.: Assessing the causal impact of covid-19 related policies on outbreak dynamics: A case study in the us. In: Proc. of ACM Web Conference. pp. 2678--2686 (2022)
\bibitem{montgomery2000measuringproxies}
Montgomery, M.R., Gragnolati, M., Burke, K.A., Paredes, E.: Measuring living standards with proxy variables. Demography \textbf{37}(2), 155--174 (2000)
\bibitem{pearl2012measurementpearlproxies}
Pearl, J.: On measurement bias in causal inference. arXiv preprint arXiv:1203.3504 (2012)
\bibitem{pearl2016causalprimer}
Pearl, J., Glymour, M., Jewell, N.P.: Causal Inference in Statistics: A Primer (2016)
\bibitem{rezende2014stochasticvae}
Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approximate inference in deep generative models. In: Proc. of ICML. pp. 1278--1286 (2014)
\bibitem{rissanen2021criticalcevaefault}
Rissanen, S., Marttinen, P.: A critical look at the consistency of causal estimation with deep latent variable models. Proc. of NeurIPS \textbf{34}, 4207--4217 (2021)
\bibitem{rosenbaum1987ipw}
Rosenbaum, P.R.: Model-based direct adjustment. Journal of the American statistical Association \textbf{82}(398), 387--394 (1987)
\bibitem{rubin1984bayesianlypotentialoutcome}
Rubin, D.B.: Bayesianly justifiable and relevant frequency calculations for the applied statistician. The Annals of Statistics pp. 1151--1172 (1984)
\bibitem{schwab2020learningdrnet}
Schwab, P., Linhardt, L., Bauer, S., Buhmann, J.M., Karlen, W.: Learning counterfactual representations for estimating individual dose-response curves. In: Proc. of AAAI. vol.~34, pp. 5612--5619 (2020)
\bibitem{tarnet}
Shalit, U., Johansson, F.D., Sontag, D.: Estimating individual treatment tffect: Generalization bounds and algorithms. In: Proc. of ACM ICML. pp. 3076--3085 (2017)
\bibitem{weinstein2013cancertcga}
Weinstein, J.N., Collisson, E.A., Mills, G.B., Shaw, K.R., Ozenberger, B.A., Ellrott, K., Shmulevich, I., Sander, C., Stuart, J.M.: The cancer genome atlas pan-cancer analysis project. Nature Genetics \textbf{45}(10), 1113--1120 (2013)
\bibitem{yao2021survey}
Yao, L., Chu, Z., Li, S., Li, Y., Gao, J., Zhang, A.: A survey on causal inference. ACM Transactions on Knowledge Discovery from Data (TKDD) \textbf{15}(5), 1--46 (2021)
\end{thebibliography}