EconBase
← Back to paper

Educational Effects in Mathematics: Conditional Average Treatment Effect depending on the Number of Treatments

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.

16,154 characters · 0 sections · 9 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.

Educational Effects in Mathematics: Conditional Average Treatment Effect depending on the Number of Treatments

abstractThis study examines the educational effect of the Academic Support Center at Kogakuin University. Following the initial assessment, it was suggested that group bias had led to an underestimation of the Center's true impact. To address this issue, the authors applied the theory of causal inference. By using T-learner, the conditional average treatment effect (CATE) of the Center's face-to-face (F2F) personal assistance program was evaluated. Extending T-learner, the authors produced a new CATE function that depends on the number of treatments (F2F sessions) and used the estimated function to predict the CATE performance of F2F assistance.

\@startsection{section}{1}{\z@} {2ex}{1ex}{\reset@font}{Introduction} The Academic Support Center at Japan's Kogakuin University (hereafter, the Center) offers {\em face-to-face} (F2F) personal assistance, as well as remote personal assistance, along with basic classes and a variety of on-demand educational materials available on its Learning Management System. These support options are open to all students. The university operates on a four-quarter system (1Q to 4Q) and conducts a proficiency test for all students at the time of admission in addition to regular examinations for credit-certified subjects at the end of each quarter. (Appendix (ref))

We studied the educational effect of the Center, and found that the average deviation in the 1Q Differentiation (hereafter, Diff.) regular examinations for F2F assistance users was 46.8, or 3.5 points lower than the 50.3 average of the nonusers group. At first glance, it appeared that F2F assistance had no positive educational effect on a student's regular examination performance. However, the fact that the average proficiency test deviation for F2F users was lower than that for nonusers suggests the possibility of underestimation due to bias. To address this issue, we applied the framework of causal inference, setting the use of F2F as a treatment (Appendix (ref)).

To estimate a Conditional Average Treatment Effect (CATE) rubin, meta-learners have been introduced in a binary treatment setting kunzel. Extending the binary setting, multiple and multi-level treatments models have been proposed harada, hu, lech, lin,yutas1. Meta-learners for multi-valued treatments makes it possible to identify the impact of the number of possible treatments on CATE performance acha. A more extensive discussion of related studies is presented in the Appendix (ref).

The main contributions of the present study can be summarized as follows tnagai3. (i) Using T-learner, we estimate the CATE of F2F assistance on the 1Q Diff. regular examination. (ii) Extending T-learner, we propose a new CATE estimator that depends on the number of treatments, which is learned with variables that include the number of treatments. (iii) We predict CATE performance depending on the number of F2F sessions.

\@startsection{section}{1}{\z@} {2ex}{1ex}{\reset@font}{Causal Inference Setup } Because our decision tree Breiman results showed that the F2F branches were the most common, we chose F2F as our main focus (Appendix (ref)). In this study, we adopted the potential outcome framework ney, rubin2 and T-learner kunzel. The treatment corresponds to the use of F2F in 1Q. The potential outcome corresponds to the deviation value of 1Q Diff. regular examination. Regarding input variable ${\bf X}$ for the estimators, we considered two types. $X_1$ represents the proficiency test deviation value at the time of admission, which we deal with as a covariate. $X_2$ represents the number of F2F sessions attended by the student in 1Q, which, for us, corresponds to the number of treatments. The number of treatments was included as an input variable (Appendix (ref)).

The data from $N=$ 1,389 students who took 1Q Diff. examination were analyzed. Of this group, 91 students used F2F assistance, while the remaining 1,298 students did not. To estimate CATE, we used random forests forest, implemented by Python scikit-learn.

figure[figure omitted — 178 chars of source]

\@startsection{section}{1}{\z@} {2ex}{1ex}{\reset@font}{Results (Input Variable ${\bf X}, {\bf x} \in \mathbb{ R}^1$)} Initially, the input variable ${\bf X}$ was the one-dimensional vector ${\bf X}=(X_1)$. We estimated the CATE of F2F as a function of $X_1=x_1$ (Appendix (ref)). Fig. (ref) shows the estimated values of deviations of 1Q Diff. with $X_1=x_1$. The value of $\hat{\mu}_0(x_1)$ is the estimated deviation value of 1Q Diff. if F2F is not used ($\textcolor{orange}{\blacktriangle}$ in Fig. (ref)), and $\hat{\mu}_1(x_1)$ is the estimated value if F2F is used ($\textcolor{blue}{\blacklozenge}$). The CATE estimator $\hat{\tau} (x_1)$ ($\textcolor{red}{\bullet}$) is defined by the difference $( \hat{\mu}_1(x_1)-\hat{\mu}_0(x_1))$. For $x_1 \le 35 $, $\hat{\tau} (x_1)$ ranges from 4 to 5. This indicates that using F2F assistance improves the regular examination deviation value in the range of 4 to 5 points. $\hat{\tau} (x_1)$ becomes negative for $37 \le x_1 \le 48$ but becomes positive again for $x_1\ge 50$, showing a value of approximately 2.

Average Treatment Effect ($ATE$) was found to be 0.48 (Appendix (ref)). This means that using F2F assistance results in an average improvement of 0.48 in the 1Q Diff. deviation values. Thus, the effect of F2F was found to be a 0.48 deviation increase rather than the 3.5 deviation decrease described in Section (ref). Average Treatment Effect on the Treated ($ATT$) was 1.00, which is compared to the results in Section (ref).

\@startsection{section}{1}{\z@} {2ex}{1ex}{\reset@font}{Estimation and Results $({\bf X}, {\bf x} \in \mathbb{R}^2)$} Multiple regression analysis was used to estimate the potential effect of the number of F2F sessions on test performance (Appendix (ref)). Accordingly, we explicitly added the number of F2F sessions $X_2$ as an input variable. The input variable ${\bf X}$ is thus the 2-dimensional vector ${\bf X}=(X_1,X_2)$ (Appendix (ref)).

Extending the definition of CATE in T-learner, we propose a new CATE estimator $\varphi (x_1,x_2 )$ with proficiency test deviation value $X_1=x_1$ and the number of F2F sessions $X_2=x_2$, defined as follows:

eqnarray[eqnarray omitted — 203 chars of source]

Here, $k$ indicates the $k$-th out of $N$ students, and $S_{x_1}$ is the set of students whose $X_1=x_1$. $|S_{x_1}|$ is the number of students in $S_{x_1}$. $X_1^{k, obs}$ and $X_2^{k, obs}$ are the $k$-th student observations of $X_1^{k}$ and $X_2^{k}$.

figure[figure omitted — 166 chars of source]

Fig. (ref) shows the CATE estimator $\varphi (x_1,x_2 )$ in Eq. (ref) with the number of F2F sessions (see also 3D surface plot in Appendix (ref)). For proficiency test deviation value $x_1 \leq $ 35, $\varphi (x_1, x_2 )$ decreases from $x_2$=1 time ($\textcolor{orange}{\blacktriangle}$ in Fig. (ref)) to $x_2$=2 times ($\textcolor{blue}{\blacklozenge}$). However, $\varphi (x_1, x_2 )$ increases as $x_2$ increases from $x_2$=3 times ($\textcolor{red}{\bullet}$), 5 times ($\textcolor{green}{\times}$), 10 times ($\textcolor{cyan}{+}$), and 14 times ($\textcolor{purple}{-}$). That is to say, we have $\varphi (x_1,2) < \varphi (x_1,3) < \varphi (x_1,5) \cdots$. It is predicted that three or more F2F sessions contribute to a greater increase in the deviation value of 1Q Diff. as the number of F2F sessions increases. While for $x_1 \geq 50$, $\varphi (x_1,x_2 )$ is roughly 2 if $x_2 \leq 3$ times ($\textcolor{red}{\bullet}$), $\varphi (x_1, x_2)$ becomes approximately 1.4 after 5 sessions ($\textcolor{green}{\times}$). Although a single F2F session leads to a positive $\varphi(x_1, 1)$, it is not predicted that the deviation value of 1Q Diff. increases as the number of F2F sessions increases.

Let $ATT_2$ be the average treatment effect for users with two variables (Appendix (ref)). From the relation of $\hat{\mu}_0 (X_1^{k,obs},X_2^{k,obs})=\hat{\mu}_0 (X_1^{k,obs},0)$, $ATT_2$ should be equal to $ATT$. In fact, we obtained $ATT_2=1.00$, which is equal to $ATT$=1.00.

\@startsection{section}{1}{\z@} {2ex}{1ex}{\reset@font}{Conclusion} In this study, we investigated the CATE of the F2F personal assistance offered by Kogakuin University's Academic Support Center on the deviation value of the 1Q Diff. regular examination by using T-learner. We proposed a CATE estimator that depends on the number of F2F sessions engaged in by students. We optimized the estimator with two variables, including the number of F2F sessions. Using the CATE estimator, we predicted CATE performance.

\@startsection{section}{1}{\z@} {2ex}{1ex}{\reset@font}*{Acknowledgment} This research was supported by the Japan Society for the Promotion of Science KAKENHI Grant Number JP24K06289.

thebibliography{00} \bibitem{Ala}Ahmed M. Alaa and Mihaela van der Schaar. 2017. Bayesian inference of individualized treatment effects using multi-task gaussian processes. In {\it Proceedings of the 31st International Conference on Neural Information Processing Systems} NIPS'17 (2017), 3427-3435. \bibitem{acha} Naoufal Acharki, Ramiro Lugo, Antoine Bertoncello, and Josselin Garnier. 2023. Comparison of meta-learners for estimating multi-valued treatment heterogeneous effects. In {\it Proceedings of the 40th International Conference on Machine Learning} PMLR 202 (2023), 91-132. \bibitem{ba}Berardino Barile, Marco Forti, Alessia Marrocco, and Angelo Castaldo. 2024. Causal impact evaluation of occupational safety policies on firms' default using machine learning uplift modelling. {\it Scientific Reports} 14 (2024), article number: 10380. \bibitem{Breiman}Leo Breiman, Jerome H. Friedman, Richard A. Olshen, and Charles J. Stone. 1984. {\it Classification and Regression Trees}. Wadsworth International Group. \bibitem{forest}Leo Breiman. 2001. Random Forests. {\it Machine Leaning} 45, 1 (2001), 5-32. \bibitem{caronb} Alberto Caron, Gianluca Baio, and Ioanna Manolopoulou. 2020. Estimating individual treatment effects using non-parametric regression models: a review. {\it Journal of Royal Statistical Society Series A} 185, 3 (2022), 1115-1149. \bibitem{fujihara}Sho Fujihara. 2019, Causal inference in studies of sociology of education. {\it Sociological Theory and Methods} 34, 1 (2019), 65-77. (in Japanese) \bibitem{harada}Shonosuke Harada and Hisashi Kashima. 2021. Graphite: estimating individual effects of graph-structured treatments. In {\it Proceedings of the 30th ACM International Conference on Information and Knowledge Management}. 659-668. \bibitem{hu}Liangyuan Hu, Chenyang Gu, Michael Lopez, Jiayi Ji, and Juan Wisnivesky. 2020. Estimation of causal effects of multiple treatments in observational studies with a binary outcome. {\it Statistical methods in medical research} 29, 11 (2020),3218-3234. \bibitem{imbens}Guido W. Imbens. 2000. The role of propensity score in estimating dose-response function. {\it Biometrika} 87, 3 (2000), 706-710. \bibitem{rubin}Guido W. Imbens and Donald B. Rubin. 2015. {\it Causal inference in statistics, social, and biomedical sciences}. Cambridge University Press. \bibitem{kell}Bryan Keller, Zach Branson. 2024. Defining, identifying, and estimating causal effects with the potential outcomes framework: A review for education research. {\it Asia Pacific Education Review} May 2024. https://doi.org/10.1007/s12564-024-09957-2. \bibitem{knu}Michael C. Knaus. 2022. Double machine learning-based programme evaluation under unconfoundedness. {\it The Econometrics Journal} 25, 3 (2022), 602-627. \bibitem{knu2}Michael C. Knaus, Michael Lechner, and Anthony Strittmatter. 2022. A Heterogeneous employment effect of job search programmes: A machine learning approach. {Journal of Human Resources} 57, 2 (2022), 597-636. \bibitem{kunzel} S\"{o}ren R. K\"{u}nzel, Jasjeet S. Sekhon, Peter J. Bickel, and Bin Yu. Metalearners for estimating heterogeneous treatment effects using machine learning. In {\it Proceedings of the national academy of sciences} 116, 10 (2019), 4156-4165. \bibitem{lech}Michael Lechner. 2001. Identification and estimation of causal effects of multiple treatments under the conditional independence assumption. In {\it M. Lechner and F. Pfeiffer (eds.) Econometric Evaluation of Labour Market Policies, Heidelberg: Physica}. 43-58. \bibitem{lin}Lin Lin, Yeying Zhu, and Liang Chen. 2019. Causal inference for multi-level treatments with machine-learned propensity scores. {\it Health Services and Outcomes Research Methodology} 19, 2 (2019), 106-126. \bibitem{luo}Huishi Luo, Fuzhen Zhuang, Ruobing Xie, Hengshu Zhu, Deqing Wang, Zhulin An, and Yongjun Xu. 2024. A survey on causal inference for recommendation. {\it The Innovation} 5, 2 (2024), 100590. https://doi.org/10.1016/j.xinn.2024.100590 \bibitem{tnagai1}Tomoko Nagai, Motoki Kino, Tetsuo Hosoya, Hirohisa Takahashi, Takanao Tsuyuki, Takayuki Muto. 2022. Analysis of educational effects at academic support center in Kogakuin university based on decision tree method. {\it Physics Education in University} 28, 2 (2022), 94-97.(in Japanese) \bibitem{tnagai2} Tomoko Nagai, Kensaku Kinjo, Kengo Kawamura, Tomoya Nakamura, Takayuki Okuda, Yuichiro Sato, Jun-ichi Mukuno, Shin Kikuta, and Naoto Kumano-go. 2023. Educational effects by the division of mathematics of Kogakuin university academic support center: decision tree analysis. {\it Japanese Society for Engineering Education} 71, 3 (2023), 112-116.(in Japanese) \bibitem{tnagai3}Tomoko Nagai, Takayuki Okuda, Tomoya Nakamura, Yuichiro Sato, Yusuke Sato, Kensaku Kinjo, Kengo Kawamura, Shin Kikuta, and Naoto Kumano-go. 2024. Evaluation of Educational effects in mathematics by means of T-learner based upon causal inference. {\it Japanese Society for Engineering Education} in press. (in Japanese) \bibitem{ney} Jerzy Splawa-Neyman, D. M. Dabrowska, T. P. Speed. 1990. On the application of probability theory to agricultural experiments. Essay on principles. Section 9. {\it Statistical Science} 5 (1990),465-472. \bibitem{yutas1} Yuta Saito, Hayato Sakata, and Kazuhide Nakata. 2020. Cost-Effective and Policy Optimaization Algorithm for Uplift Modeling with Multiple Treatments. In {\it Proceedings of the 2020 SIAM International Conference on Data Mining}. 406-414. \bibitem{yutas3}Yuta Saito. 2024. {\it CounterFactual Machine Learning}. Gijutuhyoronsya. (in Japanese) \bibitem{rubin2} Donald B. Rubin. 1974. Estimating causal effects if treatment in randomized and nonrandomized studies. {\it Journal of Educational Psychology} 66, 5 (1974), 688-701.