Causal Machine Learning for Cost-Effective Allocation of Development Aid

TL;DR

A causal machine learning framework for predicting heterogeneous treatment effects of aid disbursements to inform effective aid allocation is developed, suggesting that the total number of new HIV infections could be reduced by up to 3.3% compared to the current allocation practice.

Abstract

The Sustainable Development Goals (SDGs) of the United Nations provide a blueprint of a better future by 'leaving no one behind', and, to achieve the SDGs by 2030, poor countries require immense volumes of development aid. In this paper, we develop a causal machine learning framework for predicting heterogeneous treatment effects of aid disbursements to inform effective aid allocation. Specifically, our framework comprises three components: (i) a balancing autoencoder that uses representation learning to embed high-dimensional country characteristics while addressing treatment selection bias; (ii) a counterfactual generator to compute counterfactual outcomes for varying aid volumes to address small sample-size settings; and (iii) an inference model that is used to predict heterogeneous treatment-response curves. We demonstrate the effectiveness of our framework using data with official development aid earmarked to end HIV/AIDS in 105 countries, amounting to more than USD 5.2 billion. For this, we first show that our framework successfully computes heterogeneous treatment-response curves using semi-synthetic data. Then, we demonstrate our framework using real-world HIV data. Our framework points to large opportunities for a more effective aid allocation, suggesting that the total number of new HIV infections could be reduced by up to 3.3% (~50,000 cases) compared to the current allocation practice.

Figures (21)

• Figure 1: Causal structure of our problem setup. Our aim is to predict the effect of development aid ($A$) on the reduction in the HIV infection rate ($Y$), while controlling for various country characteristics such as socioeconomic, macroeconomic, and health-related covariates ($X_{1}, X_{2}, \ldots, X_{p}$) as potential confounders. We have a cross-sectional setting without time dependencies.
• Figure 2: Overview of our machine learning framework. The aim of our CG-CT is to predict heterogeneous treatment effects of aid disbursements on SDG outcomes. For this, our CG-CT proceeds along three components. In the first component, country characteristics (e.g., socioeconomic, macroeconomic, or health-related controls) are embedded in a lower dimension using a balancing autoencoder. Here, we also address treatment selection bias by learning a representation of covariates that is not predictive of the treatment (implemented via gradient reversal layer). In the second component, counterfactual outcomes are generated for varying treatment values, and then combined with the observational data. In the third component, an inference model is used together with the previous data to learn a relationship between outcome ($Y$) and treatment ($A$) for a given representation of covariates ($Z$). This gives the treatment--response curve.
• Figure 3: Results for predicted aid--response curves.(a) Results from the experiments with semi-synthetic data. (b) Results from the experiments with real-world data. (c) Predictions for aid--response curves in six example countries. The predicted aid--response curves for the remaining countries are provided in supplements in our GitHub repository. Vertical dashed line denotes the actual volume of development aid as observed in 2017. The $x$-axis is set to $A_\text{obs}$$\pm$$\hat{\sigma}_{A}$ (with a cut-off at zero to prevent negative values for Burundi and India), where $A_\text{obs}$ is the observed development aid and $\hat{\sigma}_{A}$ is the estimated standard deviation of development aid in 2017.
• Figure 4: Suggested vs. current aid allocation.(a) Reduction in the expected number of new HIV infections worldwide under the suggested allocation vs. the current allocation of development aid. (b) Change in the aid volume between the suggested and the current aid allocation (in USD millions). Gray: countries that were not aid recipients in 2017.
• Figure S1: Overview of HIV infections and aid disbursements.(a) HIV infection rate (i.e., the annual number of new HIV infections per 1,000 uninfected people in the population) per country in 2017. The subplot in the bottom-left corner shows five countries with the largest number of new HIV infections in 2017. (b) Volume of HIV development aid in USD millions, in 2017. The subplot in the bottom-left corner shows five countries that were the largest recipients of development aid earmarked to end the HIV epidemic in 2017. Gray: countries that were not aid recipients in 2017.