Abstract
Accurate poverty targeting in developing nations often relies on the Proxy Means Test (PMT), which estimates household expenditure from observable characteristics. However, PMT models frequently suffer from distribution shifts over time and inconsistencies between regression and ranking tasks. This paper introduces TrAdaBoost.JQL, a novel domain adaptation method that optimizes a joint loss function. By simultaneously addressing regression and ordinal classification, the model ensures that estimated expenditures align with poverty status rankings. Experimental results on Indonesian socioeconomic data demonstrate that TrAdaBoost.JQL significantly reduces targeting errors and outperforms existing domain adaptation techniques.
Introduction
Identifying the "poor" for social assistance programs is a critical challenge for governments. In many developing nations, direct income verification is difficult, leading to the widespread use of the Proxy Means Test (PMT). PMT models use observable household characteristics—such as housing materials, asset ownership, and demographics—to estimate monthly per capita expenditure (pcexp).
Two major hurdles impede the effectiveness of standard PMT models. First, distribution shift occurs when models trained on historical data (source) are applied to current populations (target) where the relationship between assets and expenditure has changed. Second, task inconsistency arises because PMT is a dual-task problem: it requires both precise value estimation (regression) and accurate household ranking (ordinal classification). Standard models optimized for regression may produce small numerical errors that lead to significant "Inclusion Errors" (providing aid to the non-poor) or "Exclusion Errors" (denying aid to the poor).
Methodology
The TrAdaBoost.JQL Framework
The proposed method, TrAdaBoost.JQL (Joint Quantile Loss), extends the TrAdaBoost algorithm by incorporating a multi-objective loss function. The core innovation lies in the Joint Quantile Loss ($e_{JQL}$), which balances two distinct error types:
- Regression Error ($\tilde{e}_R$): The standard normalized absolute error of expenditure estimation.
- Quantile Ranking Error ($\tilde{e}_Q$): A measure of the discrepancy between the predicted expenditure quantile and the true quantile.
A hyperparameter, $\lambda$, allows practitioners to tune the balance between these two objectives. By integrating this joint loss into a boosting mechanism, the algorithm iteratively reweights data. It decreases the influence of unhelpful historical source data while increasing the weight of a small "guide" set of current target data, forcing the model to learn the most relevant patterns for both estimation and ranking.
As shown in the sensitivity analysis, the choice of $\lambda$ significantly impacts performance. A balanced $\lambda$ (typically around 0.4 to 0.5) yields the highest average reduction in both Root Mean Squared Error (RMSE) and Inclusion/Exclusion Errors (IE/EE) across various eligibility thresholds.
Experimental Evaluation
Dataset and Setup
The study utilized the SUSENAS dataset, Indonesia’s National Socioeconomic Survey, spanning 2016 to 2020. The evaluation covered 16 districts (12 rural, 4 urban) and included 60 features ranging from housing conditions to employment status. To simulate real-world concept drift, the authors tested scenarios using 2016 data as the target (with 2017–2020 as source) and vice versa.
Performance Analysis
TrAdaBoost.JQL was compared against baseline XGBoost and seven other domain adaptation methods, including deep learning approaches like WANN and IWN. The results indicated that traditional reweighting methods often outperform modern deep learning DA on tabular poverty data, likely due to the relatively small sample sizes and the specific nature of socioeconomic features.
Key results include:
- RMSE Reduction: TrAdaBoost.JQL reduced RMSE by up to 22% in urban districts and 21% in rural districts compared to the baseline.
- Targeting Accuracy: The method reduced critical Inclusion and Exclusion Errors by up to 36% for the 2016 target set at a 20% eligibility threshold.
- Consistency: Unlike volatile alternative methods, TrAdaBoost.JQL consistently ranked as the top or near-top performer across all 16 districts, regardless of urban/rural classification or the specific target year.
Discussion and Conclusion
The success of TrAdaBoost.JQL highlights the importance of treating ranking as "rich" information. By incorporating ordinal classification error, the model retains "useful" instances that standard regression-only models might discard. Furthermore, the method proves robust even with very small "guide" samples (as low as 1.25% of the source data), making it highly practical for regions where new data collection is expensive.
Future work may involve exploring the sensitivity of $\lambda$ to tighter poverty thresholds (e.g., the bottom 5%) and integrating explicit outlier detection to further refine performance in districts with high economic variance. Overall, TrAdaBoost.JQL provides a robust, consistent framework for improving the accuracy of social assistance targeting in the face of shifting economic landscapes.
Summary updated