Listening to the Philippines
Listening to the Philippines
When income data is unavailable, household assets provide an alternative. Five statistical methods, one welfare index, built for poverty targeting in the Philippines.
L2Phl CAPI Baseline Survey · Sep–Oct 2025
The L2Phl CAPI baseline posed an immediate challenge: the income module (M04) contained translation errors that rendered monetary values unusable for analysis. Rather than abandon poverty measurement, we turned to an established alternative, observable household assets and living conditions.
This approach has precedent in the literature. Assets, housing quality, and infrastructure access are commonly used as proxies for long-term household welfare, particularly when income data from a single survey round is unreliable. These indicators are less subject to recall bias than self-reported annual earnings.
We applied five statistical methods, each using 40 to 77 household indicators, to rank 2,470 households from the lowest quintile (Q1) to the highest (Q5). Geography, education, and asset ownership were consistently associated with welfare ranking across all five methods.
Method 5 (Ridge PMT) was selected as the primary approach. It follows the World Bank Proxy Means Test standard, used by DSWD to target the Pantawid Pamilyang Pilipino Program (4Ps). Its linear coefficients are transparent and show which household characteristics predict welfare scores.
*M3 overfits to v2 responses. Used only as sensitivity check. PMT preferred for policy use.
"The Ridge PMT produces a transparent, cross-validated welfare prediction from 77 observable indicators, following the standard approach for poverty targeting."
Each method approaches welfare measurement from a different statistical tradition. Methods 1 and 2 are unsupervised: they rank households without reference to any self-reported welfare variable. Methods 3 and 5 are supervised: they use the M14 v2 self-reported welfare scale as a training target. Method 4 finds natural groupings in the indicator space. The table below summarizes each approach, its inputs, and its validation against v2.
Principal Component Analysis on 40 binary DHS-style indicators covering dwelling materials, assets, fuel, utilities, and sanitation. The first principal component (PC1) captures the single dimension of greatest variation across households. PC1 explains 10.2 percent of total indicator variance, consistent with the 10 to 25 percent range observed in DHS surveys across 90 countries.
Indicators load on PC1 as expected: car ownership, concrete walls, piped water, and flush toilets load positively (associated with higher welfare), while nipa roofing, surface water, and open defecation load negatively.
This method is location-free: no region dummies are included, making the resulting index internationally comparable. It can be benchmarked against the Philippines DHS 2022 wealth index, which uses a near-identical indicator set.
Unsupervised. No welfare target used in estimation. DHS-comparable across countries.
Adds 15 indicators to the M1 core set: household head's education level, employment status, written contract, financial inclusion (bank account, mobile money, savings capacity, emergency fund), OFW remittances, agricultural income, food sourcing (supermarket vs. wet market), dependency ratio, and disability status. Total: 55 indicators.
PC1 explains 6.3 percent of variance in the extended set, lower than Core PCA because the additional variables introduce dimensions (education, labour market) that do not perfectly align with asset wealth. The correlation with M1 is 0.935 (Spearman), and 74 percent of households are assigned to the same quintile. The correlation with v2 is marginally lower at 0.293 vs. 0.302 for Core PCA.
The extended set is useful when the analysis requires a broader welfare concept that includes human capital and financial access, not only physical assets.
Unsupervised. Broader welfare concept. Near-identical quintile assignment to M1.
A supervised ensemble of 500 decision trees, each trained on a bootstrap sample of households and a random subset of 25 indicators per split. The target variable is M14 v2 (self-reported welfare on a 1 to 5 scale). Each tree partitions the indicator space into regions where households report similar welfare levels. The final prediction averages across all 500 trees.
Random Forest achieves the highest correlation with v2 by far: Spearman ρ = 0.894, with an OOB RMSE of 0.805. Wall construction (concrete vs. nipa) is the most important splitting variable (14.5 percent relative importance), followed by geographic indicators and sanitation type.
The high ρ is partly mechanical: the model is trained on the same welfare variable it is validated against. This limits its use as an independent welfare measure. It is best treated as a sensitivity check, showing what an unconstrained nonlinear learner would produce from the same indicator set.
Supervised. Overfits to v2 by design. Used as sensitivity check only.
K-medians clustering applied to the first 10 principal components from M1, which together explain 34 percent of total indicator variance. The algorithm partitions households into K groups by minimizing the sum of absolute deviations from each cluster median, a method more robust to outliers than K-means.
The optimal number of clusters is K = 3 (Calinski-Harabasz index = 364.6), indicating three natural welfare classes in the data. For comparability with the other methods, results are mapped to K = 5 quintiles. Cluster sizes range from 15.8 to 23.1 percent of households, indicating unequal natural groupings.
The correlation with v2 is the lowest of all methods (ρ = 0.233), but the correlation with M1 PCA is 0.796. K-medians identifies structural asset groupings that do not closely correspond to how households perceive their own welfare, consistent with the literature showing that asset-based clusters capture long-run wealth rather than current subjective wellbeing.
Unsupervised. Finds natural groupings. Mapped to quintiles for comparability.
A supervised linear regression with L2 (ridge) regularization, predicting M14 v2 from 77 standardized household indicators plus 21 region dummies. The ridge penalty (λ = 100, selected by 5-fold cross-validation) shrinks all coefficients toward zero without setting any exactly to zero, producing a stable, interpretable linear model.
This approach follows the World Bank Proxy Means Test (PMT) framework used by DSWD for 4Ps beneficiary identification. The PMT methodology has been applied in poverty targeting programs across 30 countries. Its key advantage is coefficient transparency: each predictor's contribution to the welfare score can be inspected and explained to policymakers.
The 5-fold CV Spearman ρ is 0.272 (out-of-sample), while the quintile-level ρ against v2 is 0.410. The top coefficient is head's tertiary education (+0.084), followed by Central Visayas region dummy (-0.078, negative) and car ownership (+0.074). Six of the top 15 coefficients are region dummies, indicating that geographic location is as strongly associated with welfare as individual household characteristics.
Supervised. PMT standard. Transparent coefficients. Primary method for L2Phl analysis.
| Method | Type | Features | ρ vs v2 | Best Use |
|---|---|---|---|---|
| M1 Core PCA | Unsupervised | 40 | 0.302 | DHS-comparable, cross-country benchmarking |
| M2 Extended PCA | Unsupervised | 55 | 0.293 | Broader welfare concept (+ education, finance) |
| M3 Random Forest | Supervised | 39 | 0.894* | Sensitivity check (overfits to v2) |
| M4 K-Medians | Unsupervised | 10 PCs | 0.233 | Natural welfare groupings (optimal K = 3) |
| M5 Ridge PMT | Supervised | 77 + 21 | 0.410 | Primary: social protection targeting (4Ps) |
*M3 overfits to v2 by construction. Used as upper bound only.
The Ridge PMT assigns a single linear welfare score to each household. The top 15 coefficients indicate which characteristics are most strongly associated with welfare ranking. Head's tertiary education has the largest coefficient (+0.084). Six of the top 15 predictors are region dummies, indicating that geographic location is associated with welfare as strongly as individual household characteristics. The remaining top predictors are infrastructure and assets: car ownership, flush toilets, piped water, electricity, and the ability to save.
Central Visayas, Davao Region, and North Mindanao have negative coefficients, indicating lower predicted welfare scores in these regions, controlling for all other factors. In contrast, Negros Island and Eastern Visayas have positive coefficients. This regional variation in the PMT coefficients informs the spatial dimension of welfare targeting.
"6 of the top 15 predictors are region dummies, indicating that geographic location is as strongly associated with welfare as household asset ownership."
Households accumulate assets steadily across quintiles. Car ownership shows the largest gradient: the rate was 89.8 percent in Q5 compared with 9.3 percent in Q1. Access to piped water rises from 35.3 percent in Q1 to 83.1 percent in Q5. Electricity access, while high across all quintiles, still differs by 16.5 percentage points between Q1 (83.5 percent) and Q5 (100 percent).
These asset gradients are correlated with cumulative economic outcomes. Households in the lowest quintile tend to lack multiple assets simultaneously, indicating that asset deprivation clusters across dimensions.
Car ownership shows the largest gradient: the rate was 89.8 percent in Q5 compared with 9.3 percent in Q1. Electricity access differs by 16.5 percentage points between Q1 (83.5 percent) and Q5 (100 percent). Across indicators, households tend to rank similarly, consistent with asset clustering.
The geographic dimension of the welfare index is large. NCR households score +0.739 standard deviations above the national mean, while Visayas households score -0.413 SD below, a difference of 1.15 standard deviations. For comparison, the urban-rural gap is 0.65 SD, indicating that regional location explains more welfare variation than urbanization status in these data.
Region dummies in the PMT model are associated with these differences. Central Visayas (reg_7) and Davao Region (reg_11) have negative coefficients, while Eastern Visayas (reg_8) and Negros Island (reg_18) have positive coefficients. These regional patterns persist after controlling for individual household characteristics.
In terms of quintile composition, 77 percent of Q5 households are in Luzon (including NCR), while 61 percent of Q1 households are in Visayas and Mindanao.
"The NCR-to-Visayas gap of 1.15 SD exceeds the urban-rural gap of 0.65 SD, indicating that regional location is more associated with welfare variation than urbanization in these data."
Five different statistical approaches produced five different welfare rankings. Method 1 (Core PCA) and Method 2 (Extended PCA) agree 74 percent of the time, placing the same household in the same quintile. The consensus quintile (median across all five methods) correlates at 0.938 with M1 Core PCA and at 0.392 with self-reported welfare (v2).
At least 3 of the 5 methods assign the same quintile for 73.5 percent of households. For 34.1 percent, 4 or more methods agree. For 13.2 percent, all 5 methods assign the same quintile. Random assignment would yield 20 percent agreement at the 3-method level, so the observed 73.5 percent indicates substantial convergence across methods.
The consensus quintile, defined as the median across all five methods, correlates at 0.938 with M1 Core PCA, indicating that the welfare ranking is consistent across methods.
We recommend Method 5 (Ridge PMT) as the primary welfare ranking for the L2Phl analysis. It offers transparent linear coefficients, is regularized against overfitting through cross-validation, includes geographic variables needed for spatial targeting, and follows World Bank Proxy Means Test standards used by DSWD for 4Ps beneficiary selection.
Each of the five methods serves a purpose. Core PCA (M1) is location-free and internationally comparable, enabling benchmarking against DHS surveys and other countries' welfare indices. The Random Forest (M3) can identify nonlinear patterns as a sensitivity check, though it tends to overfit. The consensus quintile offers a median estimate when uncertainty is high.
Method 1: Core PCA Location-free, approximately 40 indicators, comparable to Philippines DHS and 90+ country surveys.
Method 5: Ridge PMT 77 features + 21 region dummies. Transparent linear coefficients. Standard for 4Ps targeting. Correlation with self-reported welfare: 0.410.
Consensus Quintile Median of all 5 methods. Households assigned to the same quintile by 3 or more methods are classified with high confidence. Suitable for sensitivity analysis.
"For the L2Phl poverty analysis, Method 5 (Ridge PMT) is the primary welfare ranking. Its transparent coefficients, cross-validated regularization, and inclusion of geographic variables make it suitable for poverty targeting and policy analysis."
Five methods, 77 indicators, and one welfare ranking constructed from household assets, housing conditions, and geographic location. The result is transparent and suitable for policy analysis, based on observable household characteristics from the L2Phl CAPI baseline.