TL;DR: Machine learning models built on satellite imagery can create detailed poverty maps, but their predictions are systematically “shrunk” toward the average. This shrinkage quietly weakens estimated treatment effects in impact evaluations. Our paper introduces two post-hoc debiasing methods—Linear Calibration Correction (LCC) and Tweedie’s correction—that remove most of this bias without collecting new survey data or retraining models, enabling a true “one map, many trials” workflow for global development research.

Why do we use satellite imagery to map poverty in the first place?

Satellite-based machine learning models can predict household wealth at high spatial resolution, even in regions with limited survey data. These models generate “wealth maps” that approximate indices like the International Wealth Index using information from Earth observation imagery.

For policymakers and researchers in development economics, global health, and related fields, these maps are incredibly valuable. They make it possible to study poverty patterns, track inequality over time, and evaluate interventions (like aid projects or infrastructure programs) across places where ground-truth data are sparse or outdated.

What goes wrong when we reuse ML predictions for causal inference?

Standard machine learning training focuses on minimizing overall prediction error, not on preserving the full spread of the outcome distribution. As a result, model predictions tend to shrink toward the mean: very poor areas are predicted as slightly richer, and very rich areas are predicted as slightly poorer.

When these “shrunk” predictions are used as outcomes in impact evaluations, they introduce attenuation bias. Estimated treatment effects (for example, the impact of an aid program on local wealth) are systematically biased toward zero: real 5–10% gains can show up as only 2–3% in the data, increasing the risk of falsely concluding there is “no effect.”

• Overestimation in poor regions → predicted poverty is too low.
• Underestimation in rich regions → predicted prosperity is too low.
• Net result: treated vs. control differences are compressed, and causal effects are underestimated.

This shrinkage has been observed in real data as well: calibration plots using Demographic and Health Survey (DHS) clusters across Africa show exactly this pattern, with model predictions overestimating the poorest clusters and underestimating the wealthiest ones.

What have existing debiasing methods tried, and why aren’t they enough?

Several existing approaches can correct or reduce attenuation bias, but they come with important practical limitations. Most notably, they often require fresh labeled data downstream or changes to the original training pipeline.

• Prediction-Powered Inference (PPI): adjusts estimates using a “rectifier” term derived from additional ground-truth labels collected specifically for the downstream causal study. This can work well, but it assumes that new high-quality survey data are available when researchers run each new trial.
• Loss-function adjustment (Ratledge et al.): modifies the model’s training objective to penalize group-level (e.g., quintile-level) bias in predictions. This reduces shrinkage, but it requires retraining the model, careful tuning of new hyperparameters, and can reduce overall predictive performance.
• Sampling-aware debiasing (e.g., DSL): uses known sampling probabilities and doubly robust estimation to build corrected pseudo-outcomes, but relies on detailed knowledge of the data-generating and sampling process, which is often unavailable.

In practice, especially in low-resource settings, it is often not realistic to collect new surveys or rerun expensive ML training pipelines every time a new impact evaluation is proposed.

What problem does the paper solve in one sentence?

The paper answers this central question: How can we debias machine learning predictions used for causal inference so that a single, pre-trained poverty map can support many downstream trials without new ground-truth data or retraining?

What new debiasing methods do we propose?

We introduce two practical, post-hoc debiasing methods that operate directly on model predictions:

1. What is Linear Calibration Correction (LCC)?

LCC fits a simple linear relationship between true outcomes and model predictions on a held-out calibration split, then inverts that relationship to rescale predictions back onto the correct scale.

How it works in practice:

• Use part of the original training data as a calibration set.
• Fit a linear model of the form: prediction = k × truth + m.
• Invert that relationship to obtain a corrected outcome: corrected = (prediction − m) / k.

When the shrinkage is approximately linear (which we see in our data), this simple transformation restores the appropriate spread in the outcome distribution and removes most attenuation in estimated treatment effects, without any new labels or retraining.

2. What is Tweedie’s correction in this context?

Tweedie’s correction uses tools from empirical Bayes to locally “de-shrink” predictions by combining an estimate of the noise variance with the density of the prediction distribution.

Intuition:

• Modern ML models often behave like they apply Berkson-type error: they predict a smooth conditional mean and then residual noise is added around that mean.
• When this happens, predictions are systematically pulled toward dense, central regions of the distribution (e.g., the global mean or local modes), especially in rare or noisy regions.
• Tweedie’s formula says that, under a Gaussian noise assumption, you can locally correct this shrinkage by subtracting a term proportional to the derivative of the log-density of predictions (the score function).

In practice, Tweedie’s correction:

1. Estimates the prediction noise level σ² from residuals on the training data.
2. Fits a smooth density model (e.g., via kernel density estimation) to the out-of-sample predictions.
3. Computes a corrected pseudo-outcome: corrected = prediction − σ² × d/dy log p(prediction).

This pseudo-outcome has the key property that its conditional mean matches the true outcome, which is exactly what we want for unbiased group-level and causal estimates.

How did we evaluate these debiasing methods?

We benchmarked LCC and Tweedie’s correction against existing alternatives in both controlled simulations and real-world data from Africa. The comparisons included:

• Naive predictions (no debiasing).
• Prediction-Powered Inference (PPI) with different amounts of fresh downstream labels.
• Ratledge-style loss adjustment trained with quintile-bias penalties.
• Our two post-hoc methods: LCC and Tweedie’s correction.

We focused on two key tasks:

1. Simulated causal experiments, where the true treatment effect is known and we can precisely measure bias and error.
2. Real-world applications using Demographic and Health Survey (DHS) data and geo-referenced aid projects from the World Bank and China, where we compare prediction-based estimates against ground-truth survey outcomes.

What did we learn from simulations?

In simulations where both the true causal effect and the data-generating process are known, Tweedie’s correction consistently delivered the best overall performance.

• Tweedie’s correction achieved the lowest mean absolute error (MAE) for treatment effects and produced a near-perfect calibration slope (essentially, predicted effects matched true effects on average).
• LCC also performed extremely well, recovering the correct scale of treatment effects with only a minimal loss relative to Tweedie’s method.
• Naive predictions and Ratledge-style models had strong shrinkage: their slopes were significantly below 1, indicating systematic underestimation of effect sizes.
• PPI improved estimates as more fresh labels were included, but its performance was tightly constrained by the size and cost of new labeled datasets.

The bottom line from simulations is that post-hoc corrections can nearly eliminate attenuation bias without touching the original training pipeline or collecting new labels.

How well do the methods work on real African DHS and aid data?

We then applied our methods to a large DHS-based dataset of wealth across 30 African countries (around 69,000 clusters from 2009–2020) and to real aid interventions funded by the World Bank and Chinese government.

Descriptive tasks: estimating regional wealth

We first looked at descriptive statistics, such as average wealth by administrative region (e.g., district-level means). Using naive predictions:

• Poor regions were overpredicted and rich regions underpredicted.
• Group-level estimates showed classic shrinkage patterns.

After applying Tweedie’s correction (and similarly LCC):

• The relationship between predicted and true regional means became close to 1:1.
• Mean absolute error decreased, especially in regions with sufficient sample size (e.g., ≥100 clusters).
• Shrinkage was essentially removed in large-sample regions.

Causal tasks: evaluating real aid interventions

Next, we treated funder-sector pairs (e.g., “World Bank, Health” or “China, Transport and Storage”) as separate “trials” and estimated the effect of interventions on local wealth using:

1. Observed wealth from DHS as the benchmark.
2. Debiased ML predictions as surrogate outcomes.

The key finding is that Tweedie’s correction produced treatment-effect estimates that nearly perfectly aligned with the benchmark from observed data, both in terms of scale and ranking of sectors. Naive predictions, by contrast, showed clear attenuation, systematically underestimating the magnitude of impacts.

When should researchers use LCC vs. Tweedie’s correction?

Both methods are practical and lightweight, but they suit slightly different needs.

Use Linear Calibration Correction (LCC) when:

• You have a reasonably sized calibration split from the original training data.
• The predicted vs. true relationship looks approximately linear (slope < 1 but not wildly nonlinear).
• You want a simple, easily interpretable scaling correction.

Use Tweedie’s correction when:

• You want a more flexible, local adjustment that doesn’t assume global linearity.
• You suspect errors are more complex near the tails of the distribution (very poor or very rich regions).
• You can estimate the predictive noise variance and density of predictions with reasonable accuracy.

In our experiments, Tweedie’s correction generally outperformed LCC in terms of both bias and error, but LCC was surprisingly competitive given its simplicity.

What does “one map, many trials” mean in practice?

The “one map, many trials” idea is simple but powerful: a single upstream team builds a high-quality, debiased wealth map that can be safely reused by many downstream evaluation teams, across many different interventions, without retraining or recollecting outcomes.

This has several practical benefits:

• Cost efficiency: expensive household surveys and ML training only need to be done once.
• Reproducibility: multiple studies rely on the same data product with transparent corrections.
• Scalability: one global map can support dozens of randomized trials, quasi-experiments, and observational studies.

Our open-source package, unshrink, implements these correction methods, and the accompanying Africa-wide wealth maps are released so that other researchers can plug them directly into their own causal analyses.

What are the key takeaways for policy and research?

Our results suggest three main takeaways for anyone using ML predictions as outcomes in causal inference:

1. Check for shrinkage and attenuation. Even high-R² models can substantially attenuate treatment effects; calibration plots and group-level comparisons are essential diagnostics.
2. Use post-hoc debiasing when retraining or new labels are not feasible. LCC and Tweedie’s correction can restore nearly unbiased treatment effects using only the original training data.
3. Invest once, reuse many times. Properly debiased outcome maps can unlock a true “one map, many trials” paradigm for development economics, global health, and environmental policy.

As satellite-driven ML systems and other surrogate-based outcome measures become more common, having reliable, well-calibrated methods to debias predictions for causal inference will be critical for ensuring that policy decisions reflect real-world impact rather than artifacts of the modeling pipeline.

FAQ: Debiasing ML Predictions for Causal Inference

Do LCC and Tweedie’s correction require new survey data?

No. Both methods are designed to work without any new ground-truth data at the downstream stage. They rely only on the original training labels and out-of-sample predictions.

Can these methods be used beyond satellite-based poverty mapping?

Yes. The same attenuation and shrinkage issues arise whenever surrogate or ML-predicted outcomes are used in place of direct measurements. This includes settings like air pollution indices, population density, and indicators derived from language models. The methods extend to any similar scenario.

Do these corrections hurt pure predictive performance?

They can slightly increase prediction variance, especially in the tails, which may raise traditional prediction errors (like MSE) in some cases. However, for causal inference and group-level estimates, the gain in calibration and unbiasedness is often worth this trade-off.

Where can I find the code and data to apply these methods?

The paper is accompanied by an open-source package, unshrink, as well as Africa-wide wealth maps and replication code. These resources make it straightforward to apply LCC and Tweedie’s correction in your own impact evaluations and descriptive studies.