A scalable, streaming-ready alternative to classical entropy balancing for survey research and causal inference.

In applied survey research and causal inference, entropy balancing is used to reweight observational data so that covariate distributions match known or desired population margins. The standard approach solves a convex optimization problem that minimizes KL divergence from base weights while enforcing exact moment-matching constraints.

However, real-world applications increasingly demand:

1. Streaming data support - Reweighting data that arrives in batches (e.g., polling waves, rolling panels)
2. Dynamic target adjustment - Adjusting to updated calibration targets (e.g., revised Census benchmarks)
3. High-dimensional scalability - Scaling to thousands of covariates from rich administrative data

The classic entropy balancing method (typically implemented via BFGS optimization) has significant limitations:

• Not adaptive — requires solving from scratch each time targets or data change
• Not scalable — struggles or fails when the number of covariates is large (d > 100)
• Not streaming-ready — assumes the full dataset is available all at once

We replace closed-form optimization with a Multiplicative Weights Update algorithm that:

1. Maintains weights via iterative exponential updates - Simple, stable weight adjustments
2. Operates in batch or mini-batch mode - Supports both traditional and streaming workflows
3. Adapts on the fly to changing target moments - No restart required when targets change

This transforms entropy balancing into a no-regret learning process over marginal constraints.

Given sample covariates from n observations with d features and target population moments, find weights such that:

• Weighted sample moments match population targets exactly
• All weights are positive and sum to one

Instead of solving a constrained optimization problem directly, MWU:

1. Initializes weights uniformly across observations
2. Computes current moment error (difference between weighted sample and targets)
3. Updates weights exponentially based on each observation's contribution to error
4. Normalizes weights to maintain valid probability distribution
5. Repeats until convergence

The learning rate parameter controls the step size, balancing convergence speed with stability.

• Batch Mode: Updates using complete dataset each iteration (fastest convergence)
• Streaming Mode: Processes mini-batches over multiple epochs (memory efficient)
• Adaptive Mode: Continues from current state when targets change (no restart cost)

We tested on synthetic datasets with n=3000 observations:

Key findings: MWU matches classical entropy balancing accuracy, adapts instantly to target shifts, and is an order-of-magnitude faster in high dimensions.

• Flexibility: Handles target changes without restarting optimization
• Scalability: Runtime scales linearly in sample size and dimensions; no matrix inversions required
• Streaming-ready: Same exponential update works in mini-batch or online settings
• Interpretability: Weights remain strictly positive and controlled by learning rate parameter

MWU offers a drop-in alternative to classical entropy balancing when you need:

• Adaptive reweighting for changing calibration targets
• Streaming data processing for large datasets or real-time applications
• High-dimensional performance with hundreds or thousands of covariates
• Memory efficiency for datasets that don't fit in memory

The MWU update is a no-regret algorithm that converges to the classical entropy balancing solution under standard mirror-descent guarantees. This ensures theoretical soundness while providing practical advantages.

• Survey research: Dynamic panel reweighting as new waves arrive
• Market research: Real-time demographic balancing for online panels
• Clinical trials: Adaptive treatment assignment balancing
• Administrative data: Efficient processing of large government datasets

• Hainmueller, J. (2012). "Entropy Balancing for Causal Effects: A Multivariate Reweighting Method to Produce Balanced Samples in Observational Studies." Political Analysis, 20(1), 25-46.
• Arora, S., Hazan, E., and Kale, S. (2012). "The Multiplicative Weights Update Method: A Meta-Algorithm and Applications." Theory of Computing, 8(1), 121-164.