TL;DR
This paper introduces a new "calibrated regression estimation" method for combining data from multiple sources, a common challenge in fields like economics and medicine. Their approach uses empirical likelihood to robustly estimate relationships, outperforming existing methods by being less sensitive to extreme values and more resilient to model misspecification.
What Problem Does This Paper Address?
Imagine you're trying to understand a complex phenomenon, but the data you need is scattered across different surveys or datasets. For example, you might have household spending data in one survey and wealth data in another. This is the "data fusion" problem. Unlike typical "missing data" problems where some information is missing for a subject, in data fusion, no single subject has all the information. Existing methods, like "doubly robust estimators," try to handle this by using models for how data is collected (propensity scores) and how missing values are filled in (imputation models). However, these methods often rely on just one of each type of model, which can be problematic if that single model is wrong. The authors tackle the challenge of building more reliable estimators when we can't be sure which of our initial models are correct.
Methods
The core of the proposed method is a "calibration approach" that leverages empirical likelihood. Instead of relying on a single propensity score model and a single imputation model, the authors propose using multiple candidate models for each. This is like having a backup plan for your backup plan!
Here's a breakdown:
- Multiple Nuisance Models: They consider a set of potential propensity score models (how likely a subject is to be in the primary dataset) and a set of potential imputation models (how to fill in missing covariate information).
- Empirical Likelihood: This statistical technique allows them to construct estimators without relying on inverse probability weighting, which can be unstable with extreme propensity scores. It essentially uses the data itself to define the likelihood function.
- Calibration Weights: The method constructs "calibration weights" for subjects in both the primary and auxiliary datasets. These weights are derived by maximizing the product of these weights subject to constraints that ensure the models are balanced. This process is inspired by techniques used in missing data analysis but adapted for the data fusion setting.
- Doubly Robust Property: The resulting estimator is "doubly robust," meaning it's consistent (gives the right answer in the long run) if at least one of the propensity score models is correct OR if at least one of the imputation models is correct. This provides significant protection against model misspecification.
Datasets
The paper uses an assembled dataset combining information from the U.S. Consumer Expenditure Survey (CEX) and the Survey of Consumer Finance (SCF) to estimate the effect of household net worth on consumption. This dataset is not newly introduced but is a standard benchmark for this type of analysis.
Key Insights and Ideas
- Strength in Numbers (of Models): The paper demonstrates that using multiple candidate models for propensity scores and imputation provides substantial robustness. It's not about finding the perfect model, but about having enough good candidates to protect against errors.
- Empirical Likelihood as a Stabilizer: By using empirical likelihood, the method avoids the instability that can arise from extreme propensity score estimates, a common issue with inverse probability weighting.
- "Doubly Robust" Gets Stronger: The paper extends the concept of double robustness to a scenario with multiple nuisance models, offering a more powerful safeguard against misspecification.
- Counter-Intuitive Efficiency Gain: The authors show that even if the propensity score model is known, modeling it with multiple candidates can still improve efficiency. This is a surprising but well-supported finding in robust statistics.
Key Results
- Simulation Studies: The proposed "CAL" (calibration) estimators consistently outperformed existing "DR" (doubly robust) estimators in simulations, especially when models were misspecified. The CAL estimators showed lower bias and root mean squared error (RMSE). For instance, when one propensity score and one imputation model were used, CAL-1001 (correct propensity, incorrect imputation) was more efficient than DR-1001.
- Application Results: In the analysis of U.S. household expenditure data, the proposed CAL-1111 estimator (using all four working models) provided results that were robust and comparable to other well-specified models. The paper highlights that when the doubly robust estimator used misspecified models (DR-0101), the results were significantly different and potentially misleading.
Limitations and Future Work
- Parametric Models: The current approach relies on parametric models for propensity scores and imputation. The authors suggest that incorporating machine learning techniques could further enhance robustness.
- Mean Regression: The paper focuses on mean regression. Extending the methodology to quantile regression problems in data fusion settings is proposed as a future research direction.
- Model Selection: While the method is robust to having some incorrect models, the performance can still be affected if all models are misspecified. The paper doesn't delve deeply into optimal strategies for selecting the initial set of candidate models.
Summary updated