Borrowing Information from an Unidentifiable Model: Guaranteed Efficiency Gain with a Dichotomized Outcome in the External Data

1. Problem Statement

The paper addresses the challenge of data integration (or data fusion) in scenarios where two distinct data sources are available but possess mismatched outcome scales:

• Target Data ($T$): Contains a continuous outcome $Y$ and covariates $X$. The goal is to estimate the regression parameter $\beta$ in a linear model $Y = \beta^T X + \epsilon$.
• External Data ($E$): Contains the same covariates $X$ but only a dichotomized version of the outcome, $Z = \mathbb{I}(Y \leq c)$, where $c$ is a known cutoff.

Key Challenges:

1. Non-Identifiability: The parameter $\beta$ is not identifiable using the external data alone. The model $P(Z=1|X) = F(c - \beta^T X, X)$ involves both the unknown parameter $\beta$ and the unknown cumulative distribution function (CDF) $F(\cdot)$. Without the continuous $Y$ from the target data, $\beta$ cannot be uniquely determined.
2. Model Misspecification: In practice, the error distribution $f(\epsilon|X)$ in the target data is rarely known. Standard estimators (like Weighted Least Squares, WLS) often rely on a "working" model for the error distribution which may be misspecified.
3. Efficiency vs. Robustness: The authors seek to leverage the large external dataset to improve the efficiency of estimating $\beta$ without sacrificing the robustness of the estimator if the error distribution is misspecified.

2. Methodology

The authors develop a semiparametric framework to combine the two data sources.

A. Theoretical Foundation

• Likelihood Construction: They construct a joint likelihood for the combined population (Target + External) using a binary indicator $R$ (1 for Target, 0 for External).
• Efficient Score Derivation: Using semiparametric theory (Bickel et al., 1993; Tsiatis, 2006), they derive the nuisance tangent space orthogonal complement ($\Lambda^\perp$). This space contains all valid estimating equations.
• The Efficient Score ($S_{eff}$): They derive the optimal efficient score function for $\beta$. This score is a weighted average of:
  1. The efficient score from the target data (involving the continuous $Y$).
  2. A score component derived from the external data (involving the dichotomized $Z$).
     Crucially, the weight depends on the unknown error distribution and the probability of being in the target population.

B. Proposed Estimators

The paper proposes two distinct estimators:

1. Estimator 1: $\hat{\beta}^*_{eff}$ (Locally Efficient Score)

   • Approach: This estimator uses the derived efficient score $S_{eff}$ but substitutes the true, unknown error distribution $f(\epsilon, x)$ with a potentially misspecified working model $f^*(\epsilon, x)$.
   • Properties: It remains asymptotically consistent even if the working model is wrong. However, it does not guarantee an efficiency gain over the standard WLS estimator (which uses only target data) if the working model is severely misspecified.

2. Estimator 2: $\tilde{\beta}^*$ (Guaranteed Efficiency Gain)

   • Approach: To overcome the limitation of the first estimator, the authors propose a convex combination (weighted average) of the standard WLS estimator ($\hat{\beta}^*_{ls}$) and the first estimator ($\hat{\beta}^*_{eff}$):
     $$\tilde{\beta}^* = \hat{\beta}^*_{ls} + W (\hat{\beta}^*_{eff} - \hat{\beta}^*_{ls})$$
   • Weight Matrix ($W$): $W$ is chosen to minimize the asymptotic variance of the resulting estimator. Specifically, $W = -\text{cov}(\hat{\beta}^*_{ls}, \hat{\beta}^*_{eff} - \hat{\beta}^*_{ls}) \times [\text{var}(\hat{\beta}^*_{eff} - \hat{\beta}^*_{ls})]^{-1}$.
   • Key Theoretical Result: This estimator guarantees an efficiency gain (lower asymptotic variance) over the standard WLS estimator, regardless of whether the working error model is correctly specified. It effectively "borrows" information from the unidentifiable external model without risking efficiency loss.

3. Key Contributions

1. Solving the Unidentifiable Integration Problem: The paper demonstrates how to extract useful information from an external dataset where the outcome is dichotomized (making the parameter unidentifiable in isolation) and combine it with a target dataset to improve estimation.
2. Robustness to Misspecification: The proposed methods do not require the correct specification of the error distribution $f(\epsilon, x)$. They utilize a working model but maintain consistency even if that model is wrong.
3. Guaranteed Efficiency: Unlike many data fusion methods that risk efficiency loss if assumptions are violated, the second proposed estimator ($\tilde{\beta}^*$) mathematically guarantees that the variance will be no larger (and typically smaller) than using the target data alone.
4. Practical Implementation: The authors provide a weighted bootstrap procedure to estimate the complex covariance matrices required for inference, making the method computationally feasible.

4. Results

Simulation Studies

The authors conducted extensive simulations under two scenarios:

• Scenario 1: Target and external populations share the same covariate distribution.
• Scenario 2: Target and external populations have different covariate distributions (requiring estimation of the propensity score $P(R=1|X)$).

Findings:

• Consistency: All proposed estimators were asymptotically unbiased regardless of whether the error distribution was correctly specified or misspecified (e.g., using a Normal or Logistic working model when the truth was a mixture or heavy-tailed $t$-distribution).
• Efficiency:
  • $\hat{\beta}^*_{eff}$ generally outperformed the target-only WLS estimator, but in some misspecification cases, the gain was marginal.
  • $\tilde{\beta}^*$ consistently demonstrated the lowest variance (highest efficiency) across all scenarios, confirming the theoretical guarantee of efficiency gain.
• Benchmark Failure: Naive benchmarks (e.g., dichotomizing the target data and running logistic regression, or simple meta-analysis) resulted in substantial bias and poor coverage rates, highlighting the necessity of the proposed integrated approach.

Real Data Application (NHANES)

The method was applied to the National Health and Nutrition Examination Survey (NHANES) to study factors affecting Body Mass Index (BMI).

• Setup: Target data had continuous BMI; external data had a binary "Overweight" indicator ($BMI \geq 25$).
• Outcome: The proposed estimator \tilde{\beta}^ identified significant associations (e.g., between physical activity and BMI) that were not statistically significant when using only the target data.
• Validation: The results aligned with medical knowledge, whereas the benchmark methods produced estimates contradicting established science (e.g., suggesting negative associations where positive ones are expected).

5. Significance

This paper makes a significant contribution to the fields of biostatistics and data fusion by:

• Bridging Data Gaps: It provides a rigorous statistical solution for integrating high-quality continuous data with abundant, lower-resolution (dichotomized) data, a common scenario in public health and genomics.
• Safety in Integration: It resolves the fear that integrating "noisy" or "misaligned" external data might degrade the quality of the primary analysis. The "Guaranteed Efficiency Gain" property ensures that researchers can safely incorporate external data to boost power without risking efficiency loss.
• Semiparametric Flexibility: By avoiding strict parametric assumptions on the error distribution, the methods are robust and applicable to complex real-world datasets where distributional assumptions are often violated.

In summary, the paper offers a theoretically sound and practically robust framework for leveraging external data sources that would otherwise be unusable for parameter estimation, thereby enhancing statistical power and scientific insight in the era of big data.