📊 statistics

Analysis of Linked Files: A Missing Data Perspective

This paper frames record linkage as a missing data problem to categorize, summarize, and evaluate the assumptions and performance of likelihood, imputation, and weighting methods for addressing linkage errors in data analysis.

Original authors: Gauri Kamat, Roee Gutman

Published 2026-04-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Gauri Kamat, Roee Gutman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but you don't have a single, perfect ID card for the people you are investigating. Instead, you have two different notebooks filled with clues.

Notebook A has names, addresses, and a list of crimes committed.
Notebook B has names, addresses, and a list of medical treatments received.

Your goal is to figure out: "Do the people who committed crimes in Notebook A also have medical issues in Notebook B?"

To do this, you have to link the records. You look at the names and addresses to guess which person in Notebook A is the same person in Notebook B. This is called Record Linkage.

The Problem: The "Guessing Game"

The problem is that names get misspelled, addresses change, and people move. Your linking algorithm is like a detective with a slightly blurry magnifying glass. Sometimes, it misses a match (thinking two people are different when they are the same). Sometimes, it makes a mistake and links two different people together.

If you just take your "linked" list and start doing math on it, your results will be wrong. It's like trying to calculate the average height of a basketball team, but you accidentally included a few toddlers because you misread their names. Your average will be too low, and your conclusions will be garbage.

The Paper's Big Idea: The "Missing Puzzle Piece"

The authors of this paper, Gauri Kamat and Roee Gutman, say: "Stop pretending the links are perfect. Treat the uncertainty like a missing puzzle piece."

They propose looking at record linkage through the lens of Missing Data.

In statistics, when data is missing, we don't just ignore it; we use special techniques to guess what it might be and account for the fact that we aren't 100% sure.
Here, the "missing data" is the truth about whether two records are actually the same person. We never truly know for sure without a magic ID card.

The paper organizes all the ways statisticians try to fix this mess into three main "toolkits":

1. The "Super-Computer" Toolkit (Likelihood & Bayesian Methods)

The Analogy: Imagine a super-smart detective who doesn't just look at the clues once. They run thousands of simulations in their head. They say, "Okay, if I assume these two are the same, what does the rest of the data look like? What if I assume they are different? Let's try both and see which story makes more sense."
How it works: These methods use complex math to create a "probability cloud" of all possible ways the files could be linked. They don't pick one answer; they keep all the possibilities in mind and average the results.
Pros: Very accurate if the computer can handle the heavy lifting.
Cons: It's computationally expensive (takes a lot of time and power).

2. The "Fill-in-the-Blanks" Toolkit (Imputation Methods)

The Analogy: Imagine you have a crossword puzzle with some missing words. Instead of guessing just once, you fill in the blanks with three different plausible words. You solve the puzzle three times. Then, you look at the three answers and combine them to get a final, more reliable solution.
How it works: This is called Multiple Imputation. You create several "fake" versions of the linked file, each with slightly different guesses about who is linked to whom. You analyze each version separately, then combine the results. This naturally accounts for the fact that your guesses might be wrong.
Pros: Easier to use than the Super-Computer method; works well with standard statistical software.
Cons: You have to be careful that your "guessing rules" match the "answering rules," or you might get biased results.

3. The "Weighted Scale" Toolkit (Weighting Methods)

The Analogy: Imagine you are weighing fruit on a scale, but you know the scale is a bit wobbly. Instead of trying to fix the scale, you just give more weight to the fruits you are sure about and less weight to the ones you are unsure about.
How it works: If the algorithm says, "I'm 90% sure these two records match," you give that data point a heavy weight in your final calculation. If it says, "I'm only 50% sure," you give it a light weight. This corrects the bias without needing to re-link the files.
Pros: Fast and simple.
Cons: Only works well in specific situations (like simple linear equations) and requires knowing how "wobbly" your scale is.

What Did They Find? (The Simulation)

The authors ran thousands of computer experiments to see which toolkit works best under different conditions. Think of it as a stress test for their tools.

The "Perfect Match" Scenario: If the files overlap a lot and the names are clear, all three toolkits work reasonably well.
The "Messy Data" Scenario: When names are misspelled or the files don't overlap much, the simple methods (like just linking and ignoring errors) fail miserably.
The "Secret Clue" Scenario: The most interesting finding is about why the mistakes happen.
- If mistakes happen randomly (like a typo), the tools can fix it.
- If mistakes happen because of a hidden pattern (e.g., the algorithm is bad at linking people with certain names and those people also have a specific medical condition), all the tools struggle. This is like a detective who is bad at recognizing people with red hats, and red hats happen to be worn by the criminals. The detective will miss the criminals systematically.

The Takeaway

The paper tells us that you cannot ignore the errors made when linking data.

If you are a researcher or a data analyst, you need to choose your toolkit based on your situation:

If you have the computing power and want the most robust answer, use the Bayesian/Super-Computer approach.
If you want a practical, flexible solution that fits into standard workflows, use Imputation (Fill-in-the-blanks).
If you are doing a quick analysis on simple data and know the error rates, use Weighting.

Most importantly, the paper warns us: If the reason you are making mistakes is related to the answer you are looking for, no amount of math will save you. You have to be honest about the limitations of your data.

In short: Don't just link the files and hope for the best. Acknowledge the uncertainty, use the right statistical "safety net," and your conclusions will be much more trustworthy.

1. Problem Statement

In many healthcare, social science, and economic applications, researchers must analyze associations between variables dispersed across multiple data files (e.g., File A and File B) that lack unique identifiers (like Social Security Numbers). Researchers rely on record linkage (using semi-identifying variables like names and addresses) to merge these files.

However, probabilistic linkage algorithms are imperfect. They introduce two types of errors:

False Links: Incorrectly linking records from different entities.
False Non-Links: Failing to link records from the same entity.

Standard post-linkage analysis often treats the linked file as "ground truth," ignoring these errors. This leads to:

Bias: Specifically, attenuation bias in regression coefficients (estimates biased toward zero) due to false links.
Selection Bias: Due to false non-links, where specific subgroups are systematically excluded.
Over-precision: Confidence intervals that are too narrow because linkage uncertainty is ignored.

The paper argues that record linkage should be viewed fundamentally as a missing data problem, where the linkage status ( $\Delta$ ) is a discrete latent variable that is never fully observed.

2. Methodology and Framework

2.1 Theoretical Framework: Linkage Mechanisms

The authors formalize the relationship between the linkage structure ( $\Delta$ ) and the data variables ( $L_A, L_B, Y_A, X_B$ ) by defining four Linkage Mechanisms, drawing parallels to Rubin's missing data mechanisms:

Strongly Non-Informative Linkage (SNL): Linkage depends only on linking variables ( $L_A, L_B$ ), not on outcomes ( $Y_A, X_B$ ). Analogous to Missing At Random (MAR).
Non-Informative Linkage (NL): Linkage depends on linking variables and covariates ( $X_B$ ), but not the outcome ( $Y_A$ ). Analogous to MAR.
Weakly Non-Informative Linkage (WNL): Linkage depends on all observed variables ( $L_A, L_B, Y_A, X_B$ ). Analogous to MAR.
Informative Linkage (IL): Linkage depends on unobserved variables. Analogous to Missing Not At Random (MNAR). This is the most problematic scenario, likely inducing severe bias.

2.2 Classification of Inferential Methods

The paper categorizes methods for analyzing linked files into three broad classes based on the missing data literature:

A. Likelihood and Bayesian Methods

Approach: Model the complete-data likelihood $L_c$ factoring into a linkage component ( $L_l$ ) and an analysis component ( $L_a$ ).
Techniques:
- Linkage Averaging: Approximating the posterior of parameters by averaging over samples of the linkage structure (e.g., Sadinle, 2017).
- Joint Modeling (Feedback): Simultaneously sampling the linkage structure ( $\Delta$ ) and analysis parameters ( $\beta$ ) using Data Augmentation (DA) or Gibbs sampling. This allows the analysis model to inform the linkage process (the "feedback effect"), which is particularly effective under WNL.
- Pseudo-Likelihood: Two-step approaches where linkage probabilities are estimated first and treated as fixed, then used in a likelihood maximization (e.g., Hof and Zwinderman, 2012).

B. Imputation Methods

Approach: Treat the linkage structure as missing data to be imputed.
Technique: Multiple Imputation (MI). Generate $M$ plausible linkage structures from the posterior distribution. Analyze each imputed dataset separately using standard methods, then pool results using Rubin's rules.
Advantage: Computationally simpler than full Bayesian joint modeling; allows for complex downstream models (e.g., clustering) that might be difficult to fit in a single joint model.
Requirement: The imputation model must be "congenial" to the analysis model (i.e., the imputation model should be at least as complex as the analysis model).

C. Weighting Methods

Approach: Correct the bias in estimating equations by weighting observations based on the probability of being a true link.
Technique: Uses the matrix of linkage probabilities ( $Q$ ). If $Q$ is known (or estimated), one can solve adjusted estimating equations (e.g., Lahiri-Larsen estimator) to obtain unbiased estimates under the NL assumption.
Limitation: Typically requires the assumption of complete linkage (File A is a subset of File B) and relies on summary statistics of linkage quality (e.g., overall match rate $\lambda$ ) in secondary analysis settings.

3. Key Contributions

Unified Missing Data Perspective: The paper provides a rigorous theoretical unification of record linkage and missing data, defining specific linkage mechanisms (SNL, NL, WNL, IL) that dictate the validity of different inferential methods.
Comprehensive Taxonomy: It systematically organizes existing and new methods into Likelihood/Bayesian, Imputation, and Weighting categories, clarifying their underlying assumptions and interrelationships.
Extensive Simulation Study: The authors conduct a large-scale factorial simulation study covering:
- Primary Analysis: Where the analyst has access to raw files and linkage variables.
- Secondary Analysis: Where the analyst only has the pre-linked file (common in privacy-restricted environments).
- Variables: Overlap rates, blocking strategies, discriminatory power of variables, measurement error levels, and the four linkage mechanisms.
Demonstration of Complex Modeling: The paper demonstrates the application of Multiple Imputation to model-based clustering (Gaussian Mixture Models) with linked data, showing that MI can handle complex multivariate structures beyond simple regression.

4. Simulation Results

The simulations evaluated seven methods (including Fellegi-Sunter, various Bayesian models, weighting estimators, and imputation) across various scenarios.

Key Findings in Primary Analysis:

Linkage Mechanism is Critical: Performance degrades significantly as the mechanism moves from SNL/NL to WNL and IL.
Feedback Effect: Under WNL (where outcomes inform linkage), Joint Modeling methods (e.g., Kamat et al., 2023; Gutman et al., 2013) outperform others by leveraging the "feedback" from the analysis model to correct linkage errors.
Overlap Matters: As the overlap between files decreases, the bias of all methods increases. However, the Steorts (2015) model (ST) remained robust with low bias and near-nominal coverage even with medium/low overlap under SNL/NL.
Weighting Methods: Generally produced valid coverage but often had larger standard errors (less efficient) compared to Bayesian methods. They struggled under WNL and IL.
Measurement Error: High measurement error exacerbated biases, particularly for weighting and pseudo-likelihood methods.

Key Findings in Secondary Analysis:

Naive Estimation: Ignoring linkage errors resulted in massive negative bias and near-zero coverage.
Weighting (Chambers et al.): Under ELE (Exchangeable Linkage Error) and NL mechanisms, weighting methods performed well. However, under WNL and IL, their performance collapsed (high bias, low coverage).
Mixture Models (Gutman et al., Slawski et al.): These methods (GT, SLW) generally showed lower bias than weighting methods under complex mechanisms (WNL/IL) but suffered from lower coverage rates (under-coverage) in high-error scenarios.

General Trends:

Bayesian/Imputation methods generally offered the best trade-off between bias and efficiency under SNL/NL.
Joint Modeling was superior when linkage depended on the outcome (WNL).
Informative Linkage (IL) remains a "worst-case" scenario where most methods fail to provide valid inferences without strong external assumptions.

5. Significance and Future Directions

Practical Impact: The paper provides a roadmap for researchers to choose the appropriate analysis method based on the nature of their data and the likely linkage mechanism. It highlights that "one-size-fits-all" approaches (like naive linkage) are dangerous.
Methodological Advancement: By framing linkage as missing data, the authors open the door to applying the vast literature of sensitivity analysis and robust inference from missing data to record linkage.
Future Research:
- Extending these methods beyond Generalized Linear Models (GLMs) to more complex multivariate techniques (e.g., factor analysis, survival analysis).
- Developing sensitivity analysis tools specifically for record linkage to assess robustness against the untestable Informative Linkage (IL) assumption.
- Investigating the validity of Multiple Imputation when the imputation and analysis models are uncongenial in complex, high-dimensional settings.

In conclusion, the paper establishes that treating record linkage as a missing data problem is not just a theoretical exercise but a necessary framework for obtaining unbiased, reliable statistical inferences from linked data.