📊 statistics

Beyond Single-Score Matching: A Two-Dimensional Propensity Score Method for Mixed Covariate Types

This paper introduces a two-dimensional propensity score matching (2D-PSM) method that separately estimates scores for categorical and continuous confounders to achieve superior multivariate balance in observational studies with mixed covariate types and complex interactions compared to traditional single-score approaches.

Original authors: Kostiantyn Botnar, Justin T. Nguyen, Kamil Khanipov, George Golovko

Published 2026-08-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Kostiantyn Botnar, Justin T. Nguyen, Kamil Khanipov, George Golovko

Original paper licensed under CC BY 4.0 (https://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: Did a new medicine actually help patients, or did they just happen to get better on their own? In the perfect world of science, you'd run a "Randomized Clinical Trial," where you flip a coin to decide who gets the medicine and who gets a sugar pill. This ensures the two groups are identical twins in every way, so any difference in health must be the medicine's fault. But in the real world, we can't always flip coins. Sometimes it's too expensive, too dangerous, or just too slow. So, scientists turn to "observational studies," looking at massive digital records of patients who already took the medicine versus those who didn't.

The problem is that these real-world groups aren't twins. The people who took the medicine might have been younger, richer, or sicker to begin with. This is called "confounding bias." To fix this, scientists use a clever trick called "Propensity Score Matching" (PSM). Think of it like a dating app for data. The app calculates a single "compatibility score" for every person based on their age, gender, medical history, and more. Then, it tries to pair up a treated patient with an untreated patient who has the exact same score. If the scores match, the scientists assume the two people are similar enough to compare fairly.

But here's the catch: real life is messy. People have all kinds of different traits. Some are simple categories (like "male" or "female," "smoker" or "non-smoker"), while others are sliding scales (like "age," "blood pressure," or "weight"). The old way of doing this matching forces all these different types of information into one single number. It's like trying to describe a whole pizza by just measuring its total weight; you lose the details about the crust, the cheese, and the pepperoni. The paper you are about to read suggests that this "one-number" approach might be missing the nuance needed to make fair comparisons in complex medical data.

The Two-Dimensional Matchmaker

In this study, a team of researchers from the University of Texas Medical Branch at Galveston decided to upgrade the dating app. They introduced a new method called Two-Dimensional Propensity Score Matching (2D-PSM). Instead of squishing all a patient's traits into one single score, they split the job into two separate dimensions: one for categorical traits (the "yes/no" or "type" stuff) and one for numerical traits (the "how much" stuff).

Imagine you are trying to match two people for a camping trip. The old method would give you a single "Camping Compatibility Number" based on everything they like. The new 2D method gives you two separate scores: a "Gear Score" (do they have a tent? a sleeping bag?) and a "Nature Score" (do they like hiking? do they hate bugs?). The researchers then try to match people who are close in both dimensions at the same time, using a special rule called an "elliptical caliper."

Think of this "elliptical caliper" as a stretchy, oval-shaped net. If two people are slightly different in their "Gear Score," the net can stretch to let them match, as long as they are very close in their "Nature Score." This flexibility allows the system to find better matches without throwing away too many people, which is a common problem when the rules are too strict.

The Great Experiment

To see if this new method actually works, the team didn't just guess; they put it to the test. They ran a massive experiment using two types of data:

Real-World Data: They grabbed five different sets of actual medical records from hospitals, involving anywhere from 460 to nearly 62,000 patients. These records covered everything from kidney disease and burn victims to pulmonary embolisms.
Synthetic Data: They also created 21 fake datasets using computers. These were perfect "test labs" where they knew exactly how the matching should work. They built these fake worlds with different levels of complexity, adding in tricky things like nonlinear relationships (where a little bit of a drug helps, but a lot hurts) and interactions between variables.

They tested their new 2D-PSM method against the traditional "one-score" method using five different machine learning algorithms (smart computer programs that learn from data). They tried different "net sizes" (calipers) to see how strict or loose the matching rules should be.

What They Found

The results were pretty exciting for the new method. When they looked at how well the groups were balanced (meaning, did the treated and untreated groups actually look alike after matching?), the 2D-PSM method was a clear winner in most cases.

The Big Picture: In the fake, complex datasets, the new method performed better than the old one in 85% of the experiments, especially when they used slightly looser matching rules. Even in the real-world hospital data, it did just as well as the old method, and often better, without losing any patients from the study.
The "Shape" of the Match: Here is the most interesting part. The old method was really good at matching the average age or weight of the groups (the "mean"). But it often messed up the spread or "shape" of the data. It was like matching two groups of people who had the same average height, but one group was all giants and dwarfs, while the other was all average-sized people. The new 2D-PSM method was much better at keeping the "shape" of the groups similar, ensuring that the spread of ages and weights was balanced too.
The Complexity Factor: The more complicated the data was (with lots of interactions between different variables), the more the new method shined. When the relationships between variables were simple and straight-line, both methods did okay. But when the data got messy and nonlinear, the 2D-PSM method pulled ahead, suggesting that splitting the problem into two dimensions helps the computer understand the data better.

Why It Matters

The authors suggest that by treating categorical and numerical data as two separate dimensions, researchers can get a much fairer comparison in medical studies. It's like realizing that to find a perfect partner, you shouldn't just look at a single "compatibility score," but should check if you match on your hobbies and your values separately.

This study didn't prove that the new method fixes every problem in medical research, and it didn't test the final medical outcomes (like who actually lived longer). But it strongly suggests that for the tricky, mixed-up data we find in electronic health records, the old "one-score" way might be too simple. The new two-dimensional approach offers a more flexible and accurate way to level the playing field, helping doctors and scientists trust their findings a little bit more.

Technical Summary: Beyond Single-Score Matching

Problem Statement
Propensity score matching (PSM) is a standard technique for reducing confounding bias in observational studies, particularly those utilizing electronic health records (EHRs). However, traditional PSM relies on a "single-score" approach, where all covariates—regardless of whether they are categorical (e.g., race, sex) or continuous (e.g., age, lab values)—are compressed into a single scalar probability of treatment. This method assumes linear covariate-treatment relationships and often fails to adequately balance heterogeneous data types common in clinical settings. Furthermore, existing evaluations of PSM often rely on synthetic datasets with homogeneous covariate structures or focus solely on univariate balance metrics (like the Absolute Standardized Mean Difference, ASMD), potentially overlooking distributional imbalances (variance) and complex nonlinear interactions.

Methodology
The authors propose Two-Dimensional Propensity Score Matching (2D-PSM), a framework that decomposes the covariate space into two distinct dimensions: categorical and numerical.

Dual Propensity Score Estimation: Instead of a single model, 2D-PSM estimates separate propensity scores for categorical confounders and numerical confounders. The study utilized five machine learning classifiers (Bernoulli Naïve Bayes, Logistic Regression, Linear Support Vector Machine, Random Forest, and XGBoost) to estimate these scores, allowing for dimension-specific model selection (e.g., excluding Bernoulli Naïve Bayes for numerical data).
Elliptical Caliper Constraint: Matching is performed in a two-dimensional space defined by the two propensity scores. The algorithm uses a greedy nearest-neighbor approach without replacement, constrained by an elliptical caliper defined as:
$\left(\frac{PS_{ct} - PS_{cc}}{C_c}\right)^2 + \left(\frac{PS_{nt} - PS_{nc}}{C_n}\right)^2 \leq 1$
Where $PS_{ct/cc}$ and $PS_{nt/nc}$ represent the scores for treated and control cohorts in categorical and numerical spaces, respectively, and $C_c$ and $C_n$ are adaptive calipers calculated as fractions of the interquartile range (IQR). This geometry allows for compensatory balance: a larger difference in one dimension can be offset by a smaller difference in the other.
Experimental Design: The method was evaluated across 26 datasets:
- 5 Real-World (RW) EHR datasets (n = 460 to 61,926) with varying covariate compositions.
- 21 Synthetic datasets (n = 10,000) generated with systematically varied treatment-assignment complexities (ranging from linear to high-order interactions) and covariate ratios (numerical vs. binary).
Evaluation Metrics: Balance was assessed using univariate metrics (ASMD for continuous, Standardized Difference in Proportions for categorical, and Variance Ratio) and multivariate metrics (Mahalanobis Balance).

Key Results

Multivariate Balance: 2D-PSM achieved superior or equivalent multivariate balance compared to traditional single-score PSM in 73.1% of conditions using the strictest caliper (0.1×IQR), increasing to 88.5% with the largest caliper (0.5×IQR). The performance gain was most pronounced in datasets with complex treatment-assignment mechanisms (nonlinearities and interactions).
Variance Ratio Preservation: A critical finding was the trade-off between mean balance (ASMD) and distributional balance (Variance Ratio, VR). While traditional PSM often achieved greater ASMD reduction, it frequently caused significant degradation in VR (worsening variance balance). 2D-PSM provided a more balanced outcome, preserving VR while still improving ASMD, resulting in simultaneous improvements in both metrics in the majority of cases.
Model Selection: No single machine learning combination dominated all scenarios. However, pairing Linear SVM for both categorical and numerical dimensions was the most frequent top performer (20/104 conditions). The results suggest that model performance is dataset-dependent, reinforcing the need for researchers to evaluate multiple combinations.
Coverage: The 2D-PSM approach maintained consistent treated-cohort coverage (approximately 70–73%) across all caliper widths, indicating that the geometric constraints did not penalize the ability to find matches.

Significance and Claims
The paper claims that 2D-PSM addresses a fundamental limitation of the single-score paradigm: the loss of information and balance optimization when compressing heterogeneous covariate types into a unidimensional space. By explicitly separating categorical and numerical dimensions, the method allows for:

Dimension-specific optimization: Tailoring the propensity score model to the specific statistical properties of each data type.
Flexible matching geometry: The elliptical caliper enables compensatory balance, offering a more nuanced approach than rigid single-axis constraints.
Improved distributional balance: The method mitigates the common issue where mean balance is achieved at the expense of variance balance, a factor critical for the validity of downstream causal inference.

The authors conclude that while improved covariate balance is a necessary condition for unbiased treatment effect estimation, 2D-PSM provides a robust framework for handling the complexity of mixed-type clinical data, particularly when treatment assignment involves nonlinear interactions. They note that future work is required to validate whether these balance improvements translate directly to reduced bias in estimated causal effects and to explore extensions to higher-dimensional spaces (e.g., separating ordinal, nominal, and continuous variables).

The Two-Dimensional Matchmaker

The Great Experiment

What They Found

Why It Matters

Technical Summary: Beyond Single-Score Matching

More like this