An international team led by Lehigh University's Taeho Kim has developed the Maximum Agreement Linear Predictor (MALP), a new method that aligns predictions closely with real-world values. By maximizing the Concordance Correlation Coefficient, MALP outperforms traditional approaches in agreement, particularly in health and biology applications. Tests on eye scans and body fat data demonstrate its advantages over least-squares methods.
Mathematicians from Lehigh University and international collaborators have introduced the Maximum Agreement Linear Predictor (MALP), a technique designed to improve forecasting in fields like health research, biology, and social sciences. Led by assistant professor Taeho Kim, the method focuses on maximizing the Concordance Correlation Coefficient (CCC), which measures how closely predicted and observed values align along a 45-degree line in a scatter plot, emphasizing both precision and accuracy.
Traditional methods, such as the least-squares approach, primarily aim to minimize average errors but may not ensure strong alignment with actual values. "Sometimes, we don't just want our predictions to be close -- we want them to have the highest agreement with the real values," Kim explains. He notes that while Pearson's correlation detects linear relationships, it does not specifically target 45-degree alignment, unlike CCC, introduced by Lin in 1989.
To test MALP, the researchers used simulated data and real-world measurements. In an ophthalmology study, they compared predictions for Stratus OCT readings from Cirrus OCT data, using images from 26 left eyes and 30 right eyes. MALP produced predictions that matched true Stratus values more closely, though least-squares slightly reduced average errors better, illustrating a tradeoff.
A similar pattern emerged with body fat data from 252 adults, incorporating measurements like weight and abdomen size to estimate body fat percentage. MALP again excelled in agreement, while least-squares minimized errors more effectively. Kim emphasizes selecting tools based on priorities: MALP for alignment, traditional methods for error reduction.
The work, detailed in an arXiv preprint dated September 5, 2025 (DOI: 10.48550/arXiv.2304.04221), could enhance predictions in medicine, public health, economics, and engineering. Future extensions aim to broaden beyond linear predictors to a general Maximum Agreement Predictor.