Putting Psychology to the Test: Rethinking Model Evaluation Through Benchmarking and Prediction

Roberta Rocca, Tal YarkoniAdvances 2021 · DOI

TL;DR

Psychology's current model evaluation practices are often idiosyncratic and lack standardized benchmarks, making it difficult to compare findings and assess true predictive power. This paper proposes adopting machine learning's approach of using common, real-world benchmarks to drive cumulative progress, improve model reliability, and emphasize practical utility.

What Problem Does This Paper Address?

The authors identify a critical issue in psychology: the lack of standardized, objective metrics for evaluating scientific models and theories. Unlike fields like machine learning or physics, psychology often relies on idiosyncratic evaluation methods that can be susceptible to "overfitting" to specific datasets or experimental contexts. This makes it challenging to reliably distinguish between robust, generalizable findings and those that are merely artifacts of the particular study design. This "evaluative freedom" hinders cumulative progress and makes it difficult to assess the true predictive utility of psychological models.

Methods

The paper advocates for a shift towards benchmarking in psychology, drawing heavily on practices from machine learning. This involves:

  • Standardized Data Sets: Using common, publicly available datasets that represent real-world challenges.
  • Out-of-Sample Evaluation: Emphasizing performance on unseen data (e.g., through cross-validation) rather than solely relying on in-sample fit. This directly addresses the problem of overfitting.
  • Objective Performance Metrics: Employing clear, quantifiable metrics that directly measure predictive accuracy and generalization.
  • Multidimensional Benchmarks: Utilizing a suite of tasks and metrics to provide a comprehensive view of a model's strengths and weaknesses, rather than relying on a single criterion.

The authors also discuss the importance of interpretability techniques from machine learning to understand complex models without sacrificing predictive power.

Datasets

The paper doesn't introduce new datasets but discusses existing ones and their potential for benchmarking:

  • ImageNet: Used as a prime example of how a large-scale dataset and associated challenge (ILSVRC) drove significant progress in computer vision.
  • Many Labs Project: Reanalyzed data from this project to illustrate the difference between in-sample and out-of-sample performance, highlighting how inflated in-sample R² can be misleading.
  • Kaggle: Mentioned as a platform that hosts data science challenges, demonstrating the power of competition-driven benchmarking.
  • Various large-scale data initiatives in psychology: The paper lists numerous examples like the Psychological Science Accelerator, Human Connectome Project, and various neuroimaging and lexical databases, suggesting these could be repurposed or expanded for benchmarking.

Key Insights and Ideas

  • Psychology's "Deceptive Convergence": While psychology appears to have common methods, there's a "realm of nearly unbounded evaluative freedom" that prevents true discrimination between good and bad models.
  • The Danger of In-Sample Fit: Relying solely on metrics like R² within the training data can lead to overly optimistic assessments of a model's predictive capacity, as demonstrated by the "sunk cost" effect example.
  • Benchmarking as a Catalyst for Progress: Just as ImageNet spurred advances in computer vision, well-designed benchmarks in psychology could accelerate progress by providing clear targets and enabling direct comparison of models.
  • Prediction and Understanding are Not Mutually Exclusive: The paper argues that emphasizing prediction doesn't necessitate abandoning the goal of scientific understanding. Instead, it can lead to more robust and practically useful explanations.
  • Domain Expertise vs. ML Expertise: The rise of data-driven approaches might shift the balance of required expertise, potentially favoring machine learning specialists over traditional domain experts in some areas.

Key Results

The paper is primarily theoretical and argumentative, rather than presenting new empirical results. However, it uses reanalyses and examples to support its claims:

  • Sunk Cost Effect Example: Demonstrates how in-sample R² can be positive and statistically significant, while out-of-sample R² is negative, indicating the model is essentially useless for prediction on new data. This highlights the critical need for out-of-sample evaluation.
  • ImageNet's Impact: Illustrates how a well-defined benchmark and competition led to a dramatic reduction in error rates in computer vision, surpassing human performance.
  • GWAS and Polygenic Scores: Shows how standardized metrics and large datasets in statistical genetics have led to significant progress in predicting complex traits like educational achievement, explaining up to 20% of variance.

Limitations and Future Work

  • Generalization is Hard: The authors acknowledge that achieving perfect generalization is challenging, and even robust metrics like cross-validation have limitations. Factors like non-representative samples and methodological biases can still affect generalizability.
  • Goodhart's Law: There's a risk that benchmarks, when treated as targets, can cease to be good measures. This can lead to optimizing for the benchmark itself rather than the underlying scientific goal. The authors suggest multidimensional benchmarks and continuous updating as mitigations.
  • Ethical Considerations: The paper briefly touches on the need to consider ethical implications, particularly regarding biases in data sets and models, and the environmental costs of large models.
  • Domain Expertise: The authors acknowledge the concern that achieving benchmark success might require less traditional domain expertise, potentially shifting the landscape of required skills.
  • Future Work: The paper calls for the development of specific benchmarks across various subfields of psychology, the creation of robust data sharing infrastructure, and the promotion of competition-oriented approaches to incentivize adoption. They emphasize the need for community-wide efforts to establish these new evaluation standards.

Summary updated

Abstract

Consensus on standards for evaluating models and theories is an integral part of every science. Nonetheless, in psychology, relatively little focus has been placed on defining reliable communal metrics to assess model performance. Evaluation practices are often idiosyncratic and are affected by a number of shortcomings (e.g., failure to assess models' ability to generalize to unseen data) that make it difficult to discriminate between good and bad models. Drawing inspiration from fields such as machine learning and statistical genetics, we argue in favor of introducing common benchmarks as a means of overcoming the lack of reliable model evaluation criteria currently observed in psychology. We discuss a number of principles benchmarks should satisfy to achieve maximal utility, identify concrete steps the community could take to promote the development of such benchmarks, and address a number of potential pitfalls and concerns that may arise in the course of implementation. We argue that reaching consensus on common evaluation benchmarks will foster cumulative progress in psychology and encourage researchers to place heavier emphasis on the practical utility of scientific models.
BibTeX
@inproceedings{rocca2021putting,
  author = {Roberta Rocca and Tal Yarkoni},
  title = {{Putting Psychology to the Test: Rethinking Model Evaluation Through Benchmarking and Prediction}},
  booktitle = {Advances},
  year = {2021},
  doi = {10.1177/25152459211026864},
}
Figure