Microcomputer-based estimation of psychophysical thresholds: The Best PEST

Harris R. Lieberman, Alex PentlandBehavior 1982 · DOI

Introduction

In experimental psychology, determining a "threshold"—the point at which a subject can just barely detect a stimulus, such as a faint light or a soft sound—is a fundamental challenge. Traditionally, finding this point accurately required a high volume of trials, which proved exhausting for subjects and time-consuming for researchers.

Before the introduction of the "Best PEST" algorithm in 1982, researchers primarily relied on "staircase" methods. These methods adjusted stimulus intensity based on the subject's immediate response: if the subject detected the stimulus, the next trial was made harder; if they failed, it was made easier. While simple to implement, these ad hoc methods were inefficient because they did not utilize the full history of previous trials to optimize the next stimulus level. This paper introduces a rigorous maximum-likelihood estimation (MLE) approach designed to run on early microcomputers like the Apple II, significantly improving threshold estimation speed and reliability.

Methodology

The authors developed Best PEST (Parameter Estimation by Sequential Testing), which treats threshold search as a statistical estimation problem rather than a set of fixed heuristic rules. The algorithm operates on several core principles:

  1. Model Assumption: It assumes the subject's response curve (the psychometric function) follows a known shape, typically a sigmoid or "logit" function.
  2. Likelihood Tracking: The system maintains a probability distribution representing the likelihood that the threshold exists at various stimulus levels.
  3. Dynamic Updating: After every response, the algorithm updates this entire probability distribution using all accumulated data from the session.
  4. Optimal Sampling: The next stimulus is placed at the current maximum-likelihood estimate of the 50% threshold. Mathematically, this point provides the maximum information gain regarding the threshold's true location.

To ensure accessibility for the research community of the time, the authors provided the full source code in Applesoft BASIC.

Experimental Setup

To evaluate the algorithm's performance, the authors utilized Monte Carlo simulations rather than human subjects to ensure statistical consistency. These simulations involved:

  • Simulated Observers: Computer models programmed to respond according to a standard symmetrical logit function.
  • Benchmarking: Best PEST was compared against the "Wetherill tracking" method (a popular optimized staircase procedure) and standard up-down staircases.
  • Statistical Reliability: Each data point in the results was derived from at least 50 simulation trials.

Results

The simulations demonstrated that Best PEST significantly outperforms traditional methods in both precision and efficiency.

Comparison of setting accuracy across different methods: Best PEST (crosses) consistently achieves lower error (higher accuracy) with fewer trials than standard staircase methods (circles).

Key findings include:

  • Superior Accuracy: Best PEST was approximately 20% more accurate than the optimized Wetherill tracking method.
  • Significant Speedup: To reach a target level of accuracy, Best PEST required one-third fewer trials than the best-case scenarios for traditional staircase methods.
  • Reduced Parameter Sensitivity: Unlike Wetherill tracking, which fluctuates in performance based on the initial "step size" chosen by the experimenter, Best PEST remains consistently efficient regardless of starting conditions.

Discussion and Practical Implications

The core innovation of Best PEST is treating every trial as a search for information. By sampling at the current best estimate, the algorithm minimizes the variance of the estimate as quickly as possible. This makes the method highly robust to "priors"—even if the experimenter's initial guess about the subject's sensitivity is incorrect, the algorithm converges rapidly.

While the method assumes a fixed psychometric function shape, its practical benefits are substantial. In 1982, the primary hurdle was the computational overhead of performing MLE on 8-bit hardware, which the authors solved using log-probabilities to prevent underflow errors. Today, these principles remain relevant, particularly in clinical settings or trials involving children, where minimizing trial counts and maximizing robustness to unknown initial parameters are critical for success.

Summary updated

BibTeX
@inproceedings{lieberman1982microcomputer,
  author = {Harris R. Lieberman and Alex Pentland},
  title = {{Microcomputer-based estimation of psychophysical thresholds: The Best PEST}},
  booktitle = {Behavior},
  year = {1982},
  doi = {10.3758/bf03202110},
}
Figure