← Back to Knowledge Graph

A graduate student aced a diagnostic interview — one of the best performances the evaluator had seen. How well should we predict they'll perform in the program? If you answer "brilliantly," you've just made the most common prediction error in psychology.

The Framework

Kahneman's four-step regression correction is a method for making predictions that account for regression to the mean. Most predictions are too extreme because System 1 generates them through intensity matching (great interview → great performance) without accounting for the imperfect correlation between the predictor and the outcome. The four steps force the correction:

  • Start with the base rate (average outcome for the reference class). For graduate students, the average GPA might be 3.3.
  • Form an impression of the evidence. The interview was outstanding — top 5%.
  • Estimate the correlation between the evidence and the outcome. Interview performance might correlate with graduate performance at r=0.3 (Kahneman's typical estimate for unstructured interviews).
  • Move from the base rate toward your impression, but only by the correlation fraction. If the impression suggests the student is at the 95th percentile, and the correlation is 0.3, move 30% of the way from the base rate (50th percentile) toward the 95th: predicted rank = 50 + 0.3 × (95−50) = 63.5th percentile. Not the 95th.

    • Where It Comes From

    Kahneman presents the four-step correction in Chapter 18 of Thinking, Fast and Slow as the practical method for "taming intuitive predictions." The key insight: System 1 naturally produces predictions that match the extremeness of the evidence (intensity matching), ignoring the correlation between the evidence and the outcome. The correction forces the prediction back toward the mean in proportion to the unreliability of the evidence.

    > "Intuitive predictions tend to be overconfident and overly extreme." — Thinking, Fast and Slow, Ch 18

    Cross-Library Connections

    The method directly addresses the planning fallacy discussed in Chapter 23: inside-view estimates are extreme (high correlation assumed between plan quality and actual outcome). Reference class forecasting adjusts back toward the base rate — functionally identical to Step 4 of the regression correction.

    The Implementation Playbook

    Hiring: After an impressive interview, apply the correction. Base rate: average employee performance. Impression: outstanding. Correlation between interviews and performance: ~0.3. Predicted performance: moderately above average, not outstanding. This prevents overpaying based on an impressive interview that may not predict actual performance.

    Sales Forecasting: A new lead looks extremely promising. Base rate: your average close rate (say 20%). Impression: 90% likely. Correlation between initial impression and actual close: maybe 0.4. Corrected prediction: 20% + 0.4 × (90%−20%) = 48%. Still promising, but not the slam-dunk your gut says.

    Investment: A startup's metrics look amazing (top 5% of peer group). Base rate: startup success rate (~10%). Correlation between current metrics and ultimate success: maybe 0.3. Predicted success: 10% + 0.3 × (95%−10%) = 35.5%. Much better than average, but nowhere near the 95% your excitement suggests.

    Personal Predictions: Apply the correction to any prediction you make about yourself or others. "I'm going to have a great quarter" → base rate + (correlation × deviation from base rate). The correction feels pessimistic. It's actually realistic.

    Key Takeaway

    The four-step regression correction is the formula for converting intuitive predictions into calibrated ones. The key variable is the correlation between the evidence and the outcome — and that correlation is almost always lower than System 1 assumes. When in doubt, use 0.3. Your predictions will feel insufficiently extreme. They will also be more accurate.

    Continue Exploring

    [[Regression to the Mean]] — The statistical principle the four-step method implements

    [[Planning Fallacy]] — The specific prediction error the method corrects

    [[Intensity Matching]] — The System 1 mechanism that produces extreme (uncorrected) predictions

    📚 From Thinking, Fast and Slow by Daniel Kahneman — Get the book