Inverse Prediction (Calibration) from a Fitted Curve

Description

Use a fitted curve backwards: from a newly observed outcome, estimate the input value that produced it, with an interval.

Most regression questions run forwards, predicting an outcome from an input. Calibration runs backwards: having fitted how a measure responds to a known input, estimate the input behind a new observation. The point estimate solves the fitted curve for the observed response and is usually biased. An interval collects every input value for which the gap between the observation and the fitted curve, divided by its standard error (residual plus fitted-value variance), is within the t limit; it can be a finite interval, two half-lines or the whole line when the curve is flat. For a few predictions Bonferroni adjustment suffices; for a calibration curve used indefinitely, intersect a simultaneous band for the curve with a bound on observation scatter. An empirical Bayes predictor applies when the calibration inputs were a random sample of future cases.

Attributes

Loss type
Latent
Stage
Evaluation
Effort
Days

Steps

  1. Fit the calibration curve to pairs of known input and observed response, by linear or nonlinear least squares.
  2. For a new response, solve the fitted curve for the input value to get the point estimate.
  3. Find all input values where the squared gap between the response and the fitted curve is within t squared times (s squared plus the variance of the fitted value); report the end points.
  4. For k predictions divide alpha by k; for open-ended routine use, apply the simultaneous-band construction.
  5. Flag unbounded intervals and collect more calibration data where the curve is flat.

Sources

  • Nonlinear Regression (Seber & Wild, 1989)

Also known as

  • Inverse prediction
  • Inverse regression
  • Calibration curve inversion
  • Discrimination interval
  • Reverse prediction

Categories

  • Analytics & Behaviour
  • Statistics & Modelling