Multiphase Regression with Estimated Changepoints

Description

Fit two or more lines or simple curves that join at estimated changepoints on a predictor, test whether a change exists and get an interval for where it occurs.

Some relationships have a few simple regimes: satisfaction flat until load time passes a threshold, then falling; engagement rising with team size, then levelling. Multiphase regression fits separate segments that meet at an unknown changepoint, with or without continuity, and smooth-transition variants where a parameter controls how sharply one line bends into the next. Because the model is linear once the changepoint is fixed, it is fitted by profiling: search a grid of changepoints, fit constrained regressions at each, and keep the one with the smallest residual sum of squares. Testing whether any change exists is non-standard: the likelihood ratio has no chi-squared limit, so simulate the maximum F for the actual design or use Worsley's bound. An interval for the changepoint comes from inverting the F test; asymptotic normal intervals are poor.

Attributes

Loss type
Unmigrated
Effort
Days

Steps

  1. Plot the outcome against the predictor and decide whether segments must join (continuous) or may jump.
  2. For each candidate changepoint on a grid, fit the constrained segment regressions and record the residual sum of squares.
  3. Take the changepoint with the smallest residual sum of squares and refit to obtain segment slopes and intercepts.
  4. Test one segment against two using the maximum F over splits, with its null distribution simulated for the same predictor values or Worsley's bound.
  5. Form the changepoint interval as all values whose residual sum of squares exceeds the minimum by less than s squared times the F quantile.
  6. If the bend looks gradual and data exist near it, fit a smooth-transition model by nonlinear least squares from the two-line solution.
  7. Allow for autocorrelated errors when observations are sequential.

Sources

  • Nonlinear Regression (Seber & Wild, 1989)

Also known as

  • Two-phase regression
  • Segmented regression with unknown breakpoint
  • Broken-stick regression
  • Piecewise linear regression
  • Bent-cable regression
  • Bacon-Watts smooth transition model
  • Switching regression

Categories

  • Analytics & Behaviour
  • Statistics & Modelling