Nonlinear Least Squares Fitting of Mechanistic Curve Models

Description

Fit a curve whose parameters have meaning (asymptote, rate, half-point) by iterative least squares, then check, compare and report it.

Learning curves, adoption curves and saturation effects are naturally described by functions nonlinear in their parameters, such as exponential rise to an asymptote or Michaelis-Menten. Linearising transformations distort the errors, and polynomials and splines give no interpretable parameters. Nonlinear least squares fits the curve directly by Gauss-Newton iteration: linearise at the current guess, solve for an increment, halve the step until the sum of squares falls, and repeat until the residual is nearly orthogonal to the tangent plane (relative offset below 0.001). Good starting values come from reading asymptotes and half-points off a plot, linearised fits and conditional linearity. Seber and Wild add that if the data never approach the plateau only the initial slope is estimable, so design the x range to cover early rise, bend and plateau.

Attributes

Loss type
Unmigrated
Effort
Days

Steps

  1. Choose the curve: a mechanistic form if theory exists, otherwise one matching the data's shape (for example zero at zero, levelling to an asymptote).
  2. Plan or check the x range so it covers the early rise, the bend and the plateau; with no plateau, prefer a simpler empirical curve.
  3. Get starting values from the plot (asymptote, half-point, initial slope) or from a linearised fit.
  4. Fit by Gauss-Newton or a library routine with checked derivatives, halving steps until the sum of squares falls, and declare convergence by the relative offset criterion (below 0.001).
  5. Assess the fit: sensible parameters, parameter correlations above about 0.99 (overparameterisation), random residuals against fitted values, order and predictors, and a lack-of-fit F test if there are replicates.
  6. Compare groups with incremental parameters on 0/1 indicators and decide by an extra sum of squares F test between full and reduced models.
  7. Report estimates with standard errors and correlations, the fitted curve over the data, and profile-based intervals where the linear approximation is doubtful.

Sources

  • Nonlinear Regression Analysis and Its Applications (Bates & Watts, 1988)
  • Nonlinear Regression (Seber & Wild, 1989)

Also known as

  • Nonlinear regression
  • Nonlinear least squares
  • Gauss-Newton curve fitting
  • Mechanistic model fitting
  • Fitting asymptotic, exponential and saturating curves

Categories

  • Analytics & Behaviour
  • Statistics & Modelling