Description
Check whether the usual plus-or-minus intervals for a fitted nonlinear curve can be trusted, and replace them with likelihood-based intervals when they cannot.
Standard errors from nonlinear least squares rely on a linear approximation that assumes a flat expectation surface with evenly spaced parameter lines. Bates and Watts found in 67 real model-data pairs that intrinsic curvature was usually small but parameter-effects curvature usually large, so linear intervals are often badly wrong. The profile t function re-fits the model with one parameter fixed at a sequence of values; plotted against the studentised parameter it is a straight unit-slope line for a linear model, bending reveals nonlinearity, and intervals are read where it crosses the t quantile. Profile traces show how other parameters adjust and reveal near-unidentifiable combinations. Root-mean-square curvatures times root F should be under 0.3. Seber and Wild add Box's bias (more than 1% of the estimate signals trouble), asymmetry measures and simulation of about 1,000 refits.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Evaluation, Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Fit the model by nonlinear least squares and record estimates, standard errors and residual standard deviation s.
- For each parameter, step it away from the estimate in fractions of a standard error in both directions, re-minimising over the other parameters at each step.
- Compute the profile t statistic at each step and plot it against the parameter with a unit-slope reference line.
- Read likelihood intervals where the profile crosses plus or minus the t quantile, noting skewed or unbounded intervals.
- Plot profile traces for each parameter pair; curved or coincident traces signal distorted joint regions or redundant parameters.
- Optionally compute RMS curvatures (flag if either times root F exceeds 0.3), Box's bias (flag above 1%) or simulate about 1,000 refits.
- If nonlinearity is severe, reparametrise, collect data where the profiles show the data are uninformative, and report likelihood intervals on the meaningful parameters.
Sources
- Nonlinear Regression Analysis and Its Applications (Bates & Watts, 1988)
- Nonlinear Regression (Seber & Wild, 1989)
Also known as
- Profile t plot
- Profile likelihood intervals for nonlinear parameters
- Likelihood profile traces
- Curvature measures
- Intrinsic and parameter-effects curvature
- Bates-Watts curvature
Categories
- Analytics & Behaviour
- Statistics & Modelling