Statistics Glossary → Nonlinear least squares and starting values
Nonlinear least squares and starting valuesconcept
A method that fits a model that is nonlinear in its parameters by iteratively minimizing the sum of squared differences between observed and predicted values; the solution is searched for starting from given initial values.
In linear regression the coefficients come from a single formula. In models such as logistic growth K/(1 + e^(-r(x - x₀))) or Michaelis-Menten Vmax·x/(Km + x), however, the parameters enter the formula nonlinearly, so there is no closed-form solution. Algorithms such as Gauss-Newton and Levenberg-Marquardt linearize the model around the current parameter values, take a step that reduces the sum of squares, and repeat until the change is small enough. This process needs a starting point; if the starting values are far from the true values, the algorithm may fail to converge, get stuck in a local minimum or wander off to meaningless values.
Good starting values can be read from a plot of the data or from what the parameters mean: on a logistic curve, for example, K is close to the largest observed value and x₀ to the point where the response reaches half of it. Parameter bounds (e.g. a positive growth rate) can make estimation stable. Standard errors and confidence intervals rely on a linear approximation at the solution and can mislead when the curvature is large; profile likelihood or bootstrap intervals are then more reliable. In nonlinear models R² does not carry its usual interpretation, and different functional forms are compared with criteria such as AIC.
An everyday example
A biologist follows the growth of a bacterial culture for 24 hours and wants to fit a logistic curve. Looking at the plot, she guesses a carrying capacity of K = 2.0 (the highest optical density), a half-growth time of x₀ = 10 hours and a rate of r = 0.5 as starting values; the algorithm converges in a few iterations to K = 1.93, x₀ = 9.6 and r = 0.58. Had she entered r = 50 at the start, the algorithm would very likely not have converged.
In MerQur
Available in nonlinear regression as the 'Function form' (exponential, logistic, Gompertz, Michaelis-Menten, Weibull, power, custom formula), 'Model formula and Start values' and 'Parameter bounds' options; for the built-in forms starting values are derived from the data, and when bounds are given the trust-region reflective (TRF) algorithm is used.