A statistical model is linear with respect to a parameter if an
estimate of that parameter can be obtained in a single iteration.
That is, the maximum-likelihood estimate of the parameter can
be defined analytically by an equation derived from the deterministic
model structure and error specification
that requires only the observed data for input.
A statistical model is non-linear with respect to a parameter
if the maximum-likelihood estimate of that parameter requires, in
addition to the deterministic model
structure, the error specification
and the observed data, the value of at least one other parameter
in the model.
Since initially the value of that parameter is usually a guess,
parameter estimation requires more than one iteration to sequentially
update the parameter values until their values converge to their
maximum-likelihood estimates.
Linear statistical models are rarely robust enough to adequately
describe biological or ecological systems.
Consequently most realistic biological and ecological models are
non-linear in design and require non-linear methodologies for parameter
estimation and interpretation of their uncertainty.
Note that simple linear regression is a linear model and gets part
of its name from that fact, and not from the straight line defined
by the deterministic model.
Likewise, a quadratic linear regression is a linear model despite
the curvature of the deterministic
model, hence such models are often referred to as curvilinear models.
|