The error specification of a stochastic
model is equal in importance to the deterministic
formulation.
Using simple linear regression as an example, it is imperative
to realize that all inferences regarding the estimates for a and
b, and their uncertainty, stem from the assumptions made concerning
the error term (epsilon) in the following stochastic
model.
For simple linear regression the assumption is that each of the
error terms has a Gaussian (normal) distribution with a mean of
zero (0) and finite variance.
It is also assumed that each of the error terms (epsilons) are
independent of each other.
Thus in general, statistical model results can heavily depend upon
the nature of model error (e.g., random sampling error, measurement
error, process error) and the probability distributions chosen to
represent it (e.g., Gaussian, Poisson, binomial, etc.).
Clearly, carefully consideration must be given to the mode error
specification when constructing a statistical model.
Too frequently error specification gets much less consideration
by analysts than it warrants.
There are three basic elements of error that may enter a statistical
model: (1) measurement error - the data are imperfect; (2) process
error - the model is imperfect; and (3) Bernoulli (and binomial
or multinomial) error - stochastic frequency outcomes such as are
observed when coin flipping.
Typically, Gaussian (normal) error is associated with continuous
variables, while binomial or multinomial error is associated with
frequency data.
However, many biological models and processes may demand alternative
simple, complex or correlated error structures that include all
of measurement, process and/or Bernoulli family errors.
|