SIMSTAT: Generalized Linear Models
Generalized Linear Models (GLM)
This option is often the technique of choice
when it can be assumed that the errors follow either a normal, binomial,
Poisson or gamma distribution and there is a link function
giving a good fit, e.g. binary logistic regression assuming
a binomial distribution and logistic link.
Appropriate links such as power,
identity, reciprocal, square root, log, logistic, probit or
complementary log-log are provided and the canonical links are
indicated as defaults. There are many options for controlling
the Simfit GLM interface. Try the test files glm.tf1 (normal),
glm.tf2 (binomial), glm.tf3 (Poisson) and glm.tf4 (gamma). The
data format is as for multilinear regression, except that an
extra column is needed for the sample sizes when binomial errors
are assumed. In the case of binary logistic regression, the y values
would all be 0 (failure) or 1 (success),
and the extra column for the number of
trials would be set equal to 1. If a factor with m levels is
represented in a logistic regression by m (0,1) dummy indicator
variables x_1, x_2, ..., x_m, and a constant term
is permitted in the regression but x_1 is suppressed, then the
estimated coefficients can be interpreted as log odds ratios for
the factor levels with respect to level 1. There is a simplified
interface to GLM which should be used for logistic regression,
binary logistic regression, polynomial logistic regression,
survival analysis (using exponential, Weibull or Cox models),
and dose response curves (using probits, logits or log-logs).
Back to Help Menu or End Help