Regression methods available in CanoDraw
When you want to summarize relations between variables
by a regression model, you can fit the model with either one predictor
(in this case the model is represented by the fitted line or curve) or
with two predictors (the fitted model is then represented by a contour
plot). CanoDraw allows you to fit following types of regression models:
-
Generalized linear models (GLM)
- this is now a widely used family of models, providing support for linear
models where the response variables represent counts, relative frequencies,
probabilities, or physical measurements like dimensions or weight. Such
variables cannot be assumed to have a constant variance - an assumption
of the classical linear regression. CanoDraw supports the standard distribution
families (Poisson, Binomial, Gamma, Gaussian) and supports the canonical
link functions for each of them (in addition, log link function selection
is available for Gamma distribution). The effects of predictor(s) can be
specified as purely linear or as a second or third degree polynomial. CanoDraw
also offers a simple form of stepwise selection of model complexity, using
either the classical tests based on analysis of deviance or model selection
based on AIC statistics.
-
Generalized additive models (GAM)
- generalized additive models replace the linear terms of individual predictors,
as used in GLMs, with smooth terms, where only the complexity, not
the exact "shape" of predictor influence is specified. CanoDraw supports
cubic-spline based smooth terms and standard distribution / link function
assumptions - the same as for GLMs. Stepwise selection of smooth term(s)
complexity, using the AIC statistics, is also supported.
-
Loess smoother - represents
an universal, reliable smoothing method, based on local, weighted linear
(or second-order polynomial) regression. CanoDraw makes available some
enhancement known from other software (namely the S and R statistical systems)
- dropping the square term and making a conditionally parametric fit.
For all the fitted regression models, CanoDraw supports
creation of simple regression diagnostic diagrams - the residuals (optionally
their absolute value) can be plotted either against the fitted values or
against the predictors.
