To do GLM, you specify a family and link function in the conventional way:
fit = mcp(model, data = df, family = gaussian("log"))
fit = mcp(model, data = df, family = binomial("identity"))Currently supported families and link functions are:
- Default:
gaussian(link = "identity"). Alternatives:gaussian(link = "log") - Default:
binomial(link = "logit"). Alternatives:binomial(link = "probit"),binomial(link = "identity") - Default:
bernoulli(link = "logit"). Alternatives:bernoulli(link = "probit"),bernoulli(link = "identity") - Default:
poisson(link = "log"). Alternatives:poisson(link = "identity") - Default:
negbinomial(link = "log", link_shape = "log")
See the “Families” menu above for more details on using GLM with
mcp. The hardest part about adding new families/links is
specifying default priors. Read more in the article on default
priors.
Non-default link functions
Some link functions are default in GLMs for good reasons:
they are computationally convenient and interpretable. A non-default
link can imply impossible values, at which point mcp
errors. For example, a bernoulli("identity") model
prob ~ 1 + x places a slope directly on
P(y = TRUE), which can easily fall below 0 or exceed 1.
Informative priors, for example
through truncation, can prevent the sampler from visiting invalid
values.
In short: think carefully and proceed at your own risk.