VGAM (version 0.8-1)

binomialff: Binomial Family Function

Description

Family function for fitting generalized linear models to binomial responses, where the dispersion parameter may be known or unknown.

Usage

binomialff(link = "logit", earg = list(), dispersion = 1, mv = FALSE,
           onedpar = !mv, parallel = FALSE, zero = NULL)

Arguments

link

Link function. See Links for more choices.

earg

Extra argument optionally used by the link function. See Links for more information.

dispersion

Dispersion parameter. By default, maximum likelihood is used to estimate the model because it is known. However, the user can specify dispersion = 0 to have it estimated, or else specify a known positive value (or values if mv

mv

Multivariate response? If TRUE, then the response is interpreted as $M$ independent binary responses, where $M$ is the number of columns of the response matrix. In this case, the response matrix should have zero/one values only.

If <

onedpar

One dispersion parameter? If mv, then a separate dispersion parameter will be computed for each response (column), by default. Setting onedpar=TRUE will pool them so that there is only one dispersion parameter to be estimat

parallel

A logical or formula. Used only if mv is TRUE. This argument allows for the parallelism assumption whereby the regression coefficients for a variable is constrained to be equal over the $M$ linear/additive predictors.

zero

An integer-valued vector specifying which linear/additive predictors are modelled as intercepts only. The values must be from the set {1,2,...,$M$}, where $M$ is the number of columns of the matrix response.

Value

• An object of class "vglmff" (see vglmff-class). The object is used by modelling functions such as vglm, vgam, rrvglm, cqo, and cao.

Warning

With a multivariate response, assigning a known dispersion parameter for each response is not handled well yet. Currently, only a single known dispersion parameter is handled well.

The maximum likelihood estimate will not exist if the data is completely separable or quasi-completely separable. See Chapter 10 of Altman et al. (2004) for more details, and safeBinaryRegression. Yet to do: add a sepcheck=TRUE, say, argument to detect this problem and give an appropriate warning.

Details

If the dispersion parameter is unknown, then the resulting estimate is not fully a maximum likelihood estimate (see pp.124--8 of McCullagh and Nelder, 1989).

A dispersion parameter that is less/greater than unity corresponds to under-/over-dispersion relative to the binomial model. Over-dispersion is more common in practice.

Setting mv = TRUE is necessary when fitting a Quadratic RR-VGLM (see cqo) because the response is a matrix of $M$ columns (e.g., one column per species). Then there will be $M$ dispersion parameters (one per column of the response matrix).

When used with cqo and cao, it may be preferable to use the cloglog link.

References

McCullagh, P. and Nelder, J. A. (1989) Generalized Linear Models, 2nd ed. London: Chapman & Hall.

Altman, M. and Gill, J. and McDonald, M. P. (2004) Numerical Issues in Statistical Computing for the Social Scientist, Hoboken, NJ: Wiley-Interscience.

Ridout, M. S. (1990) Non-convergence of Fisher's method of scoring---a simple example. GLIM Newsletter, 20(6).

See Also

quasibinomialff, Links, rrvglm, cqo, cao, betabinomial, zibinomial, dexpbinomial, mbinomial, seq2binomial, amlbinomial, simplex, binomial, safeBinaryRegression.

Examples

quasibinomialff()
quasibinomialff(link="probit")

fit = vgam(agaaus ~ poly(altitude, 2), binomialff(link=cloglog), hunua)
with(hunua, plot(altitude, agaaus, col="blue", ylab="P(agaaus=1)",
                 main="Presence/absence of Agathis australis", las=1))
ooo = with(hunua, order(altitude))
with(hunua, lines(altitude[ooo], fitted(fit)[ooo], col="red", lwd=2))


# Shows that Fisher scoring can sometime fail. See Ridout (1990).
ridout = data.frame(v = c(1000, 100, 10), r = c(4, 3, 3), n = c(5, 5, 5))
(ridout = transform(ridout, logv = log(v)))
# The iterations oscillates between two local solutions:
glm.fail = glm(r/n ~ offset(logv) + 1, weight=n,
               binomial(link=cloglog), ridout, trace=TRUE)
coef(glm.fail)
# vglm()'s half-stepping ensures the MLE of -5.4007 is obtained:
vglm.ok = vglm(r/n ~ offset(logv) + 1, weight=n,
               binomialff(link=cloglog), ridout, trace=TRUE)
coef(vglm.ok)