I want to obtain the full distribution of a Gamma (or Inverse Gaussian) distributed $y_i$ given a vector of $\bar x_i$ that have been used in the linear predictor of a coefficient. Suppose also for the Gamma GLM I have used the log-link instead of the canonical link.

Since both distributions are bi-parametric I know I can get the estimation of the mean parameter by predict(glmObj, ..., type="response") whichever distribution whichever link I have used. I'm not sure about the conditional variance. I know that $var\left(y_i\right)=\phi*V\left(\mu_i\right)$. My questions are:

  1. Is it correct to estimate $\phi$, the dispersion parameter, as the square root of glmObj$deviance/glmObj$df.residual, regardless of the distribution and canonical link?

  2. Is $V\left(\mu_i\right)$ dependent on the canonical link?

  • $\begingroup$ You can get "standard errors" (not prediction errors) from predict using the argument se.fit=TRUE. It doesn't give you their covariance, however. Any two predicted values are a function of a linear combination of regression parameters. These usually aren't independent. It is easy to compute by hand, however. $\endgroup$ – AdamO Jun 18 '13 at 15:53
  • $\begingroup$ The estimate of $\phi$ from GLMs is generally not ML (it is in some cases, but usually not); indeed in some cases the ML estimate doesn't even make sense (Poisson regression for example). The package MASS includes functions for ML estimation of the parameter corresponding to $\phi$ in one or two cases (such as the Gamma). As for 'correct' -- it depends on what you mean by 'correct'. $\endgroup$ – Glen_b Jul 18 '13 at 16:47
  • $\begingroup$ If you want to estimate the dispersion parameter $\phi$, where $\phi\ne 1$ then u need to use the quasi-likelihood approach. For example for the poisson glm, u need something like: glm(y ~ x, family =quasipoisson(link = "log")). When the $\phi=1$, then the quasi-likelihood will be the ML estimate. $\endgroup$ – Stat Dec 16 '13 at 2:01

1- the estimation of phi can be obtained by Over_disp <- sum(residuals(glmobj, type="pearson")^2)/glmobj$df.res which is different from the deviance

2- given Y~EF{b(theta) , a(phi)} d( d( b(theta) ))/dtheta^2 if a canonical link is used gives the variance function V(mu)= V(d(b(theta))/dtheta) = V(linear predictor) i.e. it does not depend upon the link.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.