[Rd] nobs(lm(...)) != nobs(glm(...)) when there are 0s in weights

[William Dunlap]

What is the rationale for nobs.lm omitting observations with
zero weights while nobs.glm includes them?

> df <- data.frame(x1=log(1:10), x2=1/(1:10), y=1:10, wt=c(0,2,0,4,0,6,7,8,9,10))
> nobs(lm(data=df, y~x1+x2, weights=wt))
[1] 7
> nobs(glm(data=df, y~x1+x2, weights=wt))
[1] 10

The anova methods for lm and glm seem to agree on the number
of degrees of freedom here, although anova.glm issues a message
about it:

> anova(lm(data=df, y~x1+x2, weights=wt))
Analysis of Variance Table

Response: y
          Df  Sum Sq Mean Sq  F value    Pr(>F)    
x1         1 196.682 196.682 1034.648 5.569e-06 ***
x2         1  11.514  11.514   60.572   0.00147 ** 
Residuals  4   0.760   0.190                       
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 
> anova(glm(data=df, y~x1+x2, weights=wt))
Analysis of Deviance Table

Model: gaussian, link: identity

Response: y

Terms added sequentially (first to last)


     Df Deviance Resid. Df Resid. Dev
NULL                     6    208.957
x1    1  196.682         5     12.275
x2    1   11.514         4      0.760
Warning message:
In summary.glm(object, dispersion = dispersion) :
  observations with zero weight not used for calculating dispersion


Bill Dunlap
Spotfire, TIBCO Software
wdunlap tibco.com