11

I am fitting the following model in "mgcv" package in R using option family=betar to predict a percentage cover response variable (cyano %):

    g6 = gam(cyano/100 ~ s(SEGLOWFLOW) + s(SEGJANAIRT) + 
             s(LOCHAB) + s(LOCSED) + s(T2PastoralHeavy) + 
             s(SEDO) + s(USDAYSRAIN) + s(USAVGSLOPE) + 
             s(USHARDNESS), data=nati1, 
             family = betar(link='logit'))

I get the following warning message:

    Warning message:
    In object$family$saturated.ll(G$y, wts, theta) :
      saturated likelihood may be inaccurate

The summary output gives me negative explained deviance values:

    R-sq.(adj) =  0.0871   Deviance explained = -29.9%
    -REML = -2370.5  Scale est. = 1         n = 463

Although the diagnostics plots (gam.check) look fine and the predictions make sense, I am worried about the warning message and the negative deviance explained value.

Any thoughts will be appreciated.

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
jatalah
  • 443
  • 1
  • 3
  • 8
  • I tried fixing some mistakes, but I am not sure whether you are using `betareg` anywhere. You do use `family=betar`, though. Please check the accuracy of your post and pardon if my edit might have introduced any new mistakes. – Richard Hardy Jul 15 '16 at 17:18
  • Great, thanks heaps. I only used betar family in the mgcv library. – jatalah Jul 18 '16 at 00:23
  • 1
    I'm running into the same question. Is it just that REML and dev.explained don't give sensible values with `betar` ? – puslet88 Feb 23 '19 at 14:26
  • Does your response variable contain 0's or 1's? If so, you likely need to increase `eps` within `betar()`. – cpp Aug 03 '20 at 16:35

0 Answers0