I am trying to fit a glm using the Gamma family and the "identity" link. As I am analyzing nutrient concentrations in coastal waters (that can not be negative and are unlikely to be zero) I assume the Gamma distribution would be suitable. There are no zeros and no negative values in the data. If I have more than one independent variable (year, site, dist, season) and interactions in the model
I get this error :
Error: no valid set of coefficients has been found: please supply starting values In addition: Warning message: In log(ifelse(y == 0, 1, y/mu)) : NaNs produced
I do understand, that using the identity link and therefore a linear predictor may produce negative values. As Gamma can not handle negative values it produces the error.
Can I add a shift (eg. add +4 to all data) to avoid this error? If yes, how would you choose the shift?
If I use the log-link, the results are diverted by the transformation (significant differences between the smaller values but non in in the higher values).
Thank you for your ideas!