I have response variable count data that should be treated as quasipoisson or something similar. This data also contains outliers which are important to the dataset. I cannot find an r package that will let me do robust regression with quasipoisson-type distribution. I could, however, collect weights for each model using robust regression and then try to use those weights in glm. Note those weights are always > 0. My problem is that I cannot find a clear definition of how weights are used in glm with family=quasipoisson. It sounds similar but I can't figure out if actually does what I want it to do (e.g. downweight outlier response variables to decrease their impact).
The response data are numbers of moths caught daily at pheromone traps, and the predictor variables are weather data. All variables are diffed to handle temporal autocorrelation. The outliers are high numbers of migratory moths in response to cold front passages.
I just found this reference which explains that low weights are interpreted as representing observations with a high variance, which does make sense in my case. But does that mean the effect is to reduce the impact of that observation on the regression? http://r.789695.n4.nabble.com/weights-in-glm-PR-8720-td910336.html