[Please keep r-help in the cc: list] I don't quite know how to interpret the difference between specifying effort as an offset vs. as weights; I would have to spend more time thinking about it/working through it than I have available at the moment. I don't know that specifying effort as weights is *wrong*, but I don't know that it's right or what it is doing: if I were the reviewer of a paper (for example) I would require you to explain what the difference is and convince me that it was appropriate. (Furthermore, "I want to do it this way because it gives me significant effects" is automatically suspicious.) This would be a good question for CrossValidated (https://stats.stackexchange.com), you could try posting it there (I would be interested in the answer!) cheers Ben Bolker On 2023-10-30 8:19 p.m., ??? wrote:> Dear Mr. Bolker, > > Thank you for the fast response. > > I also know?that a poisson (or negative binomial ) regression of glm is > generally modelled using an offset variable. > > In this case, when a weights term instead of the offset is used, this > gave me significant coefficients of covariance. > I understand that the weights function for exponential family > distributions in glm affects the variance of response variable. > > I was just wondering?whether my first model is a completely wrong model > and the use of offset variable?is valid in the case that > response variable is? not proportional to offset variable such as my > dataset. > > Sincerely, > > > Joon-Taek > > 2023? 10? 29? (?) ?? 3:25, Ben Bolker <bbolker at gmail.com > <mailto:bbolker at gmail.com>>?? ??: > > ? ?Using an offset of log(Effort) as in your second model is the more > standard way to approach this problem; it corresponds to assuming that > catch is strictly proportional to effort. Adding log(Effort) as a > covariate (as illustrated below) tests whether a power-law model (catch > propto (Effort)^(b+1), b!=0) is a better description of the data.? (In > this case it is not, although the confidence intervals on b are very > wide, indicating that we have very little information -- this is not > surprising since the proportional range of effort is very small > (246-258) in this data set. > > ? ?In general you should *not* check overdispersion of the raw data > (i.e., the *marginal distribution* of the data, you should check > overdispersion of a fitted (e.g. Poisson) model, as below. > > ? ?cheers > ? ? Ben Bolker > > > edata <- data.frame(Catch, Effort, xx1, xx2, xx3) > > ## graphical exploration > > library(ggplot2); theme_set(theme_bw()) > library(tidyr) > edata_long <- edata |> pivot_longer(names_to="var", cols =-c("Catch", > "Effort")) > ggplot(edata_long, aes(value, Catch)) + > ? ? ?geom_point(alpha = 0.2, aes(size = Effort)) + > ? ? ?facet_wrap(~var, scale="free_x") + > ? ? ?geom_smooth(method = "glm", method.args = list(family > "quasipoisson")) > # > > library(MASS) > g1 <- glm.nb(Catch~xx1+xx2+xx3+offset(log(Effort)), data=edata) > g2 <- update(g1, . ~ . + log(Effort)) > g0 <- glm(Catch~xx1+xx2+xx3+offset(log(Effort)), data=edata, > ? ? ? ? ? ?family = poisson) > performance::check_overdispersion(g0) > summary(g1) > summary(g2) > options(digits = 3) > confint(g2) > summary(g1) > > > > On 2023-10-28 3:30 a.m., ??? wrote: > > Colleagues, > > > > > > > > I have a dataset that includes five variables. > > > > - Catch: the catch number counted in some species (ind.) > > > > - Effort: fishing effort (the number of fishing vessels) > > > > - xx1, xx2, xx3: some environmental factors > > > > As an overdispersion test on the ?Catch? variable, I modeled with > negative > > binomial distribution using a GLM. The ?Effort? variable showed a > gradually > > decreasing trend during the study period. I was able to get the > results I > > wanted when considered ?Effort? function as a weights function in the > > negative binomial regression as follows: > > > > > > > > library(qcc) > > > > > Catch=c(25,2,7,6,75,5,1,4,66,15,9,25,40,8,7,4,36,11,1,14,141,9,74,38,126,3) > > > > > Effort=c(258,258,258,258,258,258,258,254,252,252,252,252,252,252,252,252,252,252,252,248,246,246,246,246,246,246) > > > > > xx1=c(0.8,0.5,1.2,0.5,1.1,1.1,1.0,0.6,0.9,0.5,1.2,0.6,1.2,0.7,1.0,0.6,1.6,0.7,0.8,0.6,1.7,0.9,1.1,0.5,1.4,0.5) > > > > > xx2=c(1.7,1.6,2.7,2.6,1.5,1.5,2.8,2.5,1.7,1.9,2.2,2.4,1.6,1.4,3.0,2.4,1.4,1.5,2.2,2.3,1.7,1.7,1.9,1.9,1.4,1.4) > > > > > xx3=c(188,40,2,10,210,102,117,14,141,28,48,15,220,115,10,14,320,20,3,10,400,150,145,160,460,66) > > > > # > > > > edata <- data.frame(Catch, Effort, xx1, xx2, xx3) > > > > # > > > > qcc.overdispersion.test(edata$Catch, type="poisson") > > > > # > > > > summary(glm.nb(Catch~xx1+xx2+xx3, weights=Effort, data=edata)) > > > > summary(glm.nb(Catch~xx1+xx2+xx3+offset(log(Effort)), data=edata)) > > > > > > > > I am not sure the application of the weights function to the negative > > binomial regression is correct. Also I wonder if there is a > better way > > doing this. Can anyone help? > > > >? ? ? ?[[alternative HTML version deleted]] > > > > ______________________________________________ > > R-help at r-project.org <mailto:R-help at r-project.org> mailing list > -- To UNSUBSCRIBE and more, see > > https://stat.ethz.ch/mailman/listinfo/r-help > <https://stat.ethz.ch/mailman/listinfo/r-help> > > PLEASE do read the posting guide > http://www.R-project.org/posting-guide.html > <http://www.R-project.org/posting-guide.html> > > and provide commented, minimal, self-contained, reproducible code. > > ______________________________________________ > R-help at r-project.org <mailto:R-help at r-project.org> mailing list -- > To UNSUBSCRIBE and more, see > https://stat.ethz.ch/mailman/listinfo/r-help > <https://stat.ethz.ch/mailman/listinfo/r-help> > PLEASE do read the posting guide > http://www.R-project.org/posting-guide.html > <http://www.R-project.org/posting-guide.html> > and provide commented, minimal, self-contained, reproducible code. >
peter dalgaard
2023-Nov-02 07:28 UTC
[R] weights vs. offset (negative binomial regression)
I think it is more clear-cut than so, at least if the Poisson situation is something to go by. There, you can do either of these and get equivalent results> fit.lung <- glm(cases ~ age + city, offset=log(pop),+ family=poisson, data=lungcancer)> fit.lung2 <- glm(cases/pop ~ age + city, weights=pop,+ family=poisson, data=lungcancer) There were 12 warnings (use warnings() to see them) (Except for the warnings about non-integer responses, which have annoyed some epidemiologists trying to work with non-log link fuctions.) The point is that you need to convert to rates on the LHS and then compensate for the fact that this has a smaller variance when the population is larger. Counts on the LHS combined with weights wouldn't be right. So I would expect that the weighted version of the OP's code should model Catch/Effort, although I'm not quite sure how glm.nb reacts to non-integer responses. -pd> On 31 Oct 2023, at 17:59 , Ben Bolker <bbolker at gmail.com> wrote: > > [Please keep r-help in the cc: list] > > I don't quite know how to interpret the difference between specifying effort as an offset vs. as weights; I would have to spend more time thinking about it/working through it than I have available at the moment. > > I don't know that specifying effort as weights is *wrong*, but I don't know that it's right or what it is doing: if I were the reviewer of a paper (for example) I would require you to explain what the difference is and convince me that it was appropriate. (Furthermore, "I want to do it this way because it gives me significant effects" is automatically suspicious.) > > This would be a good question for CrossValidated (https://stats.stackexchange.com), you could try posting it there (I would be interested in the answer!) > > cheers > Ben Bolker > > > On 2023-10-30 8:19 p.m., ??? wrote: >> Dear Mr. Bolker, >> Thank you for the fast response. >> I also know that a poisson (or negative binomial ) regression of glm is generally modelled using an offset variable. >> In this case, when a weights term instead of the offset is used, this gave me significant coefficients of covariance. >> I understand that the weights function for exponential family distributions in glm affects the variance of response variable. >> I was just wondering whether my first model is a completely wrong model and the use of offset variable is valid in the case that >> response variable is not proportional to offset variable such as my dataset. >> Sincerely, >> Joon-Taek >> 2023? 10? 29? (?) ?? 3:25, Ben Bolker <bbolker at gmail.com <mailto:bbolker at gmail.com>>?? ??: >> Using an offset of log(Effort) as in your second model is the more >> standard way to approach this problem; it corresponds to assuming that >> catch is strictly proportional to effort. Adding log(Effort) as a >> covariate (as illustrated below) tests whether a power-law model (catch >> propto (Effort)^(b+1), b!=0) is a better description of the data. (In >> this case it is not, although the confidence intervals on b are very >> wide, indicating that we have very little information -- this is not >> surprising since the proportional range of effort is very small >> (246-258) in this data set. >> In general you should *not* check overdispersion of the raw data >> (i.e., the *marginal distribution* of the data, you should check >> overdispersion of a fitted (e.g. Poisson) model, as below. >> cheers >> Ben Bolker >> edata <- data.frame(Catch, Effort, xx1, xx2, xx3) >> ## graphical exploration >> library(ggplot2); theme_set(theme_bw()) >> library(tidyr) >> edata_long <- edata |> pivot_longer(names_to="var", cols =-c("Catch", >> "Effort")) >> ggplot(edata_long, aes(value, Catch)) + >> geom_point(alpha = 0.2, aes(size = Effort)) + >> facet_wrap(~var, scale="free_x") + >> geom_smooth(method = "glm", method.args = list(family >> "quasipoisson")) >> # >> library(MASS) >> g1 <- glm.nb(Catch~xx1+xx2+xx3+offset(log(Effort)), data=edata) >> g2 <- update(g1, . ~ . + log(Effort)) >> g0 <- glm(Catch~xx1+xx2+xx3+offset(log(Effort)), data=edata, >> family = poisson) >> performance::check_overdispersion(g0) >> summary(g1) >> summary(g2) >> options(digits = 3) >> confint(g2) >> summary(g1) >> On 2023-10-28 3:30 a.m., ??? wrote: >> > Colleagues, >> > >> > >> > >> > I have a dataset that includes five variables. >> > >> > - Catch: the catch number counted in some species (ind.) >> > >> > - Effort: fishing effort (the number of fishing vessels) >> > >> > - xx1, xx2, xx3: some environmental factors >> > >> > As an overdispersion test on the ?Catch? variable, I modeled with >> negative >> > binomial distribution using a GLM. The ?Effort? variable showed a >> gradually >> > decreasing trend during the study period. I was able to get the >> results I >> > wanted when considered ?Effort? function as a weights function in the >> > negative binomial regression as follows: >> > >> > >> > >> > library(qcc) >> > >> > >> Catch=c(25,2,7,6,75,5,1,4,66,15,9,25,40,8,7,4,36,11,1,14,141,9,74,38,126,3) >> > >> > >> Effort=c(258,258,258,258,258,258,258,254,252,252,252,252,252,252,252,252,252,252,252,248,246,246,246,246,246,246) >> > >> > >> xx1=c(0.8,0.5,1.2,0.5,1.1,1.1,1.0,0.6,0.9,0.5,1.2,0.6,1.2,0.7,1.0,0.6,1.6,0.7,0.8,0.6,1.7,0.9,1.1,0.5,1.4,0.5) >> > >> > >> xx2=c(1.7,1.6,2.7,2.6,1.5,1.5,2.8,2.5,1.7,1.9,2.2,2.4,1.6,1.4,3.0,2.4,1.4,1.5,2.2,2.3,1.7,1.7,1.9,1.9,1.4,1.4) >> > >> > >> xx3=c(188,40,2,10,210,102,117,14,141,28,48,15,220,115,10,14,320,20,3,10,400,150,145,160,460,66) >> > >> > # >> > >> > edata <- data.frame(Catch, Effort, xx1, xx2, xx3) >> > >> > # >> > >> > qcc.overdispersion.test(edata$Catch, type="poisson") >> > >> > # >> > >> > summary(glm.nb(Catch~xx1+xx2+xx3, weights=Effort, data=edata)) >> > >> > summary(glm.nb(Catch~xx1+xx2+xx3+offset(log(Effort)), data=edata)) >> > >> > >> > >> > I am not sure the application of the weights function to the negative >> > binomial regression is correct. Also I wonder if there is a >> better way >> > doing this. Can anyone help? >> > >> > [[alternative HTML version deleted]] >> > >> > ______________________________________________ >> > R-help at r-project.org <mailto:R-help at r-project.org> mailing list >> -- To UNSUBSCRIBE and more, see >> > https://stat.ethz.ch/mailman/listinfo/r-help >> <https://stat.ethz.ch/mailman/listinfo/r-help> >> > PLEASE do read the posting guide >> http://www.R-project.org/posting-guide.html >> <http://www.R-project.org/posting-guide.html> >> > and provide commented, minimal, self-contained, reproducible code. >> ______________________________________________ >> R-help at r-project.org <mailto:R-help at r-project.org> mailing list -- >> To UNSUBSCRIBE and more, see >> https://stat.ethz.ch/mailman/listinfo/r-help >> <https://stat.ethz.ch/mailman/listinfo/r-help> >> PLEASE do read the posting guide >> http://www.R-project.org/posting-guide.html >> <http://www.R-project.org/posting-guide.html> >> and provide commented, minimal, self-contained, reproducible code. >> > > ______________________________________________ > R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide http://www.R-project.org/posting-guide.html > and provide commented, minimal, self-contained, reproducible code.-- Peter Dalgaard, Professor, Center for Statistics, Copenhagen Business School Solbjerg Plads 3, 2000 Frederiksberg, Denmark Phone: (+45)38153501 Office: A 4.23 Email: pd.mes at cbs.dk Priv: PDalgd at gmail.com