Poisson and negative binomial regression with offset variable in SPSS (June 2019)

[@KaptenIglo]

Thank you for the instructive video. I very much appreciate you adding text to the video as well as stating the references. Keep up the great work!

[@mikecrowson2462]

Thanks, Kaptenlglo! I appreciate your feedback!

[@Jim_wabs]

Wow! You are good-you helped me sort out the issue of using offsets.

[@biraraaychew8927]

I got your video is useful and help me to understand poisson regression easily

[@jest120]

Thank you, Mike. It was a very useful explanation of this subject!

[@ЛюдмилаКшнясева-я5ш]

1) what if used the Ed-n as continious predictor? and, 2) Ln(age) - as explisit (insted offset) predictor?

[@Jim_wabs]

Hey Mike, for the Poisson you talked about the value/df values. But you did not touch on them when it came to the NB model. How do you interpret the value/df for NB, especially if it is less than 1.

[@jjk2902]

Wow, this is useful.Thank you professor!

[@mikecrowson2462]

Hi Jacob. Glad you found it useful! Thanks for checking watching.

[@mikecrowson2462]

For an updated video on Poisson and negative binomial regression (albeit without offset variable), please go here: kzbin.info/www/bejne/rXOqoaWnf6-firc

[@weihutchinson]

Hi Mike, thank you very much for this informative video. I have a question regarding offset. Is it possible to use two offset variables? I see offset as a controlled variable. In my model, I have two offsets, I am not sure how can I make two offsets work. Can you give some suggestions on this matter? Thank you in advance.

[@cristhianpacheco5306]

Well explained. Thanks!!

[@misslollipoppa]

Thanks Mike Crowson! Would a Poisson (or NegBin) also be possible with IV's that are based on Likert scale ??? So a DV = Count variables and 3 IV's based on 5 point likert scale means (scale variable)

[@mikecrowson2462]

Hi Isabelle, oftentimes Likert-type scales comprised of 5 or greater response categories are treated as continuous. (Of course, there is longstanding debate in the literature and some would argue it is better treated as ordinal. So you might get a different answer from someone who approaches things from that perspective. But the literature in psychology and education typically treats 5 points or greater as reflecting an underlying continuous variable). Also, when you refer to "likert scale means" are you referring to a composite measure of a construct? If so, then if you have 3 items with 5 response categories, then if you sum (or take the mean) of these items, then the theoretical range of values (if the points are 1 to 5) is 3 to 15. In that case, the composite measure is going to behave more like a measure of a continuous underlying variable given the larger number of possible values associated with it. In psychology and education (areas I come from), it is common to sum or average items designed to measure a given construct into composite measures for use in subsequent analyses. If this is what you are doing, then you really will only have 1 IV, which is scores on the composite measure. Hope this helps!

[@hoadaquy1984]

Please help me explain why age is needed to be transformed into ln.age (7:12). Thank you :)

[@toddpezzuti]

I can't really explain it, but Mike said that is how to treat an offset variable. Not all models need an offset variable though. As far as I understand.

[@wai-yeecheung7076]

Thank you for the brilliant explanation.

[@clemonique]

Thank you for this extremely helpful video! I'm currently analyzing longitudinal data (long format) and was wondering whether I can proceed the same way in this case or whether I have to take anything special into account. There are several companies in the data set and each of these companies obtained funding either one or several times over the years (ranging from 1-13).

[@mikecrowson2462]

Hi Clemonique, thanks for your feedback! Panel data is kind of tricky if you are working with count data. I've never had the occasion to do it, but there are possibilities for analyzing this kind of data if you have enough cases. One option is to perform panel regression using generalized estimating equations. In SPSS, you have the option for using different link functions, so theoretically you should be able to perform a Poisson regression through this route. Another option is to use HLM in SPSS. There is a Generalized Linear Mixed Models option that you can use, so you should be able to perform HLM with the longitudinal data that way. If you don't have a reasonably large number of cases (that have the repeated measures data) these approaches might not work or at least could lead you down the road of incorrect inferences. I haven't studied this issue in enough depth, though, to offer more beyond that. I hope this helps!

[@christenbroek2283]

Thanks for the awesome videos on NB and Poisson! I used all your information and guidance for performing them on my own :). I have one question. What if the main effects in the model do not significantly predict the DV, but then when adding the moderator everything becomes significant. Do you have any ideas on how to interpret that? For example, Pressure --> Failure (non-sig), Pressure --> Failure, moderated by friend support (Positive significant)

[@gezahegnterefe3513]

thank for your instructive video , I face difficulty in analyzing binary outcome using poisson regression model please prepare a video on modified poisson regression model for binary outcome in stata

[@sudembrahma7222]

Thank you so much

[@leylaisbilenyucel1004]

Perfect, thank you!

[@mikecrowson2462]

You are very welcome! Best wishes.

[@yulinliu850]

Great lecture! Thanks a lot!

[@mikecrowson2462]

Thanks Yulin for your feedback. Best wishes!

[@ewunetyibeltal6951]

great

[@pruec3766]

Thanks Prof for great video! I am having failed convergence problem with log binomial regression model with a dichotomous outcome.. Any recommendation of your videos or advice? Thanks