Can I run gamma mixed model in zelig for zero-augmented data? Urgent! - Zelig

Kosuke Imai

29 Mar 29 Mar

5:41 p.m.

Unfortunately, the gamma mixed model in the package cannot deal with zero's in the outcome variable. Kosuke -- Department of Politics Princeton University http://imai.princeton.edu On Thu, 26 Mar 2009, Xiulian Ma wrote:

...

Dear All, I try to run a gamma mixed model in zelig, but got the following error message: Error in eval(expr, envir, enclos) : non-positive values not allowed for the gamma family Besides the dependent structure among observations, a large proportion (about 75%) of my response variable is zero. I was advised that zelig gamma mixed model can handle my zero-augmented data. Is that true? The code for my model is: z.out1 <- zelig (OutDegr ~ Pr65 + Pr80 + Pr00 + Semipr65 + Semipr80 + Semipr00 + tag(1 | cityid), data = wc3, model = "gamma.mixed") Could anybody give me some help? Thank you very much! Xiulian

- Zelig Mailing List, served by Harvard-MIT Data Center Send messages: zelig(a)lists.gking.harvard.edu [un]subscribe Options: http://lists.gking.harvard.edu/?info=zelig Zelig program information: http://gking.harvard.edu/zelig/

Reply

Kosuke Imai

31 Mar 31 Mar

8:50 p.m.

Yes, certainly you can use quasi-likelihood approach to take into account for the overdispersion. Unfortunately, Zelig does not have zero-inflated models so if you want to use those models you would have to look elsewhere... Kosuke -- Department of Politics Princeton University http://imai.princeton.edu On Sun, 29 Mar 2009, Xiulian Ma wrote:

...

Dear professor Imai, Thanks a lot for the reply! I switched from poisson.mixed to poisson.mixed. But I am still a little unsure about several things, one concerning the over-dispersion issue, and another concerning the inflated-zero. My data is to examine whether a country's world system position significantly increases the importance of cities located in the country, and whether such effects increased or declined across time. So, I have two levels, country and city. Data are measured longitudinally. The importance of cities (variable name: Outdegree) can be regarded as counts--which is--0, 1, 2, 3, ....... It is heavily zero-inflated (about 70% is zero). Wsln80(wsln90, wsln00) is the world system variables (in each time point), and is log-transformed. I set the model as following: z.out <- zelig(Outdegree ~ WSln80 + WSln90 + WSln00 + Y1990 + Y2000 + tag(1|Country_Code) + tag(1 | cityid), data = wc3, model = "poisson.mixed", family = quasipoisson(link="log")) I add "family = quasipoisson" to take into account the overdispersion (for the variable outdegree, its standard deviation is 10 times of its mean). I have two questions: 1. Is my way dealing with overdispersion correct or/and sufficient? 2. My data is zero-inflated. How to take this nature of my data into account on the basis of current model? Or do I have to switch to other models? I hope you can give some advice! Thank you very much! Xiulian -----Original Message----- From: Kosuke Imai [mailto:kimai@Princeton.EDU] Sent: Sunday, March 29, 2009 3:42 PM To: Xiulian Ma Cc: zelig(a)lists.gking.harvard.edu Subject: Re: [zelig] Can I run gamma mixed model in zelig for zero-augmented data? Urgent! Unfortunately, the gamma mixed model in the package cannot deal with zero's in the outcome variable. Kosuke -- Department of Politics Princeton University http://imai.princeton.edu On Thu, 26 Mar 2009, Xiulian Ma wrote:

Dear All, I try to run a gamma mixed model in zelig, but got the following error message: Error in eval(expr, envir, enclos) : non-positive values not allowed for the gamma family Besides the dependent structure among observations, a large proportion (about 75%) of my response variable is zero. I was advised that zelig gamma mixed model can handle my zero-augmented data. Is that true? The code for my model is: z.out1 <- zelig (OutDegr ~ Pr65 + Pr80 + Pr00 + Semipr65 + Semipr80 + Semipr00 + tag(1 | cityid), data = wc3, model = "gamma.mixed") Could anybody give me some help? Thank you very much! Xiulian

- Zelig Mailing List, served by Harvard-MIT Data Center Send messages: zelig(a)lists.gking.harvard.edu [un]subscribe Options: http://lists.gking.harvard.edu/?info=zelig Zelig program information: http://gking.harvard.edu/zelig/

Reply

Xiulian Ma

2 Apr 2 Apr

2 a.m.

Dear professor Lau, Thanks a lot for the instruction. I will try the tobit model. But I still have a few concerns: 1. What do you mean by "normal regression'? You mean normal distribution of the dependent variable? My dependent variable is highly skewed, and even the transformation will not get rid of the skewness because of too many zeros in the data. (That is why I attempted to use a gamma distribution model at the beginning) 2. Can tobit model deals with multilevel issue? (I quickly checked the manual, and it looks like can, but I am not so sure). My data has two levels, city and country--cities are located in countries. At the same time, observations of city and country are repeatedly measured across time. So the data combined both a hierarchical structure and repeated measures. Can tobit model handle that? Please advice. 3. Can the tobit model do some weighting? My data is extremely unbalanced--some rich country like USA has thousands of cities included in the sample, but for a poor or small country, there is only one or two cities included in the sample. I think I need to weight the data in some way. I hope you can give me some advice(I have been struggling with these complicated model issues), and really appreciate your help! Xiulian Ps, I also appreciate any advice and help from the rest of the group! -----Original Message----- From: monkeykupo(a)gmail.com [mailto:monkeykupo@gmail.com] On Behalf Of Olivia Lau Sent: Wednesday, April 01, 2009 8:05 AM To: Kosuke Imai Cc: Xiulian Ma; zelig(a)lists.gking.harvard.edu Subject: Re: [zelig] Can I run gamma mixed model in zelig for zero-augmented data? Urgent! Xiulian, If I understand the problem, you have a continuous dependent variable that is truncated on the left at zero. Why not use a tobit model? This is a normal regression truncated at zero from the left. It's not a zero-inflated model, but it fits the shape of your data -- and is in Zelig. Yours, Olivia On Tue, Mar 31, 2009 at 8:50 PM, Kosuke Imai <kimai(a)princeton.edu> wrote:

...

Yes, certainly you can use quasi-likelihood approach to take into account for the overdispersion. Unfortunately, Zelig does not have zero-inflated models so if you want to use those models you would have to look elsewhere... Kosuke -- Department of Politics Princeton University http://imai.princeton.edu On Sun, 29 Mar 2009, Xiulian Ma wrote:

Dear professor Imai, Thanks a lot for the reply! I switched from poisson.mixed to poisson.mixed. But I am still a little unsure about several things, one concerning the over-dispersion issue, and another concerning the inflated-zero. My data is to examine whether a country's world system position significantly increases the importance of cities located in the country, and whether such effects increased or declined across time. So, I have two levels, country and city. Data are measured longitudinally. The importance of cities (variable name: Outdegree) can be regarded as counts--which is--0, 1, 2, 3, ....... It is heavily zero-inflated (about 70% is zero). Wsln80(wsln90, wsln00) is the world system variables (in each time point), and is log-transformed. I set the model as following: z.out <- zelig(Outdegree ~ WSln80 + WSln90 + WSln00 + Y1990 + Y2000 + tag(1|Country_Code) + tag(1 | cityid), data = wc3, model = "poisson.mixed", family = quasipoisson(link="log")) I add "family = quasipoisson" to take into account the overdispersion (for the variable outdegree, its standard deviation is 10 times of its mean). I have two questions: 1. Is my way dealing with overdispersion correct or/and sufficient? 2. My data is zero-inflated. How to take this nature of my data into account on the basis of current model? Or do I have to switch to other models? I hope you can give some advice! Thank you very much! Xiulian -----Original Message----- From: Kosuke Imai [mailto:kimai@Princeton.EDU] Sent: Sunday, March 29, 2009 3:42 PM To: Xiulian Ma Cc: zelig(a)lists.gking.harvard.edu Subject: Re: [zelig] Can I run gamma mixed model in zelig for zero-augmented data? Urgent! Unfortunately, the gamma mixed model in the package cannot deal with zero's in the outcome variable. Kosuke -- Department of Politics Princeton University http://imai.princeton.edu On Thu, 26 Mar 2009, Xiulian Ma wrote:

Dear All, I try to run a gamma mixed model in zelig, but got the following error message: Error in eval(expr, envir, enclos) : non-positive values not allowed for the gamma family Besides the dependent structure among observations, a large proportion (about 75%) of my response variable is zero. I was advised that zelig gamma mixed model can handle my zero-augmented data. Is that true? The code for my model is: z.out1 <- zelig (OutDegr ~ Pr65 + Pr80 + Pr00 + Semipr65 + Semipr80 + Semipr00 + tag(1 | cityid), data = wc3, model = "gamma.mixed") Could anybody give me some help? Thank you very much! Xiulian

- Zelig Mailing List, served by Harvard-MIT Data Center Send messages: zelig(a)lists.gking.harvard.edu [un]subscribe Options: http://lists.gking.harvard.edu/?info=zelig Zelig program information: http://gking.harvard.edu/zelig/

Reply