0

I want to compare the incidence rate(asum) of 100 different cities(cityID) to see if there are significant differences among them. Given that the incidence rate is following Poisson, so it is a log-linear model.

m0=glm(asum~cityID, family=poisson, data=suit)
summary(m0)

However, the result looks so clumsy and stupid. Have I done anything wrong? Or is there any better way to analyse the data for my purpose?

Call:
glm(formula = asum ~ cityID, family = poisson, data = suit)

Deviance Residuals: 
  [1]  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
 [25]  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
 [49]  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
 [73]  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
 [97]  0  0  0  0

Coefficients:
            Estimate Std. Error z value Pr(>|z|)    
(Intercept)  3.63759    0.16222  22.424  < 2e-16 ***
cityID2      1.68542    0.17662   9.543  < 2e-16 ***
cityID3      1.79614    0.17516  10.254  < 2e-16 ***
cityID4      2.17355    0.17120  12.696  < 2e-16 ***
cityID5      1.02585    0.18908   5.426 5.78e-08 ***
cityID6      2.20306    0.17095  12.887  < 2e-16 ***
:
:
:
cityID90     0.92676    0.19166   4.836 1.33e-06 ***
cityID91     1.28239    0.18334   6.994 2.66e-12 ***
cityID92     2.40267    0.16940  14.183  < 2e-16 ***
cityID93     1.39937    0.18113   7.726 1.11e-14 ***
cityID94    -1.23969    0.34238  -3.621 0.000294 ***
cityID95     0.59652    0.20201   2.953 0.003148 ** 
cityID96    -2.25129    0.52566  -4.283 1.85e-05 ***
cityID97     1.86367    0.17435  10.689  < 2e-16 ***
cityID98     1.82625    0.17479  10.448  < 2e-16 ***
cityID99     1.96453    0.17322  11.341  < 2e-16 ***
cityID100    0.93712    0.19138   4.897 9.74e-07 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for poisson family taken to be 1)

    Null deviance: 1.0608e+04  on 99  degrees of freedom
Residual deviance: 2.8200e-14  on  0  degrees of freedom
AIC: 804.18

Number of Fisher Scoring iterations: 3

UPDATE

 m2=glm(asum~cityID+rainpd+temppd+distLon+offset(I(pop/1e5)), family=poisson, data=suit2)
> summary(m2)


cityID98     1.07292    0.03568  30.071  < 2e-16 ***
cityID99     0.88466    0.03536  25.020  < 2e-16 ***
cityID100    0.70534    0.03906  18.056  < 2e-16 ***
rainpd            NA         NA      NA       NA    
temppd            NA         NA      NA       NA    
distLon           NA         NA      NA       NA    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for poisson family taken to be 1)

    Null deviance: 6.6337e+04  on 2399  degrees of freedom
Residual deviance: 3.6238e-13  on 2300  degrees of freedom
AIC: 14700
lilkaskitc
  • 77
  • 1
  • 6
  • I appreciate this is not a direct answer to your question but would you not want to include an offset to account for city size? – mdewey Apr 27 '16 at 12:19
  • @mdewey I have done that already, and the result is similar. I am looking for a more high-level solution that altered the method of analysis. – lilkaskitc Apr 27 '16 at 14:56
  • See also https://stats.stackexchange.com/questions/146907/principled-way-of-collapsing-categorical-variables-with-many-levels – kjetil b halvorsen Mar 06 '20 at 23:24

1 Answers1

1

The fact that you are getting all the deviance residuals equal to zero suggests that your model is saturated in parameters, you are fitting one parameter per observation. Another way of looking at this is that city 1 must have a count of 38 $=e^{3.63759}$ and city 2 205 $=e^{3.63759+1.68542}$ and so on.

mdewey
  • 16,541
  • 22
  • 30
  • 57
  • I can split the data (annual) to have enough observations (daily), but how can I tell whether there are significant differences among the cities? Null deviance: 13049 on 2399 degrees of freedom Residual deviance: 2441 on 2300 degrees of freedom AIC: 9269.6 – lilkaskitc Apr 27 '16 at 16:00
  • With population offset: Null deviance: 5205 on 2399 degrees of freedom Residual deviance: 2441 on 2300 degrees of freedom AIC: 9269.6 – lilkaskitc Apr 27 '16 at 16:06
  • If you change from annual to daily you will have a very sparse dataset. To answer your other question try fitting the model with no moderator (asum~1) and then compare the two with anova(model1, model2, test = "Chisq") – mdewey Apr 27 '16 at 16:06
  • Very good, thank you. ANOVA, that's exactly what I want. So, this zero residual deviance is the best that I can get? – lilkaskitc Apr 27 '16 at 16:18
  • Please see the update above. As I want to include some more continuous explanatory variables, their estimates are shown as NA. I know there are too much parameters, how could I solve it? – lilkaskitc Apr 27 '16 at 16:26
  • You could drop city from the mode as I am not sure what it is telling you. – mdewey Apr 27 '16 at 17:17
  • city is significant as shown by ANOVA, but I think dropping it is the only way to go. Could you please also have a look of my another problem? http://stats.stackexchange.com/questions/209644/offset-in-glm-for-poisson-regression – lilkaskitc Apr 27 '16 at 17:22