[Waseem Shaukat]

Yes, this looks fine.

You can graph these residuals, but I think it is hard to see what is going on when you do that. Unlike the residual plots that are sometimes done after linear regression, the residuals of a negative binomial regression are not going to be patterned in a way that is visually comprehensible. If you want to graph something, I would just plot the predicted values vs the observed values. The -predict- command will give you the predictive values. And then -graph twoway scatter- will produce the graphs. Nothing fancy or special needed here. The purpose here is not to "test assumptions" or anything like that--it's just to get a general impression of how the model looks compared with the observed values.

That makes sense. Can you share code(s) for predicted values vs the observed value scatter plot?

[Clyde Schechter]

glm ...
predict predicted, mu
graph twoway scatter mu observed

Add whatever options you like to the -graph twoway scatter- command to adjust its appearance to your taste.
I'm wondering why you asked for help with this--it's less complicated than the graphs you've already created on your own.

[Waseem Shaukat]

glm ...
predict predicted, mu
graph twoway scatter mu observed

Add whatever options you like to the -graph twoway scatter- command to adjust its appearance to your taste.
I'm wondering why you asked for help with this--it's less complicated than the graphs you've already created on your own.

Actually, I was unaware of how to incorporate observed in this graph. After running GLM, i use below commands and get error.

. predict predicted, mu

. graph twoway scatter mu observed
variable mu not found
r(111);

. graph twoway scatter predicted observed
variable observed not found
r(111);

What possibly I am doing wrong?

[Clyde Schechter]

OK, -graph twoway scatter mu observed- was my error. It should have been -graph twoway scatter predicted outcome-, as the -predict- command will create a variable predicted, and you already have a variable outcome, which is the observed value.

[Waseem Shaukat]

OK, -graph twoway scatter mu observed- was my error. It should have been -graph twoway scatter predicted outcome-, as the -predict- command will create a variable predicted, and you already have a variable outcome, which is the observed value.

Thank you so very much. I have played with this and below commands I used to generate Figure 1.

 graph twoway (scatter predicted outcome) (lfit predicted outcome)

Also, I used below command to compare predicted vs. observed in time series to generate below figure 2.

tsgraph predicted outcome

[Waseem Shaukat]

So the last question, I really hope this is last one :-):

- What is the best to report from this model. IRR with 95% CI and p-value, or it should also be accompanied with co-efficient alone or coefficient with 95% CI, and/or HAC Standard Error (and whether SE of coefficient or SE of IRR)?

Thanks.

[Clyde Schechter]

I would report the IRR with the 95% CI. I don't believe in using p-values for this kind of situation, but many people would report that, too. The coefficients are log-transforms of the incidence rate ratio, and I think they are harder for most people to understand. So I would not show the coefficients. Similarly, the standard errors are something that most people have difficulty understanding, so I would not show them either. The goal is to communicate things as understandably as possible. In my view the combination of the IRR and the 95% CI do that best.

[Waseem Shaukat]

I would report the IRR with the 95% CI. I don't believe in using p-values for this kind of situation, but many people would report that, too. The coefficients are log-transforms of the incidence rate ratio, and I think they are harder for most people to understand. So I would not show the coefficients. Similarly, the standard errors are something that most people have difficulty understanding, so I would not show them either. The goal is to communicate things as understandably as possible. In my view the combination of the IRR and the 95% CI do that best.

That makes sense. Thank you.

Please help me write the equation with Beta terms to report. Here is the final GLM model.

glm outcome predictor i.month, family(nbinomial 1) link(log) vce(hac nwest)

[Clyde Schechter]

The left hand side is ln E(Outcome). For ease of reading I would call the coefficient of Predictor B_Predictor + summation from 1 to 12 of I_{Month i}. (where I would use the great upper case sigma notation for the summation).

I do not know how to get those symbols to show in this Forum editor.

[Waseem Shaukat]

The left hand side is ln E(Outcome). The whole thing looks like

[IMG]file:///C:/Users/clyde/AppData/Local/Temp/msohtmlclip1/01/clip_image002.png[/IMG]

perhaps, the image uploaded in not visible.

[Clyde Schechter]

So I noticed. I've re-edited it as best I can (which is not good).

[Waseem Shaukat]

The left hand side is ln E(Outcome). For ease of reading I would call the coefficient of Predictor B_Predictor + summation from 1 to 12 of I_{Month i}. (where I would use the great upper case sigma notation for the summation).

I do not know how to get those symbols to show in this Forum editor.

Do you mind sending across the image via email ([email protected]) if possible. I dont want to see an error in the equation, so having your written equation would help me understand correctly. Thanks