Poisson Regression and IRR

[Hamza Khan]

Hi everyone, we are currently working on a project where we look at Infection (CLABSI) rates compared quarterly to a baseline period.
The variables I have are CLABSI, Central Line Days and CLABSIRate (which is CLABSI/Central Line days) and Quarter.
The code I use is
poisson CLABSIRATE i.Quarter, irr
However the Quarters with 0 CLABSI is showing errors like IRR 1.50e-08
Any reason this is happening? Is this the right code

[Carlo Lazzaro]

Hamza:
welcome to this forum.
It seems that your regressand has been converted in a continuous variable.
Therefore, I would go OLS.
in addition, one prediction only is really not enough to give a fair and true view of the data generating process you're investigating.

[Weiwen Ng]

However the Quarters with 0 CLABSI is showing errors like IRR 1.50e-08
Any reason this is happening? Is this the right code

your DV is the rate of central line infections. You have some quarters with 0 infections. You enter quarters into the model as a factor variable.

If you had the baseline rate, what should the IRR for a quarter with 0 infections be? It should be some number that’s nearly 0. And 1.5 * 10 to the power of -8 is very close to 0.

there’s not enough information to say if this is the correct model, but what you are seeing is completely sensible given what you told us. If you switched to OLS, you would see betas for those quarters that would sum to nearly 0 when you add the constant term.

[Hamza Khan]

Hamza:
welcome to this forum.
It seems that your regressand has been converted in a continuous variable.
Therefore, I would go OLS.
in addition, one prediction only is really not enough to give a fair and true view of the data generating process you're investigating.

Thankyou for the reply Carlo,
So the data isn't linear. The reason we wanted to compare Infection(CLABSI) rates is because of an intervention we had put in place at our hospital.
So IRR is coming out fine I believe unless the CLABSI Rate falls down to zero

I've also tried the code

poisson Clabsiquarter i.Quarter, exposure (CLdaysQuarter) irr
where Clabsiquarter is number of CLABSI(central line infections) in a quarter and CLdaysQuarter is number of days exposed to Central line

[Hamza Khan]

your DV is the rate of central line infections. You have some quarters with 0 infections. You enter quarters into the model as a factor variable.

If you had the baseline rate, what should the IRR for a quarter with 0 infections be? It should be some number that’s nearly 0. And 1.5 * 10 to the power of -8 is very close to 0.

there’s not enough information to say if this is the correct model, but what you are seeing is completely sensible given what you told us. If you switched to OLS, you would see betas for those quarters that would sum to nearly 0 when you add the constant term.

Thankyou. But why is my p-valuq close to 1? Shouldn't it be closer to 0?

[Weiwen Ng]

Thankyou. But why is my p-valuq close to 1? Shouldn't it be closer to 0?

I don’t know, but I also think it doesn’t matter. You said above that you want to test incidence rates before and after an intervention. Now, I don’t know how to do inference in a one-arm, pre-post setting. But the key variable is going to be which quarter your intervention went into place. It doesn’t seem like you have that.

[Hamza Khan]

I don’t know, but I also think it doesn’t matter. You said above that you want to test incidence rates before and after an intervention. Now, I don’t know how to do inference in a one-arm, pre-post setting. But the key variable is going to be which quarter your intervention went into place. It doesn’t seem like you have that.

The baseline is the reference (quarter 1,2,3,4) everything after that from quarter 5 like in the image above is the intervention. The p values for all my "zero incidence rates" quarters are 0.997 for some reason. If incidence rate dropped to 0 shouldn't p value be significant?

[Hamza Khan]

I don’t know, but I also think it doesn’t matter. You said above that you want to test incidence rates before and after an intervention. Now, I don’t know how to do inference in a one-arm, pre-post setting. But the key variable is going to be which quarter your intervention went into place. It doesn’t seem like you have that.

Carlo Lazzaro Does Zero Inflated poisson have anything to do with this?

[Carlo Lazzaro]

Hamza:
welcome to this forum.
It seems that your regressand has been converted in a continuous variable.
Therefore, I would go OLS.
in addition, one prediction only is really not enough to give a fair and true view of the data generating process you're investigating.

[Hamza Khan]

My question is essentially when Incidence rate is 0 why is the p value not significant when compared to baseline Incidence rate? It's showing a standard 0.997 for every zero there is.

[Carlo Lazzaro]

Hamza:
no, I don't think so.
The issue is that you do not have counts in your regressand, but a continuous dependent variable.
I also second Weiwen Ng ' s wise advice about -dataex-.