4
$\begingroup$

I'd like to be able to calculate the confidence interval for an incidence rate under the Poisson distribution. I am inclined to use a normal approximation formula (the Wald interval) which is $\text{IR} \mp 1.96\,\sqrt{\text{IR}/N}$

However, most sources on line (for example here: http://epid.blogspot.com/2012/08/how-to-calculate-confidence-interval-of.html) use $\text{IR} \mp 1.96\,\text{IR}/\sqrt{N}$, which is not the same thing and the results are quite different.

Does anyone happen to know what is this formula and which one should I use?

$\endgroup$
  • $\begingroup$ 1. Can you show your setup (what's IR, what's N? Are they the same as at the link?) and the derivation, including assumptions? 2. Take a look at the discussion in comments under your link, particularly with what's going on with the N and T, since some of that may be relevant to your issue $\endgroup$ – Glen_b -Reinstate Monica Sep 7 '17 at 1:03
  • $\begingroup$ I'm guessing you're asking about the situation where the number of incidents for an individual is $Y_i \sim \text{Pois}(\lambda T_i)$ where the incidence rate per unit time is $\lambda$, $T_i$ is the exposure time for individual $i$ and that the $Y_i$ variates are independent. Then we define the random variable $N=\sum_i Y_i$ and $T=\sum_i T_i$, in which case $\hat{\lambda}=N/T$ is the MLE of $\lambda$ (this sort of thing is what I meant by "setup and derivation including assumptions"); you seek a CI for $\hat{\lambda}$. Do I have that correct? $\endgroup$ – Glen_b -Reinstate Monica Sep 7 '17 at 1:16
3
$\begingroup$

Assuming I understand the situation correctly:

the number of incidents for an individual is $Y_i \sim \text{Pois}(\lambda \, T_i)$ where the incidence rate per unit time is $\lambda$, $T_i$ is the exposure time for individual $i$ and the $Y_i$ variates are independent. Then we define the random variable $N=\sum_i Y_i$ and the total exposure $T=\sum_i T_i$, in which case $\hat{\lambda}=N/T$ is the MLE of $\lambda$

you seek a CI for $\hat{\lambda}$

$\text{Var}(\hat{\lambda}) = \text{Var}(\frac{N}{T}) = \frac{1}{T^2}\text{Var}(N)=\frac{1}{T^2}\text{Var}(\sum_iY_i)=\frac{1}{T^2}(\sum_i\lambda T_i)=\lambda/T$

Hence the estimated standard error is $\sqrt{\hat{\lambda}/T} = \frac{\sqrt{N}}{T} = \frac{N\times \frac{1}{\sqrt{N}}}{T}=\frac{N}{T}\times \frac{1}{\sqrt{N}}=\frac{\text{IR}}{\sqrt{N}}$

So (if $N>0$), either of these two formulas is okay:

$\text{s.e.}(\text{IR}) = \sqrt{\frac{IR}{T}} = \frac{IR}{\sqrt{N}}$

and so a large sample (large-$N$) interval for the $\text{IR}$ would multiply that by a suitable $Z$ value to get the interval half-width either side of the estimate.

You seem to have perhaps muddled up the roles of $N$ and $T$ in your calculation

Actually, it looks like the link you gave explained how their formula came about pretty clearly in the couple of lines under "If $N$ is large".

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.