I want to fit a fairly "standard" Poisson model, but with an autoregressive term.

$N_i \sim \mathrm{Pois}( \lambda_i E_i)$

with $\log \lambda_i = X_i \beta + \delta$

$\delta \sim AR(1)$

$X_i$ is a vector of covariates. $\beta$ are my coefficients. $\delta$ is an autoregressive term. $E_i$ is the size of population at time t.

The idea is that the count at time step $t$ is partially dependent on the count at time step $t-1$.

Ideally, I'd like to find some R package to fit this.

Any suggestions?

  • 4
    $\begingroup$ It would be good to consider accepting answers to some of your previous questions, all of which have received multiple answers, thus giving you some choice. There is a check mark next to each answer that you can click on to indicate which one has been addressed your query. $\endgroup$ – cardinal Feb 5 '12 at 19:06
  • $\begingroup$ Can you give some more detail as to what kind of autoregressive structure you want to assume. It's a little ambiguous at the moment. Defining $E_i$ would also be helpful. Cheers. :) $\endgroup$ – cardinal Feb 5 '12 at 19:07
  • $\begingroup$ This is an epidemiological model. The dependent variable is the number of people with a disease at time t. I can fit it reasonably well with a "standard" poisson, but it was suggested that an autoregressive term might work well for this particular study. $\endgroup$ – Noah Feb 5 '12 at 19:18
  • $\begingroup$ I guess I'm wondering what sort of autoregressive formulation you want. Are you thinking of something like $\log \lambda_i= X_i \beta + \alpha \log \lambda_{i-1} + \varepsilon_i$ where $\varepsilon_i$ is some additional randomness driving the evolution of the rate parameter? And, if this is an epidemiological model, is $N_i$ some number of, say, infected individuals? If so, then it would seem $\lambda_i \ll 1$, otherwise there is nonnegligible probability of more people than exist in the population becoming infected at time $i$. But, maybe I'm misunderstanding what you're aiming for. $\endgroup$ – cardinal Feb 5 '12 at 19:44
  • $\begingroup$ You understand perfectly. Nice summary. $N_i$ is the number of people with the disease and $\lambda$ is definitely less than 1. $\endgroup$ – Noah Feb 5 '12 at 19:50

I think you are looking for the model in Brandt et al. (2000) there called PEWMA, after the forecast function. R code to fit it is available here.

The paper also has some general discussion of possible conditionally Poisson AR models. Fro more of that, chapter 7 of Cameron and Trivedi (1998) is useful.


Have you considered a Transfer Function between N and E and your other covariates which could encode changes in parameters over time , changes in error variance over time , any necessaery autoregressive structure possibly proxying unspecified seasonal drivers and other omitted structure like level shifts, seasonal pulses and local time trends. Accomplished without letting "unusual values" distort the model/parameters. I have seen approaches like this referred to in the literature as DARIMA MODELS where the D stands for discrete. Various forms of Power Transforms such as Logs might be needed to possibly decouple the expected value from the variance of the errors ( always a good idea when needed! ) These kinds of models can be useful in detecting the possible impact of changes in N as it relates to the prediction of E.

  • 1
    $\begingroup$ Without getting into too many details, I don't believe that model would work for this data set. What I have are the counts of disease along with some categorial factors for each year. (i.e. Asian, Men, 2005, 10 out of 500 had disease.) $\endgroup$ – Noah Feb 6 '12 at 19:41

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.