4
$\begingroup$

tldr; Aside from equidispersion, what are the model assumptions I should be checking for in a poisson mixed effects model that has a random intercept, group mean centered transformations of its explanatory variables, and group means included as variables?

I have a county-year dataset with 21 counties that each have 8 years of data (N = 168). I am attempting to model the number of prescription opioid related hospitalizations for this data using the number of prescription opioid pills supplied to each county in a year, the prescription rate of each county in a year, and demographic and economic (unemployment rate, median household income) variables also at the county-year level.

I am using a poisson distribution (with a log link) and the lme4 package to estimate this model with a random intercept, group mean centered transformations of each variable, and the group mean of each variable in line with the specification found in Bell and Jones, 2015 . Additionally, I have been using the DHARMa package to visually examine the relationships between my explanatory variables and the randomized quantile residuals from the model and to test for overdispersion.

I am not formally trained in mixed effects modeling and want to be as sure as one can be that the estimates from my model are not biased. What are the assumptions of a poisson mixed effects model and is there a rigorous set of steps for testing these assumptions (either by looking at residuals or any other part of the model output)?

Thank you in advance for any help!

$\endgroup$

1 Answer 1

2
$\begingroup$

If you're using R, this is a great place to start for count models: https://cran.r-project.org/web/packages/pscl/vignettes/countreg.pdf

You'll be able to see all the models/packages available and their accompanying assumptions.

For Poisson, the main model assumptions are:

  1. The variance is equal to the mean. As you have stated, overdispersion can occur when your variance is much higher than your mean. DHARMa provides effective tools to quantify overdispersion, and when it is present, you would want to consider quasi-Poisson, negative binomial, or zero-inflated models (see CRAN package link above).

  2. The response variable is non-negative integer data.

  3. The responses are independent from one another.

  4. The responses occur over fixed time or space.

When you are incorporating mixed effects, you want to make sure that your random effect is some kind of grouping variable (perhaps a county ID in your case). You specify this to account for any in-county differences that may account for the number of opioid-related hospitalizations.

An overview of linear mixed models, and how to specify random effects can be found here: https://www.youtube.com/watch?v=QCqF-2E86r0&t=95s

Hope this helps.

$\endgroup$
3
  • $\begingroup$ Thank you for this answer! This is really helpful. In regards to the independence assumption, given that I am using a random-intercept for each county, should the responses be independent from county to county or does this assumption need to hold within county as well? $\endgroup$
    – svitkin
    Jun 12, 2020 at 16:50
  • 1
    $\begingroup$ The random intercept accounts for correlated outcomes within county. But the model assumes no correlations across counties (once you account for county of hospitalization) and your covariates. If all the counties come from the same state, this is probably a safe assumption, but if they come from different states, then you would want to account for that either with a set of dummy variables for states or a state random effect. With very few states, you should probably use the dummy approach. $\endgroup$
    – Erik Ruzek
    Jun 12, 2020 at 23:14
  • $\begingroup$ Great! Thank you, they're all from the same state so I'll stick with that assumption for now. $\endgroup$
    – svitkin
    Jun 15, 2020 at 15:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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