Equivalent of incident rate ratios for continuous variables

[Rob Wood]

Hi There,

When analysing count data across subjects with varying follow-up, one of the most appropriate descriptive statistics to report is the Poisson/incident rate, whereby the sum of all events across all subjects is divided by the total person-time.

However, I would like to know if there is an equivalent appropriate statistic when assessing continuous variables across subjects with varying follow-up e.g. costs incurred? Or what the best approach is for describing these data.

Any input would be appreciated.

Thanks,

Rob.

[Carlo Lazzaro]

Rob:
if you're dealing with a continuos variable such as cost, the best way to describe it isto report mean and standard deviation (at minimum).
I am more satisfied when I can report median (since cost distributions are usually positively skewed)and range, too.
Obviously, your choice is conditional on competing factors among those towers the difference of possible dissemination means (research report; abstract to be submitted to a congress; poster/podium presentation if the abstract is accepted; article to be submitted to a target journal in your research field).

[Marcos Almeida]

However, I would like to know if there is an equivalent appropriate statistic when assessing continuous variables across subjects with varying follow-up e.g. costs incurred?

I believe Carlo gave an overarching approach to the issue.

I just wish to add that, IMHO, "the equivalent" goes in the opposite way, I mean, for example, we have "standard" linear regression for continuous variables, and its equivalent for count data (well, one of them) is the Poisson model.

The same with longitudinal data. We'd have - xtreg - when the DV is continuous, and the equivalent would be, well, among a couple of options - xtpoisson.

To end, under a model with a continuous DV, we may get coefficients, margins, elasticities, etc.

[Weiwen Ng]

Actually, Rob's question seems to deal with the fact that people could be exposed for different time periods. For example, in health care, imagine that one person was enrolled in some health insurance plan for 10 months, whereas some other person was enrolled for 12 months. Imagine that you have each person's total healthcare spending over the time they were enrolled. Those totals aren't comparable - hence, people in my field tend to calculate and analyze cost per person per month (PMPM). That's the approach I would suggest.

[Carlo Lazzaro]

Rob:
as a second thought, different follow-ups may imply quite different things: patient has passed away before the end of the study; patients has dropped out from the study; patient has completed the study.
These three possible scenarios imply different methods to deal with complete and incomplete cost data.
The literature on this topic is really wide and suggests different approaches, which different level of complexity.
You may want to take a look at:
-https://www.ncbi.nlm.nih.gov/pubmed/16336017;
- https://www.herc.ox.ac.uk/downloads/...linical-trials (Chapter 6);
- https://www.herc.ox.ac.uk/downloads/...-in-healthcare (Chapter 7).