Estimating adjusted ICC as a test-retest-reliability coefficient taking into account time between test and retest

[Christian Meyer]

Dear all!

I like to know if there is any option to adjust an ICC for covariates. We have applied a diagnostic procedure twice in each individual. However, time between both assessments is not constant and should be controlled for, because the true scores may change by time to some extent. I organised both times of assessment in long format and was able to compute the ICC (termed rho) with a one-way random effect model:

xtreg test, i(id)

My questions are:

1) how to include time between assessments in the model in order to have an adjusted ICC and a p-value for the covariate time.
2) as our dependend measure is a count of symptoms rather than a continues measure, I am interessted if there is an option to solve the problem by e.g. count data regression.

Many thanks for your help in advance!

Christian Meyer

[Marcos Almeida]

Hello Christian,

Maybe you should think about starting the modeling with - xtpoisson - command. Here(http://www.stata.com/manuals13/xtxtpoisson.pdf) you have instructions on how to deal with a time variable as well.

Best,

Marcos

[Christian Meyer]

Thank you for advice Marcos!

However, I still have no clew about how to specify the model, because my aim is to calculate a specific ICC variant.

ad 1)

I can generate identical one-way random effect icc with:

loneway test id
icc test id
xtreg test, i(id)

I have a two-way mixed effect model (not what I would like to have) if I include the "panel_indicator" , i.e. a variable which is 0 for the first assessment and 1 for the re-assessment:

xtreg test panel_indicator, i(id)

icc test id panel_indicator, mixed

And again different results if I try:

xtreg test days_since_first_assessment, i(id)


ad 2)

The following may provide the p-value for the covariate but how to calculate an ICC?

xtpoisson test days_since_first_assessment, i(id)


Many thanks again for any advice!

[Marcos Almeida]

Unfortunately, we don't know much about the data. That said, I wonder if you could try - xtmelogit - and, after that, estimate the ICC under the SSC user-written - xtmrho - ado file.

Hopefully that helps!

Best,

Marcos

[Christian Meyer]

... the dependent variable is the number of present symptoms ranging from 0 to 9 (mean 3.7, median 4, SD 2.3). So you think about seperately analyzing each of the nine symptoms coded present vs. not-present?

[Marcos Almeida]

Sorry, in #4, where I wrote - xtmelogit - I meant - xtmepoisson - followed by the user-written - xtmrho - command.

[Christian Meyer]

Thank you again! So I have to explore if the "Median Incidence Rate Ratio (MIRR)" can be seen as an equivalent to ICC in case of count data.

Best regards,

Christian