ICC for test-retest

[Adrian Esterman]

For a test-retest survey, could someone please explain how to set the data up and calculate the ICC?

Many thanks

[Svend Juul]

A short version (this is a complex area, and more sophisticated things may be done):

1. Your data should be in long format, i.e. with a person's first and second test result in separate observations (reshape might be needed).
2. Use the loneway command; it gives the ICC directly.

[Adrian Esterman]

Thanks Svend. However the ICC generated by loneway differs significantly from that produced by other software, eg Medcalc and SPSS. I was looking for a two-way mixed effects ICC for single measure absolute agreement. I assume that loneway is a oneway fixed effects model. I was rather hoping that something could be done with the ICC command?

[Laura Gibbons]

There may be something more recent, but when I had to do this a few years back, I ended up going back to the definitions. Here is my code for what is often called ICC(3,1).

[code]
anova y id rater

* for clarity:
local ems= `e(rmse)'^2
local bms=`e(ss_1)'/`e(df_1)'
local k=2 // number of raters

di (`bms'-`ems')/(`bms'+((`k'-1)*`ems'))

Sorry, I searched the Forum and FAQ for 20 minutes and couldn't figure out how to make this a code box.