I have currently used 3-D motion capture to conduct gait analysis on children with cerebral palsy before and after an intensive physical intervention (pre- and post-testing). I collected 4 trials for pre- and 4 trials for post-testing and only one testing session occurred for each. There are 9 subjects total, but I need to use an n-of-1 design because the group is heterogeneous and there is a large amount of variability between subjects. I am using the Standard Error Measurement (SEM) to then calculate the Minimal Detectable Change for each child. The equation for the SEM requires the Intraclass Correlation Coefficient (ICC), which SPSS outputs as a negative value for 2 of the 7 variables I am investigating. Is a negative ICC acceptable? If not, how can this be corrected?
Thank you
Theoretically, the true ICC cannot be negative, because it estimates the proportion of between subjects variance. The estimate of ICC, which is actually computed, can be negative. Personally, I set negative value to zero, because it signifies that the within subject variance (or between tests variance) is huge, and hence the measure is not reliable at all. Anyway, for the SEM estimates it doesn't matter, because in this case, your SD is the SEM (all the variance is due to measurements, you put sqrt(1-0) in the formula). I suspect that your sample size is too small for a reliable assessment of ICCs: check your confidence intervals on ICCs: very likely so large that you cannot say anything on the repeatability of your measure.
Hi,
Not understanding the question well, or therefore having an answer :), I asked my son who is a researcher at Stanford and USC for a comment on this question. I hope that it helps:
Theoretically no, but stat programs only estimate ICC so they can spit out negative values.
It's much more likely when you have very small sample sizes. It basically means, I think, that the variation within the group is really big, so you can treat it like 0 anyways for further analyses if desired.
That said, it does suggest that you might not be calculating the ICC efficiently.
I've has this happen to me too, but I had thousands of clusters (schools) and about 10 spit out negative ICC's, which I interpreted to mean that the variation in these schools was relatively massive. I seem to recall it happening in particular when the cluster (school) was small and I also clustered standard errors at the same level, but I could be mis-remembering that.
- Badges
- Science topic
- Similar topics
- Correction