I'm currently developing a protocol for an observational study looking at changes in naive animals subjected to repeated behavioural testing - particularly looking at individual phenotype stability.
I'm used to doing a power/sample size calculation at this stage to determine the group sizes, however I'm a bit stuck here as I will be using only 1 group, so have no difference in means to work from.
Any thoughts?
You've highlighted an important issue. Depending on how much longitudinal data you have for each individual, you'll likely have more data that you need for temporal variables. It seems you're in the realm of linear mixed-models, possibly with some sort of temporal autocorrelation structure. For these models, the power analysis has to be done a bit differently than traditional power analysis. Here's one resource that advocates using simulations to calculate power.
http://glmm.wikidot.com/power-analysis
If this didn't hit the mark, then just add more details.
Very interesting answer, even if the link is a wee bit over my head! I like the sound of what you are saying though - it makes sense.
So far my only suggestion has been to use data from a previous experiment, comparing a test group with naive, but that doesn't seem logical to me.
Some more details:
Proposed experiment looks at effects of repeated open field exposure on naive behaviour at an individual level (or small cluster, say around baseline min/max) over three sessions. Previous experiments have only examined single exposure and looked for differences at a group level between naive and test/sham groups. I have used quartile analysis (pooling all data across groups, determining 75 grouping, and see how these proportions vary between test groups) which "looks interesting" (not big enough n's to do stats, apparently).
I don't feel like I'm looking at group differences - I more want to see if initial quartile groups stay constant to their baseline grouping in subsequent testing.
I think I'm probably thinking too much about this, and am only going to end up deeper in the murky world of stats!
Just to get my head around this:
- You want to compare some metric (behavior, etc.) before and after the three sessions of exposure. You'll no doubt be able to test whether something changed or not, but this won't be attributable to anything since you don't seem to have a predictor. Are you planning on having different levels of "repeated" field exposure (e.g. 1,2,or 3 seasons), in order to build upon previous experiments (that looked at singular exposures)? If so, these would constitute your "groups" in your model.
- I like the idea of determining if an individual changed which quartile it is in over time, as a function of repeated exposure. I'm wondering, though, if such reduction in data resolution is necessary. Presumably, you will measure several behaviors for each individual, so why not model consistency in these variables over time? You could also reduce factors by Principal Components Analysis or Independent Components Analysis in order to simply analysis (but which might also confuse the inference).
In any case, a traditional power analysis (like they teach in intro stats) is unlikely to give you what you need for an analysis like this. That's my opinion, at least. Others may have other perspectives.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size