[Primer-E] Behaviours- PCA then ANOVA on factor scores, or PERMANOVA on original data

[biologynerd]

Hello, I am hoping someone could give me some advice on an approach to analysing my data.

I have a set of behavioural data (39 behavioural variables such as time spent foraging, number of foraging bouts, time spent vigilant etc), which were measured for 34 sites over 3 nights. Each of the 34 sites had one of 5 different treatments (so evidently I don't have completely equal numbers of replicates in each treatment, but close). The first of the 3 nights was a baseline night without any treatment.

So I basically want to see if the animals changed their behaviour on nights 1 and 2 with comparison to baseline, and if that differed between treatments.

OPTION 1: would be to attempt a PERMANOVA on the original scores, and the design would be something like:

1 Between-Subjects factor (Treatment, 5 levels)
1 Within-Subjects/Repeated Measures factor (Night, 2 levels)
1 Random factor (Site, 34 levels) that is nested within Treatment
1 covariate matrix (Baseline measurements) <- there is a lot of between-site variance so this is important

39 dependent variables (behaviours)

I tried this in Primer-E, and came up with some strange 'cannot compute' type answers in the read out - so I assume I have something wrong somewhere.

Do I need to normalise my data for this? I have transformed it. Does it matter than several behaviours are highly correlated (e.g. number of bouts of vigilance, duration of bouts of vigilance), or that there are several different measurement scales (e.g. proportion of time in sight, duration (in milliseconds), number of bouts (count))

The other issue here is telling which behaviours differentiate the groups - (we would expect it was an anxiety-like behaviour) - I know you can do this with ANOSIM and then SIMPER in Primer-E, but there doesn't seem to be a SIMPER equivalent for PERMANOVA.

OPTION 2: Calculate change scores for each night by subtracting baseline values (so the depvars become change from baseline values). The result is that we don't need a covariate anymore. Complete PERMANOVA using a similar model. Do I need to normalise my data for this? I have transformed it. Same issues with correlated depvars, different measurement scales of depvars.

1 Between-Subjects factor (Treatment, 5 levels)
1 Within-Subjects/Repeated Measures factor (Night, 2 levels)
1 Random factor (Site, 34 levels) that is nested within Treatment

OPTION 3: Transform, normalise behavioural scores and the perform PCA to reduce them down to a limited number of factors. Use factor scores to calculate how each animal did on a new behavioural score that hopefully is meaningful (e.g. 'anxiety'), based on a meaningful principle component, by summing the scores that contribute to that PC for each animal. Then perform a repeated measures ANOVA (still with site nested in Treatment) on how the animals in each treatment change their score on this new 'anxiety' measure in each treatment, over two nights. This is an approach I saw in another paper, but they didn't have a repeated measure, so I'm not clear if it is ok to do the PCA on the full data set with all 3 nights, or just the change scores, or what other approach to use. I have read about PARAFAC but I'm not sure where to find it/how to implement it.

Phew. I'd really appreciate any advice, these all seem like pretty complicated approaches and I'd like to know if one of them makes more sense than the others.

[bugman]

I tried this in Primer-E, and came up with some strange 'cannot compute' type answers in the read out - so I assume I have something wrong somewhere

I'd like to see the readout,it sounds like a PERMANOVA design issuse in the way you have nested your factors

...or that there are several different measurement scales (e.g. proportion of time in sight, duration (in milliseconds), number of bouts (count))

If you plan to use Euclidean distances, you probably should normalise it.

The other issue here is telling which behaviours differentiate the groups - (we would expect it was an anxiety-like behaviour) - I know you can do this with ANOSIM and then SIMPER in Primer-E, but there doesn't seem to be a SIMPER equivalent for PERMANOVA.

You still can. Run the PERMANOVA. If you find differences between your treatment groups, return to PRIMER and set up you SIMPER as normal.

Im going to think about this a bit more and get back to you a bit later. I'd like to have a better idea of what your repsonses are.

[biologynerd]

Update: I decided to split the data into three datasets (one for each measurement type), to avoid the problem of needing to normalise, and avoiding the problem of having repeat measures of essentially the same thing (e.g. number of bouts of vigilance AND mean duration of vigilance AND mean proportion of time in sight spent vigilant, all in the same data set). So there is now one dataset for all the mean number of bouts variables, a second dataset for all the mean duration of bouts variables, and a third dataset for all the mean proportion of time in site variables. I decided to use change scores instead of trying to get a baseline covariate worksheet to work (simplify things)

Then I ran a PERMANOVA with the following design:
Treatment
Site(Treatment) <- a random factor
Night

I had to follow the example given in the manual quite carefully to get all my labels right in the data worksheet, so it could figure out that night was a repeated measure, and then it gave the right number of levels for night, and spat out sensible answers.

So, if anyone reads this in future, that's how i did it!

Only problem is now I am looking at the pairwise comparisons, and the ones that are signigicant make sense to me, but if i run it again and test for main effects, then the main effects aren't significant.

I think this is ok, since for a normal ANOVA a significant omnibus F test is not a prerequisite for pairwise comparisons, but I wouldn't mind confirmation if you have an idea about this, bugman? Or anyone else?

Oh, I also wouldn't mind some advice on whether I need to run the PERMDISP function to test for the multivariate analog of homegeneity of variances... I'm gonna read up on this and will post back with what I find if noone has any advice.

Cheers

[bugman]

I think this is ok, since for a normal ANOVA a significant omnibus F test is not a prerequisite for pairwise comparisons, but I wouldn't mind confirmation if you have an idea about this, bugman? Or anyone else?

this is really only legit if you have predefined contrasts as far as Im aware.

Oh, I also wouldn't mind some advice on whether I need to run the PERMDISP function to test for the multivariate analog of homegeneity of variances... I'm gonna read up on this and will post back with what I find if noone has any advice.

Need? Not as such, but I often use it because high variability can be a sign of ecological stress and can be useful in its own right. It can't hurt to apply it as a check of your assumptions.

[bugman]

Can you just clarify your three measurement types?

[biologynerd]

Hi Bugman,

Thanks for the replies.

My three types of measurement are:

(1) number of occurrences (e.g. number of bouts of vigilance in a 60s video)
(2) duration of bouts (e.g. mean duration of bouts of vigilance were 6.4s)
(3) proportion of time in sight spent on a particular behaviour (e.g. rats spent 0.40 or 40% of their time on camera being vigilant)

[biologynerd]

this is really only legit if you have predefined contrasts as far as Im aware.

would you mind helping me understand this a bit better?

I have five treatments, of which four are different predator odours (dog, cat, quoll, fox) and one is a control (no odour).

I am really only interested in whether or not the treatments are different to the control (as I am trying to determine whether the rats respond to each predator) - so it doesn't matter so much whether they respond 'more' or 'less' to say cat compared to dog, only whether or not they are responding to dog (so compared to control) or cat (compared to control).

Does this mean they qualify as predefined contrasts? And does this mean something different to 'planned comparisons' or are they different words for the same thing - i.e. a relationship you were interested in when you designed the study, not after you saw the data?

Thanks for your time

[bugman]

would you mind helping me understand this a bit better?


Does this mean they qualify as predefined contrasts? And does this mean something different to 'planned comparisons' or are they different words for the same thing - i.e. a relationship you were interested in when you designed the study, not after you saw the data?

Thanks for your time

You are right they are twosidesof the same coin. You should set these up in PERMANOVA so once you run your test, even if your omnibus test is not less than your alpha, you can look at your contrasts.