Post date: Aug 14, 2017 9:39:38 PM
Patrik conducted an experiment testing for NFDS back in March 2017. This is his text describing the experiment and my notes from an e-mail to him summarizing what I did (it doesn't look like I posted on this before):
We implemented a field transplant experiment testing for NFDS. A total of 1000 individuals were transplanted, collected from March 21-24, 2017. Individuals stemmed from the following populations: green-unstriped morphs, PRCN 220, OUTA 50, HVC 140, HVA 90; green-striped morphs, PRCN 30, OUTA 100, HVA 280, HVC 90 (see OSM for population details). Individual were kept in groups of 10 and each group was randomly assigned to one of two treatments: striped individuals common (40 striped and 10 unstriped individuals) versus striped individuals rare (10 striped and 40 unstriped individuals). Each of these groups of 50 individuals was then randomly assigned to one of 20 experimental bushes (in the general area of latitude 34.51 and longitude 119.80). Each bush was cleared of existing Timema by sampling it each day March 21-24. Past work demonstrates that this clears bushes of the overwhelming majority of Timema. Nonetheless, as an additional measure for ensuring accurate identification of experimental animals, each transplanted individual was marked with fine tip sharpie on the underbelly. This marked allowed us to distinguish experimental animals from any remaining residents, and the marks are not visible when Timema are resting on leaves. Individuals were released on March 26th between 9am and 3pm, working sequentially from bush one to bush 20. Each individual was released with tweezers onto an experimental plant and checked to cling well to their transplanted host. Individuals were recaptured using visual surveys and sweep nets on March 31st, as in past work, and scored as striped or unstriped.
I fit a simple GLM with a treatment effect, a beta-binomial model with treatment effects, and a hierarchical beta-binomial model with treatment effects and different stripe relative survival probabilities for each bush. The results for all three were virtually identical. Normally I would prefer the last one, but in this case there is so little variability among bushes that the individual bush estimates (particularly their precision) is influenced by the upper bound of one of the priors and they very closely follow the treatment mean (it is essentially collapsing to the simpler treatment only model). Given this, I think the simple beta-binomial model is the best one to use (though again, none of this matters as the treatment effect is virtually identical for all).
I calculated the treatment effect both in terms of the mean stripe recapture proportion for each treatment and the difference between this and the starting conditions (i.e., frequency change). Here are the posteriors:
2.5% 25% 50% 75% 97.5%
change-20% 0.16843 0.22484 0.255565 0.28634 0.34769
change-80% -0.09283 -0.03438 -0.006446 0.01943 0.06334
stripe-20% 0.36843 0.42484 0.455565 0.48634 0.54769
stripe-80% 0.70717 0.76562 0.793554 0.81943 0.86334
And similarly, here are the posterior probs. for change > 0:
change-20%, pp > 0.9999
change-80%, pp = 0.43
I also looked into the other variables, specifically plant volume and the various source population variables. Overall these are not significantly related to total recapture or stripe frequency for recaptures. There was one exception, green HVA was associated with the number of bugs recaptured (see attached plot), but not significantly associated with % striped (there was an almost marginal trend with p = 0.1, but given the exploratory nature of this I wouldn't take it too seriously). In general I don't think this is anything to worry about, particularly as the association was just with total number recaptured. And with the various combination I tried at least one was almost bound to be significant by chance.