Dear Marguerite and everyone, thank you very much for your considerate postings. I have reconsidered my analyses and excluded the zero values, because they indicate absence of the trait rather than are part of the continuum of values. Instead, I analyzed the data as a 1) discrete trait: presence/absence of trait, and 2) continuous trait, where trait present. In the absence of zeros, the continuous data log-transform to normality where sufficient variability exists.
Thank you again, Nina On 2012-04-25, at 9:13 PM, Marguerite Butler wrote: > Hi Nina and everyone, > > One thing to consider is that not all zero data are the same. Zeros under a > model of continuous trait evolution with a gaussian process as assumed under > Brownian motion and OU processes would occasionally cross zero, maybe go > negative, etc. For example if you were modeling something like the deviation > from average height. You may have a lot of individuals that are at zero > because they are all average, but their offspring will quickly move off zero > as some will be taller, some shorter than average. > > On the other hand, many of us measure traits which disappear, for example > scale counts on fish. Numbers of scale rows vary while there are scales, but > once they disappear, those lineages will be at zero, perhaps for a very long > time. In this case it is no longer behaving as a gaussian process with small > changes expected every time period. We usually think of these more as a > "threshold trait". Maybe there is some hormone or something else (a hidden > variable) underlying the determination of scales or no scales, and once it > goes below threshhold the scales disappear. The hidden variable may fit model > assumptions, but not the scale counts (what we can see and measure). For > example, the scale counts can never go negative. With a value like height, it > also never goes negative, but usually we are far away from that zero boundary > so we can casually ignore that problem:). Or we can do a log-transform, or > something else that transforms the data onto a different scale. Anyway, the! absorbing boundary zeros are, I think, an example of what Ted is talking about. > > So it depends on the nature of your data. On the other hand, for the other > variable, if there is just not much variation, but it's not stuck on any > particular value (doesn't appear to have any absorbing boundary), I think > that's less of a problem. > > HTH, > Marguerite > > > On Apr 25, 2012, at 7:17 AM, Theodore Garland Jr wrote: > >> Read over the Blomberg et al. (2003) paper. >> K is intended for continuous-valued traits and/or those evolving similar to >> Brownian motion. >> You could report it if you wished, but I would add that caveat if you do. >> >> The randomization test should be robust in any case. >> >> Cheers, >> Ted >> >> >> From: Nina Hobbhahn [[email protected]] >> Sent: Wednesday, April 25, 2012 9:19 AM >> To: Theodore Garland Jr >> Cc: Alejandro Gonzalez; Hunt, Gene; Enrico Rezende; [email protected] >> Subject: Re: [R-sig-phylo] Normality requirement for assessment of lambda >> with phylosig (phytools) and fitContinuous (geiger) >> >> Thanks all for your helpful contributions! I will use phylosignal. >> >> Ted, I'm not sure I understand your last comment, "when the data are not >> though of as continuous-valued and/or evolving similar to Brownian motion". >> What do you mean by that? Also, are you suggesting that I report the >> presence/absence of phylogenetic signal, but not the value of the K >> statistic? >> >> Many thanks again, >> >> Nina >> >> >> On 2012-04-25, at 5:54 PM, Theodore Garland Jr wrote: >> >>> However, calculating a K statistic is strange when the data are not thought >>> of as continuous-valued and/or evolving similar to Brownian motion. The >>> randomization test is OK, however. >>> >>> Cheers, >>> Ted >>> [[alternative HTML version deleted]] _______________________________________________ R-sig-phylo mailing list [email protected] https://stat.ethz.ch/mailman/listinfo/r-sig-phylo
