Jarrod Hadfield

@adam_t_clark @jebyrnes @JNGriffy @LauraEllenDee In the current simulation all observations in a block are identical, generating numerical problems. If you add errors, of any magnitude, Gaussian or not, independent or not, what is the outcome of the simulations?

@IebJarrod @jebyrnes @JNGriffy @LauraEllenDee The "residual variation" arises from differences between the linear model & quadratic generating process. I know it's "typical" to add in normally distributed resid. error, as that's an assumption for OLS - but God did not degree that error must be normal and unstructured 😉 1/2

Hey #ecology folk - when you learned #mixedModels, were you taght about the assumption of endogeneity/random effects assumption - that your random effects (i.e., site groups) should not correlate with your predictor of interest? I was not! @LauraEllenDee and I are curious!

@adam_t_clark @jebyrnes @JNGriffy @LauraEllenDee You've got no residual variation which probably generates some numerical issues. Add some residual variation, then what?

@IebJarrod @jebyrnes @JNGriffy @LauraEllenDee When I run the above example (over 1k replicates), I get 3.00 for the lm against y, 2.99 for the lm against y_re, and 2.43 for the mixed effects model. So, pretty substantial bias in the fixed effects estimate. Or am I missing something?

@adam_t_clark @jebyrnes @JNGriffy @LauraEllenDee In the simulated design (balanced, treatments vary over blocks) both give exactly the same estimates, but lm's SE is too small. If unbalanced, estimates will be different but both give unbiased linearised effect estimates. @jebyrnes's paper about collider bias is a valid concern.

@jebyrnes @JNGriffy @LauraEllenDee The first two models give us relatively good "linearised" estimates of the experimental treatment (this is probably what we usually pretend we are measuring when we use linear regressions). BUT the fixed effects estimates from the RE model are biased towards zero. 5/n

Please RT: When storing flies in EtOH for DNA extraction, I routinely place an additional paper label into the EtOH with the fly, written in pencil. If these were printed on a laser printer (ink does not visibly dissolve), *could* something leach out and interfere with the DNA?

@HuismanJisca @joel_l_pick @NatureEcoEvo @HannahLemon8 Probably make it worse - the data are then more concentrated in the region where the parent-offspring regression is steeper.

Excited to see our paper exploring the consequences of phenotypic skew for predicting selection response is now published @NatureEcoEvo 🙂 - with @HannahLemon8, Caroline Thomson and @IebJarrod nature.com/articles/s4155…

Three field assistant opportunities working in the Scottish Highlands with @Phenoweb_Scot this spring! Working on a transect monitoring blue tit breeding, invertebrates and tree phenology. For more information see phenoweb.org/news/

Please RT! One week left to apply for this quantitative genetics / statistics / bioinformatics job in Edinburgh with @IebJarrod (minor collaborative elements with @b_longdon and me) elxw.fa.em3.oraclecloud.com/hcmUI/Candidat…

@DarrenObbard @b_longdon Tough Choice?

Two-year Drosophila / DNA technician! Edinburgh & Cornwall (some time in each, but some negotiation possible. Would you rather work with @b_longdon or @DarrenObbard ?)

2yr Research Technician based in Edinburgh and/or Penryn (Cornwall) on the genetic variation of fitness in Drosophila. The position will involve collecting and maintaining flies, DNA and RNA extractions. elxw.fa.em3.oraclecloud.com/hcmUI/Candidat…

@JamieDunning @Silwood_Park We caught one in a crow trap in 2003 in front of the manor house!

3yr NERC post-doc with me, @DarrenObbard and @b_longdon on the genetic variation of fitness in Drosophila. The position will develop and implement cutting-edge methods in quantitative and population genetics to analyse large-scale pooled sequencing data: elxw.fa.em3.oraclecloud.com/hcmUI/Candidat…

@JamieDunning @jarrod_hadfield Estimating the mode from MCMC samples of a distribution is not straightforward. Kernel density estimation is used but this can behave poorly with very skewed distributions. Try adjusting the bandwidth (adjust=) and check visually.

@DarrenObbard @davidandthebees @tomhouslay @ali__wilson @b_longdon Time trees and divergence trees were so similar we had no power to distinguish which was a better fit in terms of trait evolution. Anyway, with complete genomes doesn't a straight-up relatedness matrix makes sense; didn't Felsenstein say history is bunk with complete genomes?

@davidandthebees @tomhouslay @ali__wilson @b_longdon Fit both, and see which model is better. IIRC @IebJarrod has compared time trees versus divergence trees in this way.

Using the “pick the brains of people more intelligent than you” function of Twitter: do any of you have any intuition of what the best method for building a between sample relatedness matrix for SARS-CoV-2 genomes would be? @tomhouslay @ali__wilson @DarrenObbard @b_longdon

@Ben_Sheldon_EGI @WJSmith97 @britishbirds Melodious Warbler?

Nice paper with many interesting insights into local adaptation, particularly at the leading edge of a range expansion. onlinelibrary.wiley.com/doi/10.1111/ev…

@Ben_Sheldon_EGI Other estimates have 10x more blue tits in the UK than there are Glaucous Gulls in the world....

There was a bug in MCMCglmm when fitting covu models. Analyses of data sets where some levels in the random term were not observed for the residual term will need to be rerun with v.2.32. All (2) published analyses are correct. Sorry!

@joelmcglothlin @ali__wilson @joel_l_pick @CareauVincent @DingemanseLab @PetriNiemel2 @LoeskeKruuk @KateLaskowski @DFofFreedom @JGAMartin @BioBipbip @HowardRundle4 Dam variance would be (if sire model used) Va/4+Vm + direct-maternal genetic covariance so could be negative. Might also be something like asymmetric sib-competition....

@ali__wilson @joel_l_pick @CareauVincent @DingemanseLab @PetriNiemel2 @LoeskeKruuk @KateLaskowski @DFofFreedom @JGAMartin @BioBipbip @HowardRundle4 @IebJarrod I'm not certain either, but my guess is that if you don't include the covariance and it's strongly negative, you could see a negative dam variance

I found a sig -ve group level variance. I have a biological interpretation(negative cov among group mates from comp) but how do I define VP for estimating h2, R etc. Thoughts? @PetriNiemel2 @LoeskeKruuk @DingemanseLab @joelmcglothlin @KateLaskowski @DFofFreedom @JGAMartin

@joel_l_pick @TimotheeBonnet @LoeskeKruuk @joelmcglothlin @DingemanseLab @JGAMartin @ali__wilson @LukeHolman_Evo @PetriNiemel2 @KateLaskowski @DFofFreedom @erik_postma For balanced 1-way design, residual variance in a bounded model underestimates Vp by (as n becomes large) the covariance divided by the number of groups, so with negative covariance Vp is overestimated. Bias should be small if the data are worth publishing!

@TimotheeBonnet @LoeskeKruuk @joelmcglothlin @DingemanseLab @JGAMartin @ali__wilson @LukeHolman_Evo @PetriNiemel2 @KateLaskowski @DFofFreedom @IebJarrod @erik_postma Is the residual variance then reduced in the mcmc model through absorbing that negative variance? Presumably it goes somewhere?