[R-meta] variance explained by fixed & random effects

Tue Mar 26 09:57:44 CET 2019

Hi all,

A paper from Nakagawa & Santos (2012) seems to answer my question (specifically the section "Heterogeneity - Quantifying Heterogeneity"). 

Nakagawa, S. & Santos, E.S.A. Evol Ecol (2012) 26: 1253. https://doi.org/10.1007/s10682-012-9555-5

Theresa

> On March 14, 2019 at 5:13 PM Theresa Stratmann <theresa.stratmann using senckenberg.de> wrote:
> 
> 
> Dear Wolfgang,
> 
> Thank-you for the fast reply! 
> 
> So the code is easy enough, but I am struggling to understand some of the nuance/ how to properly think about the different numbers. I have added some questions, between the dashed lines, below your answers. If they are within the scope of the mailing list and not to cumbersome, I would be curious for some explanations to help clear up my confusion. 
> 
> Thanks!
> 
> Theresa
> 
> 
> 
> One way of estimating an R^2-type measure for the fixed effects is to fit the reduced model without the predictor and then compute the proportional reduction in the variance components:
> 
> res0 <- rma.mv(yi, vi, random = ~ 1 | district/school, data=dat) 
> pmax(0, (res0$sigma2 - res$sigma2) / res0$sigma2) * 100
> max(0, (sum(res0$sigma2) - sum(res$sigma2)) / sum(res0$sigma2)) * 100
> 
> One can compute this per component or overall. There is no guarantee that the value is >= 0, so I use pmax()/max() to set negative values to 0. One could debate whether it even makes sense here to estimate how much of the school-level heterogeneity is accounted for by 'year'(since this is a district-level predictor), but that's another issue.
> 
> -----------------------------------------------------------------------------------------------------------------------------------------
> ***
> So we have: 
> 
> sigma^2 for res0 .... heterogeneity in district & school:
> 0.06506194 0.03273652
> sigma^2 for res1 .... heterogeneity in district & school, accounting for year:
> 0.11482670 0.03208976
> 
> (res0$sigma2 - res$sigma2) / res0$sigma2
> -0.76488265  0.01975659
> 
> So it looks like accounting for year increases the heterogeneity in district.
> 
> Acknowledging and moving on from the fact that year is not the best fixed effect... What does it mean if the value is negative/zero? 
> There is more variance in the random intercepts when you include the fixed effect? 
> Adding this particular fixed effect does not improve the model? 
> If I was writing this up, could I say: "Accounting for year does not explain any additional variation in the estimated variance of the random intercepts."
> More specifically: 
> "The heterogeneity accounted for by year is 0% for district level variance, and 1.98% for school level variance."
> "The overall heterogeneity accounted for by year is 0%." ???
> 
> Thinking about marginal and conditional R^2 for mixed-effects models, there is a separation of fixed and random effects: fixed effects account for X% of the variation
> and random effects accounting for Y% of the variation and Z% residual variance. But here the fixed effect is somehow included 
> in the sigma^2 (estimated variance of the random intercepts)?  This just confuses me a bit/ I missed the thought process. 
> 
> ***
> -----------------------------------------------------------------------------------------------------------------------------------------
> 
> I am not sure how to think of the question how much variance is explained by a random effect here. One could ask how much of the total heterogeneity 
> is due to district- and school-level variance, which would be:
> 
> res0$sigma2 / sum(res0$sigma2) * 100
> 
> Or one could ask how much of the unaccounted for heterogeneity (so heterogeneity not accounted for by 'year') is due to district- and school-level variance, which would be:
> 
> res$sigma2 / sum(res$sigma2) * 100
> 
> -----------------------------------------------------------------------------------------------------------------------------------------
> ***
> So, to double check, the heterogeneity accounted for by year is, what you showed above: 
> 
> max(0, (sum(res0$sigma2) - sum(res$sigma2)) / sum(res0$sigma2)) * 100
> 
> We also know that: 
> 
> res0$sigma2 / sum(res0$sigma2) * 100  = 66.52655; 33.47345
> 
> max(0, (sum(res0$sigma2) - sum(res$sigma2)) / sum(res0$sigma2)) * 100 = 0
> 
> Since these have the same denominator... is it true that: 
> [max(0, (sum(res0$sigma2) - sum(res$sigma2)) / sum(res0$sigma2)) * 100] + sum(res0$sigma2 / sum(res0$sigma2) * 100) = 100
> (I check on my data and the answer seems to be no...)
> ***
> -----------------------------------------------------------------------------------------------------------------------------------------
> 
> Or one could ask how much of the total variance (so heterogeneity + sampling variance) is due to district- and school-level variance, which would be:
> 
> k <- res$k
> wi <- 1/dat$vi
> s2 <- (k-1) * sum(wi) / (sum(wi)^2 - sum(wi^2))
> 
> res0$sigma2 / (sum(res0$sigma2) + s2) * 100
> 
> (here I am using the Higgins & Thompson, 2002, definition of a 'typical' sampling variance).
> 
> Or one could ask how much of the unaccounted for variance (so unaccounted for heterogeneity [heterogeneity not accounted for by 'year'] + sampling variance)
> is due to district- and school-level variance, which would be:
> 
> res1$sigma2 / (sum(res1$sigma2) + s2) * 100
> 
> -----------------------------------------------------------------------------------------------------------------------------------------
> ***
> Can one ask here how much of the total variance (estimate of variance in true effects + sample variation) is accounted for by year? This is really what I want. 
> ***
> -----------------------------------------------------------------------------------------------------------------------------------------
> 
> 
> > On March 13, 2019 at 6:00 PM "Viechtbauer, Wolfgang (SP)" <wolfgang.viechtbauer using maastrichtuniversity.nl> wrote:
> > 
> > 
> > Dear Theresa,
> > 
> > You forgot some parts in your code, so to make this example reproducible, let's use this model:
> > 
> > dat <- dat.konstantopoulos2011 
> > res <- rma.mv(yi, vi, mods = ~ factor(year), random = ~ 1 | district/school, data=dat)
> > 
> > And I would consider schools nested within districts here, but this isn't pertinent to the issue.
> > 
> > One way of estimating an R^2-type measure for the fixed effects is to fit the reduced model without the predictor and then compute the proportional reduction in the variance components:
> > 
> > res0 <- rma.mv(yi, vi, random = ~ 1 | district/school, data=dat) 
> > pmax(0, (res0$sigma2 - res$sigma2) / res0$sigma2) * 100
> > max(0, (sum(res0$sigma2) - sum(res$sigma2)) / sum(res0$sigma2)) * 100
> > 
> > One can compute this per component or overall. There is no guarantee that the value is >= 0, so I use pmax()/max() to set negative values to 0. One could debate whether it even makes sense here to estimate how much of the school-level heterogeneity is accounted for by 'year' (since this is a district-level predictor), but that's another issue.
> > 
> > I am not sure how to think of the question how much variance is explained by a random effect here. One could ask how much of the total heterogeneity is due to district- and schol-level variance, which would be:
> > 
> > res0$sigma2 / sum(res0$sigma2) * 100
> > 
> > Or one could ask how much of the unaccounted for heterogeneity (so heterogeneity not accounted for by 'year') is due to district- and schol-level variance, which would be:
> > 
> > res$sigma2 / sum(res$sigma2) * 100
> > 
> > Or one could ask how much of the total variance (so heterogenity + sampling variance) is due to district- and schol-level variance, which would be:
> > 
> > k <- res$k
> > wi <- 1/dat$vi
> > s2 <- (k-1) * sum(wi) / (sum(wi)^2 - sum(wi^2))
> > res0$sigma2 / (sum(res0$sigma2) + s2) * 100
> > 
> > (here I am using the Higgins & Thompson, 2002, definition of a 'typical' sampling variance).
> > 
> > Or one could ask how much of the unaccounted for variance (so unaccounted for heterogenity + sampling variance) is due to district- and schol-level variance, which would be:
> > 
> > res1$sigma2 / (sum(res1$sigma2) + s2) * 100
> > 
> > These are at least some possibilities.
> > 
> > Best,
> > Wolfgang
> > 
> > -----Original Message-----
> > From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces using r-project.org] On Behalf Of Theresa Stratmann
> > Sent: Wednesday, 13 March, 2019 17:29
> > To: r-sig-meta-analysis using r-project.org
> > Subject: [R-meta] variance explained by fixed & random effects
> > 
> > Dear Meta-Analysis Community,
> > 
> > I am using the metafor package to help me summarize some data on habitat selection. I have a question whose answer I have not found yet on the useful metafor website, and was hoping that someone could help me (or point me to an online resource I missed). 
> > 
> > As part of the description of my meta-analysis I would like to explain how much of the variation in habitat selection is due to my fixed effects (season) versus my random effects (individual & year). 
> > 
> > For a normal mixed-effects model I know how to compute the marginal and conditional R^2 (Nakagawa & Schielzeth 2013), but I am wondering how to do this correctly for a meta-analysis, or maybe, better said, talk about variance explained by the fixed and random effects. 
> > 
> > First I looked into I^2 and the post on this for random effect models. But then I read more from Borenstein et al. (https://www.meta-analysis-workshops.com/download/common-mistakes1.pdf) and was not quiet sure anymore if this is what I want (I^2 = proportion of observed variance that reflects variance in the true effects and not the sampling error ... so this seems more of a summary for the entire model (?), which also needs to be treated with care and presented properly). 
> > 
> > Then I found this question on this mailing list: https://stat.ethz.ch/pipermail/r-sig-meta-analysis/2017-September/000232.html  ...but cannot really get this code to produce anything sensible, because I guess I am not clear on the formula for res0 vs. res1. 
> > 
> > So I will give a different example, which closely approximates what I am trying to do (a model with fixed & random effects): 
> > 
> > dat <- dat.konstantopoulos2011 
> > res <- rma.mv(yi, vi, mods = ~ factor(year), list( ~ 1 | district , ~ 1|school)  data=dat) 
> > res 
> > 
> > How would you calculate / correctly describe variance explained by the fixed effects (year)? By the random effects (district, school)? I am not tied to a particular statistic, just want to make sure that whatever I do is an honest representation of my results and gives readers (& myself) the information they need to interpret the results. 
> > 
> > I am new to meta-analyses and this is just a small part of my project, so I have been able to do some reading, but not an in-depth dive. Therefore I would appreciate any help you can offer.
> > 
> > Many thanks,
> > 
> > Theresa Stratmann 
> > 
> > B.S. Ecology, The University of Georgia 
> > M.S. Wildlife & Fisheries Biology, Clemson University 
> > 
> > PhD Student 
> > Goethe University 
> > and 
> > Senckenberg Biodiversity and Climate Research Center
> 
> _______________________________________________
> R-sig-meta-analysis mailing list
> R-sig-meta-analysis using r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-meta-analysis