[R-sig-eco] population model with proportions as responses
Hi to all, This is my first post on r-sig-ecology mailing list and it concerns the difficulties I have in fomulating a certain ecological model for my data. My data consist of 567 individual roe deer crania, with 50 linear characters measured across 12 populations. I want to test the influence of habitat structure on cranial variability in this sample. Habitat structure is expressed as the ratio of forest to meadowland to plowland and all ratios add up to one (0.10 forest - 0.20 meadow - 0.70 plowland, etc...). I know that individual approach is not good since all individuals from the same populations have the same values of this habitat ratio. The idea is to use PCA or DA to reduce the dimensionality and to extract population scores that can be used in the model. My question is how to design such a model, that will include population scores and habitat ratios (maybe even two more independent variables, population density and the mean body weight)? Simple lm models are either insignificant or run out of degrees of freedom since only 12 variables are present. I tried summarizing habitat structure with the diversity index (thanks to Seth from r-sig-mixed effect list), but all such indices score ecotonal habitats the most (the ones with the most even ratio) and lm`s are again insignificant. I know that variability is constrained by the extreme habitats (the ones with the most plowland or forest, while meadowland contributes to the area of favorable foraging sites present in both extreme habitat types, but is never the most dominant habitat since it has the lowest ratios. Best regards, Milos Blagojevic paulideali...@aol.com [[alternative HTML version deleted]] ___ R-sig-ecology mailing list R-sig-ecology@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
Re: [R-sig-eco] population model with proportions as responses
Hi Milos, I'm not sure I understand the model you want to fit. Maybe you could clarify. It sounds like habitat structure is the variable that is a proportion, but it is a predictor, not a response. Is there another proportion variable that is your response? Modelling proportions is not easy. I looked up Dirichlet regression on rseek.com and found this that could help if that's really what you want http://cran.r-project.org/web/packages/DirichletReg/vignettes/DirichletReg-vig.pdf Instead, it sounds like cranial characteristics are the responses you want to model. You say I want to test the influence of habitat structure on cranial variability in this sample. Are you sure variability and not the average trend is what you're interested in? I'm guessing you might want to do a linear mixed model (a separate one for each of the 50 cranial characters) with cranial character as the response; habitat structure, diversity index, and body weight as fixed effects; and population as a random effect. If some cranial character is your response, then you have 567 data points. That should be plenty of data. Something like m1=lmer(cranial_width ~ forest_percent + meadow_percent + plowland_percent + diversity_index + body_weight + (1|population), data=cranial_data) It could be a problem that the predictors (forest_percent + meadow_percent + plowland_percent) are non-independent, but I don't know if it's a big problem. Maybe someone else could weigh in on that. I hope this helps clarify. I could be totally misunderstanding the problem. It might help even more if you could talk to someone locally. cheers, Mollie Mollie Brooks, PhD Postdoctoral Researcher, Population Ecology Research Group, Ozgul Lab, http://www.popecol.org Institute of Evolutionary Biology Environmental Studies, University of Zürich On 8 May 2013, at 8:18 AM, Milos Blagojevic paulideali...@aol.com wrote: Hi to all, This is my first post on r-sig-ecology mailing list and it concerns the difficulties I have in fomulating a certain ecological model for my data. My data consist of 567 individual roe deer crania, with 50 linear characters measured across 12 populations. I want to test the influence of habitat structure on cranial variability in this sample. Habitat structure is expressed as the ratio of forest to meadowland to plowland and all ratios add up to one (0.10 forest - 0.20 meadow - 0.70 plowland, etc...). I know that individual approach is not good since all individuals from the same populations have the same values of this habitat ratio. The idea is to use PCA or DA to reduce the dimensionality and to extract population scores that can be used in the model. My question is how to design such a model, that will include population scores and habitat ratios (maybe even two more independent variables, population density and the mean body weight)? Simple lm models are either insignificant or run out of degrees of freedom since only 12 variables are present. I tried summarizing habitat structure with the diversity index (thanks to Seth from r-sig-mixed effect list), but all such indices score ecotonal habitats the most (the ones with the most even ratio) and lm`s are again insignificant. I know that variability is constrained by the extreme habitats (the ones with the most plowland or forest, while meadowland contributes to the area of favorable foraging sites present in both extreme habitat types, but is never the most dominant habitat since it has the lowest ratios. Best regards, Milos Blagojevic paulideali...@aol.com [[alternative HTML version deleted]] ___ R-sig-ecology mailing list R-sig-ecology@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-ecology [[alternative HTML version deleted]] ___ R-sig-ecology mailing list R-sig-ecology@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
Re: [R-sig-eco] population model with proportions as responses
Dear Milos, It could be a problem that the predictors (forest_percent + meadow_percent + plowland_percent) are non-independent, but I don't know if it's a big problem. Maybe someone else could weigh in on that. Such strong linear relationship between predictors (collinearity) is a big problem. If all predictor included, the model becomes ill-defined. I suggest including only two of them. Choose the pair of predictor that lead to most easily interpretable model (probably choosing the two extreme habitat is the best solution). Best wishes Zoltan -- Botta-Dukát Zoltán Ökológiai és Botanikai Intézet Magyar Tudományos Akadémia Ökológiai Kutatóközpont 2163. Vácrátót, Alkotmány u. 2-4. tel: +36 28 360122/157 fax: +36 28 360110 botta-dukat.zol...@okologia.mta.hu www.okologia.mta.hu Zoltán BOTTA-Dukát Institute of Ecology and Botany Hungarian Academy of Sciences Centre for Ecological Research H-2163 Vácrátót, Alkomány u. 2-4. HUNGARY Phone: +36 28 360122/157 Fax..: +36 28 360110 botta-dukat.zol...@okologia.mta.hu www.okologia.mta.hu ___ R-sig-ecology mailing list R-sig-ecology@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
[R-sig-eco] interpretation of interaction between explanatory variables
Dear list members, I want to analyse the impact of a competitor community (i.e. community abundances on the one hand and community species diversity on the other hand) on mosquito larval populations of species A and B. Each variable on its own has a negative impact on mosquitoes - but when both variables are interacting, there is a positive impact... How can I interpret that? For mosquito A only C.diversity has a significant impact - but the interaction between C.abundances and C.diversity is significant? What does that mean? I used the model: lm (mosquito ~ C.abundances * C.diversity) output Mosquito A: Estimate Std. Error t value Pr(|t|) (Intercept) -0.2120 0.1159 -1.829 0.074855 . C.abundances -0.1277 0.1616 -0.790 0.434067 C.diversity -0.5787 0.1385 -4.178 0.000155 *** C.abundances:C.diversity 0.4096 0.1712 2.393 0.021476 * Output Mosquito B: Estimate Std. Error t value Pr(|t|) (Intercept) -0.2900 0.1220 -2.377 0.02233 * C.abundances -0.3856 0.1701 -2.266 0.02891 * C.diversity -0.3470 0.1458 -2.381 0.02213 * C.abundances:C.diversity 0.5367 0.1801 2.980 0.00489 ** Thanks a lot for your help! Iris [[alternative HTML version deleted]] ___ R-sig-ecology mailing list R-sig-ecology@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
[R-sig-eco] multivariate continuous response and ordinal predictor(s)
Dear all, I need to run a model with multivariate continuous responses and one (or more) ORDINAL (i.e. 1,2,3, etc.) predictor variables; these are not factors because are ordinal; the more intuitive solution could be to apply a standard lm() but I ask you if some more appropriate strategies can be adopted. I saw VGAM package but it does not seem to do what I want Thankyou in advance for any advice best paolo ___ R-sig-ecology mailing list R-sig-ecology@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-ecology