Please enable JavaScript.

Coggle requires JavaScript to display documents.

GLMM (Different Packages (glmmPQL - (This function fitted the GLMM using…

- - - - PERK: when dealing with overdispersion (as it allows for quasilikelihood) or small means (<5) in Poisson.
    - - model1<-glmmPQL(y~x, random =~1|randomEffectsVar, family=binomial)
  - - - model1<-glmmML(y~x, cluster = randomEffectsVar, family=binomial)
  - - - model1 <- MCMCglmm(y~x, random =~us(randomEffectsVar), family="poisson")
  - - - PQL is the simplest and most widely used but has problems when standard deviations of random effects are large or with binary data.
        
        works poorly for Poisson data wit low number of counts per treatment and with low expectation number for binomial
      - The Laplace approximation estimates the true GLMM likelihood instead of the quasilikelihood
      - (GHQ) is even more accurate but slower than Laplace.
        
        not feasible with analysis with more than 2-3 random effects
- - - - If you are dealing with count data (Poisson errors), the mean number of counts per treatment combination is less than 5.
      - If it is proportion data, whether the expectation of successes and expectations of failures for each observation are lower than 5.
      - Whether your dependent variable is binary.
  - - - You can use families quasibinomial and quasipoisson as usual
- - - - using information-theoretic
    - - Wald Z, Chi-square, t and F tests use the estimated standard errors and compare the statistic to zero.
        
        Chi-square tests are suitable when there is not overdispersion.
        
        t and F tests will be the tests of our choice when there is overdispersion.
        
        (LR) test is intended for nested models. It is not recommended for fixed effects in GLMM. It is unreliable for small to moderate sample sizes. If used to test fixed effects they require ML instead of REML, but even then the sample size per fixed effects levels and the number of random effects should be high.
        
        After fitting the model we check for overdispersion:
        sigma(model1)
        
        anova(model1, model2)
- - - - cor(Koalas[, 6:22], method = "spearman")
      - Glm_testCollinear<- glm(presence~pprim_ssite+psec_ssite+
        phss_1km+phss_5km+phss_2.5km+
        pm_5km+pm_2.5km+pm_1km+
        pdens_5km+pdens_2.5km+pdens_1km+
        rdens_5km+rdens_2.5km+rdens_1km+
        edens_5km+edens_2.5km+edens_1km,
        data=Koalas2,family=binomial)
        library(car)
        vif(Glm_testCollinear)
        
        clear problems of collinearity. There are two options: we would normally do the first which is to remove variables; the second is to group variables through linear combinations of highly collinear variables
    - - Our next worry is spatial correlation. Using a new approach that works with GLMs, let’s look at the correlogram of the GLM model Pearson residuals where we have ignored spatial correlation.
        
        install.packages("ncf")
        
        Correlog_Glm_2 <- spline.correlog(x = Koalas[, "easting"],
        y = Koalas[, "northing"],
        z = residuals(Glm_testCollinear2,
        type = "pearson"), xmax = 10000)
        
        plot.spline.correlog(Correlog_Glm_2)plot.spline.correlog(Correlog_Glm_2)
        
        correlogram shows correlation as a function of distance. If there is none we should see it all around zero.
      - Let’s construct our GLMM. The clear candidate as a random effects is "site". We use the function glmer. This is a good candidate because it produces AIC scores which we want to use for our information-theoretic approach for model selection.
        
        library(lme4)
        
        Combinations of variables of the global model will be used to generate candidate models using the function "dredge()".