Extension tecdat|R language uses Rshiny to explore lme4 generalized linear mixed model (GLMM) and linear mixed model (LMM)

Original link: http://tecdat.cn/?p=3138

Original source: Tuoduan Data Tribe Public Account

Using Generalized Linear Mixed Models (GLMM) and Linear Mixed Models (LMM) has become easier as software packages have advanced. As we find ourselves using these models more and more in our work, we have developed a suite of R shiny tools to simplify and speed up the common lme4 tasks of interacting with objects.

shiny apps and demos

The easiest way to demonstrate the capabilities of this app is to use the Shiny app, start some metrics here to help explore the model.

On the first tab, the function displays prediction intervals for the data selected by the user. The function quickly computes prediction intervals by sampling from simulated distributions of fixed-effect and random-effects terms and combining these simulated estimates to produce a predicted distribution for each observation.

On the next tab, the distributions of fixed effects and group-level effects are displayed on a confidence interval plot. These are useful for diagnostics and provide a way to examine the relative magnitudes of various parameters.

There are some convenience methods on the third tab that show the effect's impact or extent predictInterval. For each case, up to 12, in the selected data type, the user can view the effect of changing the fixed effects. This allows users to compare effect sizes between variables, as well as effect sizes between models on the same data.

predict

Predictions like this.

predict(m1, newdata

#>        1        2        3        4        5        6        7        8

#> 3.146336 3.165211 3.398499 3.114248 3.320686 3.252670 4.180896 3.845218

#>        9       10

#> 3.779336 3.331012

Predict lm and glm:

predInte(m1, newdata = Eval\[1:10, \], n.sims = 500, level = 0.9,

#>         fit      lwr      upr

#> 1  3.074148 1.112255 4.903116

#> 2  3.243587 1.271725 5.200187

#> 3  3.529055 1.409372 5.304214

#> 4  3.072788 1.079944 5.142912

#> 5  3.395598 1.268169 5.327549

#> 6  3.262092 1.333713 5.304931

The prediction interval is slower because it is a simulation calculation.

visualization

Visually inspect the functionality of an object. The easiest is to get the posterior distributions for the fixed and random effects parameters.

head(Sim)

#>          term        mean      median         sd

#> 1 (Intercept)  3.22673524  3.22793168 0.01798444

#> 2    service1 -0.07331857 -0.07482390 0.01304097

#> 3   lectage.L -0.18419526 -0.18451731 0.01726253

#> 4   lectage.Q  0.02287717  0.02187172 0.01328641

#> 5   lectage.C -0.02282755 -0.02117014 0.01324410

We can draw like this:

pltsim(sim(m1, n.sims = 100), level = 0.9, stat = 'median'

We can also quickly make a plot of random effects:

head(Sims)

#>   groupFctr groupID        term        mean      median        sd

#> 1         s       1 (Intercept)  0.15317316  0.11665654 0.3255914

#> 2         s       2 (Intercept) -0.08744824 -0.03964493 0.2940082

#> 3         s       3 (Intercept)  0.29063126  0.30065450 0.2882751

#> 4         s       4 (Intercept)  0.26176515  0.26428522 0.2972536

#> 5         s       5 (Intercept)  0.06069458  0.06518977 0.3105805

plotR((m1, n.sims = 100), stat = 'median', sd = TRUE

Sometimes random effects can be difficult to interpret

 Rank(m1, groupFctr = "d")

head(ranks)

#>      d (Intercept) (Intercept)_var       ER pctER

#> 1 1866   1.2553613     0.012755634 1123.806   100

#> 2 1258   1.1674852     0.034291228 1115.766    99

#> 3  240   1.0933372     0.008761218 1115.090    99

#> 4   79   1.0998653     0.023095979 1112.315    99

#> 5  676   1.0169070     0.026562174 1101.553    98

#> 6   66   0.9568607     0.008602823 1098.049    97

effect simulation

Interpreting the results of LMM and GLMM models is difficult, especially the relative effects of different parameters on the predicted results.

impact(m1, Eval\[7, \], groupFctr = "d", breaks = 5,

n.sims = 300, level = 0.9)



#>   case bin   AvgFit     AvgFitSE nobs

#> 1    1   1 2.787033 2.801368e-04  193

#> 2    1   2 3.260565 5.389196e-05  240

#> 3    1   3 3.561137 5.976653e-05  254

#> 4    1   4 3.840941 6.266748e-05  265

#> 5    1   5 4.235376 1.881360e-04  176

The results show the change in yhat from the first to the fifth quantile in terms of the magnitude of the group factor coefficients according to the newdata we provided.

ggplot(impSim, aes(x = factor(bin), y = AvgFit, ymin = AvgFit - 1.96*AvgFitSE,

ymax = AvgFit + 1.96*AvgFitSE)) +

Programs Wiki