lmer.output-fun.R 17.7 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
#### function to analyse lmer output


library(lme4)


read.lmer.output <- function(file.name){
  tryCatch(readRDS(file.name), error=function(cond)return(NULL))   # Choose a return value in case of error
}



summarise.lmer.output <- function(x){
14
15
16
17
18
 list( nobs = nobs(x),
       R2m =Rsquared.glmm.lmer(x)$Marginal,
       R2c =Rsquared.glmm.lmer(x)$Conditional,
       AIC = AIC(x),
       deviance = deviance(x),
19
       conv=x@optinfo$conv,
20
21
22
23
24
       effect.response.var=variance.fixed.glmm.lmer.effect.and.response(x),
       fixed.coeff.E=fixef(x),
       fixed.coeff.Std.Error=sqrt(diag(vcov(x))),
       fixed.var=variance.fixed.glmm.lmer(x))
}
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41


summarise.lmer.output.list <- function(f ){
    output.lmer <- read.lmer.output(f)
    if(!is.null(output.lmer)){
	res <- list(files.details=files.details(f),
		   lmer.summary=summarise.lmer.output( output.lmer))
    }else{
        res <- NULL
    }
    return(res)
}



files.details <- function(x){
	s <- data.frame(t(strsplit(x, "/", fixed=TRUE)[[1]]), stringsAsFactors= FALSE)
42
	names(s)  <- c("d1", "d2", "set", "ecocode", "trait", "filling", "model", "file" )
43
44
45
46
47
48
49
50
51
52
53
54
	s[-(1:2)]
}



#### R squred functions

# Function rsquared.glmm requires models to be input as a list (can include fixed-
# effects only models,but not a good idea to mix models of class "mer" with models
# of class "lme") FROM http://jslefche.wordpress.com/2013/03/13/r2-for-linear-mixed-effects-models/

Rsquared.glmm.lmer <- function(i){
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
# Get variance of fixed effects by multiplying coefficients by design matrix
      VarF <- var(as.vector(fixef(i) %*% t(i@pp$X)))
      # Get variance of random effects by extracting variance components
      VarRand <- colSums(do.call(rbind,lapply(VarCorr(i),function(j) j[1])))
      # Get residual variance
      VarResid <- attr(VarCorr(i),"sc")^2
      # Calculate marginal R-squared (fixed effects/total variance)
      Rm <- VarF/(VarF+VarRand+VarResid)
      # Calculate conditional R-squared (fixed effects+random effects/total variance)
      Rc <- (VarF+VarRand)/(VarF+VarRand+VarResid)
      # Bind R^2s into a matrix and return with AIC values
      Rsquared.mat <- data.frame(Class=class(i),Family="Gaussian",Marginal=Rm,
                              Conditional=Rc,AIC=AIC(i))
      return(Rsquared.mat)
}


variance.fixed.glmm.lmer <- function(i){
73

74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
# Get variance of for each fixed effects by multiplying coefficients by design matrix
var.vec <- apply(fixef(i) * t(i@pp$X),MARGIN=1,var)
# Get variance of fixed effects by multiplying coefficients by design matrix
VarF <- var(as.vector(fixef(i) %*% t(i@pp$X)))
# Get variance of random effects by extracting variance components
VarRand <- colSums(do.call(rbind,lapply(VarCorr(i),function(j) j[1])))
# Get residual variance
VarResid <- attr(VarCorr(i),"sc")^2
var.vec <- var.vec/(VarF+VarRand+VarResid)
names(var.vec) <- paste(names(var.vec),"VAR",sep=".")
return(var.vec)
}

variance.fixed.glmm.lmer.effect.and.response <- function(i){
if(sum(c("sumTfBn","sumTnBn") %in% names(fixef(i)))==2){
# Get variance of for each fixed effects by multiplying coefficients by design matrix
var.vec <- var(as.vector(fixef(i)[c("sumTfBn","sumTnBn")] %*% t(i@pp$X[,c("sumTfBn","sumTnBn")])))
# Get variance of fixed effects by multiplying coefficients by design matrix
VarF <- var(as.vector(fixef(i) %*% t(i@pp$X)))
# Get variance of random effects by extracting variance components
VarRand <- colSums(do.call(rbind,lapply(VarCorr(i),function(j) j[1])))
# Get residual variance
VarResid <- attr(VarCorr(i),"sc")^2
var.vec <- var.vec/(VarF+VarRand+VarResid)
}else{
var.vec <- NA
}    
names(var.vec) <- paste("effect.response","VAR",sep=".")
return(var.vec)
}
104
105


Georges Kunstler's avatar
Georges Kunstler committed
106
107
108
109


## function to turn lmer output from list to DF
fun.format.in.data.frame <- function(list.res,names.param){
110
111
112
113
dat.t <- data.frame(t(rep(NA,3*length(names.param))))
names(dat.t) <-  c(names.param,paste(names.param,"Std.Error",sep=".")
                   ,paste(names.param,"VAR",sep="."))
dat.t[,match(names(list.res$lmer.summary$fixed.coeff.E),names(dat.t))] <-
Georges Kunstler's avatar
Georges Kunstler committed
114
    list.res$lmer.summary$fixed.coeff.E
115
dat.t[,length(names.param)+match(names(list.res$lmer.summary$fixed.coeff.E),names(dat.t))] <-
Georges Kunstler's avatar
Georges Kunstler committed
116
    list.res$lmer.summary$fixed.coeff.Std.Error
117
118
dat.t[,match(names(list.res$lmer.summary$fixed.var),names(dat.t))] <-
    list.res$lmer.summary$fixed.var
119
res <- data.frame(list.res$files.details,list.res$lmer.summary[1:7],dat.t)
Georges Kunstler's avatar
Georges Kunstler committed
120
121
return(res)
}
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214


#################################################################
#################################################################
##### FUNCTION TO analyse the results AIC effect size

## function to compute delat R2
fun.compute.criteria.diff <- function(i,DF.results,criteria.selected){
select.simple.compet <- DF.results$id==DF.results$id[i] &
                    DF.results$trait==DF.results$trait[i] &
                    DF.results$filling==DF.results$filling[i] &
                    DF.results$model=='lmer.LOGLIN.simplecomp.Tf'
select.no.compet <- DF.results$id==DF.results$id[i] &
                    DF.results$trait==DF.results$trait[i] &
                    DF.results$filling==DF.results$filling[i] &
                    DF.results$model=='lmer.LOGLIN.nocomp.Tf'
if(sum(select.simple.compet)==1){
diff.criteria.simple.compet <- DF.results[[criteria.selected]][i] - DF.results[[criteria.selected]][
                                                                                select.simple.compet]
}else{
diff.criteria.simple.compet <- NA
}
if(sum(select.no.compet)==1){
diff.criteria.no.compet <- DF.results[[criteria.selected]][i] - DF.results[[criteria.selected]][
                                                                              select.no.compet]
}else{
diff.criteria.no.compet <- NA
}

df.res <- data.frame(diff.criteria.simple.compet,diff.criteria.no.compet)
names(df.res) <- paste(criteria.selected,c('simplecomp','nocomp'),sep=".")

return(df.res)
}



fun.compute.delta.AIC <- function(i,DF.results){
select.model.trait.fill <- DF.results$id==DF.results$id[i] &
                    DF.results$trait==DF.results$trait[i] &
                    DF.results$filling==DF.results$filling[i]
if(sum(select.model.trait.fill)>0){
delta.AIC <- DF.results[['AIC']][i] - min(DF.results[['AIC']][select.model.trait.fill])
if (sum( DF.results$nobs[select.model.trait.fill]!=DF.results$nobs[i])>0)
    stop('no same number of observation')
}else{
delta.AIC <- NA
}
df.res <- data.frame(delta.AIC=delta.AIC)
return(df.res)
}

## function to compute ratio of variance explained by a trait variable over the variance explained by the BATOT
fun.ratio.var.fixed.effect <- function(DF.results){
mat.ratio <- DF.results[,c('sumTnTfBn.abs.VAR','sumTfBn.VAR','sumTnBn.VAR','effect.response.var')]/
    DF.results[,'sumBn.VAR']
names(mat.ratio) <- c('abs.dist','Response','Effect','Effect.Response')
return(mat.ratio)
}

### FUNCTION TO REPORT BEST MODEL PER ECOREGION AND TRAITS
fun.AIC <- function(id2.one,DF.results){
 models <- c('lmer.LOGLIN.nocomp.Tf', 'lmer.LOGLIN.simplecomp.Tf','lmer.LOGLIN.E.Tf','lmer.LOGLIN.R.Tf','lmer.LOGLIN.ER.Tf','lmer.LOGLIN.AD.Tf')
 best <- as.vector(DF.results[DF.results$id2==id2.one,c('id2','trait','set','ecocode','filling','MAT','MAP','model')])[which.min(DF.results$AIC[DF.results$id2==id2.one]),]
 AIC.all <- as.vector(DF.results[DF.results$id2==id2.one,c('AIC')])
 names(AIC.all) <- as.vector(DF.results[DF.results$id2==id2.one,c('model')])
 AIC.all <- AIC.all[models]-min(AIC.all)
 res <- data.frame((best),t(AIC.all))
 names(res) <- c('id2','trait','set','ecocode','filling','MAT','MAP','best.model',paste('AIC',models,sep='.'))
 return(res)
}

fun.AICc <- function(id2.one,DF.results){
 models <- c('lmer.LOGLIN.nocomp.Tf', 'lmer.LOGLIN.simplecomp.Tf','lmer.LOGLIN.E.Tf','lmer.LOGLIN.R.Tf','lmer.LOGLIN.ER.Tf','lmer.LOGLIN.AD.Tf')
 Deviance.all <- DF.results[DF.results$id2==id2.one,'deviance']
 names(Deviance.all) <- DF.results[DF.results$id2==id2.one,'model']
 Deviance.all <- Deviance.all[models]
 nobs.all <- DF.results[DF.results$id2==id2.one,'nobs']
 names(nobs.all) <- DF.results[DF.results$id2==id2.one,'model']
 nobs.all <- nobs.all[models]
 n.param <- c(2,3,4,4,5,4)
 AICc <-  Deviance.all+2*n.param*(nobs.all)/(nobs.all-n.param-1)
 id2.n <- unique(DF.results[DF.results$id2==id2.one,c('id2')])
 res <- data.frame(id2.n,models[which.min(AICc)],t(AICc),row.names=NULL)
 names(res) <- c('id2','best.model',models)
 return(res)
}

#################################333

### function to get the data for a given model with criteria to select
fun.select.ecoregions.trait <- function(DF.results,trait.name,model.selected,
                                      nobs.min=1000,filling.selected="species",
215
                                      threshold.delta.AIC){
216
217
218
219
220
221
222
223
224
225
DF.results[DF.results$nobs>nobs.min &
           DF.results$filling==filling.selected &
           DF.results$trait==trait.name &
           DF.results$model %in% model.selected &
           DF.results$delta.AIC<threshold.delta.AIC,]
}

### function to get the data for a given model with criteria to select only site with competition
fun.select.ecoregions.trait.compet <- function(DF.results,trait.name,model.selected,
                                      nobs.min=1000,filling.selected="species",
226
                                      threshold.delta.AIC){
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
DF.results[DF.results$nobs>nobs.min &
           DF.results$filling==filling.selected &
           DF.results$trait==trait.name &
           DF.results$model %in% model.selected &
           DF.results$sumBn < 0.0 &
           ## DF.results$delta.AIC==0,]
           DF.results$delta.AIC<threshold.delta.AIC,]
}




#########################
##### FUNCTIONS FOR PLOTS
fun.plot.lmer.res.x.y <- function(model.selected,trait.name,DF.results,var.x,var.y,threshold.delta.AIC,...){
242
df.selected <- fun.select.ecoregions.trait(DF.results,trait.name=trait.name,model.selected=model.selected,threshold.delta.AIC=10000000)
243
244
245
246
plot(df.selected[[var.x]],df.selected[[var.y]],...)
}

fun.plot.lmer.res.x.y.2 <- function(model.selected,trait.name,DF.results,var.x,var.y,col.vec,pch.AIC=TRUE,cex.AIC=TRUE,col.set=TRUE,...){
247
df.selected <- fun.select.ecoregions.trait(DF.results,trait.name=trait.name,model.selected=model.selected,threshold.delta.AIC=10000000)
248
249
if(pch.AIC) {pch.vec <- c(1,16)[as.numeric(df.selected[['delta.AIC']]==0)+1]}else{pch.vec <- 1}
if(cex.AIC) {cex.vec <- c(1,1.5)[as.numeric(df.selected[['delta.AIC']]==0)+1]}else{cex.vec <- 1}
250
if(col.set) {col.vec2 <- col.vec[unclass(df.selected[['set']])]}else{col.vec2 <- 1}
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
plot(df.selected[[var.x]],df.selected[[var.y]],
     pch=pch.vec,
     cex=cex.vec,
     col=col.vec2,...)
}

fun.plot.lmer.res.boxplot <- function(model.selected,trait.name,DF.results,var.y,...){
df.selected <- fun.select.ecoregions.trait(DF.results,trait.name=trait.name,model.selected=model.selected)
boxplot(df.selected[[var.y]],...)
}

fun.plot.lmer.res.boxplot.compare.model <- function(model.selected,trait.name,DF.results,var.y,...){
df.selected <- fun.select.ecoregions.trait(DF.results,trait.name=trait.name,model.selected=model.selected)
boxplot(df.selected[[var.y]]~df.selected[['model']],...)
## compute percentage of model beteer than null
print(tapply(df.selected[['AIC.simplecomp']]<0,
       INDEX=df.selected[['model']],
       FUN=sum)/tapply(df.selected[['AIC.simplecomp']]<0,
                       INDEX=df.selected[['model']],
                       FUN=length))
}




276
fun.plot.param.error.bar <- function(df.selected,var.x,param,small.bar,col.vec){
277
segments( df.selected[[var.x]],df.selected[[param]] - 1.96*df.selected[[paste(param,"Std.Error",sep=".")]],
278
          df.selected[[var.x]],df.selected[[param]] + 1.96*df.selected[[paste(param,"Std.Error",sep=".")]],col=col.vec)
279
segments( df.selected[[var.x]]-small.bar,df.selected[[param]] - 1.96*df.selected[[paste(param,"Std.Error",sep=".")]],
280
          df.selected[[var.x]]+small.bar,df.selected[[param]] - 1.96*df.selected[[paste(param,"Std.Error",sep=".")]],col=col.vec)
281
segments( df.selected[[var.x]]-small.bar,df.selected[[param]] + 1.96*df.selected[[paste(param,"Std.Error",sep=".")]],
282
          df.selected[[var.x]]+small.bar,df.selected[[param]] + 1.96*df.selected[[paste(param,"Std.Error",sep=".")]],col=col.vec)
283
284
}

285
fun.plot.all.param.x.y.c <- function(model.selected,trait.name,DF.results,var.x,params,small.bar,threshold.delta.AIC,col.vec,col.set=TRUE,...){
286
df.selected <- fun.select.ecoregions.trait(DF.results,trait.name=trait.name,model.selected=model.selected,threshold.delta.AIC=threshold.delta.AIC)
287
288
289
if(col.set) {col.vec2 <- col.vec[unclass(df.selected[['set']])]}else{col.vec2 <- 1}
plot(df.selected[[var.x]],df.selected[[params[1]]],col=col.vec2,...)
fun.plot.param.error.bar(df.selected,var.x,param=params[1],small.bar,col=col.vec2)
290
291
}

292

293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
fun.plot.all.param.boxplot <- function(model.selected,trait.name,DF.results,params,small.bar,...){
df.selected <- fun.select.ecoregions.trait(DF.results,trait.name=trait.name,model.selected=model.selected)
 if(length(params)>1){
DF.t <- data.frame(param=rep(names(params),each=nrow(df.selected)),value=c(df.selected[[params[1]]],df.selected[[params[2]]]))
boxplot(DF.t[['value']]~DF.t[['param']],...)
}else{
boxplot(df.selected[[params[1]]],...)
}
}

fun.plot.all.param.er.diff.MAP <- function(model.selected,trait.name,DF.results,...){
df.selected <- fun.select.ecoregions.trait(DF.results,trait.name=trait.name,model.selected=model.selected)
plot(df.selected[['MAP']],df.selected[['sumTnBn']]-df.selected[['sumTfBn']],...)
}



310
fun.plot.panel.lmer.res.x.y <- function(models,traits,DF.results,var.x,var.y.l,express,...){
311
312
313
314
315
316
317
318
319
320
    ncols = length(traits)
    nrows = length(models)
    list.models <- as.list(names(models))
    names(list.models) <- rep('model',length(list.models))
    DF.results$set <- factor(DF.results$set)
    col.vec <- niceColors(n=nlevels(DF.results$set))
    par(mfrow = c(nrows, ncols), mar = c(2,2,1,1), oma = c(4,4,4,1) )
    for(i in 1:nrows)
        for(j in 1:ncols){
              fun.plot.lmer.res.x.y.2(models[i],traits[j],
321
                          DF.results,var.x,var.y=var.y.l[[i]],col.vec,...)
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
              abline(h=0,lty=3)
            if(i==1 )
                mtext(traits[j], side=3, line =1)
            if(i==nrows)
                mtext(var.x, side=1, line =4)
            if(j==1)
                mtext(paste('Effect size',list.models[i]), side=2, line =4,cex=0.9)
            if(i==nrows & j==ncols)
                legend('topright',legend=levels(DF.results$set),pch=16,
                       col=col.vec,bty='n',ncol=2)
        }
}


fun.plot.panel.lmer.res.boxplot <- function(models,traits,DF.results,var.y,express,...){
    ncols = length(traits)
    nrows = length(models)
    list.models <- as.list(names(models))
    names(list.models) <- rep('model',length(list.models))
    par(mfrow = c(nrows, ncols), mar = c(2,2,1,1), oma = c(4,4,4,1) )
    for(i in 1:nrows)
        for(j in 1:ncols){
              fun.plot.lmer.res.boxplot(models[i],traits[j],
                          DF.results,var.x,var.y,...)
              abline(h=0,lty=3)
            if(i==1 )
                mtext(traits[j], side=3, line =1)
            if(j==1)
                mtext(paste("Effect size", list.models[i],sep=" "), side=2, line =4,cex=0.9)
        }
}


fun.plot.panel.lmer.res.boxplot.compare <- function(models,traits,DF.results,var.y,express,...){
    ncols = length(traits)
    list.models <- as.list(names(models))
    names(list.models) <- rep('model',length(list.models))
    par(mfrow = c(1, ncols), mar = c(2,2,1,1), oma = c(4,4,4,1) )
    for(j in 1:ncols){
              fun.plot.lmer.res.boxplot.compare.model(models,traits[j],
                          DF.results,var.y,names=names(models),...)
              abline(h=0,lty=3)
                mtext(traits[j], side=3, line =1,cex=2)
            if(j==1 )
                mtext("Effect size", side=2, line =4,cex=1.5)
        }
}

fun.plot.panel.lmer.parameters.c <- function(models,traits,DF.results,var.x,list.params,threshold.delta.AIC,small.bar=10,...){
    ncols = length(traits)
    nrows = length(models)
    par(mfrow = c(nrows, ncols), mar = c(2,2,1,1), oma = c(4,4,4,1) )
374
375
    DF.results$set <- factor(DF.results$set)
    col.vec <- niceColors(n=nlevels(DF.results$set))
376
377

## ### TO COMPARE THE PARAMTERS WE NEED TO DIVIDE THE ABS>DIST per two
378
    ## DF.results$sumTnTfBn.abs <- DF.results$sumTnTfBn.abs/2
379
380
    for(i in 1:nrows)
        for(j in 1:ncols){
381
382
383
              fun.plot.all.param.x.y.c(models[i],traits[j],DF.results,var.x,params=list.params[[i]],
                                       small.bar=small.bar,
                                       threshold.delta.AIC=threshold.delta.AIC,col.vec=col.vec,col.set=TRUE,...)
384
385
386
387
388
389
390
391
392
393
394
              abline(h=0,lty=3)
              if(length(list.params[[i]])>1)
                  legend("topright",names(list.params[[i]]),
                         pch=rep(1,length(list.params[[i]])),
                         col=1:length(list.params[[i]]),bty='n',cex=1)
            if(i==1 )
                mtext(traits[j], side=3, line =1)
            if(i==nrows)
                mtext(var.x, side=1, line =4)
            if(j==1)
                mtext(paste('param',names(models)[i]), side=2, line =4,cex=1)
395
396
397
            if(i==nrows & j==ncols)
                legend('topright',legend=levels(DF.results$set),pch=16,
                       col=col.vec,bty='n',ncol=2)
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
        }
}



fun.plot.panel.lmer.parameters.boxplot <- function(models,traits,DF.results,list.params,small.bar=10,...){
    ncols = length(traits)
    nrows = length(models)
    par(mfrow = c(nrows, ncols), mar = c(2,2,1,1), oma = c(4,4,4,1) )
### TO COMPARE THE PARAMTERS WE NEED TO DIVIDE THE ABS>DIST per two
    ## DF.results$sumTnTfBn.abs <- DF.results$sumTnTfBn.abs/2
    for(i in 1:nrows)
        for(j in 1:ncols){
              fun.plot.all.param.boxplot(models[i],traits[j],DF.results,params=list.params[[i]],small.bar=small.bar,...)
              abline(h=0,lty=3)
            if(i==1 )
                mtext(traits[j], side=3, line =1)
            if(j==1)
                mtext(paste('param',names(models)[i]), side=2, line =4,cex=1)
        }
}