library("robumeta")
library("metafor")
library("dplyr")
library("meta")
library(xlsx) #to allow for loading excel sheet
library("varhandle") #to allow for handling factors
library(multcomp)
library(dplyr)
library(tidyr)
library(tidyverse)


#read data and only include the ones with "Y" included
data<-read.xlsx("LENA_Prediction.xlsx",1,rm.na=TRUE)
data<-subset(data,Included =="Y")

#Calculated weighted effect szies for repeated measures
weighted<-data %>% group_by(Study.Index,Population,Measure) %>% 
    summarize(Weightedr = sum(r*N)/sum(N),WeightedN = mean(N), WeightedRecAge=sum(AgeRecording..m.*N)/sum(N),WeightedGap=sum(Agegap*N)/sum(N))

#subset data to create unique variables
data1 <- subset(data, 
                subset = !duplicated(data[c("Study.Index", "Population","Measure")]),
                select = c("Study.Index", "Population","Type", "Female.N","Language","Measure","N","Country"))

#prepare data for analysis
final<-merge(data1, weighted, by = c("Study.Index", "Population","Measure"))

final<-read.xlsx("LENA_Mod.xlsx",1,rm.na=TRUE) # all data
Gap<-subset(final,Study.Index !="Gilkerson et al. (2018)") #data without Gilkerson 2015 for the moderator analysis of gap
Population<-subset(final,Population !="Mixed") #data without the "mix" group for the moderator analysis of developmental status
Languagemodality<-subset(final,Languagemodality !="Composite")  #data without the "composite" measure for the moderator analysis of language modality
Testtype<-subset(final,Testtype !="Composite") #data without the "composite" measure for the moderator analysis of test methods

#ran meta-analsys on the correaltion, using Measure (AWC, CV, CT) as subgroups; create funnel and forest plots.
#Replace the measure of interest to obtain correlation values for other factors. 
metacor(final$Weightedr,round(final$WeightedN,digits=0),final$Study.Index,sm="ZCOR",byvar=final$Measure,comb.fixed=F)->mainALL
summary(mainALL)
pdf("main4.pdf",width=9,height=13)
funnel(mainALL)
forest.meta(mainALL,pooled.totals=F,pooled.events=F,hetstat=F, backtransf=FALSE,digits.weight=2)
dev.off()
# Snesitivity analyses in whcih each unique sample size was excluded one at a time did not substantially alter the results, suggesting that
#ni single effect size unduly influened the overall mean effect size
metainf(mainALL, pooled="random")

#Moderator analyses, plots.
##Replace data=Testtype and ~Type with other data and moderators to obtain results.
dat <- escalc(measure="ZCOR", ri=Weightedr, ni=WeightedN, data=Testtype, slab=paste(Study.Index, Population, sep=", ")) 
res <- rma(yi, vi, data=dat) 
res.modage <- rma(yi, vi, mods = ~ Testtype, data=dat,method = "REML") 
res.modage
anova(res.modage)


#Figure 4, examing the effect of the moderator "gap" on the correlation between lena automated measures and language
preds <- predict(res.modage, newmods = cbind(0:120),transf = exp)
wi <- 1/sqrt(dat$vi)
size <- 0.5 + 3 * (wi - min(wi))/(max(wi) - min(wi))
#plots for Fig 4
plot(dat$WeightedGap, exp(dat$yi), pch = 19, cex = size,
     xlab = "Weighted Gap", ylab = "Effect size transformed",
     las = 1, bty = "l", log = "y")
lines(0:120, preds$pred,lty = "dashed",col = 'red')
lines(0:120, preds$ci.lb, lty = "dashed")
lines(0:120, preds$ci.ub, lty = "dashed")
abline(h = 1, lty = "dotted")