R/StanDINO.script.R

In JihongZ/StanDCM: Diagnostic Classification Model using Stan

#' @title Generate Stan code and Run the estimation for DINO model
#'
#'
#' @description
#' The \code{StanDINO.run} Function allows to automate Stan code geneartion for DINO model with binary resposnes
#'
#' @usage
#' StanDINA.run<-function(Qmatrix,response.matrix,script.path=NA,save.path=getwd(),
#' save.name="DINA_uninf",iter=1000,warmup = floor(iter/2),
#' chain.num=3,init.list='random',control.list=NA)
#'
#' @param Qmatrix A required matrix
#' @param response.matrix save the .stan file to somewhere; the default path is getwd()
#' @param script.path save the .stan file to somewhere; the default path is getwd()
#' @param save.name name the .stan
#' @param iter name the .stan
#' @param chain.num name the .stan
#' @param init.list name the .stan
#' @param control.list name the .stan
#'
#' @return StanDINA returens an object of class StanDINA. Methods for StanDINA objects include
#' @importFrom plyr failwith id summarize count desc mutate arrange rename summarise
#' @import stringr CDM
#'
#'
#' @author {Zhehan Jiang, University of Alabama, \email{zjiang17@@ua.edu} \cr Jihong Zhang, University of Iowa, \email{jihong-zhang@uiowa.edu}}
#'
#' @export
#' @examples
#' \dontrun{
#' #----------- DINO model-----------#
#' mod1<-StanDINO.run(Qmatrix, respMatrix, iter=20, init.list='cdm', chain.num = 3)
#' summary(mod1)
#' }


StanDINO.script<-function(Qmatrix,save.path=getwd(),save.name="DINO_uninf"){

  #Load packages
  # Install.package("plyr")
  # Install.package('stringr')

  nc<-ncol(Qmatrix)
  nr<-nrow(Qmatrix)
  temp.table.col<-unique(apply(combn(rep(c(0,1),nc),nc),2,function(x){paste(x,collapse = "")}))
  temp.table.col<-temp.table.col[order(temp.table.col)]
  temp.table<-matrix(0,nr,length(temp.table.col))
  colnames(temp.table)<-temp.table.col
  rownames(temp.table)<-paste('item',c(1:nr),sep='')
  temp.table<-as.data.frame(temp.table)
  for (i in 1:nr){
    temp.table[i,]<-paste('l',i,'_0',sep='')
  }
  intercept<-temp.table[,1]

  #Generate attribute combinations
  comb.generator<-function(x.vector){
    if(length(x.vector)>1){
      temp.attr<-x.vector
      temp.attr.sav<-NULL
      for(i in 1:length(temp.attr)){
        temp.1<-combn(temp.attr,i)
        temp.2<-apply(temp.1,2,function(x){paste(x,collapse = "")})
        temp.attr.sav<-c(temp.attr.sav,temp.2)
      }
    }
    if(length(x.vector)==1){temp.attr.sav<-x.vector}
    temp.attr.sav
  }
  #vectors needed for combination.generator
  Item.load.id<-list()
  for ( i in 1:nr){
    Item.load.id[[i]]<-grep('1',Qmatrix[i,])}

  Attr.load.id<-list()
  attr.load.id<-matrix(0,length(temp.table.col),nc)
  for ( i in 1:length(temp.table.col)){
    attr.load.id[i,]<-unlist(strsplit(temp.table.col[i],split=''))
    Attr.load.id[[i]]<-grep('1',attr.load.id[i,])
  }

  #Generate Combination for both Item.load and Attr.load
  Item.Comb<-list()
  for ( i in 1:nr){
    Item.Comb[[i]]<-comb.generator(Item.load.id[[i]])
  }
  Attr.Comb<-list()
  for ( i in 2:length(temp.table.col)){
    Attr.Comb[[1]]<-0
    Attr.Comb[[i]]<-comb.generator(Attr.load.id[[i]])
  }
  constraints.list<-list()
  nway.inter.list<-list()
  for(i in 1:nr){
    for(a in 2:length(temp.table.col)){
      ifzero<-as.numeric(paste(Item.Comb[[i]][Item.Comb[[i]]%in%(Attr.Comb[[a]])],collapse=''))
      if((!is.na(ifzero))){
        temp.table[i,a]<-paste(c(temp.table[i,a],
                                 paste("S","l",i,"_",nchar(Item.Comb[[i]][Item.Comb[[i]]%in%(Attr.Comb[[a]])]),Item.Comb[[i]][Item.Comb[[i]]%in%(Attr.Comb[[a]])],sep='',collapse='')
        ),collapse='')
        if(a==length(temp.table.col)){
          nway.inter.list[[i]]<-nchar(Item.Comb[[i]][Item.Comb[[i]]%in%(Attr.Comb[[a]])])
          constraints.list[[i]]<-paste("l",i,"_",nchar(Item.Comb[[i]][Item.Comb[[i]]%in%(Attr.Comb[[a]])]),Item.Comb[[i]][Item.Comb[[i]]%in%(Attr.Comb[[a]])],sep='')
        }
      }
    }
  }

  #Create Lambda Table
  Lamda.Table<-temp.table
  for(i in 1:nr){
    for(a in 1:length(Lamda.Table)){
      t.ref<-unique(as.character(Lamda.Table[i,]))
      pos<-c(1:length(t.ref))[Lamda.Table[i,a]==t.ref]
      temp.table[i,a]<-paste("t",i,"_",pos,sep='')}}

  #Generate LCDM specification
  out<-list()
  out[[1]]<-Lamda.Table
  out[[2]]<-temp.table
  out[[3]]<-constraints.list
  out[[4]]<-nway.inter.list
  out[[5]]<-intercept
  OUTPUT<-out
  nclass<-ncol(OUTPUT[[1]]);Nc<-nclass

  #Produce kernel expressions across items and attributes
  Kernel.exp<-OUTPUT[[1]]
  Kernel.exp.detect<-OUTPUT[[1]] #052719updates
  Kernel.exp.dino<-OUTPUT[[1]] #052719updates
  for (i in 1:nrow(OUTPUT[[1]])){
    for ( j in 1:ncol(OUTPUT[[1]])){
      if(sum(grep('S',OUTPUT[[1]][i,j]))!=0){Kernel.exp[i,j]<-gsub('S','+',OUTPUT[[1]][i,j])
      Kernel.exp.detect[i,j]<-NA} #052719updates
    }
  }
  for (i in 1:nrow(OUTPUT[[1]])){ #052719updates
    theClosestEffect<-which(is.na(Kernel.exp.detect[i,]))[1] #052719updates
    useToReplaceLonger<-Kernel.exp[i,theClosestEffect] #052719updates
    Kernel.exp.dino[i,is.na(Kernel.exp.detect[i,])]<-useToReplaceLonger #052719updates
  } #052719updates

  #Monotonicity constraint in terms of the interaction terms of the item effects
  Constrain.List1<-NULL
  name.inter<-unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])>=2]
  numway.inter<-unlist(OUTPUT[[4]])[unlist(OUTPUT[[4]])>=2]
  subname.inter<-substr((unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])>=2]), (nchar(unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])>=2])-unlist(OUTPUT[[4]])[unlist(OUTPUT[[4]])>=2]+1),
                        nchar(unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])>=2]))

  if(length(name.inter)!=0){
    for (inter in 1: length(name.inter)){
      temp.nw<-numway.inter[inter]
      temp.nm<-name.inter[inter]
      temp.subnm<-strsplit(subname.inter[inter],split='')[[1]]
      temp.sel<-paste(unlist(strsplit(temp.nm,split = '_'))[1],"_",(1:(temp.nw-1)),sep='')
      first.sel<-unlist(OUTPUT[[3]])[grep(paste((temp.sel),collapse="|"),unlist(OUTPUT[[3]]))]
      second.sel<-sub(".*_.", "", first.sel)
      for (sel in 1:length(temp.subnm)){
        SEL<-second.sel[sel]
        Constrain.List1<-rbind(
          paste(temp.nm,">-(0", paste("+",first.sel[grep(SEL,second.sel)],
                                      sep='',collapse=''),")",sep=''),Constrain.List1)
      }
    }
    Constrain.List1<-as.character(Constrain.List1)
  }else{
    Constrain.List1<-NULL
  }

  itemParmName<-c(unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==1],unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==2],
                  unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==3],
                  unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==4],
                  unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==5],
                  unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==6],
                  unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==7],
                  unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==8],
                  unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==9],
                  unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==10],
                  unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==11],OUTPUT[[5]])
  numMainEffect<-length(unlist(OUTPUT[[3]])[unlist(OUTPUT[[4]])==1])
  Constrain.List<-paste('  real<lower=0>',itemParmName[1:numMainEffect],';\n ')
  Unconstrain.List<-paste('  real',itemParmName[-(1:numMainEffect)],';\n ')
  Reparm<-as.data.frame(matrix(0,nr,nclass))

  #############################################################
  ###########052719update: Only one main effect is needed #####
  keep.oneMainEffect<-rep(1,nr)   #052719update:
  zero.list<-constraints.list
  fixparm.vec<-NULL #052719update: later they will be fixed to zero

  for(i in 1:nr){
    keep.oneMainEffect[i]<-OUTPUT[[3]][[i]][max(which(OUTPUT[[4]][[i]]==1))]#052719update:
    if(length(zero.list[[i]])==1){zero.list[[i]]=NA}else{#052719update:
      zero.list[[i]]<-constraints.list[[i]][-1]#052719update:
      fixparm.vec<-c(fixparm.vec,zero.list[[i]])#052719update:
    }
  } #intercept, hi.interaction,zero.list/fixparm.vec are what we need
  fixparm.vec<-unlist(constraints.list)[!unlist(constraints.list)%in%keep.oneMainEffect]
  Constrain.List<-paste('  real<lower=0>',keep.oneMainEffect,';\n ')#052719update:
  Unconstrain.List<-paste('  real',intercept,';\n ')
  #############################################################
  #############################################################

  #Produce Stan code for PImat parameter
  for(loopi in 1:nr){
    for( loopc in 1:nclass){
      Reparm[loopi,loopc]<-paste('  PImat[',loopi,',',loopc,']=inv_logit(',paste(Kernel.exp.dino[loopi,loopc]),');\n',sep='')
    }
  }

  #########052719update:create g and s parameters
  gParm<-rep(0,nr)
  sParm<-rep(0,nr)
  for(loopi in 1:nr){
    gParm[loopi]<-paste('  gParm[',loopi,']=inv_logit(',paste(Kernel.exp.dino[loopi,1]),');\n',sep='')
    sParm[loopi]<-paste('  sParm[',loopi,']=1-inv_logit(',paste(Kernel.exp.dino[loopi,nclass]),');\n',sep='')
  }


  Modelcontainer<-paste('   vector[Nc] contributionsC;\n','    vector[Ni] contributionsI;\n\n',sep='')
  Parmprior<-paste(c(paste('   //Prior\n'),paste('   ',itemParmName,'~normal(0,15)',';\n',sep=''),paste('   Vc~dirichlet(rep_vector(2.0, Nc));',sep='')))
  #############################################################
  ###########052719update: Only one main effect is needed #####
  update.Parmprior<-Parmprior
  fix.Parmprior<-NULL
  for(i in 1:length(Parmprior)){
    if(grepl(paste(fixparm.vec, collapse = "|"), Parmprior[i])){
      update.Parmprior[i]<-""
    }
  }
  update.Parmprior<-update.Parmprior[update.Parmprior!='']
  fix.Parmprior<-c(paste('  real',fixparm.vec,';\n '),
                   paste(' ',fixparm.vec,"=0",';\n ')
  )
  ##therefore we can use: fix.Parmprior,update.Parmprior
  #############################################################
  #############################################################

  #Likelihood Stan code
  Likelihood<-'
  \n
  //Likelihood
  for (iterp in 1:Np){
    for (iterc in 1:Nc){
      for (iteri in 1:Ni){
        if (Y[iterp,iteri] == 1)
          contributionsI[iteri]=bernoulli_lpmf(1|PImat[iteri,iterc]);
        else
          contributionsI[iteri]=bernoulli_lpmf(0|PImat[iteri,iterc]);
      }
      contributionsC[iterc]=log(Vc[iterc])+sum(contributionsI);
    }
  target+=log_sum_exp(contributionsC);
  }
  '


  #Data Specification
  data.spec<-'
  data{
  int Np;
  int Ni;
  int Nc;
  matrix[Np, Ni] Y;
  }
  '
  #Parameter Specification
  parm.spec<-paste(c('
  parameters{
  simplex[Nc] Vc;\n ',paste0(Constrain.List),paste0(Unconstrain.List),
                     '}\n'),collapse='')

  #Reparameter Specification
  transparm.spec<-paste(c('
  transformed parameters{
  matrix[Ni, Nc] PImat;
  vector[Ni] gParm;
  vector[Ni] sParm;\n',
                          gParm, #052719update
                          sParm, #052719update
                          paste0(unlist(Reparm)),'}\n'),collapse='')

  #Model Specification update052619
  model.spec<-paste(c('\nmodel {\n',paste(c(Modelcontainer,update.Parmprior,Likelihood),sep=''),'\n}',sep=''))
  model.spec<-model.spec[!startsWith(str_remove_all(model.spec," "),"~")]
  #Generated Quantities Specification
  generatedQuantities.spec<-'
  \n
  generated quantities {
  vector[Ni] log_lik[Np];
  vector[Ni] contributionsI;
  matrix[Ni,Nc] contributionsIC;
  //Posterior
  for (iterp in 1:Np){
    for (iteri in 1:Ni){
      for (iterc in 1:Nc){
        if (Y[iterp,iteri] == 1)
          contributionsI[iteri]=bernoulli_lpmf(1|PImat[iteri,iterc]);
        else
          contributionsI[iteri]=bernoulli_lpmf(0|PImat[iteri,iterc]);
          contributionsIC[iteri,iterc]=log(Vc[iterc])+contributionsI[iteri];
      }
      log_lik[iterp,iteri]=log_sum_exp(contributionsIC[iteri,]);
    }
  }
  }
  '

  if (.Platform$OS.type == "unix") {
    filename = paste(paste(save.path,save.name,sep='/'),'.stan',sep='')
  }else{
    filename = paste(paste(save.path,save.name,sep='\\'),'.stan',sep='')
  }

  sink(file=filename, append=FALSE)
  cat(
    paste(c('   ',
            data.spec,parm.spec,transparm.spec,model.spec,generatedQuantities.spec)
    ))
  sink(NULL)

}