Pooling correlation matrix

[kikabradford]

I've been trying to apply the mifa analysis on my data with ordinal variables, using polr method on mice. I ran into some challenges on mifa, and I thought it was easier to run the MI using mice separately - which worked out well. Now, I meant to follow the path took for mifa https://github.com/vahidnassiri/mifa/blob/master/R/mifa.cov.R, adapting the combination of the MIs into a single averaged correlation matrix (instead of covariance) - getting errors.
I can't seem to troubleshoot the issues by myself (I am new in R), any insights are highly appreciated - The R code is below (the blue text is a note about it). Any thoughts?

Thanks,
Kika

####### MULTIPLE IMPUTATION
#reading my dataset

icmbio=read_csv("icmbio5-R.csv") %>%

  mutate_all(as.factor)%>%
select(3:75)

#Specifying the mice method per column
icmbio_init = mice(icmbio, maxit=0)
meth=icmbio_init$method
meth[c(1:67)]="polr"        #Proportional odds model for ordered variables (Likert scale items)
meth[c(1:67)]="logreg"    #Logistic regression for binary variables (sex and education)
meth[c(70:73)]="lda"         # unordered categorical variables -  Linear discriminant analysis (Wu et al 2015)

#imputing (mice)
icmbio1= mice(icmbio, method=meth, m=5, remove_collinear = FALSE)          # skip removal of collinear variables

# checking if everything is imputed
colSums(is.na(complete(icmbio1,1)))

# extracting imputed datasets (worked)
comp.mice=NULL
mi.na=rep(0,5)
for (i in 1:5){
  comp.mice[[i]]=complete(icmbio1,i)
  mi.na[i]=sum(is.na(comp.mice[[i]]))
}

### End: Implementing the imputation

### CORRELATION
# Now estimating correlation matrix based on imputed values,
cor.mice.imp=NULL
prop.exp=rep(0,5)
for (i in 1:5){
  cor.tmp=polychoric(comp.mice[[i]])  # polychoric correlations considering the categorical nature of the data
  cor.mice.imp[[i]]=cor.tmp
} 

##### ERRORS 

I managed to do these steps above, extract and save the complete and each imputed dataset. I tried using your code for extracting the datasets and estimating the correlation matrixes (instead of covariances - adapting it to a correlation function), but was not able to extract separate matrix, although was able to run the code and calculate it using polychoric(imp.icmbio1)$rho.

## Combining the estimated correlation from different imputations (DID NOT WORK)

cor.combined=Reduce('+',cor.mice.imp)/5
Error in f(init, x[[i]]) : non-numeric argument to binary operator

##I also tried the micombine.cor function of mice 
cor.comb= miceadds::micombine.cor(mi.res=icmbio1)
Error in stats::cor(dat_ii, dat_jj, method = method, use = "pairwise.complete.obs") : 
  'x' must be numeric

[jszczypinski]

Hey, I am not sure if you are still looking for an answer, but this is what I come up based on mifa.cov function from mifa package. I used POLYCHORIC_R() function from EFA.dimensions package to compute the correlations. "cols_to_cor" argument is used to subset the columns for correlation and exclude other columns that might have been useful for imputation, such as sex and age.
Hit me up will you have any questions!

# Computing covariance matrix of incomplete data using multiple imputation by
# Multivariate Imputation by Chained Equations (MICE) method.
mifa.cor <- function(data.miss,M,maxit.mi = 5,method.mi='pmm', cols_to_cor){
  # Please make sure the 'mice' package is installed. Use install.packages("mice")
  # to install it.
  # The input variable:
  # data.miss: the dataset contaning missing values, the missing values should be
  # shown with NA.
  # M: the number of generated imputations, for more information see
  # R documentations for mice package.
  # maxit.mi: a scalar giving the number of iterations for each imputation,
  # for more information see R documentations for mice package. The default is
  # set as 5.
  # method.mi : the method which should be used for imputation. It can be a string
  # or a vector of strings of the size equal to number of items. for more information see
  # R documentations for mice package. The default is set is 'pmm'.
  # cols_to_cor: columns that will be use to construct the correlation matrix
  N=dim(data.miss)[1]
  require(mice)
  require(EFA.dimensions)
  require(data.table)
  imputed_mice = mice(data.miss, m=M, maxit=maxit.mi,method=method.mi,print=FALSE)
  ### Begin: Implementing the sequantial imputation:
  # checking if everything is imputed
  method.levels.mi=levels(imputed_mice$loggedEvents$meth)
  if ('constant' %in% method.levels.mi){
    stop('Probably at least one column with constant observed part.')
  }
  # extracting imputed datasets
  comp.mice=NULL
  mi.na=rep(0,M)
  for (i in 1:M){
    comp.mice[[i]]=complete(imputed_mice,i)
    mi.na[i]=sum(is.na(comp.mice[[i]]))
  }
  # implementing sequential imputations in case that some of the columns are not
  # imputed due to collinearity, etc.
  while (sum(mi.na)>0){
    for (i in 1:M){
      imp.tmp=mice(comp.mice[[i]], m=1, maxit=maxit.mi,method=method.mi,print=FALSE)
      comp.mice[[i]]=complete(imp.tmp,1)
      mi.na[i]=sum(is.na(comp.mice[[i]]))
    }
  }
  ### Eend: Implementing the sequantial imputation:
  # Now estimating covariance matrix based on imputed values
  cor.mice.imp=NULL
  prop.exp=rep(0,M)
  for (i in 1:M){
    data.tmp <- as.data.table((comp.mice[[i]]))
    data.tmp <- data.tmp[, ..cols_to_cor]
    cor.tmp <- POLYCHORIC_R(data.tmp)
    cor.mice.imp[[i]] <- cor.tmp(verbose = FALSE)
  }
  # Combining the estimated covariance from different imputations
  cor.mice=Reduce('+',cor.mice.imp)/M
  return(cor.mice)
}