Fit Correlation Matrix

gen_corr_draw.R

# function written by HuffPost Pollster's Natalie Jackson for the 2016 election
# ibraries and import --------

library(dplyr)
library('Matrix')

seat_data <- read.csv("seat_partisanship_demog.csv")  # read in a 435 
seat_data <- seat_data %>%
  select(clinton2016, obama2012, house2016, pct_nonwhite)

# make a correlations matrix for simulation --------
seat_cor <- as.data.frame(t(merged))

names(seat_cor) <- gsub("-","",gseats$Code)
corr <- cor(seat_cor, use="complete.obs")
row.names(corr) <-  gsub("-","",gseats$Code)

head(corr)

write.csv(corr,"Tracker/Data/corr_matrix.csv",row.names=FALSE)



# this function reads the correlation matrix and corrects values to work properly when generating draws --------
buildCorrelationConstant <- function() {
  # Read correlations from file
  frame <- read.csv("Tracker/Data/corr_matrix.csv")
  rawMatrix <- as.matrix(frame)

  # Make the matrix the nearest positive definite, but still correlated,
  # because CorrelatedRnorm() won't work with the negative eigenvalues the
  # original generates.
  pdMatrix <- as.matrix(nearPD(rawMatrix, corr=TRUE)$mat)

  ev <- eigen(pdMatrix, symmetric=TRUE)

  return(ev$vectors %*% diag(sqrt(ev$values)) %*% t(ev$vectors))
}

kCorrelationConstant <- buildCorrelationConstant()

# here, we actually draw some random numbers ----------
CorrelatedRnorm <- function() {
  n <- nrow(kCorrelationConstant)
  random <- rnorm(n)
  ret <- kCorrelationConstant %*% random
return(drop(ret))

# and here's a plot to compare clinton share and the generated error ------
data.frame("corr"=CorrelatedRnorm(),"clinton"=seat_data$clinton2016)  %>% 
  ggplot(aes(clinton,corr))+
  geom_point()