In CJangelo/COA34: Psychometric Analyses Aligned with FDA COA Guidances 3 and 4

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

Evaluate Test-Retest Reliability

Recently, the FDA has recommended ICC(A,1). Refer to this article: https://link.springer.com/article/10.1007/s11136-018-2076-0

COA34::compute_iccA1 is just a wrapper for the icc function in the irr R package. This R function just makes it easier to compute the values and output the tables. The underlying code is available here: https://github.com/cran/irr/blob/master/R/icc.R

TODO

1. Figure out what the generating value is - is it the residual correlation? Sim study needed!

2. Explore further how MAR/MNAR drop-out severely attenuates the ICC(A,1) value

print.dir = "C:/Users/ciaconangelo/OneDrive - OPEN Health/Documents/misc_R_table_output"

Required R packages

library(COA34)
# In addition, you'll need the following
library(ggplot2)
library(grid)
library(gridExtra)

Generate data

  set.seed(12162021)

  sim.out <- COA34::sim_pro_dat_v2(N=1000,
                            number.timepoints = 3,
                            Beta.PRO = NULL,
                            number.of.anchor.groups = 5,
                            polychor.value = 0.7,
                            corr = 'ar1',
                            cor.value = 0.8,
                            #var.values = c(7.136))
                            var.values = c(2))
  dat <- sim.out$dat

Implement drop-out

dat <- COA34::dropout(dat = dat,
               type_dropout  = c('mcar', 'mar', 'mnar'),
               prop.miss = 0.5,
               stochastic.component = 0.2)

Simulate Items

 # Simulate IRT items
 # use the scores generated as the theta in the IRT model 
  sim.out2 <- COA34::sim_irt_item(dat = dat, J = 5, K = 4, latent.variable = 'Y_comp')
    str(sim.out2)
    dat <- sim.out2$dat

# Score the PRO - just take a simple sum score here:
  dat$PRO.score <- apply(dat[, grep('Item', colnames(dat))], 1, sum)

# Create the same PRO score, but with MAR drop-out:
  dat$PRO.score_mar <- dat$PRO.score
  dat$PRO.score_mar[is.na(dat$Y_mar)] <- NA
# Note that you've just set the PRO score to missing wherever the Y_mar variable is missing

# Now for the other missing types:
    dat$PRO.score_mcar <- dat$PRO.score_mnar <- dat$PRO.score
    dat$PRO.score_mcar[is.na(dat$Y_mcar)] <- NA
    dat$PRO.score_mnar[is.na(dat$Y_mnar)] <- NA

  aggregate(cbind(PRO.score, PRO.score_mcar, PRO.score_mar, PRO.score_mnar) ~ Time,
                  function(x) mean(x, na.rm = T), 
                  data = dat, 
                  na.action = na.pass)

Compute PGIS delta

# You already have the PGIS_bl and PGIS_delta in the generated data,
# you wouldn't have that in a real dataset, so drop that first
dat <- dat[, !(colnames(dat) %in% c('PGIS_bl', 'PGIS_delta'))]
# This makes this more realistic
# use the function below:

dat <- COA34::compute_anchor_delta(dat = dat,
                                   subject.id = 'USUBJID',
                                   time.var = 'Time',
                                   anchor = 'PGIS')

Compute PRO Score delta

dat <- COA34::compute_change_score(dat = dat,
                                   subject.id = 'USUBJID',
                                   time.var = 'Time',
                                   score = c('PRO.score', 'PRO.score_mcar', 
                                             'PRO.score_mar', 'PRO.score_mnar'))

# The function creates variables with the same name plus "_delta":
str(dat)

Compute Test-Retest Reliability

Compute the ICC(A,1) values and output them to a table:

icc <- COA34::compute_iccA1(dat = dat,
                            PRO.score = c('PRO.score', 
                                          'PRO.score_mcar', 
                                          'PRO.score_mar', 
                                          'PRO.score_mnar'),
                            time.var = 'Time',
                            subject.id = 'USUBJID',
                            anchor = 'PGIS_delta',
                            stable.score = 0, 
                            first.timepoint = 'Time_1', 
                            second.timepoint = 'Time_2')

Print the table out:

library(R2Word)

R2Word::dump_df_mat_to_file(out = icc$icc.A1,
                            decimals = c(2, 0, 0), 
                            table.title = 'ICC(A,1) - Vignette Illustration', 
                            table.footnote = '**All data simulated',
                            file.name = 'trtr_vig_iccA1', 
                            print.dir = print.dir)

Shell Tables

If you're putting together shell tables, use the function shell_table, specifying that you want columns 2 and 3 to be converted to x instead of the numeric values.

st <- R2Word::shell_table(out = icc$icc.A1, decimals = c(2, 0), cols = c(2, 3))

R2Word::dump_df_mat_to_file(out = st,
                            decimals = 0, 
                            table.title = 'ICC(A,1) - Vignette Illustration', 
                            table.footnote = '**All data simulated',
                            file.name = 'trtr_vig_iccA1_shell_tables', 
                            print.dir = print.dir)

TODO: What is the generating value of ICC(A,1)?

Is it the correlation between the two timepoints? That is equal to 0.80.

sim.out$cor.mat

CJangelo/COA34 documentation built on June 23, 2022, 12:10 p.m.