In zzhou0721/MCID: Estimating the Minimal Clinically Important Difference

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "README-"
)

MCID

'MCID' is an R package used to provide the point and interval estimation on the minimal clinically important difference (MCID) at the population and individual level. For population level, it produces a constant value for the estimation of MCID. For individual level, the MCID is defined as a linear function of patient's clinical characteristics, and the estimated linear coefficients can be obtained.

Installation

You can install the released version of MCID from GitHub with:

# install.packages("devtools")
devtools::install_github("zzhou0721/MCID")

Example

Generate data

rm(list = ls())
library(MCID)
n <- 500
lambdaseq <- 10 ^ seq(-3, 3, 0.1)
deltaseq <- seq(0.1, 0.3, 0.1)
a <- 0.1
b <- 0.55
c <- -0.1
d <- 0.45

set.seed(115)
p <- 0.5
y <- 2 * rbinom(n, 1, p) - 1
z <- rnorm(n, 1, 0.1)
y_1 <- which(y == 1)
y_0 <- which(y == -1)
x <- c()
x[y_1] <- a + z[y_1] * b + rnorm(length(y_1), 0, 0.1)
x[y_0] <- c + z[y_0] * d + rnorm(length(y_0), 0, 0.1)

Determine MCID at the population level

To determine MCID at the populaton level, we first need to select an optimal value for the tuning parameter δ, which is used to control the difference between 0-1 loss and surrogate loss.

sel <- cv.pmcid(x = x, y = y, delseq = deltaseq, 
                k = 5, maxit = 100, tol = 1e-02)
delsel <- sel$'Selected delta'
delsel

Then with selected optimal value of δ, we can determine the point and interval estimation of MCID at the population level. The confidence interval is constructed based on the asymptotic normality. In our simulated data, the true population MCID is 0.5.

result <- pmcid(x = x, y = y, n = n, delta = delsel, 
                maxit = 100, tol = 1e-02, alpha = 0.05)
result$'Point estimate'
result$'Standard error'
result$'Confidence interval'

Determine MCID at the individual level

To determine MCID at the individual level, we first need to select a combination of the optimal values for the tuning parameters δ and λ. δ is used to control the difference between 0-1 loss and surrogate loss. λ is the coefficient of the penalty term used to avoid the overfitting issue.

sel <- cv.imcid(x = x, y = y, z = z, lamseq = lambdaseq, 
                delseq = deltaseq, k = 5, maxit = 100, tol = 1e-02)
lamsel <- sel$'Selected lambda'
delsel <- sel$'Selected delta'
lamsel
delsel

Then with selected δ and λ, we can determine the point and interval estimation for the linear coefficients of the individualized MCID function. The confidence intervals are constructed based on the asymptotic normalities. In our simulated data, the true linear coefficients of the individualized MCID are β₀ = 0 and β₁ = 0.5

result <- imcid(x = x, y = y, z = z, n = n, lambda = lamsel, 
                delta = delsel, maxit = 100, tol = 1e-02, alpha = 0.05)
result$'Point estimates'
result$'Standard errors'
result$'Confidence intervals'