Introduction to the pimeta package

Kengo Nagashima

2019-09-17

The pimeta package is easy. Load your data and then pass it the pima function!

library("pimeta")
library("ggplot2")
## Warning: package 'ggplot2' was built under R version 3.6.1
data(sbp, package = "pimeta")

# a parametric bootstrap prediction interval
piboot <- pima(
  y        = sbp$y,      # effect size estimates
  se       = sbp$sigmak, # within studies standard errors
  B        = 25000,      # number of bootstrap samples
  seed     = 14142135,   # random number seed
  parallel = 4           # multi-threading     
)
piboot
## 
## Prediction & Confidence Intervals for Random-Effects Meta-Analysis
## 
## A parametric bootstrap prediction and confidence intervals
##  Heterogeneity variance: DerSimonian-Laird
##  Variance for average treatment effect: Hartung (Hartung-Knapp)
## 
## No. of studies: 10
## 
## Average treatment effect [95% prediction interval]:
##  -0.3341 [-0.8807, 0.2240]
##  d.f.: 9
## 
## Average treatment effect [95% confidence interval]:
##  -0.3341 [-0.5613, -0.0985]
##  d.f.: 9
## 
## Heterogeneity measure
##  tau-squared: 0.0282
##  I-squared:  70.5%
plot(piboot, base_size = 10, studylabel = sbp$label)

Several type of methods ("HTS", "HK", "SJ", …) are supported.

# Higgins-Thompson-Spiegelhalter prediction interval
pima(sbp$y, sbp$sigmak, method = "HTS")
## 
## Prediction & Confidence Intervals for Random-Effects Meta-Analysis
## 
## Higgins-Thompson-Spiegelhalter prediction and confidence intervals
##  Heterogeneity variance: DerSimonian-Laird
##  Variance for average treatment effect: approximate
## 
## No. of studies: 10
## 
## Average treatment effect [95% prediction interval]:
##  -0.3341 [-0.7598, 0.0917]
##  d.f.: 8
## 
## Average treatment effect [95% confidence interval]:
##  -0.3341 [-0.5068, -0.1613]
##  d.f.: 9
## 
## Heterogeneity measure
##  tau-squared: 0.0282
##  I-squared:  70.5%

The convert_bin() function converts binary outcome data to effect size estimates and within studies standard errors vectors. A data set of 13 placebo-controlled trials with cisapride that was reported by Hartung and Knapp (Stat Med., 2001, doi:10.1002/sim.1009) was analyzed below. Estimated values on the logarithmic scale can be back-transformed to the original scale with the trans option (in print and plot).

m1 <- c(15,12,29,42,14,44,14,29,10,17,38,19,21)
n1 <- c(16,16,34,56,22,54,17,58,14,26,44,29,38)
m2 <- c( 9, 1,18,31, 6,17, 7,23, 3, 6,12,22,19)
n2 <- c(16,16,34,56,22,55,15,58,15,27,45,30,38)
dat <- convert_bin(m1, n1, m2, n2, type = "logOR")
head(dat, n = 3)
##          y        se
## 1 2.098986 0.9847737
## 2 3.357026 1.0165653
## 3 1.565232 0.5747840
pibin <- pima(dat$y, dat$se, seed = 2236067, parallel = 4)
print(pibin, digits = 3, trans = "exp")
## 
## Prediction & Confidence Intervals for Random-Effects Meta-Analysis
## 
## A parametric bootstrap prediction and confidence intervals
##  Heterogeneity variance: DerSimonian-Laird
##  Variance for average treatment effect: Hartung (Hartung-Knapp)
## 
## No. of studies: 13
## 
## Average treatment effect [95% prediction interval]:
##  4.141 [0.533, 33.692]
##  d.f.: 12
##  Scale: exponential transformed
## 
## Average treatment effect [95% confidence interval]:
##  4.141 [2.224, 7.820]
##  d.f.: 12
##  Scale: exponential transformed
## 
## Heterogeneity measure
##  tau-squared: 0.718
##  I-squared:  69.9%
binlabel <- c(
   "Creytens", "Milo", "Francois and De Nutte", "Deruyttere et al.",
   "Hannon", "Roesch", "De Nutte et al.", "Hausken and Bestad",
   "Chung", "Van Outryve et al.", "Al-Quorain et al.", "Kellow et al.",
   "Yeoh et al.")
plot(pibin, digits = 2, base_size = 10, studylabel = binlabel, trans = "exp")

Save a plot to PNG:

png("forestplot.png", width = 500, height = 300, family = "Arial")
plot(piboot, digits = 2, base_size = 18, studylabel = sbp$label)
dev.off()

or

p <- plot(piboot, digits = 2, base_size = 10, studylabel = sbp$label)
ggsave("forestplot.png", p, width = 5, height = 3, dpi = 150)