mixmeta (version 1.2.0)

alcohol: Alcohol Intake and Colorectal Cancer

Description

The dataset contains the data on 8 cohort studies participating in the Pooling Project of Prospective Studies of Diet and Cancer. A total of 3,646 cases and 2,511,424 person-years were included in the analysis. Each study estimated the incidence relative rate in different categories of alcohol intake while controlling for a set of potential confounders, using non-drinkers as the reference. The categories where then converted in a dose by assigning to each the median value of individual consumptions, with studies reporting estimates at different levels in a continuous scale.

Usage

alcohol

Format

A data frame with 48 observations on the following 7 variables:

• id: label for each study, derived from the first author's name.

• type: code for study design (cohort estimating incidence rate).

• dose: assigned dose level (gr/day of alcohol intake).

• cases: number of cases for each dose category.

• peryears: amount of person-time for each dose category.

• logrr: estimated logarithm of the incidence relative rate.

• se: standard error of the estimates.

Details

The data are stored in a long format, with each record reporting the information for each dose categories and studies including multiple records. The reference category for each study included, although the log-RR is fixed to 0 with no standard error (comparing the category with itself). The information on these reference categories is needed to compute the approximate correlations between estimates in the same study.

Examples

# NOT RUN {
### REPRODUCE THE RESULTS IN CRIPPA ET AL (2016) AND ORSINI ET AL (2012)

# LOAD THE PACKAGE dosresmeta AND splines
library(dosresmeta) ; library(splines)

# COMPUTE THE WITHIN-STUDY CORRELATIONS EXCLUDING THE REFERENCE
addS <- lapply(split(alcohol, alcohol$id), function(x)
  covar.logrr(y=logrr, v=se^2, cases=cases, n=peryears, type=type, data=x))
sub <- subset(alcohol, !is.na(se))

# NOT ACCOUNTING FOR WITHIN-STUDY CORRELATIONS
nocor <- mixmeta(logrr ~ 0 + dose, S=se^2, random= ~ 0 + dose|id, data=sub,
  method="ml")
summary(nocor)

# ACCOUNTING FOR WITHIN-STUDY CORRELATIONS
lin <- mixmeta(logrr ~ 0 + dose, random= ~ 0 + dose|id, data=sub, method="ml",
  control=list(addSlist=addS))
summary(lin)

# ALLOWING NON-LINEARITY IN BOTH FIXED AND RANDOM PARTS
nonlin <- mixmeta(logrr ~ 0 + ns(dose, knots=c(10,25)), data=sub, 
  random= ~ 0 + ns(dose, knots=c(10,25))|id, method="ml",
  control=list(addSlist=addS))
summary(nonlin)

# SIMPLIFY THE MODEL BY ASSUMING LINEARITY IN THE RANDOM PART
nonlin2 <- update(nonlin, random= ~ 0 + dose|id)
summary(nonlin2)

# FIXED-EFFECTS MODEL (TRICK: random TO DEFINE THE GROUPING, THEN FIX IT TO 0)
nonlinfix <- mixmeta(logrr ~ 0 + ns(dose, knots=c(10,25)), random= ~ 1|id,
  data=sub, method="ml",bscov="fixed", control=list(addSlist=addS, Psifix=0))
summary(nonlinfix)

# COMPARE THE MODELS
AIC(nocor, lin, nonlin, nonlin2, nonlinfix)

# PREDICT THE RR FOR 12g/day FOM TWO MODELS
exp(predict(nocor, newdata=data.frame(dose=12), ci=TRUE))
exp(predict(lin, newdata=data.frame(dose=12), ci=TRUE))

# PREDICT (RECREATE SPLINES FOR EASY CODING)
predlin <- exp(predict(lin, newdata=data.frame(dose=0:60), ci=TRUE))
prednonlin <- exp(predict(nonlin, newdata=data.frame(dose=0:60), ci=TRUE))

# DISPLAY THE NON-LINEAR EFFECT
col1 <- do.call(rgb, c(as.list(col2rgb("blue") / 255), list(0.2)))
col2 <- do.call(rgb, c(as.list(col2rgb("green") / 255), list(0.2)))
plot(0:60, predlin[,1], type="l", ylim=c(0.85,1.9), ylab="RR",
  xlab="Alcohol intake (gr/day)", main="Dose-response")
polygon(c(0:60,60:0), c(predlin[,2], rev(predlin[,3])), col=col1, border=NA)
lines(0:60,prednonlin[,1], lty=5)
polygon(c(0:60,60:0), c(prednonlin[,2],rev(prednonlin[,3])), col=col2, border=NA)
# }