# NOT RUN {
## Convert 'modlist' into 'replist'.
ml1 <- modlist(reps, model = l4)
rl1 <- replist(ml1, group = gl(7, 4))
plot(rl1)
summary(rl1[[1]])
## Optimize model based on Akaike weights.
rl2 <- replist(ml1, group = gl(7, 4), opt = TRUE,
optPARS = list(crit = "weights"))
plot(rl2)
# }
# NOT RUN {
## Remove kinetic outliers,
## use first replicate name for output.
ml3 <- modlist(reps, model = l4)
rl3 <- replist(ml3, group = gl(7, 4), check = "uni1",
remove = "KOD", names = "first")
plot(rl3, which = "single")
## Just aggregation and no refitting.
ml4 <- modlist(reps, model = l4)
rl4 <- replist(ml4, group = gl(7, 4), doFit = FALSE)
plot(rl4, which = "single")
## Scenario 1:
## automatic removal of runs that failed to
## fit during 'modlist' by using 'testdat' set.
ml5 <- modlist(testdat, model = l5)
rl5 <- replist(ml5, gl(6, 4))
plot(rl5, which = "single")
## Scenario 2:
## automatic removal of runs that failed to
## fit during 'replist':
## samples F3.1-F3.4 is set to 1.
dat1 <- reps
ml6 <- modlist(dat1)
ml6[[9]]$DATA[, 2] <- 1
ml6[[10]]$DATA[, 2] <- 1
ml6[[11]]$DATA[, 2] <- 1
ml6[[12]]$DATA[, 2] <- 1
rl6 <- replist(ml6, gl(7, 4))
plot(rl6, which = "single")
# }
Run the code above in your browser using DataLab