In philipdelff/NMdata: Preparation, Checking and Post-Processing Data for PK/PD Modeling

Subject-level variables

In a proper PK/PD analysis we need to explore the data at multiple variability levels. What we looked at above is at dosing and sampling level (one row per dosing or sampling event). NMdata provides very useful functions to extract information at other levels of variability. Before extracting the subject-level information we will add an individual exposure measure to the dataset. We will use the empirical Bayes' estimate of the individual maximum concentration. This is derived as the maximum prediction across the sample times - it may be better to simulate the model at a richer time scale to get better precision. Again, we make use of data.table but feel free to use what you prefer. findCovs - like other NMdata functions - will work on data.frame's, tibble's and other data.frame-like classes (classes that inherit from data.frame).

res1 <- NMscanData(system.file("examples/nonmem/xgxr001.lst", package="NMdata"),
                   col.row="ROW",merge.by.row=TRUE,quiet=TRUE,as.fun="data.table")
res1$trtact <- reorder(res1$trtact,res1$DOSE)
## with data.table, create a new column representing ID-level Cmax
res1[,Cmax:=max(IPRED),by=.(ID)]
## findCovs picks the columns that do not vary within cols.id. One row
## per value of cols.id.
res1.id <- findCovs(res1,by="ID")
dim(res1.id)
ggplot(res1.id,aes(WEIGHTB,Cmax/DOSE,colour=trtact))+
    geom_point()+
    labs(x="Bodyweight at baseline (kg)")

by can be longer than one. Include a study column, a treatment column to support cross-over, or a model column if you are comparing model fits (see below how to do this very easily).

If your model includes occasion variability, you probably also want to look at

## we have no occasion variability in this data
## res1.id.occ <- findCovs(res1,by=c("ID","OCC"))

The simplest use of findCovs is to get variables that are constant across the whole dataset:

findCovs(res1)

Let's take a look at what is in the res1.id generated above. It is a mix of variables that vary at subject level and variables that are constant across the full dataset.

dim(res1.id)
head(res1.id,2)

findCovs has a counterpart in findVars which finds variables that do vary within constant values of optional by columns. To get only the ones that vary within the dataset (i.e. they are truely subject-level variables and not say a study number if we only have one study), we can do

res1.id2 <- findVars(res1.id)
dim(res1.id2)
head(res1.id2,2)

findVars supports the by argument too. So you can use findVars(res1,by="ID") to find variables that are non-constant within (at least one value of) ID. by can be of arbitrary length.

Of course, we most often know what covariates or other subject-level variables to look at, and we would not search for them after running a nonmem model. But in the situation where you are looking at someone else's work or you are doing a meta analysis across models where the data has been coded slightly differently, these simple tools can be very useful. Or if you like writing generic wrapper functions, you may find this very handy. Also, remember that if a variable is returned by these functions, you know that they fullfill the variability requirement. We know that variables in res1.id above do not vary within ID, meaning that no ID has more than one value of the variable. NA counts as any other value, so if a value is returned for subject in res1.id, this subject does not have NA's in res1.

philipdelff/NMdata documentation built on March 5, 2025, 12:20 a.m.