Data Preparation Tools

This vignette is still under development. Please make sure to see latest version available here.

Introduction

Getting data ready for modeling is a crucial and often underestimated task. Mistakes during the process of combining data sets, defining time variables etc. can lead to difficulties during modeling, need for revisiting data set preparation, and in worst case wasted time working with an erroneos data set. Avoiding those mistakes by integrating checks into the data preparation process is a key element in an efficient and reliable data preparation work flow.

Furthermore, Nonmem has a number of restrictions on the format of the input data, and problems with the data set is a common reason for Nonmem not to behave as expected. When this happens, debugging can be time-consuming. NMdata includes some simple functions to prevent these situations.

This vignette uses data.table syntax for the little bit of data manipulation performed. However, you don’t need to use data.table at all to use these or any tool in NMdata. The data set is a data.table:

pk <- readRDS(file = system.file("examples/data/xgxr2.rds", package = "NMdata"))
class(pk)
#> [1] "data.table" "data.frame"

If you are not familiar with data.table, you can keep reading this vignette and learn what NMdata can do. In all brevity, data.table is a powerful enhancement to the data.frame class. The syntax differ from data.frame, and the few places where this affects the examples provided here, explanations will be given.

Data assembly

Compare presence and class of columns across data sets

When stacking (rbind) and merging, it is most often necessary to check if two or more data sets are compatible for the operation. compareCols compares columns across two or more data sets.

To illustrate the output of compareCols, a slightly modified version of the pk dataset has been created. One column (CYCLE) has been removed, and AMT has been recoded to character. compareCols tells us about exactly these two differences:

compareCols(pk, pk.reduced)
#> Dimensions:
#>          data nrows ncols
#> 1:         pk  1502    24
#> 2: pk.reduced   751    23
#> 
#> Columns that differ:
#>    column      pk pk.reduced
#> 1:  CYCLE integer       <NA>
#> 2:    AMT integer  character

Before merging or stacking, we may want to recode AMT in one of the datasets to get the class we need, and decide what to do about the CYCLE column which is missing in one of the datasets (add information or fill with NA?).

Rename columns based on contents

The model estimation step is heavily dependent (and in Nonmem almost entirely based) on numeric data values. The source data will often contain character variables, i.e. columns with non-numeric data values.

If the column names reflect whether the values are numeric, double-checking can be avoided. renameByContents renames columns if a function of their contents returns TRUE.

pk.renamed <- renameByContents(data = pktmp, fun.test = NMisNumeric,
    fun.rename = tolower, invert.test = TRUE)

We make use of the function NMisNumeric which tests if Nonmem can interpret the contents as numeric. If say the subject ID is of character class, it can be valid to Nonmem. Subject ID "1039" will be a numeric in Nonmem, "1-039" will not. NMisNumeric will return TRUE if and only if all elements are either missing or interpretable as numeric. We invert the condition (invert.test=TRUE), and the names of the columns that Nonmem cannot interpret as numeric become lowercase. We use compareCols to illustrate that three columns were renamed:

compareCols(pktmp, pk.renamed)
#> Dimensions:
#>          data nrows ncols
#> 1:      pktmp  1502    23
#> 2: pk.renamed  1502    23
#> 
#> Columns that differ:
#>      column     pktmp pk.renamed
#> 1:   EVENTU character       <NA>
#> 2:     NAME character       <NA>
#> 3: TIMEUNIT character       <NA>
#> 4:   eventu      <NA>  character
#> 5:     name      <NA>  character
#> 6: timeunit      <NA>  character

We can now easily see that if we wish to include the information contained in eventu, pktmp, and pk.renamed, we have to modify or translate their contents first.

Automated checks of merge results

Merges are a very common source of data creation bugs. As simple as they may seem, merges can leave you with an unexpected number of rows, some repeated and/or some omitted. Often, we can impose restrictions on the merge operation that allows for automated validation of the results.

Imagine the very common example that we have a longitudinal PK data set (called pk), and we want to add subject-level covariates from a secondary data set (dt.cov). We want to merge by ID, and all we can allow to happen is columns to be added to pk from dt.cov. If rows disappear or get repeated, or if columns get renamed, it’s unintended and should return an error.

We merge the two data sets and the check of the dimensions raises no alarm - the number of rows is unchanged from pk to pk2, and one of two columns in dt.cov was added.

pk2 <- merge(pk, dt.cov, by = "ID")
dims(pk, dt.cov, pk2)
#>      data nrows ncols
#> 1:     pk  1502    24
#> 2: dt.cov   150     2
#> 3:    pk2  1502    25

What we didn’t realize is that we now have twice as many rows for subject 31.

pk[ID == 31, .N]
#> [1] 10
pk2[ID == 31, .N]
#> [1] 20

Using mergeCheck, we get an error. This is because mergeCheck compares the actual rows going in and out of the merge and not just the dimensions:

try(mergeCheck(pk, dt.cov, by = "ID"))
#> Rows disappeared during merge.
#> Rows duplicated during merge.
#> Overview of dimensions of input and output data:
#>         data nrows ncols
#> 1:        pk  1502    25
#> 2:    dt.cov   150     2
#> 3: merged.df  1502    26
#> Overview of values of by where number of rows in df1 changes:
#>     ID N.df1 N.result
#> 1:  31    10       20
#> 2: 180    10        0
#> Error in mergeCheck(pk, dt.cov, by = "ID") : 
#>   Merge added and/or removed rows.

Notice that mergeCheck tells us for which values of ID the input and output differ so we can quickly look into the data sets and make a decision how we want to handle this. In this case we discard the covariate value for subject 31 and use all.x=TRUE argument to get NA for subjects 31 and 180:

dt.cov2 <- dt.cov[ID != 31]
pk2.check <- mergeCheck(pk, dt.cov2, by = "ID", all.x = TRUE)
#> The following columns were added: COV

To ensure the consistency of rows before and after the merge, you could use merge(...,all.x=TRUE) and then check dimensions before and after (yes, both all.x=TRUE and the dimension check are necessary). This is not needed if you use mergeCheck.

mergeCheck does not try to reimplement merging. Under the hood, the merging is performed by data.table::merge.data.table to which most arguments are passed. What mergeCheck does is to add the checks that the results are consistent with the criteria outlined above. data.table::merge.data.table is generally very fast, and even if there is a bit of extra calculations in mergeCheck, it should never be slow.

Another problem the programmer may not realize during a merge is when column names are shared across x1 and x2 (in addition to columns that are being merged by). This will silently create column names like col.x and col.y in the output. mergeCheck will by default give a warning if that happens (can be modified using the fun.commoncols argument). Also, there is an optional argument to tell mergeCheck how many columns are expected to be added by the merge, and mergeCheck will fail if another number of columns are added. This can be useful for programming.

The row order is by default maintained by mergeCheck. Apart from this, there is only one difference from the behavior of the merge.data.frame function syntax, being that the by argument must always be supplied to mergeCheck. Default behavior of merge.data.frame is to merge by all common column names, but for coding transparency, this is intentionally not allowed by mergeCheck.

In summary, mergeCheck verifies that the rows that result from the merge are the exact same as in one of the existing datasets, only columns added from the second input dataset. You may think that this will limit your merges, and that you need merges for inner and outer joins etc. You are exactly right - mergeCheck is not intended for those merges and does not support them. When that is said, the kind of merges that are supported by mergeCheck are indeed very common. All merges in the NMdata package are performed with mergeCheck.

Exclusion flags

It is good practice not to discard records from a dataset but to flag them and omit them in model estimation. When reporting the analysis, we also need to account for how many data records were discarded due to which criteria. A couple of functons in NMdata help you do this in a way that is easy to integrate with Nonmem.

The implementation in NMdata is based on sequentially checking exclusion conditions. This means we can summarize how many records and subjects were excluded from the analysis due to the different criteria. The information is represented in one numerical column for Nonmem, and one (value-to-value corresponding) character column for the rest of us in the resulting data.

Assign and count flags

For use in Nonmem, the easiest is that inclusion/exclusion is determined by a single column in data - we call that column FLAG here, but any column name can be used. FLAG obviously draws on information from other columns such as TIME, DV, and many others, depending on your dataset and your way of working.

The function that applies inclusion/excluasion rules is called flagsAssign, and it takes a dataset and a data.frame with rules as arguments.

dt.flags <- fread(text = "FLAG,flag,condition
10,Below LLOQ,BLQ==1
100,Negative time,TIME<0")

pk <- flagsAssign(pk, tab.flags = dt.flags, subset.data = "EVID==0")
#> Coding FLAG = 100, flag = Negative time
#> Coding FLAG = 10, flag = Below LLOQ
pk <- flagsAssign(pk, subset.data = "EVID==1", flagc.0 = "Dosing")

fread is used to create a data.table (like read.csv to create a data.frame) for readability, one line for each row in the data.table created. flagsAssign applies the conditions sequentially and by decreasing value of FLAG. FLAG=0 means that the observation is included in the analysis. You can use any expression that can be evaluated within the data.frame. In this case, BLQ has to exist in pk.

Finally, flags are assigned to EVID==1 rows. Here, no flag table is used. This means that all EVID==1 rows will get FLAG=0.

In Nonmem, you can include IGNORE=(FLAG.NE.0) in $DATA or $INFILE.

Again, the omission will be attributed to the first condition matched. Default is to apply the conditions by the order of decreasing numerical flag value. Use flags.increasing=TRUE if you prefer the opposite. However, what cannot be modified is that 0 is the numerical value for rows that are not matched by any conditions.

What rows to omit from a dataset can vary from one analysis to another. Hence, the aim with the chosen design is that the inclusion criteria can be changed and applied to overwrite an existing inclusion/exclusion selection. For another analysis we want to include the observations below LLOQ. We have two options. Either we simply change the IGNORE statement given above to IGNORE=(FLAG.LT.10), or you create a different exclusion flag for that one. If you prefer to create a new set of exclusion flags, just use new names for the numerical and the character flag columns so you don’t overwrite the old ones. See help of flagsAssign and flagsCount for how.

Summarize data exclusions

An overview of the number of observations disregarded due to the different conditions is then obtained using flagsCount:

tab.count <- flagsCount(data = pk[EVID == 0], tab.flags = dt.flags)
print(tab.count)
#>                  flag N.left Nobs.left N.discard N.disc.cum Nobs.discard
#> 1: All available data    150      1352        NA          0           NA
#> 2:      Negative time    150      1350         0          0            2
#> 3:         Below LLOQ    131       755        19         19          595
#> 4:       Analysis set    131       755        NA         19           NA
#>    Nobs.disc.cum
#> 1:             0
#> 2:             2
#> 3:           597
#> 4:           597

flagsCount includes a file argument to save the the table right away.

Finalize data format and write to file

Once the dataset is in place, NMdata provides a few useful functions to ensure the formatting of the written data is compatible with Nonmem. These functions include checks that Nonmem will be able to interpret the data as intended, and more features are under development in this area.

Automated ordering of columns

The order of columns in Nonmem is important for two reasons. One is that a character in a variable read into Nonmem will make the run fail. The other is that there are restrictions on the number of variables you can read into Nonmem, depending on the version. NMorderColumns tries to put the used columns first, and other or maybe even unusable columns in the back of the dataset. It does so by a mix of recognition of column names and analysis of the column contents.

Columns that cannot be converted to numeric are put in the back, while column bearing standard Nonmem variable names like ID, TIME, EVID etc. will be pulled up front. You can of course add column names to prioritize to front (first) or back (last). See ?NMorderColumns for more options.

pk <- NMorderColumns(pk)

We may want to add MDV and rerun NMorderColumns.

Writing data to files

For the final step of writing the dataset, NMwriteData is provided. Most importantly, it writes a csv file with appropriate options for Nonmem to read it as well as possible. It can also write an rds for R with equal contents (or RData if you prefer), but with the rds including all information (such as factor levels) which cannot be saved in csv. If you should use NMscanData to read Nonmem results, this information can be used automatically. It also provides a proposal for text to include in the $INPUT and $DATA sections of the Nonmem control streams.

NMwriteData(pk)
#> Data _not_ witten to any files.
#> For NONMEM:
#> $INPUT ROW ID NOMTIME TIME EVID CMT AMT DV FLAG STUDY BLQ CYCLE DOSE
#> PART PROFDAY PROFTIME WEIGHTB eff0
#> $DATA <data file>
#> IGN=@
#> IGNORE=(FLAG.NE.0)

Notice, NMwriteData detected the exclusion flag and suggests to include it in $DATA.

If a file name had been provided, the data would have been written, and the path to the data file would have been included in the message written back to the user. There are several arguments that will affect the proposed text for the Nonmem run, see ?NMwriteData.

Updating Nonmem control streams to read new data file

I may be the only one, but sometimes during the modeling stage, I want to go back and change or add something to the data creation step. Then once I have written a new data file, my Nonmem $INPUT sections no longer match the data file. In NMwriteData you can use a last argument to get columns pushed towards the back so the Nonmem runs should still work, but maybe you need the column in your nonmem runs, and so you have no way around updating the control streams. And that can be quite a lot of control streams.

NMdata has a couple of functions to extract and write sections to Nonmem control streams called NMreadSection and NMwriteSection. We are not going into detail with what these functions can do, but let’s stick to the example above. We can do

NMwriteSection("run001.mod", "INPUT", "$INPUT ROW ID TIME EVID CMT AMT DV FLAG STUDY BLQ CYCLE DOSE FLAG2 NOMTIME PART PROFDAY PROFTIME WEIGHTB eff0")

But in fact, we can go a step further and take the information straight from NMwriteData

text.nm <- NMwriteData(pk)
#> Data _not_ witten to any files.
#> For NONMEM:
#> $INPUT ROW ID NOMTIME TIME EVID CMT AMT DV FLAG STUDY BLQ CYCLE DOSE
#> PART PROFDAY PROFTIME WEIGHTB eff0
#> $DATA <data file>
#> IGN=@
#> IGNORE=(FLAG.NE.0)

NMwriteData invisibly returns a list of sections ($INPUT and $DATA). NMwriteSection can use these directly. So to write only the $INPUT section to run001.mod, we do the following. Please notice the single brackets in text.nm["INPUT"] which mean that we still send a list to NMwriteSection.

NMwriteSection("run001.mod", list.sections = text.nm["INPUT"])

If you run this in a loop over the control streams that use the created data set, you are all set to rerun the models as needed.