Built 2024-04-23 using NMdata 0.1.6.

Please make sure to see latest vignette version available here.


This vignettes aims at enabling you at

  • Using NMscanData to read and combine all output and input data based only on (the path to) the Nonmem list file (understanding how NMscanData prioritizes output and input data in case of redundancy)

  • Switching between combining output and input data by mimicking the Nonmem data filters (IGNORE/ACCEPT) and merging by a row identifier

  • Configuring NMdata to return the data class of your preference (say data.table or tbl) instead of data.frame which is default

  • Using automatically generated meta data to look up information on input and output tables, how they were combined, and results of checks performed by NMscanData.

  • Including input data rows that were not processed by Nonmem (ACCEPT and IGNORE)

  • Combining such data sets for multiple models

  • If available, using an rds file to represent the input data in order to preserve all data properties (e.g. factor levels) from data set preparation

  • After having checked the rare exceptions, feeling confident that NMscanData should work on all your Nonmem models


This vignette focuses on how to use NMdata to automate what needs to be trivial: get one dataset out of a Nonmem run, combining all output tables and including additional columns and rows from the input data. After scanning the Nonmem list file and/or control stream for file and column names, the data files are read and combined.

In brevity, the most important steps are:

  • Based on list file (often .lst): Identify input and output table files
  • Read and combine output tables
  • If wanted, read input data and restore variables that were not output from the Nonmem model.
  • If wanted, also restore rows from input data that were disregarded in Nonmem (e.g. observations or subjects that are not part of the analysis)

An additional complication is the potential renaming of input data column names in the Nonmem $INPUT section. NMscanData by default (but optionally) follows the column names as read by Nonmem.

This way of reading the output and input data is fully compatible with most other of the great R packages for reading data from Nonmem.

In most cases, the steps above are not too hard to do. But with the large degree of flexibility Nonmem offers, the code will likely have to be adjusted between models. The implementation in NMdata works for the vast majority of models and aims at preventing and checking for as many caveats as possible. It is fast too.

Default argument values can be configured depending on your setup (data standards, directory structure and other preferences).

Like the rest of NMdata, this functionality assumes as little as possible about how you work. It assumes nothing about the Nonmem model itself and as little as possible about the organization of data and file paths/names. This makes it powerful for meta analyses, for reading a model developed by someone else - or one written by ourselves when we used to do things slightly differently. It will work out of the box in the vast majority of cases.

We start by attaching NMdata. Also, I use data.tablefor a few post-processing steps. You can just as well use base R or dplyr if you prefer. Then ggplot2.

#> NMdata 0.1.6. Browse NMdata documentation at
#> https://philipdelff.github.io/NMdata/
## not necessary for NMdata to run, but we use thse in the examples

For the examples we will be using files that are available in the NMdata package. To type a little less, we use this shortcut function:

file.NMdata <- function(...) system.file(file.path("examples/nonmem",...), package="NMdata")

Note on file names and directory structures

Depending on your Nonmem setup, habits and preferences, you may name your control streams and list files differently than this vignette. Here, we use the NMdata default which is .mod and .lst. You can easily configure NMdata to match your preferences. See the FAQ for how. So for now, rest assured that this is easy to adjust and read on.

Get started - the basics

Try NMscanData on a control stream or a list file:

res1 <- NMscanData(file.NMdata("xgxr018.lst"))
#> Input and output data were searched for candidate unique row identifiers. None
#> found. To skip this check, please use merge.by.row=TRUE or merge.by.row=FALSE.
#> Model:  xgxr018 
#> Used tables, contents shown as used/total:
#>                  file     rows columns     IDs
#>       xgxr018_res.txt  905/905     6/6 150/150
#>  xgxr018_res_vols.txt  905/905     3/7 150/150
#>    xgxr018_res_fo.txt  150/150     1/2 150/150
#>     xgxr4.rds (input) 905/1502   21/23 150/150
#>              (result)      905    31+2     150
#> Input and output data combined by translation of
#> Nonmem data filters.
#> Distribution of rows on event types in returned data:
#>  EVID CMT output result
#>     0   2    755    755
#>     1   1    150    150
#>   All All    905    905

NMscanData tells that it has read a model called xgxr018 and how output and input data were combined. We shall see how these properties can be modified in a bit. Then follows an overview of how much data is used from the data files that were read. It used

  • one output data file (based on the $TABLE section(s) in the .lst file) from which it used all 905 rows and all 15 column, totaling 150 different values of ID.
  • the input file, but only 905 out of 1502 rows and 22 out of 23 columns. Input data did not contain any ID’s that weren’t used.

In the resulting data, 755 out of the 905 rows are EVID==0, the remaining 150 rows are EVID==1.

Let’s take a quick look at key properties of the data that was returned. It’s a data.frame with the additional NMdata class (for now, we just use it as a data.frame).

#> [1] "NMdata"     "data.frame"
#> [1] 905  33

The data used for the example is a PK single ascending dose data set, great thanks to the xgxr package authors.

The obtained dataset contains both model predictions (i.e. from output tables) and a character variable, trtact (i.e. from input data). To the .lst (output control stream) file path was supplied by us.

#> 1 31       0    0    1   1   3  0    0     1 0.1812 2307400    0   0    0 0.042
#> 2 32       0    0    1   1   3  0    0     1 0.1812 2307400    0   0    0 0.042
#> 1 0.1785   0     1    3    1       1        0  87.031 56.461 0.7245691     mg
#> 2 0.1785   0     1    3    1       1        0 100.620 45.096 0.7245691     mg
#>     NAME TIMEUNIT TRTACT   flag trtact   model nmout
#> 1 Dosing    Hours   3 mg Dosing   3 mg xgxr018  TRUE
#> 2 Dosing    Hours   3 mg Dosing   3 mg xgxr018  TRUE

You may have noticed that when reading the model, we were told that 37 columns were read while 39 columns are found in the result. The reason is the last two columns added by NMscanData called model and nmout. model obviously contains the name of the model which is by default derived from the list file name. See later in the “Recover rows” section what nmout represents.

What will NMscanData return?

Column in output data can overlap, and data can be available in both output and input data. The following main principles are followed by NMscanData:

  • Priority of data
    • Output data prevails over input data
    • Row-specific output data is preferred over ID-level (FIRSTONLY or LASTONLY) tables
    • Output tables are prioritized by their order of appearance (order of $TABLE sections)
    • The primary aim is to return the output data. If input and output cannot be meaningfully combined (very rare), output will be returned.
  • Properties of returned data if input data used (use.input=TRUE)
    • Input data column names will be default be returned as defined in the $INPUT section in Nonmem.
    • Columns that are dropped in Nonmem (DROP or SKIP) are included by NMscanData.
    • Columns that are not included in $INPUT are named as in the input data file.
    • If rows are being recovered from input data (the recover.rows argument), no information from output is merged onto these rows.

Once you have data from NMscanData, NMinfo can be used to browse meta information on what data was combined and how that was done.

Use a unique row identifier

Above, we were told that “Input and output data combined by translation of Nonmem data filters (not recommended).” Because of the very commonly used ACCEPT and IGNORE statements in Nonmem $DATA sections, the rows in output tables are often a subset of the input data rows. If no other information is available, NMscanData reads and interprets the ACCEPT or IGNORE statements and applies them to the input data before combining with the output data.

A more robust approach is using a unique row identifier in both input data and output data. NMscanData can use this for merging the data. This means that the ACCEPT or IGNORE are not interpreted at all. Even though NMscanData should work even without, it is always recommended to always include a unique row identifier in both input and output tables (in fact, we just need it in one full-length output table).

The following model happens to have such a unique row identifier in the column called ROW. The default NMscanData behavior is to use the row identifier if it can find it. The name of the column with the row identifier can be supplied using the col.row argument (and the default can be changed using the NMdataConf function). The default is to look for ROW.

All features shown below will work whether you supply col.row or not. We use col.row because it is more robust and because it allows us to easily trace a row in the analysis back to the source data. We are now told that the data was merged by ROW - that’s better.

res1.tbl <- NMscanData(file.NMdata("xgxr003.lst"),as.fun=tibble::as_tibble)
#> Model:  xgxr003 
#> Used tables, contents shown as used/total:
#>                  file     rows columns     IDs
#>       xgxr003_res.txt  905/905     7/7 150/150
#>  xgxr003_res_vols.txt  905/905     3/7 150/150
#>    xgxr003_res_fo.txt  150/150     1/2 150/150
#>     xgxr1.csv (input) 905/1502   21/24 150/150
#>              (result)      905    32+2     150
#> Input and output data merged by: ROW 
#> Distribution of rows on event types in returned data:
#>  EVID CMT output result
#>     0   2    755    755
#>     1   1    150    150
#>   All All    905    905

Get your preferred data class back from NMscanData

When reading res1.tbl, we also added the as.fun argument. the “as.” refers to as_tibble, as.data.frame, as.data.table etc. - a function applied to the data before it’s returned by NMscanData (or any other NMdata function). So now we have a tibble:

#> [1] "NMdata"     "tbl_df"     "tbl"        "data.frame"

I happen to be a data.table user so I am more comfortable working that way. Instead of using the as.fun all the time, we will change the default behavior using the NMdataConf function. Because NMdata is implemented in data.table we don’t need to pass the data.table::as.data.table function but we can (better) use the string "data.table" (again, data.table is the exception - for anything else, please pass a function):


Notice, NMdataConf will set the default value for all NMdata functions that use that argument. So when setting as.fun this way, we will get the desired class returned from all data generating NMdata functions.

We don’t want the same information about the dimensions repeated, so we use the quiet argument this time.

res1.dt <- NMscanData(file.NMdata("xgxr003.lst"),quiet=TRUE)

As expected we got a data.table this time:

#> [1] "NMdata"     "data.table" "data.frame"

Browse the metadata

An NMdata object returned by NMscanData comes with meta information about when and how what was read, and how the data was combined. The NMinfo function browses this information, and three options are available. It provides three sections of meta data:

  • “details”: A list including the function call, what options were effective (if input was included, rows recovered, if data was merged by a row identifier or combined by filters etc).

  • “tables”: Overview of the tables that were read and combined by NMscanData and properties of the different tables.

  • “columns”: Information on the columns that were treated by NMscanData (see example below).

The following shows the “columns” information as example. Remember, we are still getting a data.table because we used NMdataConf to change the configuration. We use the data.table print function to only look at first and last ten rows.

#>     variable                 file     source  level COLNUM
#>       <char>               <char>     <char> <char>  <int>
#>  1:       ID xgxr018_res_vols.txt     output    row      1
#>  2:  NOMTIME            xgxr4.rds      input    row      2
#>  3:     TIME            xgxr4.rds      input    row      3
#>  4:     EVID            xgxr4.rds      input    row      4
#>  5:      CMT            xgxr4.rds      input    row      5
#>  6:      AMT            xgxr4.rds      input    row      6
#>  7:       DV      xgxr018_res.txt     output    row      7
#>  8:     FLAG            xgxr4.rds      input    row      8
#>  9:    STUDY            xgxr4.rds      input    row      9
#> 10:       KA      xgxr018_res.txt     output    row     10
#> ---                                                       
#> 31:   trtact            xgxr4.rds      input    row     31
#> 32:    model                 <NA> NMscanData  model     32
#> 33:    nmout                 <NA> NMscanData    row     33
#> 34:       DV xgxr018_res_vols.txt     output    row     NA
#> 35:     PRED xgxr018_res_vols.txt     output    row     NA
#> 36:      RES xgxr018_res_vols.txt     output    row     NA
#> 37:     WRES xgxr018_res_vols.txt     output    row     NA
#> 38:       ID            xgxr4.rds      input    row     NA
#> 39:       DV            xgxr4.rds      input    row     NA
#> 40:       ID   xgxr018_res_fo.txt     output     id     NA

The column names are sorted by the order in the resulting dataset, the order given by the COLNUM column. The variables in the bottom that have COLNUM==NA were redundant when combining the data (the same columns were included from other sources). The file names and their source (input/output) and a “level” are given. “level” is the information level of the source. Input data and full-length output tables are “row” level, a firstonly or lastonly table is id-level. And then there is the model column added by NMscanData which is obviously model-level. nmout is the other column added by NMscanData and both model and nmout have NA file and NMscanData as source.

More options and features

Let’s have a quick look at the data we got back. The following is done with data.table. The comments in the code should make it clear what happens if you are not familiar with data.table. You can do this with base::tapply, stats::aggregate, a combination of dplyr::group_by and dplyr::summarize, or whatever you prefer.

gmPRED is calculated for sample times only and represents the geometric mean of population prediction (PRED) by dose and nominal time.

## trtact is a character. Make it a factor with levels ordered by
## numerical dose level. The := is a data.table assignment within
## res3. In dplyr, you could use mutate.
## Derive geometric mean pop predictions by treatment and nominal
## sample time. In dplyr, use group_by, summarize, and ifelse?

Notice, how little data is shown on the small doses. Remember, only 905 of the 1502 rows in the input data were used? Most of the rows excluded in the analysis are so due to observation being below the quantification limit (BLQ). The next section shows how to recover all the input data rows with NMscanData.

Recover rows

We may want to include the input data that was ignored by Nonmem. Use recover.rows=TRUE to include all rows from input data.

res2 <- NMscanData(file.NMdata("xgxr014.lst"),recover.rows=TRUE)
#> Model:  xgxr014 
#> Used tables, contents shown as used/total:
#>               file      rows columns     IDs
#>    xgxr014_res.txt   905/905   12/12 150/150
#>  xgxr2.rds (input) 1502/1502   22/24 150/150
#>           (result)      1502    34+2     150
#> Input and output data merged by: ROW 
#> Distribution of rows on event types in returned data:
#>  EVID CMT input-only output result
#>     0   1          2      0      2
#>     0   2        595    755   1350
#>     1   1          0    150    150
#>   All All        597    905   1502

Besides the model column holding the model name, NMscanData creates one other column by default. nmout is a boolean column created by NMscanData expressing whether each row was in the output data (nmout==TRUE) or they were recovered from the input data (nmout==FALSE).

We recognize these numbers from the message from NMscanData - the number of rows in output (905) and number of rows from input only (597). Since we changed the default value of as.fun with NMdataConf, res2 is a data.table.

#>     nmout     N
#>    <lgcl> <int>
#> 1:   TRUE   905
#> 2:  FALSE   597

We make use of the nmout column to only calculate gmPRED for observations (EVID==0) processed by Nonmem.

## add geometric mean pop predictions by treatment and nominal sample
## time. Only use sample records.

Obviously, we were lucky that meaningful values were assigned to DV for the BLQ and pre-dose samples in input data, so we in this case could easily plot all the data.

Combine multiple models

NMscanData by default adds a column called model for convenience when working with multiple models. You can specify both column name (which is by model default) and model name (contents of that column) as arguments in NMscanData. Using NMdataConf, You can also configure the default column name and the function that generates the model name.

The default is to derive the model name from the lst file name (say,

xgxr001.lst becomes xgxr001). In the following we use this to compare population predictions from two different models. We read them again just to show the use of the argument to name the models ourselves. Remember, we configure NMdata’s as.fun option so we are working with data.table and we easily stack with rbind (rbind.data.table) filling in NA’s. We add a couple of options to specify how input and output data are to be combined.

NMdataConf(as.fun="data.table", ## already set above, repeated for completeness
           col.row="ROW",       ## This is default, included for completeness
           merge.by.row=TRUE    ## Require input and output data to be combined by merge
res1.m <- NMscanData(system.file("examples/nonmem/xgxr001.lst", package="NMdata"),
## using a custom modelname for this model
res2.m <- NMscanData(system.file("examples/nonmem/xgxr014.lst", package="NMdata"),
                     modelname="One compartment",
## notice fill is an option to rbind with data.table (like bind_rows in dplyr)
res.mult <- rbind(res1.m,res2.m,fill=T)
## Notice, the NMdata class disappeared
#> [1] "data.table" "data.frame"

In this, we specifically wanted to rename one model for illustration of the modelname argument. We can pass a function to customize how NMscanData derives it from the list file path. This one skips the characters and leading zeros, so we just get an integer. We could pass use the modelname argument in NMdata but why not changed the default instead?

namefun <- function(path) sub("^[[:alpha:]0]+","",fnExtension(basename(path),""))
res1.m <- NMscanData(system.file("examples/nonmem/xgxr001.lst", package="NMdata"),
res2.m <- NMscanData(system.file("examples/nonmem/xgxr014.lst", package="NMdata"),
## notice fill is an option to rbind with data.table (like bind_rows in dplyr)
res.mult <- rbind(res1.m,res2.m,fill=T)
#>     model     N
#>    <char> <int>
#> 1:      1   905
#> 2:     14   905
## resetting default

NMdataConf can be used to change a lot of the default behaviour of the functions in NMdata so it fits in with your current setup and preferred work flow.

Preserve all input data properties - use rds

Return to the example above creating the dataset res22. Notice in the list of tables in the message from NMscanData, that input data was a .rds file. This is why we could sort the plots correctly on the dose level without reordering the factor levels first.

#> [1] "factor"
#> [1] "Placebo" "3 mg"    "10 mg"   "30 mg"   "100 mg"  "300 mg"

If the argument use.rds is TRUE, NMscanData will look for an rds file next to the input data file (which is a delimited text file) the exact same name as the text file except the extension must be .rds rather than say .csv (for Nonmem and NMscanData, the extension of the delimited text file doesn’t matter). If it finds the rds file, this will be used instead. No checks are done of whether the contents are similar in any way to the delimited text file which is ignored in this case.

There are three advantages of using rds files:

  • All attributes are kept. This includes column classes and factor levels.
  • Reading speed may be improved (NMdata uses fread from data.table which is extremely fast for delimited files so in many cases this difference can be small).
  • File sizes are greatly reduced from text to rds. This can be a big advantage if you are transferring files or reading over a network connection. NMdata is generally very fast (thanks to data.table) so file/network access (I/O) is likely to be the main bottleneck.

If you write Nonmem datasets with the NMdata::NMwriteData, you can get an rds file automatically, exactly where NMscanData will look for it. Preparing datasets using NMdata is described in this vignette.

You probably want to use NMdataConf to change the default behavior if you don’t want to use rds files.

Reading data without automatically combining it - the NMscaData building blocks

Each of the steps involved in reading and combining the data from a model run can be done separately.

The lst file was scanned for output tables, and they were all read (including interpreting the possible firstonly option). The input data has been used based on the $DATA and $INPUT sections of the control stream. The key steps in this process are available as independent functions.

  • NMreadTab: Read an Nonmem output table based on the path to the output table file.

  • NMscanTables: Read all output data files defined in a Nonmem run. Return a list of tables (as data.frames or data.tables).

  • NMtransInput: Read input data based on a Nonmem file. Data will be processed and named like the Nonmem model. ACCEPT and IGNORE filters can be applied as well. There are a few limitations to this functionality at this point. More about them below.

What should I do for my models to be compatible with NMscanData?

The answer to this should be as close to “nothing” as possible - that’s more or less the aim of the function. You just have to make sure that the information that you need is present in input data and output data. No need to output information that is unchanged from input, but make sure to output what you need (like IPRED, CWRES, CL, ETA1 etc which cannot be found in input). Some of these values can be found from other files generated by Nonmem but notice: NMscanData uses only input and output data.

It is recommended to always use a unique row identifier in both input and output data. This is the most robust way to merge back with input data. In firstonly tables, include the subject ID. Again, everything will most likely work even if you don’t, I personally don’t like relying on “most likely” when I can just as well have robustness.


Even if there are a few limitations to what models NMscanData can handle, there is a good chance you will never run into any of them, as they are mostly quite rare. If you do, reach out to me, and we’ll figure it out.

Input data file must exist and be unmodified since model run

If merging with input data, the input data must be available as was when the model was run. If you want to avoid this potential issue, Nonmem can be run in a wrapper script that either copies the input data, or runs NMscanData and saves the output in a compressed file format (like rds or zip).

Not all data filter statements implemented

Nested ACCEPT and IGNORE statements are not supported at this point. The resulting number of rows after applying filters is checked against row-level output table dimensions (if any available). In other words, you have to be unlucky to run into trouble without an error. But it is always recommended to use a unique row identifier in both input and output tables in order to avoid relying on interpretation of Nonmem code.

The RECORDS and NULL options in $DATA are not implemented. If using RECORDS, please use the col.row option to merge by a unique row identifier.

Character time variables not interpreted

Nonmem supports a clocktime input format for a column called TIME in input data. Based on a day counter and a character (“00:00”) clock format, Nonmem (or rather, NM-TRAN) can calculate the individual time since first record. This behaviour is not mimicked by NMscanData, and the only ways to get TIME in this case are to either include it in an output TABLE or to code the translation yourself after calling NMscanData. Of course, this is on the todo list.

Some TABLE options not supported

For now, only output tables returning either all rows or one row per subject can be merged with input. Tables written with options like FIRSTLASTONLY (two rows per subject) and OBSONLY are disregarded with a warning (you can read them with NMscanTables). LASTONLY is treated like FIRSTONLY, i.e. as ID-level information if not available elsewhere.


In this vignette you should have learned to

  • NMscanData can automatically read and combine all output and input data, only based on the path to the list (.lst) file
    • Output data is prioritized over input data
  • Switch between combining output and input data by mimicking the Nonmem data filters (IGNORE/ACCEPT) and merging by a row identifier
    • merge.by.row is the argument of interest
  • Configure NMdata to return your favorite data class
    • NMdataConf(as.fun="data.table") for data.table
    • NMdataConf(as.fun=tibble::as_tibble) for tibbles (tbl)
  • Look up meta data on input and output tables, how they were combined, and results of checks performed by NMscanData
    • Use NMinfo on the result coming out of NMscanData
  • Include input data rows that were not processed by Nonmem
    • recover.rows=TRUE
  • Combining such data sets for of a multiple models
    • By default, the column called model will hold the model name which you can use when combining (rbind) multiple model data sets
    • Use the modelname option to change the model name or how the model name is derived from the list file path.
  • Use an rds file to preserve all input data

You should have seen that NMscanData have very little limitations in what Nonmem models it can read. You should not have to change anything in the way you work to make use of NMscanData.