knitr::opts_chunk$set( collapse = TRUE, comment = "#>" )
library(tidyverse) library(magrittr) library(Tplyr) library(knitr)
When you look at a summary table within a clinical report, you can often break it down into some basic pieces. Consider this output.
Different variables are being summarized in chunks of the table, which we refer to as "layers". Additionally, this table really only contains a few different types of summaries, which makes many of the calculations rather redundant. This drives the motivation behind Tplyr. The containing table is encapsulated within the tplyr_table()
object, and each section, or "layer", within the summary table can be broken down into a tplyr_layer()
object.
tplyr_table()
ObjectThe tplyr_table()
object is the conceptual "table" that contains all of the logic necessary to construct and display the data. Tplyr tables are made up of one or more layers. Each layer contains an instruction for a summary to be performed. The tplyr_table()
object contains those layers, and the general data, metadata, and logic necessary to prepare the data before any layers are constructed.
When a tplyr_table()
is created, it will contain the following bindings:
target
- The dataset upon which summaries will be performedcount_layer_formats
- Default formats to be used on count layers in the tableshift_layer_formats
- Default formats to be used on shift layers in the tabledesc_layer_formats
- Default formats to be used on descriptive statistics layers in the tablepop_data
- The dataset containing population information. This defaults to the target
datasetcols
- A categorical variable in the target
dataset to present summaries grouped by column (in addition to the treat_var
variable)table_where
- The where
clause provided, used to subset the target
datasettreat_var
- Variable used to distinguish treatment groups in the target
dataset.header_n
- Default header N values based on treat_var
and any cols
variablespop_treat_var
- Variable used to distinguish treatment groups in pop_data
dataset (if different than the treat_var
variable in the target
dataset)layers
- The container for individual layers of a tplyr_table()
treat_grps
- Additional treatment groups to be added to the summary (i.e. Total)The function tplyr_table()
allows you a basic interface to instantiate the object. Modifier functions are available to change individual parameters catered to your analysis.
t <- tplyr_table(tplyr_adsl, TRT01P, where = SAFFL == "Y") t
tplyr_layer()
ObjectUsers of Tplyr interface with tplyr_layer()
objects using the group_<type>
family of functions. This family specifies the type of summary that is to be performed within a layer. count
layers are used to create summary counts of some discrete variable. shift
layers summarize the counts for different changes in states. Lastly, desc
layers create descriptive statistics.
vars(OutsideVariable, InsideVariable)
. This allows you to do tables like Adverse Events where you want to see the Preferred Terms within Body Systems, all in one layer. Count layers can also distinguish between distinct and non-distinct counts. Using some specified by variable, you can count the unique occurrences of some variable within the specified by grouping, including the target. This allows you to do a summary like unique subjects and their proportion experiencing some adverse event, and the number of total occurrences of that adverse event.set_format_strings()
, and you can also add your own summaries using set_custom_summaries()
. This allows you to easily implement any additional summary statistics you want presented.cnt <- group_count(t, AGEGR1) cnt dsc <- group_desc(t, AGE) dsc shf <- group_shift(t, vars(row=COMP8FL, column=COMP24FL)) shf
Everyone has their own style of coding - so we've tried to be flexible to an extent. Overall, Tplyr is built around tidy syntax, so all of our object construction supports piping with magrittr
(i.e. %>%
).
There are two ways to add layers to a tplyr_table()
: add_layer()
and add_layers()
. The difference is that add_layer()
allows you to construct the layer within the call to add_layer()
, whereas with add_layers()
you can attach multiple layers that have already been constructed upfront:
t <- tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_count(AGEGR1, by = "Age categories n (%)") )
Within add_layer()
, the syntax to constructing the count layer for Age Categories was written on the fly. add_layer()
is special in that it also allows you to use piping to use modifier functions on the layer being constructed
t <- tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_count(AGEGR1, by = "Age categories n (%)") %>% set_format_strings(f_str("xx (xx.x%)", n, pct)) %>% add_total_row() )
add_layers()
, on the other hand, lets you isolate the code to construct a particular layer if you wanted to separate things out more. Some might find this cleaner to work with if you have a large number of layers being constructed.
t <- tplyr_table(tplyr_adsl, TRT01P) l1 <- group_count(t, AGEGR1, by = "Age categories n (%)") l2 <- group_desc(t, AGE, by = "Age (years)") t <- add_layers(t, l1, l2)
Notice that when you construct the layers separately, you need to specify the table to which they belong. add_layer()
does this automatically. tplyr_table()
and tplyr_layer()
objects are built on environments, and the parent/child relationships are very important. This is why, even though the layer knows who its table parent is, the layers still need to be attached to the table (as the table doesn't know who its children are). Advanced R does a very good job at explaining what environments in R are, their benefits, and how to use them.
Notice that when you construct a tplyr_table()
or a tplyr_layer()
that what displays is a summary of information about the table or layer? That's because when you create these objects - it constructs the metadata, but does not process the actual data. This allows you to construct and make sure the pieces of your table fit together before you do the data processing - and it gives you a container to hold all of this metadata, and use it later if necessary.
To generate the data from a tplyr_table()
object, you use the function build()
:
t <- tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_count(AGEGR1, by = "Age categories n (%)") ) t %>% build() %>% kable()
But there's more you can get from Tplyr. It's great to have the formatted numbers, but what about the numeric data behind the scenes? Maybe a number looks suspicious and you need to investigate how you got that number. What if you want to calculate your own statistics based off of the counts? You can get that information as well using get_numeric_data()
. This returns the numeric data from each layer as a list of data frames:
get_numeric_data(t) %>% head() %>% kable()
By storing pertinent information, you can get more out of a Tplyr object than processed data for display. And by specifying when you want to get data out of Tplyr, we can save you from repeatedly processing data while your constructing your outputs - which is particularly useful when that computation starts taking time.
The bulk of Tplyr coding comes from constructing your layers and specifying the work you want to be done. Before we get into this, it's important to discuss how Tplyr handles string formatting.
String formatting in Tplyr is controlled by an object called an f_str()
, which is also the name of function you use to create these formats. To set these format strings into a tplyr_layer()
, you use the function set_format_strings()
, and this usage varies slightly between layer types (which is covered in other vignettes).
So - why is this object necessary. Consider this example:
t <- tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_desc(AGE, by = "Age (years)") %>% set_format_strings( 'n' = f_str('xx', n), 'Mean (SD)' = f_str('xx.xx (xx.xxx)', mean, sd) ) ) t %>% build() %>% kable()
In a perfect world, the f_str()
calls wouldn't be necessary - but in reality they allow us to infer a great deal of information from very few user inputs. In the calls that you see above:
row_label2
column are taken from the left side of each =
in set_format_strings()
f_str()
callf_str()
call determine the descriptive statistic summaries that will be performed. This is connected to a number of default summaries available within Tplyr, but you can also create your own summaries (covered in other vignettes). The default summaries that are built in include:n
= Number of observationsmean
= Meansd
= Standard Deviationvar
= Varianceiqr
= Inter Quartile Rangeq1
= 1st quartileq3
= 3rd quartilemin
= Minimum valuemax
= Maximum valuemissing
= Count of NA valuesf_str()
call, then those two summaries are formatted into the same string. This allows you to do a "Mean (SD)" type format where both numbers appear.This simple user input controls a significant amount of work in the back end of the data processing, and the f_str()
object allows that metadata to be collected.
f_str()
objects are also used with count layers as well to control the data presentation. Instead of specifying the summaries performed, you use n
, pct
, distinct_n
, and distinct_pct
for your parameters and specify how you would like the values displayed. Using distinct_n
and distinct_pct
should be combined with specifying a distinct_by()
variable using set_distinct_by()
.
tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_count(AGEGR1, by = "Age categories") %>% set_format_strings(f_str('xx (xx.x)',n,pct)) ) %>% build() %>% kable() tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_count(AGEGR1, by = "Age categories") %>% set_format_strings(f_str('xx',n)) ) %>% build() %>% kable()
Really - format strings allow you to present your data however you like.
tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_count(AGEGR1, by = "Age categories") %>% set_format_strings(f_str('xx (•◡•) xx.x%',n,pct)) ) %>% build() %>% kable()
But should you? Probably not.
As covered under string formatting, set_format_strings()
controls a great deal of what happens within a descriptive statistics layer. Note that there are some built in defaults to what's output:
tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_desc(AGE, by = "Age (years)") ) %>% build() %>% kable()
To override these defaults, just specify the summaries that you want to be performed using set_format_strings()
as described above. But what if Tplyr doesn't have a built in function to do the summary statistic that you want to see? Well - you can make your own! This is where set_custom_summaries()
comes into play. Let's say you want to derive a geometric mean.
tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_desc(AGE, by = "Sepal Length") %>% set_custom_summaries( geometric_mean = exp(sum(log(.var[.var > 0]), na.rm=TRUE) / length(.var)) ) %>% set_format_strings( 'Geometric Mean (SD)' = f_str('xx.xx (xx.xxx)', geometric_mean, sd) ) ) %>% build() %>% kable()
In set_custom_summaries()
, first you name the summary being performed. This is important - that name is what you use in the f_str()
call to incorporate it into a format. Next, you program or call the function desired. What happens in the background is that this is used in a call to dplyr::summarize()
- so use similar syntax. Use the variable name .var
in your custom summary function. This is necessary because it allows a generic variable name to be used when multiple target variables are specified - and therefore the function can be applied to both target variables.
Sometimes there's a need to present multiple variables summarized side by side. Tplyr allows you to do this as well.
tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_desc(vars(AGE, AVGDD), by = "Age and Avg. Daily Dose") ) %>% build() %>% kable()
Tplyr summarizes both variables and merges them together. This makes creating tables where you need to compare BASE, AVAL, and CHG next to each other nice and simple. Note the use of dplyr::vars()
- in any situation where you'd like to use multiple variable names in a parameter, use dplyr::vars()
to specify the variables. You can use text strings in the calls to dplyr::vars()
as well.
Count layers generally allow you to create "n" and "n (%)" count type summaries. There are a few extra features here as well. Let's say that you want a total row within your counts. This can be done with add_total_row()
:
tplyr_table(tplyr_adsl, TRT01P) %>% add_layer( group_count(AGEGR1, by = "Age categories") %>% add_total_row() ) %>% build() %>% kable()
Sometimes it's also necessary to count summaries based on distinct values. Tplyr allows you to do this as well with set_distinct_by()
:
tplyr_table(tplyr_adae, TRTA) %>% add_layer( group_count('Subjects with at least one adverse event') %>% set_distinct_by(USUBJID) %>% set_format_strings(f_str('xx', n)) ) %>% build() %>% kable()
There's another trick going on here - to create a summary with row label text like you see above, text strings can be used as the target variables. Here, we use this in combination with set_distinct_by()
to count distinct subjects.
Adverse event tables often call for counting AEs of something like a body system and counting actual events within that body system. Tplyr has means of making this simple for the user as well.
tplyr_table(tplyr_adae, TRTA) %>% add_layer( group_count(vars(AEBODSYS, AEDECOD)) ) %>% build() %>% head() %>% kable()
Here we again use dplyr::vars()
to specify multiple target variables. When used in a count layer, Tplyr knows automatically that the first variable is a grouping variable for the second variable, and counts shall be produced for both then merged together.
Lastly, let's talk about shift layers. A common example of this would be looking at a subject's lab levels at baseline versus some designated evaluation point. This would tell us, for example, how many subjects were high at baseline for a lab test vs. after an intervention has been introduced. The shift layer in Tplyr is intended for creating shift tables that show these data as a matrix, where one state will be presented in rows and the other in columns. Let's look at an example.
# Tplyr can use factor orders to dummy values and order presentation tplyr_adlb$ANRIND <- factor(tplyr_adlb$ANRIND, c("L", "N", "H")) tplyr_adlb$BNRIND <- factor(tplyr_adlb$BNRIND, c("L", "N", "H")) tplyr_table(tplyr_adlb, TRTA, where = PARAMCD == "CK") %>% add_layer( group_shift(vars(row=BNRIND, column=ANRIND), by=PARAM) %>% set_format_strings(f_str("xx (xxx%)", n, pct)) ) %>% build() %>% kable()
The underlying process of shift tables is the same as count layers - we're counting the number of occurrences of something by a set of grouping variables. This differs in that Tplyr uses the group_shift()
API to use the same basic interface as other tables, but translate your target variables into the row variable and the column variable. Furthermore, there is some enhanced control over how denominators should behave that is necessary for a shift layer.
There's quite a bit more to learn! And we've prepared a number of other vignettes to help you get what you need out of Tplyr.
vignette("table")
vignette("desc")
vignette("count")
vignette("shift")
vignette("riskdiff")
vignette("sort")
vignette("options")
vignette("styled-table")
In building Tplyr, we needed some additional resources in addition to our personal experience to help guide design. PHUSE has done some great work to create guidance for standard outputs with collaboration between multiple pharmaceutical companies and the FDA. You can find some of the resource that we referenced below.
Analysis and Displays Associated with Adverse Events
Analyses and Displays Associated with Demographics, Disposition, and Medications
Analyses and Displays Associated with Measures of Central Tendency
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.