CDISC Analysis Results Data with {cards} + {gtsummary}

蘇丹杰 (Daniel D. Sjoberg)

Introduction

Acknowledgements

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA4.0).

蘇丹杰 (Daniel D. Sjoberg)

• Senior Principal Data Scientist at Genentech/Roche

• 13+ years at Memorial Sloan Kettering Cancer Center as a Biostatistician

• Published 250+ papers in peer-reviewed journals and served on Editorial Board of European Urology

• Written and contributed to many R packages available on CRAN, R Universe, and GitHub.

danieldsjoberg.com

@ddsjoberg.bsky.social

social/@ddsjoberg

linkedin.com/in/ddsjoberg

github.com/ddsjoberg

Outline

• Analysis Results Standard (ARS)

• Analysis Results Data (ARD)

• {cards} + {cardx} R packages

• {gtsummary} R Package

Analysis Results Standard

Analysis Results Standard (ARS)

CDISC has clear guidelines on how to collect, clean, and prepare data for analysis.

But nothing about results!

The ARS is meant to fill that gap.

Analysis Results Standard (ARS)

• Emerging standard for prospectively encoding statistical analysis reporting pipeline in a machine-readable format.

• Primary objectives are to leverage analysis results metadata to drive the automation of results and support storage, access, processing, traceability and reproducibility of results.

• Logical model that describes analysis results and associated metadata.

• Focus on concepts, not layout, e.g. the summary statistics, not how the results are shown in a table.

• Learn more at https://www.cdisc.org/events/webinar/analysis-results-standard-public-review

Emphasis on the PROSPECTIVE!

• Automation

• Concepts, not layout

Analysis Results Standard (ARS)

Example ARS Flow

• Overview of the meta data

  • the data sets, the exclusions (e.g. ITT population)

  • I know the details here are sparse, but we need to get the meat of this presentation!

• Overview of ARD

  • Structured object to save summary results

  • Can be used for simply summary statistics or complex analyses

  • We can also save instructions formatting the results

• Lastly, TLGs

  • We’re all familiar with the TLGs, I presume.

  • The ARS fills the gap between ADaM and TLGs

Analysis Results Standard (ARS)

Example ARS Flow

• {cards} helps us easily build analysis results data sets

  • YES, THIS IS A DRAFT HEX STICKER

• after the re-factor is complete, gtsummary will help us build tables from ARDs

Analysis Results Data

Analysis Results Data (ARD)

• Encodes statistical analysis outcomes in a machine-readable format.

• Primary objective is to streamline the processes of automation, ensuring reproducibility, promoting reusability, and enhancing traceability.

• The ARD model specified how statistical results are saved into a structured format.

• The ARD can be used to to subsequently create tables and figures.

This is where I start getting interested!

The CDISC model is not over prescriptive how these results need to look.

We’ll review our general approach, which lends itself to the R language

Analysis Results Data (ARD)

• After the initial creation of an ARD, the results can later be re-used again and again for subsequent reporting needs.

• The same results can be used again and again.

  • Great for when we are switching context (eg CSR to a slideshow): you don’t need to re-calculate and re-check those calculations

• Re-creating a figure with only 2 of the 3 treatment arms,

• Pruning a large complex table to show only the most interesting results for a presentation (when the initial results may been prepared for regulatory purposes).

{cards}

{cards} R Package

Let’s check out a simple example

library(cards)

# create ARD with default summary statistics
ard_continuous(ADSL, variables = AGE)

{cards} data frame: 8 x 8

  variable   context stat_name stat_label   stat fmt_fn
1      AGE continuo…         N          N    254      0
2      AGE continuo…      mean       Mean 75.087      1
3      AGE continuo…        sd         SD  8.246      1
4      AGE continuo…    median     Median     77      1
5      AGE continuo…       p25  25th Per…     70      1
6      AGE continuo…       p75  75th Per…     81      1
7      AGE continuo…       min        Min     51      1
8      AGE continuo…       max        Max     89      1

ℹ 2 more variables: warning, error

• Best way to introduce is with a simple example

• one line per statistic

• includes default set of summary statistics

• Briefly review each column: stat names, stat labels, numeric results, fmt_fn

• a simple data frame result, which means it is easy to work with a modify

{cards}: ard_continuous() arguments

• by: summary statistics are calculated by all combinations of the by variables, including unobserved factor levels

• statistic: specify univariate summary statistics. Accepts any function, base R, from a package, or user-defined.

• fmt_fn: Override the default formatting functions, e.g. when you need

ADSL |> 
  ard_continuous(
    variables = AGE,
    by = ARM,                               # stats by treatment arm
    statistic = ~list(mean = \(x) mean(x)), # return the mean
    fmt_fn = ~list(mean = 0)                # format the result
  ) |> 
  apply_fmt_fn() # add a character column of rounded results

{cards} data frame: 3 x 11

  group1 group1_level variable stat_name stat_label   stat stat_fmt
1    ARM      Placebo      AGE      mean       Mean 75.209       75
2    ARM    Xanomeli…      AGE      mean       Mean 74.381       74
3    ARM    Xanomeli…      AGE      mean       Mean 75.667       76

ℹ 4 more variables: context, fmt_fn, warning, error

Using the by, statastic, and fmt_fn arguments

• fmt_fn uses integers as aliases, but can of course use a proper function

• note the new group1 and group1_level columns. NO PROTECTED NAMES

Also calling apply_fmt_fn() adds a column of formatted results

{cards}: ard_categorical()

ADSL |> 
  ard_categorical(
    by = ARM,
    variables = AGEGR1
  ) |> 
  dplyr::filter(stat_name %in% c("n", "p")) |> # keep most common stats 
  print(n = 8)

{cards} data frame: 18 x 11

  group1 group1_level variable variable_level stat_name stat_label  stat
1    ARM      Placebo   AGEGR1            <65         n          n    14
2    ARM      Placebo   AGEGR1            <65         p          % 0.163
3    ARM    Xanomeli…   AGEGR1            <65         n          n    11
4    ARM    Xanomeli…   AGEGR1            <65         p          % 0.131
5    ARM    Xanomeli…   AGEGR1            <65         n          n     8
6    ARM    Xanomeli…   AGEGR1            <65         p          % 0.095
7    ARM      Placebo   AGEGR1            >80         n          n    30
8    ARM      Placebo   AGEGR1            >80         p          % 0.349

ℹ 10 more rows

ℹ Use `print(n = ...)` to see more rows

ℹ 4 more variables: context, fmt_fn, warning, error

Any unobserved levels of the variables will be present in the resulting ARD.

Categorical summaries have a similar structure to the continuous

This time we have a new column variable_level

{cards}: Other Summary Functions

• ard_hierarchical(): similar to ard_categorical(), but built for nested tabulations, e.g. AE terms within SOC

• ard_dichotomous(): similar to ard_categorical(), and tabulates a single value of the variable

• ard_complex(): similar to ard_continuous(), but the summary functions can be more complex and accepts other arguments like the full and subsetted (within the by groups) data sets.

• ard_missing(): tabulates rates of missingness

The results from all these functions are entirely compatible with one another, and can be stacked into a single data frame.

{cards}: Other Functions

In addition to exporting functions to prepare summaries, {cards} exports many utilities for wrangling ARDs and creating new ARDs.

Constructing: bind_ard(), tidy_as_ard(), nest_for_ard(), check_ard_structure(), and many more

Wrangling: shuffle_ard(), get_ard_statistics(), replace_null_statistic(), etc.

• tidy_as_ard() is USED everywhere!

{cardx}

{cardx} R Package

• While {cards} performs basic (and very common) summaries, {cardx} exports ard_*() functions for more complex analytic results.

• The list is growing, but we have functions for t-tests, Wilcoxon tests, standardized mean differences, ANOVA (including repeated measures), survey methods, confidence intervals for proportions and centrality estimates, and more.

• Elegant solution for nearly every regression model type in the R ecosystem, where we can identify regression types (linear, logistic, Cox, etc.), identify the underlying variable names for categorical variables (different from the model terms), identify reference groups for categorical variables, and much much more.

Yes, this is also a draft hex sticker!

I could do an entire talk on the regression model solution. It’s amazing work with Joseph Larmarange.

{cardx}

Supported regression model types include:

betareg::betareg(), biglm::bigglm(), biglmm::bigglm(), brms::brm(), cmprsk::crr(), fixest::feglm(), fixest::femlm(), fixest::feNmlm(), fixest::feols(), gam::gam(), geepack::geeglm(), glmmTMB::glmmTMB(), lavaan::lavaan(), lfe::felm(), lme4::glmer.nb(), lme4::glmer(), lme4::lmer(), logitr::logitr(), MASS::glm.nb(), MASS::polr(), mgcv::gam(), mice::mira, mmrm::mmrm(), multgee::nomLORgee(), multgee::ordLORgee(), nnet::multinom(), ordinal::clm(), ordinal::clmm(), parsnip::model_fit, plm::plm(), pscl::hurdle(), pscl::zeroinfl(), rstanarm::stan_glm(), stats::aov(), stats::glm(), stats::lm(), stats::nls(), survey::svycoxph(), survey::svyglm(), survey::svyolr(), survival::cch(), survival::clogit(), survival::coxph(), survival::survreg(), tidycmprsk::crr(), VGAM::vglm().

{cardx} in Practice

In the example below, we’re adding a confidence interval around the rate of subjects who completed the study.

We’re using the default computation method (Wald) and confidence level (95%).

cards::ADSL |> 
  dplyr::mutate(COMPLETED = DCDECOD == "COMPLETED") |> 
  cardx::ard_proportion_ci(variables = COMPLETED) # using default CI method 'Wald'

{cards} data frame: 6 x 8

   variable   context  stat_name stat_label      stat fmt_fn
1 COMPLETED proporti…          N          N       254      0
2 COMPLETED proporti…   estimate   estimate     0.433      1
3 COMPLETED proporti…   conf.low   conf.low      0.37      1
4 COMPLETED proporti…  conf.high  conf.high     0.496      1
5 COMPLETED proporti… conf.level  conf.lev…      0.95      1
6 COMPLETED proporti…     method     method Wald Con…   <fn>

ℹ 2 more variables: warning, error

Notice anything strange about the stat column?

How do we have a character result mixed with numeric?

LIST COLUMNS! You could include a matrix result, an entire modeling object, anything.

{cardx} in Practice: Gone Wrong 😱

What happens when statistics are un-calculable?

It’s one of my favorite features of {cards} and {cardx}

ard_gone_wrong <- 
  cards::ADSL |> 
  cards::ard_continuous(
    variable = AGEGR1,
    statistic = ~list(kurtosis = \(x) e1071::kurtosis(x))
  )
ard_gone_wrong

{cards} data frame: 1 x 8

  variable stat_name stat_label stat   warning     error
1   AGEGR1  kurtosis   kurtosis      argument… non-nume…

ℹ 2 more variables: context, fmt_fn

print_ard_conditions(ard_gone_wrong)

• Where is the statistic? AGEGR1 is character

• Even when there are errors or warnings, we still get the ARD with the expected structure returned.

  • THIS IS BIG! There are MANY circumstances, when you are designing TLGs early in a study when you do not have all the data required to calculate every statistic.

  • This allows you to design everything up-front.

• We can also report these warnings and errors back to users. <!CLICK!>

{cards} + {cardx}

• These packages together provide robust set of tools for preparing ARDs.

• Incredibly easy to extend them for your own specific cases.

• But what’s next? I want a pretty table!

{gtsummary}

{gtsummary} Background

• Born from my time as a biostatistician at Memorial Sloan Kettering Cancer Center.

• I led a team of other statisticians and I wanted us to transition to R.

• Began writing a package that eventually became {gtsummary} to help make this transition smooth.

• Since then the package has grown and grown, and is now the most downloaded summary table package on CRAN.

• Package the American Statistical Association’s (ASA) 2021 award for Innovation in Statistical Programming and Analytics

{gtsummary} R Package

• gtsummary tables are composable, that leads to a wonderful user experience

• no more needing to memorize an enormous function with too many arguments

• Importantly, INCLUDES A GRAMMER FOR merging AND stacking. This is huge for creating bespoke tables.

{gtsummary} R Package

• a gtsummary table can be converted into any of these types, and styling and formatting functions from those packages can be used to continue to make the table exactly as you’d like.

• each of these engines have their own strengths

• there is a lot more I could share about the package: I’ve never given a three-slide summary of its features before!

{cards} + {gtsummary}

How is this related to ARD and the {cards} package? 🤔🤔

• The {cards} package does not present results and this is where the {gtsummary} package shines.

• The {gtsummary} package is currently being refactored with a {cards} backend.

• In addition to a refactor, adding new features making it easier to create common pharma tables.

{cards} + {gtsummary}

• Integrating with the most popular summary tabling package in the R ecosystem is BIG!

• With a huge user base, you can easily find solutions for your esoteric questions

  • Questions that have likely alreday been answered on stackoverflow.

Stay Tuned!

• Expect new releases of {cards}, {cardx}, and {gtsummary} soon!

• These packages working together will serve as a strong combination for ARD-first TLG creation in the pharmaceutical space.

• And I will get to these hex logos too!

• If you have suggestions, please join the conversation on GitHub

• CLICK

• Any questions??

Mock Tables

• With {cards}+{gtsummary} it’s easy to create bespoke table shells.

cards::ADSL |> 
  cards::ard_continuous(
    variables = AGE, 
    fmt_fn = ~list(everything() ~ \(x) "xx")
  ) |> 
  cards::apply_fmt_fn() |> 
  print(n = 3)

  variable   context stat_name stat_label   stat stat_fmt
1      AGE continuo…         N          N    254       xx
2      AGE continuo…      mean       Mean 75.087       xx
3      AGE continuo…        sd         SD  8.246       xx

• Pass this ARD to card_summary() and the table will be populated with "xx" placeholders.

Mock Tables

Characteristic	Placebo N = xx	Xanomeline Low Dose N = xx	Xanomeline High Dose N = xx
Age
Median (IQR)	xx (xx, xx)	xx (xx, xx)	xx (xx, xx)
Mean (SD)	xx (xx)	xx (xx)	xx (xx)
Range	xx - xx	xx - xx	xx - xx
Age Group, n (%)
<65	xx (xx.x%)	xx (xx.x%)	xx (xx.x%)
65-80	xx (xx.x%)	xx (xx.x%)	xx (xx.x%)
>80	xx (xx.x%)	xx (xx.x%)	xx (xx.x%)