ddsjoberg / gtsummary Goto Github PK

Should you be able to provide an probability estimate and get a time point back? For example provide median survival and get the time that it occurred?

Previous note from Emily: For example, model results for OS and PFS could be placed side by side

Complete the following:

• Add checks for the input
• Add sorting type to .$meta_analysis
• Sort the levels by requested type

mod1 <-   glm(vs ~ cyl,family = binomial, data =mtcars ) %>% 
          tbl_regression()


mod2 <-   glm(mpg ~ cyl,family = gaussian, data =mtcars ) %>% 
  tbl_regression()

first model runs fine second does not

Look into supporting this feature.

library(survival)
library(gtsummary)
library(dplyr)
lung2 <- mutate(lung, s2 = ifelse(sex == 1, "1\2", "2/3"))
tbl_survival(lung2, Surv(time,status)~s2, times = c(500,540))

If you look at the output you will see that the function does not print "1\2" but prints 1 and a special character.

Add message to knit_print functions. If the output type is word, inform users the correct output type is rtf_document

Decide on inputs to tbl_uvregression. Currently, the method and outcome are being passed as strings, but perhaps non-standard evaluation is more consistent with other functions? Check out the geom_smooth function and match it?

make every function that takes in a tbl_summary, tbl_regression, or tbl_uvregression object an S3 generic function. in the help file for each of these functions, we'll list every subsequent function that can be called

Each {gtsummary} function that assists in creating/modifying tables added {gt} code to a list called gt_calls. This list contains one command per entry. For example:

> t = 
+   trial %>%
+   tbl_summary()
> t$gt_calls[1:2]
$gt
[1] "gt(data = x$table_body)"

$cols_label_label
[1] "cols_label(label = md('**Characteristic**'))"

Each command is stored as a string, but think this is not best practices. Probably, each of the commands should be stored as a quoted expression using quote() or rlang::expr()? Look into updating this.

I wanted to see what would happen if you ran a model with all levels and then removed one of the levels in a separate model and then used tbl_merge to look at both models. Both the missing level and the reference group get a dashed line which makes it difficult to determine which one is the reference and which is the missing level. Should the missing level just be blank instead of dashed?

library(survival)
library(gtsummary)
library(dplyr)


lung2 <- mutate(lung, stat = ifelse(status==2,0,1),
                ph.ecog = as.character(ph.ecog))

moda <-   glm(stat ~ ph.ecog,family = binomial, data =lung2) %>% 
          tbl_regression()

lung3 <- lung2 %>% 
         filter(ph.ecog != 3) %>% 
         mutate(ph.ecog = as.character(ph.ecog))

modb <-   glm(stat ~ ph.ecog,family = binomial, data =lung3) %>% 
  tbl_regression()

gtsummary::tbl_merge(tbls = list(moda,modb), tab_spanner = c("1", "s"))

The entire dichotomous value portion needs to be revisited.

• the user should be able to pass a level that will be presented
• the function makes assumptions about which levels should be presented (which is fine), but sometimes these are not observed in the data set, and the entire variable is silently omitted from the output.
• Variables that should be assumed to be dichotomous and their value to present
  • logical, TRUE
  • yes, no factors, yes
  • yes, no characters, yes
  • 0/1 numeric, 1

1. In gt, the fmt_* prefix is used for formatting columns.
2. We need to change the names of fmt_table1, fmt_regression, and fmt_uni_regression.
3. One suggestion is tbl_summary, tbl_regression, and tbl_uregression.
4. I've never liked fmt_uni_regression and don't really like tbl_uregression either.

1. The fmt_* functions in gt format columns. We need to rename fmt_percent, fmt_pvalue, and fmt_beta. Perhaps we can call them str_percent, etc., as they convert items to strings? But str_ is already a common prefix to operate ON strings, not for making strings.
2. I don't think it is possible to use fmt_* functions from gt to format the summary statistics. I am not sure what the best method is.

• One possibility is to keep the code as it is and pass an integer in the digits argument (i.e. digits = list(age = 1). But this assumes we want every statistic rounded to the same number of places.
• We can generalize the above option slightly to accept arguments like digits = list(age = c(0, 1)). If the statistics being presented were "{mean} ({sd})", the mean would be rounded to the nearest integer and the SD to one decimal place.
• Or maybe there is another option that allows for any kind of formatting (scientific, currency, dates, etc.)

Imported biostatR issue

If you use a long dataset for analysis of time-dependent covariates and pass the model results to fmt_regression, the sample size is the number of rows in the dataset rather than the number of unique observations in the data.

I would like users to be able to pass literal strings (e.g. stat_overall = "Overall (N = {N})") as well as interpreted strings, like stat_overall = md("Overall (N = {N})")

The function currently allows you to merge two regression models that don't have the same model covariates. The results look good, but not perfect.

library(survival)
t1 <-
  glm(response ~ marker + grade + age, trial, family = binomial) %>%
  tbl_regression(
    label = list(trt = "Treatment", grade = "Grade", age = "Age"),
    exponentiate = TRUE
  )
t2 <-
  coxph(Surv(ttdeath, death) ~ stage + age, trial) %>%
  tbl_regression(
    label = list(trt = "Treatment", grade = "Grade", age = "Age"),
    exponentiate = TRUE
  )
t3 <-
  tbl_merge(
    tbls = list(t1, t2),
    tab_spanner = c("Tumor Response", "Time to Death")
  )
t3

The dashes being added are due to the code that adds "---" for the reference groups of categorical variables. This means that "---" is being differentially added for missing variables depending on whether they are categorical.

I need to update the missing gt code to only add "---" if the variable is present in a particular model. All other blanks can merely by "".

What needs to get done for the initial conversion from biostatR to gtsummary.

Priority

• What prefix should we use for our primary functions? #2
  • I like tbl_. I think it can work.
• What should we name our primary functions?
  • fmt_table1 is tbl_summary
  • fmt_regression is tbl_regression
  • fmt_uni_regression is tbl_uvregression
• What prefix should we use for our numeric formatting functions? #3
  • fmt_pvalue is style_pvalue
  • fmt_percent is style_percent
  • fmt_beta is style_sigfig
• Convert fmt_table1 to tbl_summary utilizing gt
• Convert fmt_regression to tbl_regression utilizing gt
• Convert fmt_uni_regression to tbl_uvregression utilizing gt

New Functions

• Convert inline_text to work with updated function outputs. Need to discuss and decide on best input options.
  • Keep input as they were in biostatR, cell = "<var name>:<var level (if applicable)>:<by level (if applicable)>".
  • Change to variable = "<varname>", level = <var level (if applicable)>, by = <by level (if applicable)>. I think this is more transparent for the users compared to the first option implemented in biostatR. The variable and by levels are stored as strings in the tables. A user may pass level = 2, and in the background, we'd be looking for the level that equals "2". I don't think it should be an issue, however.
    • These constructs don't allow for the user to select the p-value, q-value, or n (anything else that comes later). Perhaps we could change the by = to column = . The user could pass the name of the column they want to select (e.g. column = "stat_1" or column = "pvalue"). We can create a helper function that converts by variable levels to the appropriate column name, and users could pass something like this, column = level("Drug"), which would be equivalent column = "stat_1"
• Convert bold_labels, bold_levels, italicize_labels, and italicize_levels to gt compatible output
• Convert bold_p to gt compatible output
• Convert add_global to gt compatible output
• Convert add_q to gt compatible output
• Convert add_comparison to gt compatible output
• Convert add_overall to gt compatible output
• Convert add_n to gt compatible output
• Convert add_overall to gt compatible output
• Write as_gt function to convert output into gt objects
• Write S3 print and knit_print method functions

Function Updates

• Convert tbl_summary vignette
  • Largely complete. Needs editing and the inline_text() section when that function is ready
• Convert tbl_regression vignette
  • Rough draft complete. Model outline after the tbl_summary() vignette
• Add message to knit_print functions. If the output type is word, inform users the correct output type is rtf_document
• Decide on inputs to tbl_uvregression. Currently, the method and outcome are being passed as strings, but perhaps non-standard evaluation is more consistent with other functions? Check out the geom_smooth function and match it?
• Update cols_label_summary function that modify headers for tbl_summary
  • The function still needs some work. I would like users to be able to pass literal strings (e.g. stat_overall = "Overall (N = {N})") as well as interpreted strings, like stat_overall = md("**Overall** (N = {N})")
  • Look into adding line breaks in the column label and test for all output types. Perhaps we can make a line break function that adds appropriate line breaks depending on output type.
• Update add_global.tbl_uvregression to pass the dots to car::Anova
• Re-write all package functions and utility functions not using deprecated tidyverse functions
• Load gt package when gtsummary is loaded. I added gt to Depends in DESCRIPTION and it does indeed load. But the messages are not as cute as when tidyverse loads packages. Perhaps update?
• Add smart default headers for tbl_regression, e.g. display "Odds Ratio" in column header, when logistic regression model is passed
• Add default footnote to add_comparison indicating tests performed.
• Add default footnote to tbl_summary denoting which statistics are being displayed
• Move all gt formatting calls to their own file. Save the calls for each function in separate lists. Or move the the list to the bottom of the function in same file
  • Only for the tbl_* functions, the gt_calls were moved to below the function. The other helper functions' reference to gt code wasn't as intrusive
• Add gt code to include --- for reference groups in tbl_regression
• Add gt code to include --- for reference groups in tbl_uvregression
• Include default headers for tbl_summary

complete tbl_summary vignette

Previous note from Emily: x-year estimates and confidence intervals from Kaplan-Meier, either overall or by a predictor

This will allow people to get cumulative incidence for competing risks for each event.

library(survival)
library(gtsummary)
library(dplyr)
lung2 <- mutate(lung, s2 = ifelse(sex == 1, "1\2", "2/3"))

tbl_survival(lung2, Surv(time,status)~sex,times = c(500,550)
             ,level_label = "male female"
             )

How does this work because when you try to add your own level labels it removes the subheadings?

Some arguments for tbl_summary() take named lists where each name is a variable in the dataset, and others take named lists where the names are continuous or categorical. It would be preferable if these naming conventions were consistent. Each list can take in a named list of variable names, and ..continuous.., ..categorical.. will be reserved names and make the modification for all variable of that type.

I would like to write a function that knows the environment it's being run in a creates the appropriate line break. Want to use this the column labels, so the N=200 is on a second line.

When missing values are printed in tbl_summary objects, the text display is "Unknown".

@karissawhiting brought up some time ago the possibility of changing the default text.

Option added in 9c730f0

this is a more descriptive name for what the function does

The {gt} package is renaming one of the functions used in {gtsummary}. (from cell_styles() to cell_text()) rstudio/gt#258

When this update goes through, we'll need to update our code. As of this moment, the cell_styles() function is used in:
The bolding/italicizing of labels/levels functions,
tab_style_bold_p(),
tbl_merge(),
tbl_regression(),
tbl_summary(),
tbl_uvregression()

Consider changing the re test in add_comparison to lme4. Perhaps this will signal to the users exactly how the p-value is calculated, much like how t.test signals the t.test() function was used.

at the moment, the function accepts the literal column names from .$table_body. Secondarily, the function assesses whether the input matches one of the by variable levels and converts that to the appropriate column name.

I think there is probably a better way to pass these arguments.

Some regression model functions strip the labels when model.frame(x) is created (e.g. coxph does this for categorical variables, but not continuous variables). A simple fix to getting the labels would be to pass the data as an optional argument. If the label is stripped by model.frame() the function can secondarily look for the label in data.

I think the update will be as easy as change

var_label = map_chr(
  .data$variable, ~ 
    label[[.x]] %||% 
    attr(stats::model.frame(x)[[.x]], "label") %||%  
    .x
)

to

var_label = map_chr(
  .data$variable, ~ 
    label[[.x]] %||% 
    attr(stats::model.frame(x)[[.x]], "label") %||%  
    attr(data[[.x]], "label") %||%
    .x
)

where data is NULL by default

print message if add_global is used without first installing cars package. the cars package has many dependencies that most will not need.

Many deprecated functions were used to create the package. Re-write with the the non-deprecated versions using tidy evaluation.

Completed conversions:

• mutate_()
• filter_()
• group_by_()
• arrange_()
• count_()
• nest_()
• unnest_()
• spread_()
• gather_()
• complete_()

Re-write all package functions and utility functions not using deprecated tidyverse functions

The left column header becomes centered (rather than left justified) after adding a spanning header.

I think this may happen in {gt} rather than something I did in {gtsummary}?

library(gtsummary)
#> Loading required package: gt
trial %>%
  dplyr::select(grade, age, response, marker) %>%
  tbl_summary(by = "grade") %>%
  as_gt() %>%
  tab_spanner(label = "Tumor Grade", columns = starts_with("stat_"))

Update add_global.tbl_uvregression to pass the dots (...) to car::Anova

Add code to have headers like "OR" and "IRR" for logistic and Poisson regression models for random effects models, as we do for GLM models.

library(gtsummary)
library(lme4)
mod_glmer <-
  glmer(am ~ hp + (1 | gear), mtcars, family = binomial) %>%
  tbl_regression(exponentiate = TRUE)

mod_glm <-
  glm(am ~ hp, mtcars, family = binomial) %>%
  tbl_regression(exponentiate = TRUE)

The gt output tables don't display in the online help files.

Look at the gt package help pages to see who they display their examples. It looks like they are displaying image files.
Perhaps one solution is to add all the code to create the tables in the data-raw folder, and use code to save image of the tables. The help files will then import the images.

@karissawhiting putting the comments from #83 in this issue to keep track.

1. We should shorten the tbl_summary table examples (and perhaps others). Also, two examples should be sufficient (rather than the three we have). With fewer examples and shorter tables, the space between example code and the rest of the help file will be shorter.
2. Rather than call the section "Figures", perhaps we should call it "Example Output"? That may help people understand more easily that these are the output tables.
3. Is there a way to make a separation between two example tables?

What do you think?

First draft complete. Needs editing

By default the tbl_survival results are printed in long format (one line per stratum per time point). We want to add functionality for users to have the table print in wide format. Results can be made in two ways:

1. With each stratum variable as a column, or
2. With each time point as a column

We need to decide whether to incorporate these as options in tbl_survival or by making the modification via a new piped function. I am leaning towards making them options in the function.

We will need to return a long tibble in the results (default printing format) regardless of the chosen output format, so the inline_text() function identify the correct statistics to return.

Whatever we decide, it should be compatible with the yet to be complete function tbl_survival.cuminc (#64) AND with the survival quantile updates to tbl_survival.survfit (#61)

finish tbl_regression vignette

I added gt to Depends in DESCRIPTION and it does indeed load. But the messages are not as cute as when tidyverse loads packages. Perhaps update?