Using the table1 Package to Create HMTL Tables of Descriptive Statistics

Introduction

It is standard practice in epidemiology and related fields that the first table of any journal article, referred to as “Table 1”, is a table that presents descriptive statistics of baseline characteristics of the study population stratified by exposure. This package makes it fairly straightforward to produce such a table using R. The output format is HTML (which has the advantage of being easy to copy into a Word document; Chrome browser works well). The preferred way of using this package is in combination with knitr (for HTML generation; use R markdown or the spin format, for instance), making use of the results="asis" chunk option to pass the HTML output through untouched (this vignette is a good example). The package does allow quite a bit of flexibility to customize the table’s contents and appearance, but this does come at the cost of ease-of-use (more programming, some knowledge of CSS).

Example 1

The first example is inspired by this blog post, which is about how to accomplish a similar task using the htmlTable package. It uses the melanoma data set from the boot package for illustration, and I have copied here the code used to prepare the data:

library(boot) 

melanoma2 <- melanoma
 
# Factor the basic variables that
# we're interested in
melanoma2$status <- 
  factor(melanoma2$status, 
         levels=c(2,1,3),
         labels=c("Alive", # Reference
                  "Melanoma death", 
                  "Non-melanoma death"))

As I first attempt, we can do the following:

table1(~ factor(sex) + age + factor(ulcer) + thickness | status, data=melanoma2)

Note that the table1 package uses a familiar formula interface, where the variables to include in the table are separated by ‘+’ symbols, the “stratification” variable (which creates the columns) appears to the right of a “conditioning” symbol ‘|’, and the data argument specifies a data.frame that contains the variables in the formula.

But because we don’t have nice labels for the variables and categories, it doesn’t look great. To improve things, we can create factors with descriptive labels for the categorical variables (sex and ulcer), label each variable the way we want, and specify units for the continuous variables (age and thickness), like this:

melanoma2$sex <- 
  factor(melanoma2$sex, levels=c(1,0),
         labels=c("Male", 
                  "Female"))
 
melanoma2$ulcer <- 
  factor(melanoma2$ulcer, levels=c(0,1),
         labels=c("Absent", 
                  "Present"))

label(melanoma2$sex)       <- "Sex"
label(melanoma2$age)       <- "Age"
label(melanoma2$ulcer)     <- "Ulceration"
label(melanoma2$thickness) <- "Thickness"

units(melanoma2$age)       <- "years"
units(melanoma2$thickness) <- "mm"

table1(~ sex + age + ulcer + thickness | status, data=melanoma2, overall="Total")

This looks better, but still not quite the same as the original blog post: in the blog post the “Total” column is on the left, while we have it on the right; the two “Death” strata (Melanoma and Non-melanoma) should be grouped together under a common heading; the continuous variables Age and Thickness show only Means (SD) (with a ±), and not Median [Min, Max] like the table1 default output; most values are displayed with two significant digits rather than three (I will not concern myself with the footnote here, but it could be added as well). To achieve the same result, we need to customize the output further, and in this case that involves using the slightly more complicated “default” (i.e. non-formula) interface to table1.

First, we set up our labels differently, using a list:

labels <- list(
    variables=list(sex="Sex",
                   age="Age (years)",
                   ulcer="Ulceration",
                   thickness="Thickness (mm)"),
    groups=list("", "", "Death"))

# Remove the word "death" from the labels, since it now appears above
levels(melanoma2$status) <- c("Alive", "Melanoma", "Non-melanoma")

Next, we set up our “strata”, or column, as a list of data.frames, in the order we want them displayed:

strata <- c(list(Total=melanoma2), split(melanoma2, melanoma2$status))

Finally, we can customize the contents using custom renderers. A custom render can be a function that take a vector as the first argument and return a (named) character vector. There is also a simpler way to customize the table contents using an abbreviated code syntax instead of a render function, but it allows less control over rounding (see below). Here, for example, we specify render functions for the continuous and categorical variables as follows:

my.render.cont <- function(x) {
    with(stats.apply.rounding(stats.default(x), digits=2), c("",
        "Mean (SD)"=sprintf("%s (&plusmn; %s)", MEAN, SD)))
}
my.render.cat <- function(x) {
    c("", sapply(stats.default(x), function(y) with(y,
        sprintf("%d (%0.0f %%)", FREQ, PCT))))
}

And here is the result:

table1(strata, labels, groupspan=c(1, 1, 2),
       render.continuous=my.render.cont, render.categorical=my.render.cat)

This is now looking pretty similar to the original blog post, but admittedly there are still some differences: the sexes are inverted (the original blog post got it wrong); I added units to the continuous variables; I include the number of individuals in each column under the column heading; the percentages are different, because I think they should add to 100% within a column, and in the original blog post they add to 100% along a row (except for the Total column, which adds to 100% within the column). This last point is the most contentious. In my version, it is easier to compare the different types of outcomes with respect to variables like sex, while in the original version it is easier to compare sexes with respect to outcomes. However, this is not really the standard application for these kinds of tables (at least not the one I have in mind). Usually, the columns would represent exposure or treatment groups, not outcomes, and we want to compare those groups with respect to the distribution of baseline characteristics, and for this purpose having percentages add up to 100% within columns makes the most sense. Let’s continue with an example of that nature, using simulated data.

Example 2

For this second example, we will use simulated data. We imagine a clinical trial where subjects have been randomized in a 2:1 ratio to receive an active treatment or placebo. For simplicity, we will only consider three baseline characteristics: age, sex and weight.

f <- function(x, n, ...) factor(sample(x, n, replace=T, ...), levels=x)
set.seed(427)

n <- 146
dat <- data.frame(id=1:n)
dat$treat <- f(c("Placebo", "Treated"), n, prob=c(1, 2)) # 2:1 randomization
dat$age   <- sample(18:65, n, replace=TRUE)
dat$sex   <- f(c("Female", "Male"), n, prob=c(.6, .4))  # 60% female
dat$wt    <- round(exp(rnorm(n, log(70), 0.23)), 1)

# Add some missing data
dat$wt[sample.int(n, 5)] <- NA

label(dat$age)   <- "Age"
label(dat$sex)   <- "Sex"
label(dat$wt)    <- "Weight"
label(dat$treat) <- "Treatment Group"

units(dat$age)   <- "years"
units(dat$wt)    <- "kg"

Using the default settings, we obtain this table:

table1(~ age + sex + wt | treat, data=dat)

Note that when contains missing values (here weight), be it continuous or categorical, these are reported as a distinct category (with count and percent).

The “Overall” column can be easily removed (or relabeled):

table1(~ age + sex + wt | treat, data=dat, overall=F)

We can also have stratification by two variables, in which case they are nested. For example, to see each treatment group split by sex:

table1(~ age + wt | treat*sex, data=dat)

Or, switch the order:

table1(~ age + wt | sex*treat, data=dat)

Or, no stratification:

table1(~ treat + age + sex + wt, data=dat)

Finally, we may again consider something a bit more complicated, using the default (i.e., non-formula) interface. Suppose that instead of simply being assigned to placebo or active treatment, there were actually two doses of treatment randomized, 5 mg and 10 mg, and we want columns for each dose level separately, as well as for all treated subjects.

dat$dose <- (dat$treat != "Placebo")*sample(1:2, n, replace=T)
dat$dose <- factor(dat$dose, labels=c("Placebo", "5 mg", "10 mg"))

strata <- c(split(dat, dat$dose), list("All treated"=subset(dat, treat=="Treated")), list(Overall=dat))

labels <- list(
    variables=list(age=render.varlabel(dat$age),
                   sex=render.varlabel(dat$sex),
                   wt=render.varlabel(dat$wt)),
    groups=list("", "Treated", ""))

table1(strata, labels, groupspan=c(1, 3, 1))

Using abbreviated code to specify a custom renderer

Suppose that for continuous variables, we want to display the percent coefficient of variation (CV%) instead of the standard deviation (SD). We also want to display the geometric mean and geometric coefficient of variation. We already discussed custom render functions that could be used to accomplish this, but a simpler alternative is to use abbreviated code. This is a character string that contains certain keywords which are substituted for computed values in the table output. The list of recognized keywords comes from the output of the stats.default function and includes: N, NMISS, MEAN, SD, CV, GMEAN, GCV, MEDIAN, MIN, MAX, IQR, FREQ, PCT. Keyword matching is case insensitive, and any text other than the keywords is left untouched. We can specify a vector of character strings, in which case each result will be displayed in its own row in the table. We can use a named vector to specify labels for each row; a dot (‘.’) can be used to indicate that the abbreviated code string itself be used as the row label. Significant digits can be controlled using the digits argument (default: 3). Here is a continuation of the example from the previous section that produces the desired result:

table1(strata, labels, groupspan=c(1, 3, 1),
       render.continuous=c(.="Mean (CV%)", .="Median [Min, Max]",
                           "Geo. mean (Geo. CV%)"="GMEAN (GCV%)"))

Using CSS to control the table’s appearance

One thing I did not mention is that the output of table1() is just a bunch of HTML tags. To get the table to render, you have to open it in a web browser. If you were to just run the code above, paste the output in a file with the .html extension, and open it in your browser, it wouldn’t look the same as the output I presented. That is deliberate, because it is intended that table’s appearance be customized through CSS, which is ideally suited to this purpose (the trade-off is that you have to know some CSS to achieve the desired results).

The expected way that this package is to used is in conjunction with R Markdown, making use of the results='asis' chunk option to have the HTML output appear as is in the final document. (at least, that’s how I use it). You can embed your own CSS in the document by including this in the YAML header of your file (where style.css is the name of the file that contains the CSS):

output: 
  html_document:
    css: style.css

If style.css contains these definitions, then you will get the same output as above:

table.Rtable1 {
    font-family: "Arial", Arial, sans-serif;
    font-size: 10pt;
    border-collapse: collapse;
    padding: 0px;
    margin: 0px;
    margin-bottom: 10pt;
}
.Rtable1 th, .Rtable1 td {
    border: 0;
    text-align: center;
    padding: 0.5ex 1.5ex;
    margin: 0px;
}
.Rtable1 thead>tr:first-child>th {
    border-top: 2pt solid black;
}
.Rtable1 thead>tr:last-child>th {
    border-bottom: 1pt solid black;
}
.Rtable1 tbody>tr:last-child>td {
    border-bottom: 2pt solid black;
}
.Rtable1 th.grouplabel {
    padding-left: 0;
    padding-right: 0;
}
.Rtable1 th.grouplabel>div {
    margin-left: 1.5ex;
    margin-right: 1.5ex;
    border-bottom: 1pt solid black;
}
.Rtable1 th.grouplabel:last-child>div {
    margin-right: 0;
}
.Rtable1 .rowlabel {
    text-align: left;
    padding-left: 2.5ex;
}
.Rtable1 .firstrow {
    padding-left: 0.5ex;
    font-weight: bold;
}

By changing the CSS, you can change the appearance of the table to your liking. Two examples are given below for illustration purposes (the topclass argument allows the class attribute of the toplevel tag to be changed, which in turn allows different CSS styles to be applied).

table1(~ age + sex + wt | treat, data=dat, topclass="Rtable1 style2")

table1(~ age + sex + wt | treat, data=dat, topclass="Rtable1 style3")

Transposed table

By default, the table produced by table1 will have strata or subgroups as columns, and variables as rows. In some cases, it may be desirable to transpose the table such that each column is a variables and the rows are strata. This makes most sense when all the variables are continuous and when a compact representation is desired. It can be achieved by using the transpose = TRUE option.

An example:

dat <- expand.grid(i=1:50, group=LETTERS[1:3])
dat <- cbind(dat, matrix(round(exp(rnorm(6*nrow(dat))), 1), nrow=nrow(dat)))
names(dat)[3:8] <- paste0("V", 1:6)

Default:

table1(~ V1 + V2 + V3 + V4 + V5 + V6 | group, data=dat,
       topclass="Rtable1 style3 style4",
       render="Mean (CV%)<br/>Median [Min, Max]<br/>GMean (GCV%)")

Transposed:

table1(~ V1 + V2 + V3 + V4 + V5 + V6 | group, data=dat,
       topclass="Rtable1 style3 style4",
       render="Mean (CV%)<br/>Median [Min, Max]<br/>GMean (GCV%)",
       transpose=TRUE)