Bonus Slides: Making pretty regression tables with pipes

Madison R Users Group

Repository for this talk: https://github.com/tjmahr/MadR_Pipelines

Earlier, I covered pipelines.

Then, I used pipelines with dplyr.

Now, I will demonstrate off how I use pipelines to make pretty-printed regression tables.

library("magrittr")
library("dplyr")
library("broom")
library("stringr")
library("knitr")

broom's tidy function extracts summary values into a data-frame.

m_table <- tidy(model)
m_table %>% print(digits = 3)

         term estimate std.error statistic
1 (Intercept)    6.526     0.479     13.63
2 Sepal.Width   -0.223     0.155     -1.44
   p.value
1 6.47e-28
2 1.52e-01

knitr's kable function prints a data-frame as a Markdown table, which rmarkdown/pandoc can convert to HTML or Word.

kable(m_table, digits = 3,
      col.names = c("Param", "B", "SE", "t", "p"))

Raw Markdown output

|Param       |      B|    SE|      t|     p|
|:-----------|------:|-----:|------:|-----:|
|(Intercept) |  6.526| 0.479| 13.628| 0.000|
|Sepal.Width | -0.223| 0.155| -1.440| 0.152|

Two more things that I want.

1. APA style for printing numbers
   • Two digits of precision for estimates, errors and statistics.
   • Three digits for p-values.
   • Bounded values (correlations, p-values) don't need leading zero.
   • Tiny p-values should not be 0.000. Use < .001 instead.
2. Flexibility for many related models. (I don't know the final in-text model beforehand.)

• Build a set of generic, reusable functions for formatting numbers.
• Make a custom pipeline to format the term column.
• Assemble a pipeline to convert a model into a pretty table.

The formatC function is big and confusing, but after I figured out something that works well enough, I extracted a function. Now I don't have to worry about solving that problem again.

# Print with n digits of precision
fixed_digits <- function(xs, n = 2) {
  formatC(xs, digits = n, format = "f")
}

# Want .100 to print with two digits
c(1000.012, 0.0001, 0.100) %>% fixed_digits(2)

[1] "1000.01" "0.00"    "0.10"   

Here I do a sanity check to see if this operation is necessary.

# Don't print leading zero on bounded numbers.
remove_leading_zero <- function(xs) {
  # Problem if any value is greater than 1.0
  digit_matters <- xs %>% as.numeric %>%
    abs %>% is_greater_than(1)
  if (any(digit_matters)) {
    warning("Non-zero leading digit")
  }
  str_replace(xs, "^(-?)0", "\\1")
}
c(1.00, -0.12,  0.87, 0.82) %>% remove_leading_zero

[1] "1"    "-.12" ".87"  ".82" 

I format p-values with the previous two functions and overwrite the really small ones with the < .001 template.

# Print three digits of a p-value, but use
# the "< .001" notation on tiny values.
format_pval <- function(ps, html = FALSE) {
  tiny <- ifelse(html, "&lt;&nbsp;.001", "< .001")
  ps_chr <- ps %>% fixed_digits(3) %>%
    remove_leading_zero
  ps_chr[ps < 0.001] <- tiny
  ps_chr
}
m_table$p.value %>% format_pval

[1] "< .001" ".152"  

A few packages already solve the pretty-table problem–stargazer, texreg, xtable–and they do amazing things. But their functions usually rely on an argument where you can specify custom labels for the regression predictors. This is fine for one-off tables.

When I have a bunch of related models, I prefer to use a string-manipulation pipeline to convert R names into long-hand names.

fix_names <- . %>%
  str_replace(".Intercept.", "Intercept") %>%
  str_replace("Species", "") %>%
  # Capitalize species names
  str_replace("setosa", "Setosa") %>%
  str_replace("versicolor", "Versicolor") %>%
  str_replace("virginica", "Virginica") %>%
  # Clean up special characters
  str_replace_all(".Width", " Width") %>%
  str_replace_all(".Length", " Length") %>%
  str_replace_all(":", " x ")

1. Format the numbers
2. Rename the terms
3. Rename the column headings

two_digits <- . %>% fixed_digits(2)
table_names <- c("Parameter", "Estimate", "SE",
                 "_t_", "_p_")

format_model_table <- . %>%
  mutate_each(funs(two_digits),
              -term, -p.value) %>%
  mutate(term = fix_names(term),
         p.value = format_pval(p.value)) %>%
  set_colnames(table_names)

1. Fit a model
2. Extract values for summary table
3. Format table
4. Print as markdown

lm(..., iris) %>%
  tidy %>%
  format_model_table %>%
  kable