My R Reference Guide

Data transformation, visualization, statistics, and epidemiology

Author

Eleanor Renteria

1 How to Use This Guide

The guide moves in the order of a real project: get data in, clean it, transform it, visualize it, run statistics, apply epidemiology designs, then share results. Every section follows the same pattern:

  • What it’s for: when you would reach for it
  • Packages: which package does what
  • Worked example: commented code using a dataset built into R
  • Watch out for: common mistakes
  • Learn more: official documentation

1.1 Install everything once

Run this one time in the console (not in this file).

install.packages(c(
  "tidyverse", "janitor", "readxl", "haven", "data.table", "skimr",
  "survey", "srvyr", "broom", "performance", "lme4", "boot", "rsample",
  "epitools", "epiR", "MatchIt", "gtsummary", "tableone", "mice",
  "survival", "survminer", "forecast", "poLCA", "pwr", "gt", "here",
  "sf", "tigris", "shiny", "patchwork", "scales"
))
library(tidyverse)   # loads dplyr, tidyr, ggplot2, stringr, forcats, lubridate, purrr, readr

1.2 Built-in datasets used here

Dataset What it contains Used for
mtcars 32 cars, engine and mileage data filtering, regression
iris 150 flowers, 3 species grouping, PCA
airquality New York air readings with missing values missing data, dates
ToothGrowth guinea pig tooth growth by vitamin C dose comparing groups
infert, esoph case-control studies odds ratios, matching
lung (survival) lung cancer survival times survival analysis
AirPassengers monthly airline passengers 1949 to 1960 time series
api (survey) California school survey sample survey weighting

2 Part 1: Getting Data In and Cleaning It

2.1 Importing data

What it’s for: reading files into R as data frames.

Package Function File type
readr read_csv() .csv
readxl read_excel() .xls, .xlsx
haven read_xpt(), read_sas(), read_dta(), read_sav() SAS (NHANES), Stata, SPSS
here here() builds file paths that work on any computer
library(readxl); library(haven); library(here)
df   <- read_csv(here("data", "my_file.csv"))
xl   <- read_excel(here("data", "my_file.xlsx"), sheet = 1)
demo <- read_xpt(here("data", "DEMO_J.XPT"))   # NHANES demographics file

Watch out for: read_csv() (readr, underscore) is faster and tidier than base read.csv(). NHANES files are .XPT, so use read_xpt().

Learn more: readr, haven, here

2.2 First look and quick cleaning

What it’s for: understanding a dataset and fixing messy names before doing anything else.

library(janitor); library(skimr)
glimpse(airquality)          # column types and first values
Rows: 153
Columns: 6
$ Ozone   <int> 41, 36, 12, 18, NA, 28, 23, 19, 8, NA, 7, 16, 11, 14, 18, 14, …
$ Solar.R <int> 190, 118, 149, 313, NA, NA, 299, 99, 19, 194, NA, 256, 290, 27…
$ Wind    <dbl> 7.4, 8.0, 12.6, 11.5, 14.3, 14.9, 8.6, 13.8, 20.1, 8.6, 6.9, 9…
$ Temp    <int> 67, 72, 74, 62, 56, 66, 65, 59, 61, 69, 74, 69, 66, 68, 58, 64…
$ Month   <int> 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,…
$ Day     <int> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18,…
skim(airquality)             # summary, missing counts, mini histograms
Data summary
Name airquality
Number of rows 153
Number of columns 6
_______________________
Column type frequency:
numeric 6
________________________
Group variables None

Variable type: numeric

skim_variable n_missing complete_rate mean sd p0 p25 p50 p75 p100 hist
Ozone 37 0.76 42.13 32.99 1.0 18.00 31.5 63.25 168.0 ▇▃▂▁▁
Solar.R 7 0.95 185.93 90.06 7.0 115.75 205.0 258.75 334.0 ▅▃▅▇▅
Wind 0 1.00 9.96 3.52 1.7 7.40 9.7 11.50 20.7 ▂▇▇▃▁
Temp 0 1.00 77.88 9.47 56.0 72.00 79.0 85.00 97.0 ▂▃▇▇▃
Month 0 1.00 6.99 1.42 5.0 6.00 7.0 8.00 9.0 ▇▇▇▇▇
Day 0 1.00 15.80 8.86 1.0 8.00 16.0 23.00 31.0 ▇▇▇▇▆
messy <- tibble(`Patient ID` = 1:3, `Age (yrs)` = c(30, 30, 45))
messy |> clean_names()       # becomes patient_id, age_yrs
# A tibble: 3 × 2
  patient_id age_yrs
       <int>   <dbl>
1          1      30
2          2      30
3          3      45
messy |> get_dupes(`Age (yrs)`)   # find duplicate rows
# A tibble: 2 × 3
  `Age (yrs)` dupe_count `Patient ID`
        <dbl>      <int>        <int>
1          30          2            1
2          30          2            2
mtcars |> tabyl(cyl, am)     # quick cross-tab
 cyl  0 1
   4  3 8
   6  4 3
   8 12 2

Watch out for: clean names right after importing so every later step uses consistent snake_case.

Learn more: janitor, skimr

3 Part 2: Transforming Data

3.1 The pipe

|> means “and then.” df |> filter(x > 1) |> select(y) reads “take df, and then filter, and then select.” The older %>% from magrittr works the same way.

3.2 Selecting columns and filtering rows

Package Function Does
dplyr select() keep or drop columns
dplyr filter() keep rows that meet a condition
dplyr arrange() sort rows
dplyr distinct(), slice_max() unique rows, top rows
mtcars |>
  select(mpg, cyl, hp) |>          # keep three columns
  filter(cyl == 4, mpg > 25) |>     # comma means AND
  arrange(desc(mpg))                # highest mpg first
                mpg cyl  hp
Toyota Corolla 33.9   4  65
Fiat 128       32.4   4  66
Honda Civic    30.4   4  52
Lotus Europa   30.4   4 113
Fiat X1-9      27.3   4  66
Porsche 914-2  26.0   4  91
# Base R equivalent
mtcars[mtcars$cyl == 4 & mtcars$mpg > 25, c("mpg", "cyl", "hp")]
                mpg cyl  hp
Fiat 128       32.4   4  66
Honda Civic    30.4   4  52
Toyota Corolla 33.9   4  65
Fiat X1-9      27.3   4  66
Porsche 914-2  26.0   4  91
Lotus Europa   30.4   4 113
# Helpers
iris |> select(starts_with("Petal")) |> head(3)
  Petal.Length Petal.Width
1          1.4         0.2
2          1.4         0.2
3          1.3         0.2
mtcars |> filter(cyl %in% c(4, 6)) |> nrow()
[1] 18

Watch out for: use == to compare and = only to name arguments. filter() silently drops rows where the condition is NA.

3.3 Creating and changing columns

mtcars |>
  mutate(
    kpl = mpg * 0.425,                         # new numeric column
    power = case_when(                         # categories from rules
      hp < 100 ~ "low",
      hp < 200 ~ "medium",
      TRUE     ~ "high"                        # everything else
    ),
    heavy = if_else(wt > 3.5, "yes", "no")     # two-way choice
  ) |>
  select(mpg, kpl, hp, power, heavy) |>
  head()
                   mpg    kpl  hp  power heavy
Mazda RX4         21.0 8.9250 110 medium    no
Mazda RX4 Wag     21.0 8.9250 110 medium    no
Datsun 710        22.8 9.6900  93    low    no
Hornet 4 Drive    21.4 9.0950 110 medium    no
Hornet Sportabout 18.7 7.9475 175 medium    no
Valiant           18.1 7.6925 105 medium    no

Watch out for: case_when() checks rules top to bottom and stops at the first match, so order matters.

3.4 Grouping and summarizing

iris |>
  group_by(Species) |>
  summarise(
    n = n(),
    mean_petal = mean(Petal.Length),
    sd_petal = sd(Petal.Length)
  )
# A tibble: 3 × 4
  Species        n mean_petal sd_petal
  <fct>      <int>      <dbl>    <dbl>
1 setosa        50       1.46    0.174
2 versicolor    50       4.26    0.470
3 virginica     50       5.55    0.552
mtcars |> count(cyl, gear)      # shortcut for group_by + n()
  cyl gear  n
1   4    3  1
2   4    4  8
3   4    5  2
4   6    3  2
5   6    4  4
6   6    5  1
7   8    3 12
8   8    5  2
# data.table: very fast for millions of rows
library(data.table)
dt <- as.data.table(iris)
dt[, .(mean_petal = mean(Petal.Length)), by = Species]
      Species mean_petal
       <fctr>      <num>
1:     setosa      1.462
2: versicolor      4.260
3:  virginica      5.552

Watch out for: add .groups = "drop" or ungroup() afterward so later steps are not still grouped.

Learn more: dplyr, data.table

3.5 Factors and categories (forcats)

What it’s for: controlling category order and labels, which matters for tables, plots, and the reference group in regression.

df <- mtcars |> mutate(cyl = factor(cyl))
levels(df$cyl)
[1] "4" "6" "8"
df |> mutate(cyl = fct_relevel(cyl, "8")) |> pull(cyl) |> levels()   # 8 becomes reference
[1] "8" "4" "6"
df |> mutate(cyl = fct_recode(cyl, Four = "4", Six = "6", Eight = "8")) |> count(cyl)
    cyl  n
1  Four 11
2   Six  7
3 Eight 14
df |> mutate(cyl = fct_infreq(cyl)) |> count(cyl)    # order by frequency
  cyl  n
1   8 14
2   4 11
3   6  7
df |> mutate(cyl = fct_lump_n(cyl, n = 2)) |> count(cyl)  # rare levels become "Other"
    cyl  n
1     4 11
2     8 14
3 Other  7

Learn more: forcats

3.6 Text cleaning (stringr)

x <- c("  Type 2 Diabetes ", "type 1 diabetes", "HTN", "Hypertension")
str_trim(x)                       # remove extra spaces
[1] "Type 2 Diabetes" "type 1 diabetes" "HTN"             "Hypertension"   
str_to_lower(x)                   # lowercase
[1] "  type 2 diabetes " "type 1 diabetes"    "htn"               
[4] "hypertension"      
str_detect(x, regex("diabetes", ignore_case = TRUE))   # TRUE/FALSE
[1]  TRUE  TRUE FALSE FALSE
str_replace(x, "HTN", "Hypertension")
[1] "  Type 2 Diabetes " "type 1 diabetes"    "Hypertension"      
[4] "Hypertension"      
str_extract("ID-00452", "[0-9]+") # pull numbers out
[1] "00452"
str_c("Patient", 1:3, sep = "_")  # paste together
[1] "Patient_1" "Patient_2" "Patient_3"

Watch out for: stringr patterns are regular expressions, so . means “any character.” Use fixed(".") for a literal dot.

Learn more: stringr

3.7 Dates and times (lubridate)

aq <- airquality |> mutate(date = make_date(1973, Month, Day))
head(aq$date)
[1] "1973-05-01" "1973-05-02" "1973-05-03" "1973-05-04" "1973-05-05"
[6] "1973-05-06"
ymd("2026-11-20"); mdy("11/20/2026"); dmy("20-11-2026")
[1] "2026-11-20"
[1] "2026-11-20"
[1] "2026-11-20"
aq |> mutate(weekday = wday(date, label = TRUE), mon = month(date, label = TRUE)) |>
  select(date, weekday, mon) |> head(3)
Error in `mutate()`:
ℹ In argument: `weekday = wday(date, label = TRUE)`.
Caused by error in `wday()`:
! unused argument (label = TRUE)
# Time between dates, useful for follow-up time
interval(ymd("2020-01-15"), ymd("2023-06-01")) / years(1)
[1] 3.375342
ymd("2026-01-01") + days(90)      # 90-day window
[1] "2026-04-01"

Learn more: lubridate

3.8 Missing values and multiple imputation

What it’s for: finding, removing, filling, or imputing NAs.

colSums(is.na(airquality))                       # NA count per column
  Ozone Solar.R    Wind    Temp   Month     Day 
     37       7       0       0       0       0 
airquality |> drop_na(Ozone) |> nrow()           # drop rows missing Ozone
[1] 116
airquality |> replace_na(list(Solar.R = 0)) |> head(5)
  Ozone Solar.R Wind Temp Month Day
1    41     190  7.4   67     5   1
2    36     118  8.0   72     5   2
3    12     149 12.6   74     5   3
4    18     313 11.5   62     5   4
5    NA       0 14.3   56     5   5
mean(airquality$Ozone)                           # NA
[1] NA
mean(airquality$Ozone, na.rm = TRUE)             # ignore NA
[1] 42.12931
airquality |> fill(Ozone, .direction = "down") |> head(5)   # carry forward
  Ozone Solar.R Wind Temp Month Day
1    41     190  7.4   67     5   1
2    36     118  8.0   72     5   2
3    12     149 12.6   74     5   3
4    18     313 11.5   62     5   4
5    18      NA 14.3   56     5   5

Dropping rows can bias results when missingness is not random. Multiple imputation (mice) creates several filled-in datasets, fits the model on each, and pools the results.

library(mice)
imp <- mice(airquality, m = 5, seed = 123, printFlag = FALSE)
fit <- with(imp, lm(Ozone ~ Temp + Wind))
summary(pool(fit))
         term   estimate  std.error statistic       df      p.value
1 (Intercept) -66.004935 22.7838035 -2.897011 41.47674 5.991132e-03
2        Temp   1.769961  0.2392335  7.398466 48.30852 1.741085e-09
3        Wind  -2.961797  0.6652761 -4.451982 36.10737 7.867647e-05

Learn more: tidyr, mice, van Buuren, Flexible Imputation of Missing Data (free online)

3.9 Reshaping wide and long (tidyr)

What it’s for: “long” format (one row per measurement) is what ggplot2 and most models want. “Wide” is easier to read.

wide <- iris |> mutate(id = row_number()) |> head(3)
long <- wide |>
  pivot_longer(cols = Sepal.Length:Petal.Width,
               names_to = "measure", values_to = "cm")
long
# A tibble: 12 × 4
   Species    id measure         cm
   <fct>   <int> <chr>        <dbl>
 1 setosa      1 Sepal.Length   5.1
 2 setosa      1 Sepal.Width    3.5
 3 setosa      1 Petal.Length   1.4
 4 setosa      1 Petal.Width    0.2
 5 setosa      2 Sepal.Length   4.9
 6 setosa      2 Sepal.Width    3  
 7 setosa      2 Petal.Length   1.4
 8 setosa      2 Petal.Width    0.2
 9 setosa      3 Sepal.Length   4.7
10 setosa      3 Sepal.Width    3.2
11 setosa      3 Petal.Length   1.3
12 setosa      3 Petal.Width    0.2
long |> pivot_wider(names_from = measure, values_from = cm)
# A tibble: 3 × 6
  Species    id Sepal.Length Sepal.Width Petal.Length Petal.Width
  <fct>   <int>        <dbl>       <dbl>        <dbl>       <dbl>
1 setosa      1          5.1         3.5          1.4         0.2
2 setosa      2          4.9         3            1.4         0.2
3 setosa      3          4.7         3.2          1.3         0.2
# Split and combine columns
tibble(bp = c("120/80", "135/90")) |> separate_wider_delim(bp, "/", names = c("sys", "dia"))
# A tibble: 2 × 2
  sys   dia  
  <chr> <chr>
1 120   80   
2 135   90   

3.10 Joining tables (dplyr)

patients <- tibble(id = 1:4, age = c(34, 51, 29, 62))
labs     <- tibble(id = c(1, 2, 2, 5), a1c = c(5.4, 7.1, 6.8, 6.0))
left_join(patients, labs, by = "id")   # keep all patients
# A tibble: 5 × 3
     id   age   a1c
  <dbl> <dbl> <dbl>
1     1    34   5.4
2     2    51   7.1
3     2    51   6.8
4     3    29  NA  
5     4    62  NA  
inner_join(patients, labs, by = "id")  # only matches
# A tibble: 3 × 3
     id   age   a1c
  <dbl> <dbl> <dbl>
1     1    34   5.4
2     2    51   7.1
3     2    51   6.8
anti_join(patients, labs, by = "id")   # patients with no labs
# A tibble: 2 × 2
     id   age
  <int> <dbl>
1     3    29
2     4    62
full_join(patients, labs, by = "id")   # everyone
# A tibble: 6 × 3
     id   age   a1c
  <dbl> <dbl> <dbl>
1     1    34   5.4
2     2    51   7.1
3     2    51   6.8
4     3    29  NA  
5     4    62  NA  
6     5    NA   6  

Watch out for: if a key repeats (patient 2 has two labs), rows multiply. Check nrow() before and after every join.

3.11 Repeating steps (across and purrr)

# Same function on many columns
iris |> group_by(Species) |> summarise(across(where(is.numeric), mean))
# A tibble: 3 × 5
  Species    Sepal.Length Sepal.Width Petal.Length Petal.Width
  <fct>             <dbl>       <dbl>        <dbl>       <dbl>
1 setosa             5.01        3.43         1.46       0.246
2 versicolor         5.94        2.77         4.26       1.33 
3 virginica          6.59        2.97         5.55       2.03 
mtcars |> mutate(across(c(mpg, hp), ~ round(.x / max(.x), 2))) |> head(3)
               mpg cyl disp   hp drat    wt  qsec vs am gear carb
Mazda RX4     0.62   6  160 0.33 3.90 2.620 16.46  0  1    4    4
Mazda RX4 Wag 0.62   6  160 0.33 3.90 2.875 17.02  0  1    4    4
Datsun 710    0.67   4  108 0.28 3.85 2.320 18.61  1  1    4    1
# Same model for each group
mtcars |>
  split(~ cyl) |>
  map(\(d) lm(mpg ~ wt, data = d)) |>
  map_dbl(\(m) coef(m)[["wt"]])
        4         6         8 
-5.647025 -2.780106 -2.192438 
# Read many files at once
# files <- list.files("data", pattern = ".csv", full.names = TRUE)
# all <- map(files, read_csv) |> list_rbind()

Learn more: purrr

4 Part 3: Visualizing with ggplot2

4.1 The grammar

Every plot is built from layers joined with +:

  1. ggplot(data, aes(x, y, color, fill, size, shape)): data and mappings
  2. geom_*(): the shape (points, bars, lines)
  3. labs(): words
  4. scale_*(): colors and axes
  5. facet_*(): small multiples
  6. theme(): everything visual that is not data

4.2 Common geoms

p <- ggplot(mtcars, aes(wt, mpg))
p + geom_point()                                  # scatter
p + geom_point() + geom_smooth(method = "lm")     # with trend line
ggplot(mtcars, aes(factor(cyl))) + geom_bar()     # counts
ggplot(iris, aes(Species, Petal.Length)) + geom_boxplot()
ggplot(iris, aes(Petal.Length)) + geom_histogram(bins = 20)
ggplot(airquality, aes(Day, Temp, color = factor(Month))) + geom_line()
ggplot(iris, aes(Petal.Length, fill = Species)) + geom_density(alpha = 0.5)
ggplot(ToothGrowth, aes(factor(dose), len)) + geom_violin() + geom_jitter(width = 0.1)

geom_bar() counts rows for you. geom_col() uses a y value you already calculated.

4.3 Titles, subtitles, captions, and labels

base <- ggplot(mtcars, aes(wt, mpg, color = factor(cyl))) + geom_point(size = 3)
base + labs(
  title    = "Heavier cars get fewer miles per gallon",
  subtitle = "Motor Trend 1974 road tests",
  caption  = "Source: mtcars dataset",
  tag      = "A",                    # panel letter for figures
  x        = "Weight (1,000 lbs)",
  y        = "Miles per gallon",
  color    = "Cylinders"             # legend title
)

4.4 Text style: bold, italic, size, angle, alignment

All text is styled with element_text() inside theme().

Argument Options Meaning
face “plain”, “bold”, “italic”, “bold.italic” font style
size number points
color “navy”, “#1B4F72” color
hjust 0, 0.5, 1 left, center, right
vjust 0, 0.5, 1 bottom, middle, top
angle 0 to 360 slant the text
family “serif”, “sans”, “mono” font
margin margin(t, r, b, l) space around text
lineheight number spacing between lines

Theme pieces you can target: plot.title, plot.subtitle, plot.caption, axis.title.x, axis.title.y, axis.text.x, axis.text.y, legend.title, legend.text, strip.text (facet labels).

base +
  labs(title = "Heavier cars get fewer miles per gallon",
       subtitle = "Motor Trend 1974", caption = "Source: mtcars",
       x = "Weight", y = "MPG", color = "Cylinders") +
  theme(
    plot.title    = element_text(face = "bold", size = 16, hjust = 0.5),
    plot.subtitle = element_text(face = "italic", hjust = 0.5, color = "gray40"),
    plot.caption  = element_text(face = "italic", hjust = 0, size = 8),
    plot.title.position = "plot",              # align title with whole plot
    axis.title.x  = element_text(face = "bold", margin = margin(t = 10)),
    axis.text.x   = element_text(angle = 45, hjust = 1),   # slanted labels
    legend.title  = element_text(face = "bold")
  )

Italic inside a label (like species names or p):

ggplot(iris, aes(Sepal.Length, Petal.Length)) + geom_point() +
  labs(title = expression(paste("Petal length in ", italic("Iris"), " species")),
       y = expression(Petal~length~(cm^2)))

4.5 Legend placement

base + theme(legend.position = "bottom")          # "top", "left", "right", "none"

base + theme(legend.position = "inside", legend.position.inside = c(0.85, 0.8))

base + guides(color = guide_legend(nrow = 1, reverse = TRUE))

4.6 Annotations

base +
  annotate("text", x = 5, y = 30, label = "Light, efficient cars", fontface = "bold") +
  annotate("rect", xmin = 1.5, xmax = 2.5, ymin = 25, ymax = 35, alpha = 0.1, fill = "blue") +
  annotate("segment", x = 4.5, xend = 5.3, y = 20, yend = 15.5,
           arrow = arrow(length = unit(0.2, "cm"))) +
  geom_hline(yintercept = mean(mtcars$mpg), linetype = "dashed") +
  geom_text(data = mtcars |> rownames_to_column("car") |> filter(mpg > 30),
            aes(wt, mpg, label = car), vjust = -1, size = 3, inherit.aes = FALSE)

4.7 Colors and scales

library(scales)
base + scale_color_brewer(palette = "Set2")           # ColorBrewer palettes

base + scale_color_viridis_d()                        # colorblind friendly

base + scale_color_manual(values = c("4" = "#1B9E77", "6" = "#D95F02", "8" = "#7570B3"))

ggplot(mtcars, aes(wt, mpg, color = hp)) + geom_point() + scale_color_gradient(low = "gold", high = "red")

base + scale_x_continuous(limits = c(1, 6), breaks = 1:6) +
  scale_y_continuous(labels = label_number(suffix = " mpg"))

ggplot(mtcars, aes(factor(cyl))) + geom_bar(aes(y = after_stat(count / sum(count)))) +
  scale_y_continuous(labels = label_percent())

base + coord_flip()       # swap axes

base + scale_y_log10()    # log scale

4.8 Facets

ggplot(mtcars, aes(wt, mpg)) + geom_point() + facet_wrap(~ cyl, labeller = label_both)

ggplot(mtcars, aes(wt, mpg)) + geom_point() + facet_grid(am ~ cyl, scales = "free_y")

4.9 Themes

base + theme_minimal()
base + theme_classic()
base + theme_bw()
base + theme_void()

Set one theme for the whole document: theme_set(theme_minimal(base_size = 13)).

Other theme elements: panel.grid.minor = element_blank() removes minor grid lines, panel.background = element_rect(fill = "white"), axis.line = element_line(color = "black").

4.10 Combining plots (patchwork)

library(patchwork)
p1 <- ggplot(mtcars, aes(wt, mpg)) + geom_point()
p2 <- ggplot(mtcars, aes(factor(cyl))) + geom_bar()
(p1 | p2) + plot_annotation(title = "Two views of mtcars", tag_levels = "A")

Learn more: ggplot2 reference, Wickham, ggplot2: Elegant Graphics for Data Analysis (free online), R Graph Gallery

4.11 Maps (sf and tigris)

What it’s for: county or state maps, such as rates by county for health equity work.

library(sf); library(tigris)
options(tigris_use_cache = TRUE)
wa <- counties(state = "WA", cb = TRUE)          # downloads WA county shapes
wa$fake_rate <- runif(nrow(wa), 5, 25)           # replace with your data via left_join
ggplot(wa) +
  geom_sf(aes(fill = fake_rate), color = "white") +
  scale_fill_viridis_c(name = "Rate per 100k") +
  labs(title = "Example rate by Washington county") +
  theme_void()

Join your data to the map with left_join(wa, my_data, by = "GEOID") using the county FIPS code.

Learn more: sf, tigris, Lovelace et al., Geocomputation with R (free online)

5 Part 4: Statistics

5.1 Descriptive statistics

summary(mtcars$mpg)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
  10.40   15.43   19.20   20.09   22.80   33.90 
mtcars |> summarise(mean = mean(mpg), median = median(mpg), sd = sd(mpg),
                    iqr = IQR(mpg), min = min(mpg), max = max(mpg))
      mean median       sd   iqr  min  max
1 20.09062   19.2 6.026948 7.375 10.4 33.9
quantile(mtcars$mpg, c(0.25, 0.5, 0.75))
   25%    50%    75% 
15.425 19.200 22.800 
table(mtcars$cyl); prop.table(table(mtcars$cyl))

 4  6  8 
11  7 14 

      4       6       8 
0.34375 0.21875 0.43750 

5.2 Which test do I use?

Outcome Groups Test R function
Numeric 2 independent t-test (or Wilcoxon if skewed) t.test(), wilcox.test()
Numeric 2 paired paired t-test t.test(paired = TRUE)
Numeric 3+ ANOVA (or Kruskal-Wallis) aov(), kruskal.test()
Categorical any chi-square (Fisher if counts < 5) chisq.test(), fisher.test()
Two numerics none correlation cor.test()
t.test(len ~ supp, data = ToothGrowth)                    # two groups

    Welch Two Sample t-test

data:  len by supp
t = 1.9153, df = 55.309, p-value = 0.06063
alternative hypothesis: true difference in means between group OJ and group VC is not equal to 0
95 percent confidence interval:
 -0.1710156  7.5710156
sample estimates:
mean in group OJ mean in group VC 
        20.66333         16.96333 
wilcox.test(len ~ supp, data = ToothGrowth)

    Wilcoxon rank sum exact test

data:  len by supp
W = 575.5, p-value = 0.06366
alternative hypothesis: true location shift is not equal to 0
a <- aov(len ~ factor(dose), data = ToothGrowth); summary(a)
             Df Sum Sq Mean Sq F value   Pr(>F)    
factor(dose)  2   2426    1213   67.42 9.53e-16 ***
Residuals    57   1026      18                     
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
TukeyHSD(a)                                               # which pairs differ
  Tukey multiple comparisons of means
    95% family-wise confidence level

Fit: aov(formula = len ~ factor(dose), data = ToothGrowth)

$`factor(dose)`
        diff       lwr       upr    p adj
1-0.5  9.130  5.901805 12.358195 0.00e+00
2-0.5 15.495 12.266805 18.723195 0.00e+00
2-1    6.365  3.136805  9.593195 4.25e-05
chisq.test(table(mtcars$cyl, mtcars$am))

    Pearson's Chi-squared test

data:  table(mtcars$cyl, mtcars$am)
X-squared = 8.7407, df = 2, p-value = 0.01265
fisher.test(table(mtcars$vs, mtcars$am))

    Fisher's Exact Test for Count Data

data:  table(mtcars$vs, mtcars$am)
p-value = 0.4727
alternative hypothesis: true odds ratio is not equal to 1
95 percent confidence interval:
  0.3825342 10.5916087
sample estimates:
odds ratio 
  1.956055 
cor.test(mtcars$mpg, mtcars$wt)                           # Pearson

    Pearson's product-moment correlation

data:  mtcars$mpg and mtcars$wt
t = -9.559, df = 30, p-value = 1.294e-10
alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
 -0.9338264 -0.7440872
sample estimates:
       cor 
-0.8676594 
cor.test(mtcars$mpg, mtcars$wt, method = "spearman")      # rank based

    Spearman's rank correlation rho

data:  mtcars$mpg and mtcars$wt
S = 10292, p-value = 1.488e-11
alternative hypothesis: true rho is not equal to 0
sample estimates:
      rho 
-0.886422 
round(cor(mtcars[, 1:5]), 2)                              # correlation matrix
       mpg   cyl  disp    hp  drat
mpg   1.00 -0.85 -0.85 -0.78  0.68
cyl  -0.85  1.00  0.90  0.83 -0.70
disp -0.85  0.90  1.00  0.79 -0.71
hp   -0.78  0.83  0.79  1.00 -0.45
drat  0.68 -0.70 -0.71 -0.45  1.00

5.3 Multiple testing

Running many tests inflates false positives. Adjust p-values:

pvals <- c(0.01, 0.02, 0.03, 0.04, 0.20)
p.adjust(pvals, method = "bonferroni")   # strict
[1] 0.05 0.10 0.15 0.20 1.00
p.adjust(pvals, method = "BH")           # false discovery rate, common in genomics
[1] 0.05 0.05 0.05 0.05 0.20

5.4 Regression

Linear (numeric outcome) and logistic (yes/no outcome):

library(broom)
lin <- lm(mpg ~ wt + hp + factor(cyl), data = mtcars)
summary(lin)

Call:
lm(formula = mpg ~ wt + hp + factor(cyl), data = mtcars)

Residuals:
    Min      1Q  Median      3Q     Max 
-4.2612 -1.0320 -0.3210  0.9281  5.3947 

Coefficients:
             Estimate Std. Error t value Pr(>|t|)    
(Intercept)  35.84600    2.04102  17.563 2.67e-16 ***
wt           -3.18140    0.71960  -4.421 0.000144 ***
hp           -0.02312    0.01195  -1.934 0.063613 .  
factor(cyl)6 -3.35902    1.40167  -2.396 0.023747 *  
factor(cyl)8 -3.18588    2.17048  -1.468 0.153705    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 2.44 on 27 degrees of freedom
Multiple R-squared:  0.8572,    Adjusted R-squared:  0.8361 
F-statistic: 40.53 on 4 and 27 DF,  p-value: 4.869e-11
tidy(lin, conf.int = TRUE)       # clean table of coefficients
# A tibble: 5 × 7
  term         estimate std.error statistic  p.value conf.low conf.high
  <chr>           <dbl>     <dbl>     <dbl>    <dbl>    <dbl>     <dbl>
1 (Intercept)   35.8       2.04       17.6  2.67e-16  31.7     40.0    
2 wt            -3.18      0.720      -4.42 1.44e- 4  -4.66    -1.70   
3 hp            -0.0231    0.0120     -1.93 6.36e- 2  -0.0476   0.00140
4 factor(cyl)6  -3.36      1.40       -2.40 2.37e- 2  -6.24    -0.483  
5 factor(cyl)8  -3.19      2.17       -1.47 1.54e- 1  -7.64     1.27   
glance(lin)                      # R-squared, AIC
# A tibble: 1 × 12
  r.squared adj.r.squared sigma statistic  p.value    df logLik   AIC   BIC
      <dbl>         <dbl> <dbl>     <dbl>    <dbl> <dbl>  <dbl> <dbl> <dbl>
1     0.857         0.836  2.44      40.5 4.87e-11     4  -71.2  154.  163.
# ℹ 3 more variables: deviance <dbl>, df.residual <int>, nobs <int>
logit <- glm(am ~ wt + hp, data = mtcars, family = binomial)
tidy(logit, exponentiate = TRUE, conf.int = TRUE)   # odds ratios
# A tibble: 3 × 7
  term        estimate std.error statistic p.value conf.low conf.high
  <chr>          <dbl>     <dbl>     <dbl>   <dbl>    <dbl>     <dbl>
1 (Intercept)  1.56e+8    7.44        2.53 0.0113   4.34e+3  4.08e+17
2 wt           3.09e-4    3.07       -2.63 0.00843  3.36e-8  2.30e- 2
3 hp           1.04e+0    0.0177      2.04 0.0409   1.01e+0  1.09e+ 0

Other families: family = poisson for counts, MASS::glm.nb() for overdispersed counts.

5.5 Checking model assumptions (performance)

library(performance)
check_model(lin)                 # linearity, normal residuals, equal variance, outliers
Error in `loadNamespace()`:
! there is no package called 'see'
check_collinearity(lin)          # VIF; above 5 to 10 is a concern
# Check for Multicollinearity

Low Correlation

 Term  VIF   VIF 95% CI adj. VIF Tolerance Tolerance 95% CI
   wt 2.58 [1.45, 3.60]     1.61      0.39     [0.28, 0.69]
   hp 3.50 [1.85, 4.89]     1.87      0.29     [0.20, 0.54]

Moderate Correlation

        Term  VIF   VIF 95% CI adj. VIF Tolerance Tolerance 95% CI
 factor(cyl) 5.11 [2.71, 7.35]     1.50      0.20     [0.14, 0.37]
r2(lin)
# R2 for Linear Regression
       R2: 0.857
  adj. R2: 0.836

5.6 Mixed models for repeated measures (lme4)

What it’s for: when the same subject is measured more than once, or people are clustered (patients in clinics). A random intercept lets each subject have their own baseline.

library(lme4)
# sleepstudy: reaction time over 10 days of sleep deprivation, 18 subjects
mm <- lmer(Reaction ~ Days + (1 | Subject), data = sleepstudy)
summary(mm)
Linear mixed model fit by REML ['lmerMod']
Formula: Reaction ~ Days + (1 | Subject)
   Data: sleepstudy

REML criterion at convergence: 1786.5

Scaled residuals: 
    Min      1Q  Median      3Q     Max 
-3.2257 -0.5529  0.0109  0.5188  4.2506 

Random effects:
 Groups   Name        Variance Std.Dev.
 Subject  (Intercept) 1378.2   37.12   
 Residual              960.5   30.99   
Number of obs: 180, groups:  Subject, 18

Fixed effects:
            Estimate Std. Error t value
(Intercept) 251.4051     9.7467   25.79
Days         10.4673     0.8042   13.02

Correlation of Fixed Effects:
     (Intr)
Days -0.371
mm2 <- lmer(Reaction ~ Days + (Days | Subject), data = sleepstudy)  # random slope too
anova(mm, mm2)
Data: sleepstudy
Models:
mm: Reaction ~ Days + (1 | Subject)
mm2: Reaction ~ Days + (Days | Subject)
    npar    AIC    BIC  logLik -2*log(L)  Chisq Df Pr(>Chisq)    
mm     4 1802.1 1814.8 -897.04    1794.1                         
mm2    6 1763.9 1783.1 -875.97    1751.9 42.139  2  7.072e-10 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Learn more: lme4 vignette

5.7 Bootstrapping

What it’s for: confidence intervals without assuming a normal distribution, by resampling your data with replacement many times.

library(boot)
med_fun <- function(data, i) median(data$mpg[i])
set.seed(42)
b <- boot(mtcars, med_fun, R = 2000)
b

ORDINARY NONPARAMETRIC BOOTSTRAP


Call:
boot(data = mtcars, statistic = med_fun, R = 2000)


Bootstrap Statistics :
    original  bias    std. error
t1*     19.2 0.08595    1.255853
boot.ci(b, type = c("perc", "bca"))
BOOTSTRAP CONFIDENCE INTERVAL CALCULATIONS
Based on 2000 bootstrap replicates

CALL : 
boot.ci(boot.out = b, type = c("perc", "bca"))

Intervals : 
Level     Percentile            BCa          
95%   (16.85, 21.40 )   (15.90, 21.20 )  
Calculations and Intervals on Original Scale
# Tidyverse style
library(rsample)
set.seed(42)
bootstraps(mtcars, times = 2000) |>
  mutate(med = map_dbl(splits, \(s) median(analysis(s)$mpg))) |>
  summarise(lower = quantile(med, 0.025), upper = quantile(med, 0.975))
# A tibble: 1 × 2
  lower upper
  <dbl> <dbl>
1  16.8  21.4

5.8 Power and sample size (pwr)

Give three of four values (effect size, n, alpha, power) and pwr solves for the fourth.

library(pwr)
pwr.t.test(d = 0.5, power = 0.8, sig.level = 0.05)    # n per group for a medium effect

     Two-sample t test power calculation 

              n = 63.76561
              d = 0.5
      sig.level = 0.05
          power = 0.8
    alternative = two.sided

NOTE: n is number in *each* group
pwr.2p.test(h = ES.h(0.30, 0.20), power = 0.8)        # comparing two proportions

     Difference of proportion power calculation for binomial distribution (arcsine transformation) 

              h = 0.2319843
              n = 291.6887
      sig.level = 0.05
          power = 0.8
    alternative = two.sided

NOTE: same sample sizes
pwr.chisq.test(w = 0.3, df = 2, power = 0.8)

     Chi squared power calculation 

              w = 0.3
              N = 107.0521
             df = 2
      sig.level = 0.05
          power = 0.8

NOTE: N is the number of observations

Cohen’s guide for d: 0.2 small, 0.5 medium, 0.8 large.

Learn more: broom, performance, pwr, Cohen, Statistical Power Analysis for the Behavioral Sciences (1988)

6 Part 5: Epidemiology in R

6.1 Survey weighting (survey and srvyr)

What it’s for: NHANES, BRFSS, and similar surveys oversample some groups on purpose. Weights, strata, and clusters make estimates represent the whole population and give correct standard errors. Unweighted results from these surveys are biased.

library(survey); library(srvyr)
data(api)       # apistrat: stratified sample of California schools
des <- apistrat |> as_survey_design(strata = stype, weights = pw, fpc = fpc)
des |> summarise(mean_api = survey_mean(api00, vartype = "ci"))
# A tibble: 1 × 3
  mean_api mean_api_low mean_api_upp
     <dbl>        <dbl>        <dbl>
1     662.         644.         681.
des |> group_by(stype) |> summarise(pct_meals = survey_mean(meals))
# A tibble: 3 × 3
  stype pct_meals pct_meals_se
  <fct>     <dbl>        <dbl>
1 E          51.8         2.98
2 H          30.4         3.47
3 M          46.1         3.43
# Regression with the design
svyglm(api00 ~ ell + meals, design = des) |> tidy()
# A tibble: 3 × 5
  term        estimate std.error statistic   p.value
  <chr>          <dbl>     <dbl>     <dbl>     <dbl>
1 (Intercept)  824.        8.76      94.1  2.03e-164
2 ell           -0.506     0.388     -1.30 1.94e-  1
3 meals         -3.11      0.276    -11.3  4.85e- 23

NHANES template (variable names from the CDC tutorials):

nhanes_des <- nhanes |>
  as_survey_design(ids = SDMVPSU, strata = SDMVSTRA, weights = WTMEC2YR, nest = TRUE)
# Subset AFTER building the design, never before
adults <- nhanes_des |> filter(RIDAGEYR >= 20)

Watch out for: filtering the data before creating the design gives wrong standard errors. Pick the weight that matches your variables (interview vs exam vs fasting subsample).

Learn more: CDC NHANES tutorials, srvyr, Lumley, Complex Surveys (2010)

6.2 Study designs at a glance

Design Direction Measure Key R tools
Cohort (prospective) exposure to outcome, forward risk, risk ratio, incidence rate epitools, epiR, survival
Retrospective cohort uses existing records, still exposure to outcome same as cohort dplyr cohort building
Case-control outcome to past exposure odds ratio glm, clogit, MatchIt
Cross-sectional one point in time prevalence, prevalence ratio survey, Poisson regression

6.3 The 2x2 table

Outcome yes Outcome no
Exposed a b
Unexposed c d

Risk ratio = [a / (a + b)] / [c / (c + d)]. Odds ratio = (a × d) / (b × c).

library(epiR)
tab <- matrix(c(40, 60, 20, 80), nrow = 2, byrow = TRUE)   # a, b, c, d
epi.2by2(tab, method = "cohort.count")       # risk ratio, risk difference, attributable fraction
             Outcome+    Outcome-      Total                 Inc risk *
Exposure+          40          60        100     40.00 (30.33 to 50.28)
Exposure-          20          80        100     20.00 (12.67 to 29.18)
Total              60         140        200     30.00 (23.74 to 36.86)

Point estimates and 95% CIs:
-------------------------------------------------------------------
Inc risk ratio                                 2.00 (1.26, 3.17)
Inc odds ratio                                 2.67 (1.42, 5.02)
Attrib risk in the exposed *                   20.00 (7.60, 32.40)
Attrib fraction in the exposed (%)            50.00 (21.73, 68.56)
Attrib risk in the population *                10.00 (-0.09, 20.09)
Attrib fraction in the population (%)         33.33 (20.83, 46.65)
E-value                                        3.41 (1.84, NA)
-------------------------------------------------------------------
Uncorrected chi2 test that OR = 1: chi2(1) = 9.524 Pr>chi2 = 0.002
Fisher exact test that OR = 1: Pr>chi2 = 0.003
Wald confidence limits
CI: confidence interval
* Outcomes per 100 population units 
epi.2by2(tab, method = "case.control")       # odds ratio
             Outcome+    Outcome-      Total                 Odds
Exposed +          40          60        100  0.67 (0.45 to 1.00)
Exposed -          20          80        100  0.25 (0.14 to 0.39)
Total              60         140        200  0.43 (0.32 to 0.57)

Point estimates and 95% CIs:
-------------------------------------------------------------------
Exposure odds ratio                            2.67 (1.42, 5.02)
Attrib fraction (est) in the exposed (%)      62.50 (29.66, 79.99)
Attrib fraction (est) in the population (%)   41.67 (32.34, 56.82)
E-value                                        2.65 (1.67, NA)
-------------------------------------------------------------------
Uncorrected chi2 test that OR = 1: chi2(1) = 9.524 Pr>chi2 = 0.002
Fisher exact test that OR = 1: Pr>chi2 = 0.003
Wald confidence limits
CI: confidence interval 
library(epitools)
riskratio(matrix(c(80, 20, 60, 40), 2, byrow = TRUE))   # epitools wants unexposed in row 1, outcome-no first
$data
          Outcome
Predictor  Disease1 Disease2 Total
  Exposed1       80       20   100
  Exposed2       60       40   100
  Total         140       60   200

$measure
          risk ratio with 95% C.I.
Predictor  estimate    lower    upper
  Exposed1        1       NA       NA
  Exposed2        2 1.263006 3.167047

$p.value
          two-sided
Predictor   midp.exact fisher.exact  chi.square
  Exposed1          NA           NA          NA
  Exposed2 0.002145336  0.003191817 0.002028231

$correction
[1] FALSE

attr(,"method")
[1] "Unconditional MLE & normal approximation (Wald) CI"

6.4 Building a cohort from records

Steps for a retrospective cohort from EHR-style or All of Us data:

  1. Define the index date (first diagnosis, first prescription, enrollment)
  2. Apply inclusion and exclusion criteria
  3. Require a lookback window (no prior outcome in, say, 365 days before index)
  4. Define follow-up: from index until outcome, death, end of data, or end of study
  5. Calculate person-time and the event indicator
set.seed(1)
visits <- tibble(
  id = rep(1:6, each = 3),
  date = ymd("2019-01-01") + sample(0:1500, 18),
  code = sample(c("exposure", "outcome", "other"), 18, replace = TRUE)
)
index <- visits |> filter(code == "exposure") |>
  group_by(id) |> summarise(index_date = min(date))
prior <- visits |> inner_join(index, by = "id") |>
  filter(code == "outcome", date < index_date, date >= index_date - days(365)) |> distinct(id)
cohort <- index |>
  anti_join(prior, by = "id") |>                          # exclude prior outcome
  left_join(visits |> filter(code == "outcome"), by = "id", relationship = "many-to-many") |>
  filter(is.na(date) | date > index_date) |>
  group_by(id, index_date) |>
  summarise(event_date = suppressWarnings(min(date)), .groups = "drop") |>
  mutate(
    event = as.integer(!is.na(event_date) & is.finite(event_date)),
    end_date = if_else(event == 1, event_date, ymd("2023-12-31")),
    years = as.numeric(end_date - index_date) / 365.25
  )
cohort
# A tibble: 3 × 6
     id index_date event_date event end_date   years
  <int> <date>     <date>     <int> <date>     <dbl>
1     1 2020-11-09 2021-10-13     1 2021-10-13 0.925
2     4 2019-11-26 2022-07-24     1 2022-07-24 2.66 
3     5 2020-12-29 2022-02-02     1 2022-02-02 1.10 
cohort |> summarise(events = sum(event), person_years = sum(years),
                    rate_per_1000py = 1000 * events / person_years)
# A tibble: 1 × 3
  events person_years rate_per_1000py
   <int>        <dbl>           <dbl>
1      3         4.68            641.

6.5 Case-control analysis

# infert: matched case-control study of infertility after abortion
glm(case ~ spontaneous + induced + education, data = infert, family = binomial) |>
  tidy(exponentiate = TRUE, conf.int = TRUE)
# A tibble: 5 × 7
  term             estimate std.error statistic      p.value conf.low conf.high
  <chr>               <dbl>     <dbl>     <dbl>        <dbl>    <dbl>     <dbl>
1 (Intercept)         0.172     0.728   -2.42   0.0157         0.0372     0.676
2 spontaneous         3.33      0.212    5.67   0.0000000139   2.22       5.12 
3 induced             1.53      0.209    2.04   0.0414         1.02       2.32 
4 education6-11yrs    1.12      0.706    0.156  0.876          0.293      4.93 
5 education12+ yrs    0.976     0.704   -0.0347 0.972          0.257      4.28 
# Conditional logistic regression respects the matched sets
library(survival)
clogit(case ~ spontaneous + induced + strata(stratum), data = infert) |> summary()
Call:
coxph(formula = Surv(rep(1, 248L), case) ~ spontaneous + induced + 
    strata(stratum), data = infert, method = "exact")

  n= 248, number of events= 83 

              coef exp(coef) se(coef)     z Pr(>|z|)    
spontaneous 1.9859    7.2854   0.3524 5.635 1.75e-08 ***
induced     1.4090    4.0919   0.3607 3.906 9.38e-05 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

            exp(coef) exp(-coef) lower .95 upper .95
spontaneous     7.285     0.1373     3.651    14.536
induced         4.092     0.2444     2.018     8.298

Concordance= 0.776  (se = 0.044 )
Likelihood ratio test= 53.15  on 2 df,   p=3e-12
Wald test            = 31.84  on 2 df,   p=1e-07
Score (logrank) test = 48.44  on 2 df,   p=3e-11

Matching or propensity scores (MatchIt):

library(MatchIt)
m <- matchit(am ~ wt + hp, data = mtcars, method = "nearest", ratio = 1)
summary(m)$sum.matched[, 1:3]      # balance after matching
         Means Treated Means Control Std. Mean Diff.
distance     0.8879044     0.1114532       4.2186553
wt           2.4110000     3.4011538      -1.6048352
hp         126.8461538   148.8461538      -0.2617106
matched <- match.data(m)

6.6 Cross-sectional analysis

Prevalence ratios are often preferred over odds ratios when the outcome is common (over about 10%). Use Poisson regression with robust errors, or svyglm(family = quasipoisson) for survey data.

pr <- glm(am ~ factor(cyl), data = mtcars, family = poisson)
tidy(pr, exponentiate = TRUE, conf.int = TRUE)
# A tibble: 3 × 7
  term         estimate std.error statistic p.value conf.low conf.high
  <chr>           <dbl>     <dbl>     <dbl>   <dbl>    <dbl>     <dbl>
1 (Intercept)     0.727     0.354    -0.901  0.368    0.332      1.35 
2 factor(cyl)6    0.589     0.677    -0.781  0.435    0.129      2.04 
3 factor(cyl)8    0.196     0.791    -2.06   0.0395   0.0296     0.784

6.7 Table 1 and confounding

library(gtsummary)
trial |>                                           # built into gtsummary
  select(trt, age, grade, response) |>
  tbl_summary(by = trt, missing = "ifany") |>
  add_p() |> add_overall() |> bold_labels()
Characteristic Overall
N = 2001
Drug A
N = 981
Drug B
N = 1021
p-value2
Age 47 (38, 57) 46 (37, 60) 48 (39, 56) 0.7
    Unknown 11 7 4
Grade


0.9
    I 68 (34%) 35 (36%) 33 (32%)
    II 68 (34%) 32 (33%) 36 (35%)
    III 64 (32%) 31 (32%) 33 (32%)
Tumor Response 61 (32%) 28 (29%) 33 (34%) 0.5
    Unknown 7 3 4
1 Median (Q1, Q3); n (%)
2 Wilcoxon rank sum test; Pearson’s Chi-squared test
# Crude vs adjusted odds ratios
glm(response ~ trt + age + grade, data = trial, family = binomial) |>
  tbl_regression(exponentiate = TRUE)
Error in `check_pkg_installed()`:
! The package "broom.helpers" (>= 1.20.0) is required.

A big change between crude and adjusted estimates (often more than 10%) suggests confounding.

Learn more: epiR, MatchIt, gtsummary, Batra et al., The Epidemiologist R Handbook (free online), Rothman, Epidemiology: An Introduction

6.8 Survival analysis (time to event)

What it’s for: how long until an event (death, relapse), handling people who leave the study before the event (censoring).

library(survival); library(survminer)
lung2 <- lung |> mutate(status = status - 1, sex = factor(sex, labels = c("Male", "Female")))
km <- survfit(Surv(time, status) ~ sex, data = lung2)
summary(km, times = c(180, 365))
Call: survfit(formula = Surv(time, status) ~ sex, data = lung2)

                sex=Male 
 time n.risk n.event survival std.err lower 95% CI upper 95% CI
  180     89      49    0.644  0.0408        0.569        0.730
  365     35      36    0.336  0.0434        0.261        0.433

                sex=Female 
 time n.risk n.event survival std.err lower 95% CI upper 95% CI
  180     71      14    0.842  0.0387        0.770        0.922
  365     30      22    0.526  0.0597        0.421        0.658
ggsurvplot(km, conf.int = TRUE, risk.table = TRUE, pval = TRUE,
           xlab = "Days", title = "Kaplan-Meier survival by sex")

survdiff(Surv(time, status) ~ sex, data = lung2)        # log-rank test
Call:
survdiff(formula = Surv(time, status) ~ sex, data = lung2)

             N Observed Expected (O-E)^2/E (O-E)^2/V
sex=Male   138      112     91.6      4.55      10.3
sex=Female  90       53     73.4      5.68      10.3

 Chisq= 10.3  on 1 degrees of freedom, p= 0.001 
cox <- coxph(Surv(time, status) ~ sex + age + ph.ecog, data = lung2)
tidy(cox, exponentiate = TRUE, conf.int = TRUE)         # hazard ratios
# A tibble: 3 × 7
  term      estimate std.error statistic   p.value conf.low conf.high
  <chr>        <dbl>     <dbl>     <dbl>     <dbl>    <dbl>     <dbl>
1 sexFemale    0.575   0.168       -3.29 0.000986     0.414     0.799
2 age          1.01    0.00927      1.19 0.232        0.993     1.03 
3 ph.ecog      1.59    0.114        4.08 0.0000445    1.27      1.99 
cox.zph(cox)                                            # proportional hazards check
        chisq df    p
sex     2.305  1 0.13
age     0.188  1 0.66
ph.ecog 2.054  1 0.15
GLOBAL  4.464  3 0.22

Learn more: Zabor, Survival Analysis in R tutorial

6.9 Time series

What it’s for: a value tracked over regular time points (monthly ER visits, daily crisis calls). Watch for trend, seasonality, and autocorrelation.

library(forecast)
autoplot(AirPassengers) + labs(title = "Monthly airline passengers")

decompose(AirPassengers, type = "multiplicative") |> autoplot()

fit <- auto.arima(AirPassengers)
fit
Series: AirPassengers 
ARIMA(2,1,1)(0,1,0)[12] 

Coefficients:
         ar1     ar2      ma1
      0.5960  0.2142  -0.9819
s.e.  0.0888  0.0880   0.0292

sigma^2 = 132.3:  log likelihood = -504.92
AIC=1017.85   AICc=1018.17   BIC=1029.35
forecast(fit, h = 24) |> autoplot()

ggAcf(AirPassengers)                  # autocorrelation

Turn your own data into a time series: ts(df$count, start = c(2019, 1), frequency = 12).

Learn more: Hyndman and Athanasopoulos, Forecasting: Principles and Practice (free online)

7 Part 6: Finding Patterns

7.1 Principal component analysis (PCA)

What it’s for: squeezing many correlated numeric variables into a few components. Works on continuous variables.

pca <- prcomp(USArrests, scale. = TRUE)     # always scale when units differ
summary(pca)                                # variance explained
Importance of components:
                          PC1    PC2     PC3     PC4
Standard deviation     1.5749 0.9949 0.59713 0.41645
Proportion of Variance 0.6201 0.2474 0.08914 0.04336
Cumulative Proportion  0.6201 0.8675 0.95664 1.00000
pca$rotation                                # loadings: which variables drive each PC
                PC1        PC2        PC3         PC4
Murder   -0.5358995 -0.4181809  0.3412327  0.64922780
Assault  -0.5831836 -0.1879856  0.2681484 -0.74340748
UrbanPop -0.2781909  0.8728062  0.3780158  0.13387773
Rape     -0.5434321  0.1673186 -0.8177779  0.08902432
tibble(state = rownames(USArrests), as_tibble(pca$x)) |>
  ggplot(aes(PC1, PC2, label = state)) + geom_text(size = 3) +
  labs(title = "US states on the first two components")

screeplot(pca, type = "lines")

7.2 Latent class analysis (LCA)

What it’s for: finding hidden groups of people from categorical answers (like adversity patterns). PCA groups variables; LCA groups people.

library(poLCA)
data(values)       # 4 yes/no survey items, coded 1 and 2
f <- cbind(A, B, C, D) ~ 1
set.seed(10)
lca2 <- poLCA(f, values, nclass = 2, nrep = 5, verbose = FALSE)
lca3 <- poLCA(f, values, nclass = 3, nrep = 5, verbose = FALSE)
c(BIC_2 = lca2$bic, BIC_3 = lca3$bic)      # lower BIC is better
   BIC_2    BIC_3 
1057.313 1081.856 
lca2$P                                      # class sizes
[1] 0.7207525 0.2792475

Watch out for: poLCA needs categories coded 1, 2, 3 (no zeros). Fit several class counts and compare BIC.

Learn more: poLCA, James et al., An Introduction to Statistical Learning (free online)

8 Part 7: Sharing Your Work

8.1 Exporting plots and tables

ggsave("figure1.png", plot = base, width = 7, height = 5, dpi = 300)   # poster quality
ggsave("figure1.pdf", plot = base, width = 7, height = 5)
write_csv(cohort, "cohort.csv")
library(gt)
mtcars |> head() |> gt() |> tab_header(title = "First six cars") |> gtsave("table.html")
tbl_summary(trial) |> as_flex_table() |> flextable::save_as_docx(path = "table1.docx")

8.2 Making your own Quarto documents

A .qmd file has three parts:

  1. YAML header between --- lines: title, format, options
  2. Text written in markdown: **bold**, *italic*, # Heading, [link](url)
  3. Code chunks that start with three backticks and {r}

Chunk options go at the top of a chunk with #|:

Option Does
#| echo: false hide code, show output
#| eval: false show code, do not run
#| include: false run silently (good for setup)
#| warning: false hide warnings
#| fig-cap: "Caption" figure caption
#| fig-width: 7 figure size
#| label: fig-scatter lets you cross-reference with @fig-scatter

Output formats in the YAML:

format: html          # web page
format: docx          # Word
format: pdf           # needs tinytex: quarto install tinytex
format: revealjs      # slides

For slides, each ## heading is a new slide. Images: ![](photo.png){width=60%}. Render with the Render button or quarto render file.qmd in the terminal. Use here() for paths so the project works on any computer, and use RStudio Projects to keep each study in its own folder.

Learn more: Quarto guide, Quarto slides, Wickham et al., R for Data Science, 2e (free online)

8.3 R Shiny apps

What it’s for: interactive web apps where users pick options and plots update. Shiny needs a live R session, so this code is shown, not run. Paste it into a file called app.R and click Run App.

Every app has a UI (what users see), a server (what R calculates), and shinyApp() to join them.

library(shiny); library(ggplot2)

ui <- fluidPage(
  titlePanel("Explore mtcars"),
  sidebarLayout(
    sidebarPanel(
      selectInput("xvar", "X variable", choices = names(mtcars), selected = "wt"),
      selectInput("yvar", "Y variable", choices = names(mtcars), selected = "mpg"),
      sliderInput("size", "Point size", min = 1, max = 6, value = 3),
      checkboxInput("trend", "Add trend line", value = FALSE)
    ),
    mainPanel(plotOutput("scatter"), tableOutput("summary"))
  )
)

server <- function(input, output) {
  output$scatter <- renderPlot({
    p <- ggplot(mtcars, aes(.data[[input$xvar]], .data[[input$yvar]])) +
      geom_point(size = input$size) + theme_minimal()
    if (input$trend) p <- p + geom_smooth(method = "lm")
    p
  })
  output$summary <- renderTable({
    data.frame(stat = c("mean", "sd"),
               value = c(mean(mtcars[[input$yvar]]), sd(mtcars[[input$yvar]])))
  })
}

shinyApp(ui, server)

Inputs (input$name) come from the UI. Outputs use matching pairs: plotOutput() with renderPlot(), tableOutput() with renderTable(), textOutput() with renderText(). Share free at shinyapps.io with rsconnect::deployApp().

Learn more: Shiny tutorials, Wickham, Mastering Shiny (free online)

9 Part 8: Common Errors and Fixes

Error message Usual cause Fix
could not find function "x" package not loaded library(pkg)
there is no package called 'x' not installed install.packages("x")
object 'x' not found typo, or column used outside a dplyr verb check spelling, use df$x
non-numeric argument to binary operator numbers stored as text as.numeric()
arguments imply differing number of rows vectors of different lengths check length()
unexpected symbol missing comma or parenthesis check the line above
NAs introduced by coercion text that is not a number inspect with unique()
Result is all NA missing values na.rm = TRUE
Rows multiplied after join duplicate keys check count(df, id) |> filter(n > 1)
Error in plot.new(): figure margins too large plot pane too small enlarge pane
ggplot + error + placed at start of line put + at the end of the line

General tips: read the last line of the error first, restart R (Session, Restart R) when things act strange, and use ?function_name for help.

10 Sources

  • Wickham, Çetinkaya-Rundel, and Grolemund. R for Data Science, 2e. https://r4ds.hadley.nz
  • Wickham. ggplot2: Elegant Graphics for Data Analysis, 3e. https://ggplot2-book.org
  • Batra et al. The Epidemiologist R Handbook. https://epirhandbook.com
  • Lumley. Complex Surveys: A Guide to Analysis Using R. Wiley, 2010.
  • Hyndman and Athanasopoulos. Forecasting: Principles and Practice, 3e. https://otexts.com/fpp3/
  • van Buuren. Flexible Imputation of Missing Data, 2e. https://stefvanbuuren.name/fimd/
  • James et al. An Introduction to Statistical Learning. https://www.statlearning.com
  • Wickham. Mastering Shiny. https://mastering-shiny.org
  • CDC NHANES Tutorials. https://wwwn.cdc.gov/nchs/nhanes/tutorials/
  • Quarto documentation. https://quarto.org
  • Package documentation on CRAN and tidyverse.org, linked in each section