• This course introduces R, which is both
  • A programming language
  • A software
• We only have three hours. Therefore, you will not be manipulating data in this class. Instead, the course provides you with a toolbox
  • Researchers mostly follow the same steps in each data analysis project: import data, clean and prepare data, describe data, analyse data, plot data, export the plots and results
  • Slides represent a function that is essential for the first three steps
  • (Free) online resources are provided for the last three steps
• My aim is for you to grasp the fundamentals of R so that you can reuse the functions I demonstrate and train yourself using free online resources

• R is an important tool for many data analysts who use it to prepare, analyse and plot data, as well as create websites, reports and presentations (like this one, which was made entirely in R). But most importantly, it
  • is free
  • is open-source
  • allows anyone to reproduce your results (see, for instance, the Institute For Replication)
  • is very powerful and flexible
  • has a very engaged community that you can ask for help

• R is basically a big fancy calculator

• R is basically a big fancy calculator

• Pro tip: the first time you use RStudio, go to Tools –> Global Options… and
  1. Untick the box Restore .RData into workspace at startup
  2. Set Save workspace to .RData on exit to Never
  • I really advise you to do this so that each time you open a new R session, you start with a brand new, empty environment. Imagine still having loaded data named expe1 and not understanding why the right columns are not showing? Been there, done that

• R is essentially a big, fancy calculator
  • Do you want to find \(\hat{\boldsymbol{\beta}} = \underset{\boldsymbol{\beta}}{\arg\min} \; \sum_{i=1}^{n} \big(y_i - \mathbf{x}_i^\top \boldsymbol{\beta}\big)^2\) ? It can do that for you — if you know how to ask!
• R works with objects. Objects can be (almost) anything
  • A single number or multiple numbers
  • A single character or character string (a word, multiple words, a sentence, multiple sentences, etc.)
  • A data frame comprising numbers and character strings
  • A plot
  • The output of a Generalised Random Forest
  • And, most importantly, all of the above, nested like Russian dolls 🪆
• Combine a calculator with objects and you get the essence of R: a large calculator that manipulates objects

• Functions are objects that
  • Take an input with arguments
  • Follow a series of instructions given arguments
  • Provide an output given arguments
• For example, imagine that you want to square all the values in an object

x <- c(10, 4, 7, 12, 3, 18, 4, 12, 8) # Manually define an object x that takes multiple values

sqrt(x = x)

[1] 3.162278 2.000000 2.645751 3.464102 1.732051 4.242641 2.000000 3.464102
[9] 2.828427

• Functions are objects that
  • Take an input with arguments
  • Follow a series of instructions given arguments
  • Provide an output given arguments
• For example, imagine you want to find the lowest value in a given column, add its line number, and then multiply the result by the time of day. You could write a function to perform this task

x <- c(10, 4, 7, 12, 3, 18, 4, 12, 8) # Manually define an object x that takes multiple values

min_x <- function(x) {                          # Take the object x as an input
  i <- which.min(x)                             # Find the smallest value of x
  val <- x[i] + i                               # Add the line number to the smallest value
  val * as.numeric(format(Sys.time(), "%H"))    # Multiply the result by the hour of the day (0–23)
}

min_x(x = x)

[1] 104

• Packages are collections of functions, dataframes and documentation that come together
  • Some packages are built into R and are part of what is called base R
  • Some packages need to be downloaded. These are known as contributed packages and are extremely useful
• Very few people work in base R. Those who do so do so because they
  • Need the full flexibility of R. For instance, Offer-Westort, Rosenzweig, and Athey (2024) wrote an adaptive assignment algorithm based on balanced linear Thompson sampling in base R, and their code can be accessed online
  • Need full control and reproducibility over their analyses. National statistics organisations such as INSEE mostly use base R because they cannot afford their statistics to be altered if an update to a package changes its calculation method

install.packages("tidyverse") # Install a package
library("tidyverse")          # Load    a package

• Packages can be very powerful. Make the most of them!
  • At least, they come with a Github. Just do a Google Search, and you will find their Github. Unfortunately, the github contains more or less informations about the functions and their arguments
  • Here, the dplyr package

• Most of the data that you’ll use will be in the same formats
  • CSV or TXT. The most used ones for psychologists. Easy to share, light, easy to import/export.
  • XLSX. Excel’s format. Also easy to share and to export, although a bit more heavy. Does not have an advantaged over CSV to share a dataframe
  • SAV, or SAS, or DTA. Respectively, SPSS’s, SAS’, and Stata’s formats. You’ll learn how to use it when you need to, using the haven package
  • PARQUET, a file format specifically designed for (very) large data sets (more information here). Again, you’ll learn how to use it when you need to, using the nanoparquet package
• Why almost ready ? Because we have not yet talked about the Working directory
  • R needs to be told everything, including were to load the data from. My best advice is to work with R projects
  • R Projects are .Rproj files that make your life much easier. Essentially, it tells R what your working directory is, so that you can work with a folder (and subfolders) for each of your data analysis project

• To define a new R Project, go to File –> New Project… and follow the steps

• An alternative is to define the working directory manually. I am not fond of this approach, as it can be cumbersome and vulnerable to even minor changes in the directory file

setwd("C:/Users/jaime/OneDrive/Documents/Emplois/Enseignements/Paris-Cité/2025 - Introduction à R/Support")

• I have downloaded a dataframe and saved it is in the folder Data that is in my working directory
  • Dataframe is called penguins.csv and contains data about 344 penguins (Horst, Hill, and Gorman 2020). You can also get it there

df <- read.table(file = "Data/penguins.csv", 
                 header = TRUE, 
                 sep = ",", 
                 row.names = "X")

• datasummary_skim() describes the object, with a visual aid for numerical data

library(modelsummary)
datasummary_skim(data = df, type = "numeric")

	Unique	Missing Pct.	Mean	SD	Min	Median	Max	Histogram
bill_length_mm	165	1	43.9	5.5	32.1	44.5	59.6
bill_depth_mm	81	1	17.2	2.0	13.1	17.3	21.5
flipper_length_mm	56	1	200.9	14.1	172.0	197.0	231.0
year	3	0	2008.0	0.8	2007.0	2008.0	2009.0

• There are six different types of vector, each relating to a different kind of data
  • Numeric. Mostly, it is numbers with a decimal
  • Integer. Whole numbers (no decimal)
  • Logical. TRUE or FALSE data. Can also take values NA or not, representing Missing Values
  • Complex. Complex numbers. We won’t cover it, my brain already hurts (\(i^2 = -1 \text{ ?!}\))
  • Character. String values, whether letters, words, sentences… When ordered, characters become factors (for instance, Never; Sometimes; Always is an ordered factor, or ordered character)
  • Raw. We won’t cover it, as it is not used 99.99% of the time
• These types of data are stored in vectors, which are R’s fundamental data structure. When it comes to data, in R, everything is either a vector or an aggregate of vectors

• R is usually able to assign the correct data type to a vector. However, as you saw, it failed for the body_mass_g column, treating it as a character rather than a numeric value
  • Question: Why is this a problem?
  • Let’s reassign it. Remember when I said there are base R and contributed packages? I’ll show you two ways to do the same operation

df$body_mass_g <- as.integer(x = df$body_mass_g)

str(object = df$body_mass_g)

 int [1:344] 3750 3800 3250 NA 3450 3650 3625 4675 3475 4250 ...

df <- df %>%
  mutate(body_mass_g = as.integer(body_mass_g))
str(object = df$body_mass_g)

 int [1:344] 3750 3800 3250 NA 3450 3650 3625 4675 3475 4250 ...

summary(object = df$body_mass_g)

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
   2700    3550    4050    4202    4750    6300       2 

• Again, everything can be done in base R. But, some manipulations are highly cumbersome in base R. To ease data wrangling, a suite of packages (and functions) were made under the name tidyverse
  • The tidyverse follows a philosophy of data (Wickham 2014)
  1. Each variable forms a column
  2. Each observation forms a row (not a person: an observation)
  3. Each type of observational unit forms a table

install.packages("tidyverse") # Install a package
library("tidyverse")          # Load    a package

• R works using nested expressions from the inside out. It can get quite ridiculous somehow. To avoid doing so, a pipe operator has been created

mean(x = 
  sqrt(x = 
    abs(x = 
      log(x = 
        as.numeric(x = 
          factor(x = c(-10, 0, 10, 100), 
                 levels = c(-10, 0, 10, 100))
        )
      )
    )
  )
)

[1] 0.7645279

c(-10, 0, 10, 100) %>%                         # Create a vector
  factor(levels = c(-10, 0, 10, 100)) %>%      # convert to factor
  as.numeric() %>%                             # factor → numeric
  log() %>%                                    # take log
  abs() %>%                                    # absolute value
  sqrt() %>%                                   # square root
  mean()                                       # mean of the result

[1] 0.7645279

• The %>% operator takes the object on its left side and makes it the first argument of the function on its right side. Said differently, R can now work linearly!
• Another pipe operator |> also exists. You can find a summary of the main differences here. Personally, I use the %>% operator out of habit, and also because it is easier to use with an AZERTY keyboard

• For example, what if I prefer to count in pounds rather than grams?

df$body_mass_pounds <- df$body_mass_g / 453.59237

df <- df %>% 
  mutate(body_mass_pounds = body_mass_g / 453.59237)

• Get the mean and the standard deviation of this new column

mean(x = df$body_mass_pounds)

[1] NA

sd(x = df$body_mass_pounds)

[1] NA

• R needs to be told what to do with missing values (NAs)

mean(x = df$body_mass_pounds, na.rm = TRUE)

[1] 9.263283

sd(x = df$body_mass_pounds, na.rm = TRUE)

[1] 1.768007

• We sometimes need to pivot data in order to run a repeated measures analysis

• We sometimes need to pivot data in order to run a repeated measures analysis

df_wide %>%
  pivot_longer(
    cols = c(before,
             after),
    names_to = "time",
    values_to = "value")

• R has a built-in table() function. It won’t get you very far, but it is sometimes enough to quickly skim through your data (check how many categories a column has; whether some categories need to be merged together; the number of observations per category…)

df %>%
  select(sex, species) %>%
  table()

        species
sex      Adelie Chinstrap Gentoo
  female     73        34     58
  male       73        34     61

• You can get the percentages by adding the prop.table() function

df %>%
  select(sex, species) %>%
  table() %>%
  prop.table() * 100

        species
sex        Adelie Chinstrap   Gentoo
  female 21.92192  10.21021 17.41742
  male   21.92192  10.21021 18.31832

df %>%

  summarize(
    mean = mean(body_mass_g),
    sd = sd(body_mass_g),
    n = n(),
    se = sd / sqrt(n),
    confint_low = mean - 1.96 * se,
    confint_high = mean + 1.96 * se,
    .by = sex)

     sex     mean       sd   n       se confint_low confint_high
1   male 4545.685 787.6289 168 60.76689    4426.581     4664.788
2 female 3862.273 666.1720 165 51.86142    3760.624     3963.921
3   <NA>       NA       NA  11       NA          NA           NA

df %>%
  drop_na() %>%
  summarize(
    mean = mean(body_mass_g),
    sd = sd(body_mass_g),
    n = n(),
    se = sd / sqrt(n),
    confint_low = mean - 1.96 * se,
    confint_high = mean + 1.96 * se,
    .by = sex)

     sex     mean       sd   n       se confint_low confint_high
1   male 4545.685 787.6289 168 60.76689    4426.581     4664.788
2 female 3862.273 666.1720 165 51.86142    3760.624     3963.921

• In this case, applying the drop_na() function to the entire dataset does not alter the results. However, most of the time it will. It would be better to do the following. Why?

df %>%
  select(body_mass_g, sex) %>%
  drop_na() %>%
  summarize(
    mean = mean(body_mass_g),
    sd = sd(body_mass_g),
    n = n(),
    se = sd / sqrt(n),
    confint_low = mean - 1.96 * se,
    confint_high = mean + 1.96 * se,
    .by = sex)

     sex     mean       sd   n       se confint_low confint_high
1   male 4545.685 787.6289 168 60.76689    4426.581     4664.788
2 female 3862.273 666.1720 165 51.86142    3760.624     3963.921

• You might be familiar with good old \(Y_i = \alpha + \beta X_i + \varepsilon_i\), also called the regression by its friends. Let’s fit it, with one, two, and three predictors, with interactions, and have a look at the results

\[ \begin{align} \text{Model 1: } Y_i &= \alpha + \varepsilon_i \\ \text{Model 2: } Y_i &= \alpha + \beta \, X_{i,\text{flipper length}} + \varepsilon_i \\ \text{Model 3: } Y_i &= \alpha + \beta \, X_{i,\text{flipper length}} + \lambda \, X_{i,\text{bill depth}} + \varepsilon_i \\ \text{Model 4: } Y_i &= \alpha + \beta \, X_{i,\text{flipper length}} + \lambda \, X_{i,\text{bill depth}} + \gamma \, (X_{i,\text{flipper length}} \times X_{i,\text{bill depth}}) + \varepsilon_i \end{align} \]

lm1 <- lm(formula = body_mass_g ~ 1, data = df)
lm2 <- lm(body_mass_g ~ 1 + flipper_length_mm, df)
lm3 <- lm(body_mass_g ~ 1 + flipper_length_mm + bill_depth_mm, df)
lm4 <- lm(body_mass_g ~ 1 + flipper_length_mm + bill_depth_mm + flipper_length_mm:bill_depth_mm, df)

• ANOVA: the afex package (see its Github)
• Generalized linear model: the glm() function (see here)
• Random effects model: the lme4 package. See Bates et al. (2015) and Kuznetsova, Brockhoff, and Christensen (2017)
• Fixed effects (panel) model: fixest (Bergé 2018) or plm (Croissant and Millo 2008) packages
• Marginal means: the marginaleffects package (Arel-Bundock, Greifer, and Heiss 2024). See also the accompanying free e-book Model to Meaning, as well as Rohrer & Arel-Bundock (2025)
• Plot data: the marvellous ggplot2 package (Wickham 2016, with the free e-books ggplot2: Elegant Graphics for Data Analysis and R Graphics Cookbook)
• Sandwich (or other kind of) standard errors: the sandwich package (Zeileis, Köll, and Graham 2020), which is incorporated in modelsummary (see its vcov argument here)
• Power analysis: the pwr package (see its vignette)
• Psychometrics: the lavaan package for EFAs, CFAs, SEMs… (Rosseel 2012, with its website)
• Survey data: the survey package (Lumley 2004, with online vignettes and a nice free e-book)
• To check that your model is performing appropriately: the performance package (Lüdecke et al. 2021 and its vignettes)

• What should you do if your code isn’t working?
  • Things to avoid: never use a GPT to correct your code. Of course, it can solve and even write the code for you. However, getting the answer does not constitute learning. R coding is not easy: it requires effort and is challenging. Taking the easy road is not a solution
  • Things to do: The most important packages (such as the tidyverse) have a really nice webpage. They sometimes come with a Cheat Sheet (here for dplyr and here for tidyr). They might even have an accompanying book (Wickham, Cetinkaya-Rundel, and Grolemund 2023) which can be freely accessed online
• What to do when you do not know how to do what you are aiming to do?
  • R is a highly logical operator. First thing to do is to ask yourself: (1) Where am I? (2) Where do I want to go, very precisely? (3) What are the logical steps that I should follow to go from (1) to (2)? Doing so, you’ll break the steps down to small bits, which are more manageable
  • Finally, the vast majority of your questions have already been answered online. Try Stack Overflow, which is dedicated to programming and has a section specifically for R. You can also try Cross Validated, which is dedicated to statistics. Look for answers before asking a new question!