Data Analysis Using R: A Workflow That Runs from Start to Finish

1. From isolated scripts to an analysis project

A script that runs on its author's computer may fail elsewhere because of absolute paths, leftover Environment objects, unrecorded package versions or data edited outside code.

RStudio Projects provide stable root folders for code, data and outputs. Posit documentation recommends projects to separate working contexts and manage related files (Posit Software, PBC, n.d.-a).

2. Folder structure and immutable data

Keep raw data in data/raw without overwriting it. Scripts create analysis data in data/derived and automatically generate tables and figures in output. This clearly separates inputs, transformations and products.

Do not store passwords, API keys or personally identifying data in code repositories. Keep sensitive data outside public repositories and explain access arrangements in README.

3. Start with a clean session

RStudio can restore earlier workspaces. Although convenient during exploration, hidden objects may let scripts succeed despite missing data-creation steps. Disable automatic .RData restoration and make scripts create every required object.

RStudio documentation describes workspace and startup settings. Agree these settings within the team and document them in README (Posit Software, PBC, n.d.-b).

# Run after opening the correct RStudio Project
getwd()

required_dirs <- c(
  "data/raw",
  "data/derived",
  "output/tables",
  "output/figures"
)

for (path in required_dirs) {
  if (!dir.exists(path)) dir.create(path, recursive = TRUE)
}

set.seed(20261025)

4. Import and check the data schema

Immediately check rows, names, types, unique IDs, ranges and missing counts after import. Do not wait for a model error before returning to these checks.

The example uses base R functions so readers can run it without installing packages. Code 99 becomes NA at import because the codebook defines it as user-missing.

raw <- read.csv(
  "data/raw/survey.csv",
  na.strings = c("", "NA", "99"),
  stringsAsFactors = FALSE
)

stopifnot("id" %in% names(raw))
stopifnot(!anyDuplicated(raw$id))

str(raw)
summary(raw)
colSums(is.na(raw))

likert_items <- c(
  paste0("satisfaction", 1:4),
  paste0("trust", 1:4)
)

range_errors <- vapply(
  raw[likert_items],
  function(x) sum(!is.na(x) & !(x %in% 1:5)),
  integer(1)
)

range_errors

5. Declare data types and category labels

Numeric columns are not necessarily continuous. Gender codes 1, 2 and 3 are category labels. Converting them to factors prevents R from treating code differences as quantitative distances.

Factor ordering affects model reference groups. Specify levels and labels through code rather than relying on their order of appearance in a file.

analysis <- raw

analysis$gender <- factor(
  analysis$gender,
  levels = c(1, 2, 3),
  labels = c("Woman", "Man", "Self-described")
)

analysis$satisfaction2_r <-
  ifelse(
    is.na(analysis$satisfaction2),
    NA,
    6 - analysis$satisfaction2
  )

6. Calculate scale scores with transparent missingness rules

Using rowMeans with na.rm = TRUE can create a score from only one remaining indicator. This is often inappropriate for scales intended to cover several contents. The example requires at least three of four valid indicators.

The three-of-four rule is illustrative. Prespecify scoring from scale design, missingness and the analysis plan before examining results.

score_if_enough <- function(data, variables, min_valid) {
  values <- data[variables]
  n_valid <- rowSums(!is.na(values))
  score <- rowMeans(values, na.rm = TRUE)
  score[n_valid < min_valid] <- NA_real_
  score
}

analysis$satisfaction_mean <- score_if_enough(
  analysis,
  c("satisfaction1", "satisfaction2_r",
    "satisfaction3", "satisfaction4"),
  min_valid = 3
)

analysis$trust_mean <- score_if_enough(
  analysis,
  paste0("trust", 1:4),
  min_valid = 3
)

7. Describe data and generate diagnostic plots

Descriptive tables should show valid counts, means, standard deviations, minima and maxima. Distribution plots reveal unusual values or structures that one summary cannot represent.

Do not use plots as automatic tests. Interpret them alongside measurement design and modeling objectives.

describe_one <- function(x) {
  c(
    n = sum(!is.na(x)),
    mean = mean(x, na.rm = TRUE),
    sd = sd(x, na.rm = TRUE),
    min = min(x, na.rm = TRUE),
    max = max(x, na.rm = TRUE)
  )
}

descriptives <- sapply(
  analysis[c("satisfaction_mean", "trust_mean", "outcome")],
  describe_one
)

write.csv(
  descriptives,
  "output/tables/descriptives.csv"
)

png("output/figures/outcome_histogram.png",
    width = 1600, height = 900, res = 150)
hist(analysis$outcome,
     main = "Distribution of outcome",
     xlab = "Outcome")
dev.off()

8. Estimate models, confidence intervals and diagnostics

The illustrative regression below has a continuous outcome. This is not a default choice for all data. Formulas, control variables and missing-data handling must match the research question.

summary provides estimates and standard errors under model assumptions. confint adds confidence intervals. Diagnostic plots help inspect functional form, residuals and influential observations, but conclusions require several evidence sources.

fit <- lm(
  outcome ~ satisfaction_mean + trust_mean + age,
  data = analysis
)

summary(fit)
confint(fit, level = 0.95)

png("output/figures/model_diagnostics.png",
    width = 1600, height = 1600, res = 150)
par(mfrow = c(2, 2))
plot(fit)
dev.off()

model_table <- cbind(
  estimate = coef(fit),
  confint(fit, level = 0.95)
)

write.csv(
  model_table,
  "output/tables/model_coefficients.csv"
)

9. Record the environment and rerun everything

R and package versions form part of the method. sessionInfo records operating system, R version and loaded packages; official manuals document base R function behavior (R Core Team, n.d.). For projects with many packages, lockfiles describe dependency versions.

Reproducibility also requires a clean-session check. Wilson et al. (2017) recommend automating repeated steps and documenting analysis dependencies.

write.csv(
  analysis,
  "data/derived/survey_analysis.csv",
  row.names = FALSE,
  na = ""
)

capture.output(
  sessionInfo(),
  file = "output/session-info.txt"
)

# Final check:
# 1. Restart R
# 2. Open the project
# 3. Run scripts in the order specified in README
# 4. Compare regenerated tables and figures

10. Reporting template and handover checklist

• Scripts run from a clean R session.
• No absolute paths point to personal computers.
• Raw data are not overwritten.
• Types, ranges, IDs and missingness are checked through code.
• Scoring rules are explicit.
• Tables and figures are generated from analysis objects.
• Seeds are set for random steps.
• sessionInfo and package versions are retained.
• Sensitive data and credentials remain outside code repositories.