When you're dealing with sensitive data—think healthcare records or financial transactions—simply deleting the "Name" column is rarely enough. I've seen plenty of analysts make the mistake of thinking they've anonymized a dataset, only to realize that a combination of birth date, zip code, and gender can uniquely identify a huge chunk of the population. We need to move beyond simple deletion and look at actual obfuscation.

How do I replace IDs with something random but still keep them linkable?

If you need to perform joins across multiple tables, you can't just delete the Patient ID. You need a "pseudonym." The best way to do this in R is through hashing. I usually reach for the digest package because it's reliable and straightforward. A hash takes an input and turns it into a fixed-length string that's mathematically impossible to reverse without the original key.

library(digest)
library(dplyr)

# Let's imagine a patient dataset
patients <- data.frame(
  patient_id = c("P101", "P102", "P103"),
  name = c("Alice Smith", "Bob Jones", "Charlie Brown"),
  diagnosis = c("Flu", "Cold", "Covid")
)

# We'll create a salt—a random string added to the ID to prevent 
# "rainbow table" attacks where someone guesses the ID.
salt <- "s3cr3t_k3y_2024"

patients_anon <- patients %>%
  mutate(patient_id = sapply(patient_id, function(x) {
    digest(paste0(x, salt), algo = "sha256")
  })) %>%
  select(-name) # Now we can safely drop the name

print(patients_anon)

What's the best way to handle dates and ages without giving away the person's identity?

Exact birth dates are a goldmine for re-identification. Unless your analysis specifically requires the exact day, you should "generalize" the data. I prefer binning. Instead of saying someone is 43, say they are in the "40-49" bracket. It preserves the statistical utility of the data while adding a layer of plausible deniability.

# Using a dataset with ages
clinical_data <- data.frame(
  patient_id = c("H1", "H2", "H3", "H4"),
  age = c(22, 25, 44, 48)
)

# I use cut() to create bins. 
# This transforms a precise number into a categorical range.
clinical_data_anon <- clinical_data %>%
  mutate(age_group = cut(age, 
                         breaks = seq(0, 100, by = 10), 
                         right = FALSE, 
                         labels = c("0-9", "10-19", "20-29", "30-39", "40-49", "50-59", "60-69", "70-79", "80-89", "90-99"))) %>%
  select(-age)

print(clinical_data_anon)

How do I stop "outliers" from leaking who a person is?

This is where things get tricky. Imagine a dataset of employees in a small company. Even if you remove names, if there is only one person in the "Executive" role who is over 60, everyone knows who that record belongs to. This is a failure of k-anonymity.

The simplest fix is suppression. If a specific combination of attributes (like Role + Age Group) appears fewer than, say, 3 times in your dataset, you mask those values. It's a bit aggressive, but it's the only way to be sure.

library(dplyr)

staff_data <- data.frame(
  role = c("Dev", "Dev", "Dev", "Dev", "CEO", "HR", "HR"),
  city = c("NY", "NY", "NY", "NY", "NY", "LA", "LA")
)

# Identify combinations that are too rare (less than 2 occurrences)
staff_anon <- staff_data %>%
  group_by(role, city) %>%
  mutate(count = n()) %>%
  ungroup() %>%
  mutate(role = ifelse(count < 2, "Other", role)) %>%
  select(-count)

# Notice how the 'CEO' became 'Other' because they were the only one in that group.
print(staff_anon)

📋 Practical Task

Exercise: Scrubbing the Clinical Trial Participant Registry

You have been handed a raw dataframe containing participants from a clinical trial. Your goal is to prepare this dataset for a third-party research team by applying the three techniques discussed in this lesson.

The Dataset:

trial_data <- data.frame(
  participant_id = c("T001", "T002", "T003", "T004", "T005"),
  full_name = c("John Doe", "Jane Roe", "Baby Yoda", "Sarah Connor", "Ellen Ripley"),
  age = c(24, 28, 21, 55, 82),
  zip_code = c("10001", "10001", "10001", "90210", "90210"),
  condition = c("Mild", "Mild", "Severe", "Mild", "Severe")
)

Your Requirements:

• Pseudonymize: Use the digest package to hash the participant_id. Use the salt "trial_secret_2024".
• Generalize: Convert the age column into 10-year bins (e.g., 20-29, 80-89).
• Suppress: Find any zip_code that appears fewer than 3 times in the dataset and replace the value with "MASKED".
• Cleanup: Remove the full_name column entirely.

Produce the final trial_data_final dataframe that meets all these security criteria.