Data Preparation for Survival Analysis

Overview

Proper data preparation is crucial for reliable survival analysis. This vignette provides comprehensive guidance on preparing data for use with jsurvival, covering common data issues and their solutions.

Essential Data Components

Required Variables

Every survival analysis requires at minimum:

1. Time variable: Time from study entry to event or censoring
2. Event indicator: Binary variable indicating whether event occurred

Optional Variables

• Grouping factors: For comparing survival between groups
• Continuous predictors: For cut-point analysis or multivariable models
• Date variables: For calculating time intervals
• Stratification variables: For adjusted analyses

Time Variable Preparation

Calculating Time from Dates

Using Built-in Time Interval Calculator

The jsurvival module includes a time interval calculator:

# In jamovi: Survival → Data Preparation → Time Interval Calculator
# Assign:
# - Start Date: Study entry date (e.g., diagnosis_date)
# - End Date: Event/censoring date (e.g., death_date or last_followup_date)
# - Output Unit: months or years

# This creates a new variable with calculated time intervals

Manual Calculation in R

library(lubridate)

# Example data with dates
clinical_data <- data.frame(
  patient_id = 1:100,
  diagnosis_date = as.Date("2020-01-01") + sample(0:365, 100, replace = TRUE),
  death_date = as.Date("2020-01-01") + sample(200:1000, 100, replace = TRUE),
  last_followup_date = as.Date("2023-12-31")
)

# Calculate time to death or censoring
clinical_data$death_occurred <- !is.na(clinical_data$death_date)
clinical_data$event_date <- ifelse(clinical_data$death_occurred,
                                   clinical_data$death_date,
                                   clinical_data$last_followup_date)
clinical_data$event_date <- as.Date(clinical_data$event_date, origin = "1970-01-01")

# Time in months
clinical_data$time_months <- as.numeric(
  difftime(clinical_data$event_date, clinical_data$diagnosis_date, units = "days")
) / 30.44  # Average days per month

# Time in years
clinical_data$time_years <- as.numeric(
  difftime(clinical_data$event_date, clinical_data$diagnosis_date, units = "days")
) / 365.25  # Accounting for leap years

Handling Time Zero Issues

Left Truncation

# Patients entering study at different times
# Example: analyzing from landmark time (e.g., 3 months post-diagnosis)

# Include only patients who survived to landmark
landmark_time <- 3  # months
eligible_data <- subset(clinical_data, time_months >= landmark_time)

# Adjust time origin to landmark
eligible_data$time_from_landmark <- eligible_data$time_months - landmark_time

Negative Time Values

# Check for and handle negative times
negative_times <- clinical_data$time_months < 0
if(any(negative_times, na.rm = TRUE)) {
  print(paste("Found", sum(negative_times, na.rm = TRUE), "negative time values"))
  
  # Options:
  # 1. Exclude these patients
  clinical_data <- clinical_data[!negative_times | is.na(negative_times), ]
  
  # 2. Set to small positive value
  # clinical_data$time_months[negative_times] <- 0.001
  
  # 3. Investigate data entry errors
  # View(clinical_data[negative_times, ])
}

Event Indicator Preparation

Creating Binary Event Variables

Death as Event

# From death status variable
clinical_data$death_binary <- ifelse(clinical_data$vital_status == "Dead", 1, 0)

# From death date
clinical_data$death_binary <- ifelse(is.na(clinical_data$death_date), 0, 1)

# Verify coding
table(clinical_data$death_binary, clinical_data$vital_status, useNA = "always")

Other Events

# Disease progression
clinical_data$progression <- ifelse(clinical_data$progression_status == "Progressive Disease", 1, 0)

# Recurrence
clinical_data$recurrence <- ifelse(!is.na(clinical_data$recurrence_date), 1, 0)

# Composite endpoint (death or progression, whichever comes first)
clinical_data$death_or_progression <- pmax(clinical_data$death_binary, 
                                          clinical_data$progression, 
                                          na.rm = TRUE)

Multiple Event Types

Competing Risks Setup

# Create separate indicators for different event types
clinical_data$death_cancer <- ifelse(clinical_data$death_cause == "Cancer", 1, 0)
clinical_data$death_other <- ifelse(clinical_data$death_cause == "Other" | 
                                   clinical_data$death_cause == "Cardiovascular", 1, 0)

# For patients still alive, both should be 0
alive_patients <- clinical_data$vital_status == "Alive"
clinical_data$death_cancer[alive_patients] <- 0
clinical_data$death_other[alive_patients] <- 0

# Verify: sum should equal total deaths
table(clinical_data$death_cancer + clinical_data$death_other, 
      clinical_data$vital_status)

Data Quality Checks

Consistency Checks

# Function to perform comprehensive data quality checks
check_survival_data <- function(data, time_var, event_var) {
  
  checks <- list()
  
  # 1. Missing values
  checks$missing_time <- sum(is.na(data[[time_var]]))
  checks$missing_event <- sum(is.na(data[[event_var]]))
  
  # 2. Negative or zero times
  checks$negative_times <- sum(data[[time_var]] <= 0, na.rm = TRUE)
  
  # 3. Event indicator values
  event_values <- unique(data[[event_var]])
  checks$valid_events <- all(event_values %in% c(0, 1, NA))
  checks$event_table <- table(data[[event_var]], useNA = "always")
  
  # 4. Extreme values
  checks$extreme_times <- list(
    max_time = max(data[[time_var]], na.rm = TRUE),
    min_time = min(data[[time_var]], na.rm = TRUE),
    outliers = sum(data[[time_var]] > quantile(data[[time_var]], 0.99, na.rm = TRUE), na.rm = TRUE)
  )
  
  # 5. Event rate
  checks$event_rate <- mean(data[[event_var]], na.rm = TRUE)
  
  return(checks)
}

# Run checks
data_quality <- check_survival_data(clinical_data, "time_months", "death_binary")
print(data_quality)

Outlier Detection

# Identify potential outliers in survival times
Q1 <- quantile(clinical_data$time_months, 0.25, na.rm = TRUE)
Q3 <- quantile(clinical_data$time_months, 0.75, na.rm = TRUE)
IQR <- Q3 - Q1

# Define outliers using 1.5 * IQR rule
lower_bound <- Q1 - 1.5 * IQR
upper_bound <- Q3 + 1.5 * IQR

outliers <- clinical_data$time_months < lower_bound | clinical_data$time_months > upper_bound
print(paste("Found", sum(outliers, na.rm = TRUE), "potential outliers"))

# Review outliers
if(sum(outliers, na.rm = TRUE) > 0) {
  outlier_data <- clinical_data[outliers & !is.na(outliers), 
                               c("patient_id", "time_months", "death_binary")]
  print(outlier_data)
}

Handling Missing Data

Missing Data Patterns

# Assess missing data patterns
library(VIM)  # Visualization and Imputation of Missing values

# Create missingness indicator
clinical_data$time_missing <- is.na(clinical_data$time_months)
clinical_data$event_missing <- is.na(clinical_data$death_binary)

# Missing data summary
missing_summary <- clinical_data %>%
  summarise(
    n_total = n(),
    time_missing = sum(time_missing),
    event_missing = sum(event_missing),
    both_missing = sum(time_missing & event_missing),
    complete_cases = sum(!time_missing & !event_missing)
  )

print(missing_summary)

Strategies for Missing Data

Complete Case Analysis

# Use only patients with complete survival data
complete_data <- clinical_data[!is.na(clinical_data$time_months) & 
                              !is.na(clinical_data$death_binary), ]

print(paste("Complete cases:", nrow(complete_data), "of", nrow(clinical_data)))

Administrative Censoring

# For missing death dates, censor at last known follow-up
clinical_data$time_months_imputed <- clinical_data$time_months
clinical_data$death_binary_imputed <- clinical_data$death_binary

# If death status missing but last follow-up available, assume alive
missing_death <- is.na(clinical_data$death_binary) & !is.na(clinical_data$last_followup_date)
if(any(missing_death)) {
  clinical_data$death_binary_imputed[missing_death] <- 0
  clinical_data$time_months_imputed[missing_death] <- 
    as.numeric(difftime(clinical_data$last_followup_date[missing_death],
                       clinical_data$diagnosis_date[missing_death], 
                       units = "days")) / 30.44
}

Explanatory Variable Preparation

Categorical Variables

Factor Ordering

# Ensure proper ordering for categorical variables
clinical_data$tumor_grade <- factor(clinical_data$tumor_grade, 
                                   levels = c("Grade 1", "Grade 2", "Grade 3", "Grade 4"),
                                   ordered = TRUE)

clinical_data$stage <- factor(clinical_data$stage,
                             levels = c("Stage I", "Stage II", "Stage III", "Stage IV"),
                             ordered = TRUE)

# For unordered factors, set reference level
clinical_data$treatment <- factor(clinical_data$treatment)
clinical_data$treatment <- relevel(clinical_data$treatment, ref = "Standard")

Handling Rare Categories

# Combine rare categories to ensure adequate sample sizes
table(clinical_data$histology)

# Combine categories with <10 patients
rare_histologies <- names(table(clinical_data$histology))[table(clinical_data$histology) < 10]
clinical_data$histology_grouped <- as.character(clinical_data$histology)
clinical_data$histology_grouped[clinical_data$histology_grouped %in% rare_histologies] <- "Other"
clinical_data$histology_grouped <- factor(clinical_data$histology_grouped)

table(clinical_data$histology_grouped)

Continuous Variables

Age Preparation

# Check age distribution
summary(clinical_data$age_at_diagnosis)

# Create age groups if needed
clinical_data$age_group <- cut(clinical_data$age_at_diagnosis,
                              breaks = c(0, 50, 65, 80, Inf),
                              labels = c("≤50", "51-65", "66-80", ">80"),
                              include.lowest = TRUE)

# Age as decades
clinical_data$age_decades <- floor(clinical_data$age_at_diagnosis / 10) * 10

Biomarker Preparation

# Log transformation for skewed biomarkers
clinical_data$log_biomarker <- log10(clinical_data$biomarker_value + 1)

# Standardization (z-score)
clinical_data$biomarker_zscore <- scale(clinical_data$biomarker_value)[,1]

# Handle extreme values
biomarker_99th <- quantile(clinical_data$biomarker_value, 0.99, na.rm = TRUE)
clinical_data$biomarker_winsorized <- pmin(clinical_data$biomarker_value, biomarker_99th)

Data Validation

Cross-Validation with Source Data

# Compare calculated times with expected ranges
validation_checks <- function(data) {
  issues <- list()
  
  # Check if all events occurred before end of study
  study_end <- as.Date("2023-12-31")
  late_events <- data$event_date > study_end & data$death_binary == 1
  issues$late_events <- sum(late_events, na.rm = TRUE)
  
  # Check if follow-up times are reasonable
  unreasonable_followup <- data$time_months > 240  # > 20 years
  issues$long_followup <- sum(unreasonable_followup, na.rm = TRUE)
  
  # Check for impossible age at events
  impossible_age <- data$age_at_diagnosis < 0 | data$age_at_diagnosis > 120
  issues$impossible_age <- sum(impossible_age, na.rm = TRUE)
  
  return(issues)
}

validation_results <- validation_checks(clinical_data)
print(validation_results)

Final Dataset Preparation

# Create final analysis dataset
analysis_data <- clinical_data %>%
  # Include only complete cases for primary analysis
  filter(!is.na(time_months), !is.na(death_binary)) %>%
  # Remove patients with negative or zero follow-up time
  filter(time_months > 0) %>%
  # Remove patients with missing key covariates
  filter(!is.na(age_at_diagnosis), !is.na(stage)) %>%
  # Select relevant variables
  select(
    patient_id,
    time_months,
    death_binary,
    age_at_diagnosis,
    age_group,
    sex,
    stage,
    tumor_grade,
    treatment,
    biomarker_value,
    histology_grouped
  )

# Final data summary
cat("Final analysis dataset:\n")
cat("N =", nrow(analysis_data), "\n")
cat("Events =", sum(analysis_data$death_binary), "\n")
cat("Event rate =", round(mean(analysis_data$death_binary) * 100, 1), "%\n")
cat("Median follow-up =", round(median(analysis_data$time_months), 1), "months\n")

Export for jamovi

CSV Export

# Export cleaned data for use in jamovi
write.csv(analysis_data, "survival_analysis_data.csv", row.names = FALSE)

# Include data dictionary
data_dictionary <- data.frame(
  variable = names(analysis_data),
  description = c(
    "Patient identifier",
    "Time from diagnosis to death or censoring (months)",
    "Death indicator (1=death, 0=alive/censored)",
    "Age at diagnosis (years)",
    "Age group categories",
    "Patient sex",
    "Cancer stage",
    "Tumor grade",
    "Treatment received",
    "Biomarker value",
    "Histology type (grouped)"
  ),
  type = c(
    "Identifier",
    "Continuous",
    "Binary",
    "Continuous", 
    "Categorical",
    "Categorical",
    "Ordinal",
    "Ordinal",
    "Categorical",
    "Continuous",
    "Categorical"
  )
)

write.csv(data_dictionary, "data_dictionary.csv", row.names = FALSE)

Quality Assurance Checklist

Pre-Analysis Checklist

Before proceeding with survival analysis, verify:

• Time variable contains no negative or zero values
• Event indicator is properly coded (0/1 or No/Yes)
• Missing data patterns have been assessed and addressed
• Categorical variables have appropriate reference levels
• Continuous variables have been checked for outliers
• Data types are correct (numeric, factor, date)
• Sample size is adequate (rule of thumb: 10 events per variable)
• Follow-up time is sufficient for meaningful analysis

Common Data Preparation Mistakes

1. Incorrect time calculation: Using calendar time instead of time from study entry
2. Wrong event coding: Mixing up 0 and 1 for events and censoring
3. Including impossible values: Negative times, events after last follow-up
4. Ignoring competing risks: Treating all deaths as same event type
5. Improper handling of missing data: Excluding too many patients without justification
6. Not checking assumptions: Ignoring data distribution and outliers

Conclusion

Proper data preparation is essential for reliable survival analysis results. Key principles include:

1. Careful time calculation with appropriate handling of dates and time zero
2. Consistent event coding with validation against source data
3. Comprehensive quality checks to identify and resolve data issues
4. Appropriate handling of missing data based on missingness patterns
5. Proper preparation of explanatory variables with attention to clinical meaning

Taking time for thorough data preparation will ensure more reliable and interpretable survival analysis results in jsurvival.