Individual Patient Data (IPD) Meta-Analysis

This skill teaches IPD meta-analysis, the "gold standard" of evidence synthesis that uses raw participant-level data from multiple studies.

Overview

IPD meta-analysis analyzes the original individual-level data from each study rather than summary statistics. This enables more powerful analyses, proper handling of time-to-event data, and exploration of patient-level effect modifiers.

When to Use This Skill

Activate this skill when users:

• Have access to individual patient data from multiple trials
• Want to explore subgroup effects or treatment-effect modifiers
• Need to analyze time-to-event (survival) outcomes
• Ask about one-stage vs two-stage approaches
• Want to standardize outcomes across studies
• Need to handle missing data properly

Core Concepts to Teach

1. IPD vs Aggregate Data Meta-Analysis

Comparison:

Aspect	Aggregate Data	IPD
Data level	Study summaries	Individual patients
Subgroup analysis	Ecological bias risk	Patient-level, unbiased
Time-to-event	Requires approximations	Exact analysis
Missing data	Cannot address	Can model properly
Standardization	Limited	Full flexibility
Effort	Low	High (data collection)

Socratic Questions:

• "Why might analyzing individual data give different results than combining averages?"
• "What is ecological bias and why does it matter for subgroup analyses?"
• "When would the extra effort of IPD collection be worthwhile?"

2. One-Stage vs Two-Stage Approaches

Two-Stage Approach:

Stage 1: Analyze each study separately
         → Get study-specific estimates

Stage 2: Combine estimates using standard MA
         → Pool using random effects

One-Stage Approach:

Single model: All data in one hierarchical model
              → Accounts for clustering within studies
              → More flexible for complex analyses

When to Use Each:

Situation	Recommended Approach
Simple outcomes, many studies	Two-stage (simpler)
Few studies, sparse data	One-stage (more stable)
Complex interactions	One-stage (more flexible)
Time-to-event	One-stage (preferred)
Non-linear effects	One-stage (necessary)

3. Two-Stage IPD Meta-Analysis

Stage 1 - Study-Level Analysis:

library(dplyr)
library(broom)

# Analyze each study separately
study_results <- ipd_data %>%
  group_by(study_id) %>%
  do(tidy(glm(outcome ~ treatment + age + sex, 
              data = ., 
              family = binomial))) %>%
  filter(term == "treatment")

# Extract treatment effects and SEs
effects <- study_results %>%
  select(study_id, estimate, std.error)

Stage 2 - Meta-Analysis:

library(metafor)

# Standard random-effects MA
ma_result <- rma(
  yi = effects$estimate,
  sei = effects$std.error,
  method = "REML"
)

summary(ma_result)
forest(ma_result)

4. One-Stage IPD Meta-Analysis

Mixed-Effects Model:

library(lme4)

# One-stage with random intercepts and slopes
model <- glmer(
  outcome ~ treatment + age + sex + 
    (1 + treatment | study_id),
  data = ipd_data,
  family = binomial
)

summary(model)

Interpretation:

• Fixed effects: Overall treatment effect adjusted for covariates
• Random intercepts: Study-specific baseline risks
• Random slopes: Study-specific treatment effects (heterogeneity)

For Time-to-Event:

library(survival)
library(coxme)

# Stratified Cox model (two-stage equivalent)
cox_stratified <- coxph(
  Surv(time, event) ~ treatment + age + sex + strata(study_id),
  data = ipd_data
)

# Frailty model (one-stage)
cox_frailty <- coxme(
  Surv(time, event) ~ treatment + age + sex + (1 | study_id),
  data = ipd_data
)

5. Exploring Treatment-Effect Modifiers

Why IPD is Essential:

• Aggregate data subgroups → ecological bias
• IPD → true patient-level interactions

Interaction Analysis:

# Test treatment-covariate interaction
model_interaction <- glmer(
  outcome ~ treatment * age_group + sex + 
    (1 + treatment | study_id),
  data = ipd_data,
  family = binomial
)

# Compare with main effects model
anova(model_main, model_interaction)

Visualization:

library(ggplot2)

# Forest plot by subgroup
ggplot(subgroup_effects, aes(x = estimate, y = subgroup)) +
  geom_point() +
  geom_errorbarh(aes(xmin = ci_low, xmax = ci_high), height = 0.2) +
  geom_vline(xintercept = 0, linetype = "dashed") +
  labs(x = "Treatment Effect (log OR)", y = "Subgroup")

6. Handling Missing Data

Common Approaches:

Method	Description	Assumption
Complete case	Exclude missing	MCAR (rarely true)
Single imputation	Fill with mean/mode	Underestimates uncertainty
Multiple imputation	Create multiple datasets	MAR
Pattern mixture	Model missingness	MNAR sensitivity

Multiple Imputation with IPD:

library(mice)

# Impute within each study
imputed_data <- ipd_data %>%
  group_by(study_id) %>%
  group_modify(~ {
    mice(.x, m = 20, method = "pmm", printFlag = FALSE) %>%
      complete("long")
  })

# Analyze each imputed dataset
results <- imputed_data %>%
  group_by(.imp) %>%
  do(tidy(glmer(outcome ~ treatment + (1|study_id), 
                data = ., family = binomial)))

# Pool results using Rubin's rules
pool(results)

7. Data Harmonization

Common Challenges:

• Different outcome definitions
• Different covariate coding
• Different follow-up times
• Different measurement scales

Harmonization Steps:

# Standardize variables across studies
harmonized <- ipd_data %>%
  mutate(
    # Standardize age (z-score within study)
    age_std = (age - mean(age)) / sd(age),
    
    # Harmonize outcome timing
    outcome_6mo = case_when(
      study_id == "A" ~ outcome_week24,
      study_id == "B" ~ outcome_month6,
      TRUE ~ outcome_6months
    ),
    
    # Recode categorical variables
    sex = case_when(
      sex %in% c("M", "male", "1") ~ "Male",
      sex %in% c("F", "female", "2") ~ "Female"
    )
  )

8. Reporting IPD Meta-Analysis

PRISMA-IPD Checklist Items:

• Data collection and integrity checking
• Proportion of IPD obtained vs available
• Handling of studies without IPD
• Missing data approach
• One-stage vs two-stage justification

Assessment Questions

1. Basic: "What is the main advantage of IPD over aggregate data meta-analysis?"

   • Correct: Avoids ecological bias in subgroup analyses; enables patient-level effect modifier exploration

2. Intermediate: "When would you choose a one-stage over a two-stage approach?"

   • Correct: Few studies, sparse events, complex interactions, time-to-event outcomes

3. Advanced: "How would you handle a situation where you have IPD for 60% of studies and only aggregate data for the rest?"

   • Guide: Combined IPD + AD analysis; sensitivity analysis comparing IPD-only vs combined

Common Misconceptions

1. "IPD always gives different results than aggregate MA"

   • Reality: Often similar for main effects; differs mainly for subgroups

2. "One-stage is always better than two-stage"

   • Reality: Two-stage is often sufficient and more transparent

3. "IPD eliminates all bias"

   • Reality: Still subject to selection bias, publication bias if not all trials share data

Example Dialogue

User: "I'm coordinating an IPD meta-analysis of 8 cancer trials. How do I analyze survival outcomes?"

Response Framework:

1. Congratulate on IPD collection effort
2. Discuss one-stage vs two-stage for survival
3. Recommend stratified Cox or frailty models
4. Address censoring and follow-up differences
5. Guide through effect modifier analysis
6. Discuss PRISMA-IPD reporting

References

• Riley RD et al. IPD Meta-Analysis. BMJ 2010
• Stewart LA, Tierney JF. IPD Meta-Analysis of Randomized Trials. Cochrane Handbook
• Debray TPA et al. Get real in IPD meta-analysis. BMC Med Res Methodol 2015
• PRISMA-IPD Statement

Adaptation Guidelines

Glass (the teaching agent) MUST adapt this content to the learner:

1. Language Detection: Detect the user's language from their messages and respond naturally in that language
2. Cultural Context: Adapt examples to local healthcare systems and research contexts when relevant
3. Technical Terms: Maintain standard English terms (e.g., "IPD", "one-stage", "frailty model") but explain them in the user's language
4. Level Adaptation: Adjust complexity based on user's demonstrated knowledge level
5. Socratic Method: Ask guiding questions in the detected language to promote deep understanding
6. Local Examples: When possible, reference studies or guidelines familiar to the user's region

Example Adaptations:

• 🇧🇷 Portuguese: Reference Brazilian IPD collaborations (e.g., oncology networks)
• 🇪🇸 Spanish: Include Latin American clinical trial networks
• 🇨🇳 Chinese: Reference Chinese IPD initiatives and data sharing policies

Related Skills

• meta-analysis-fundamentals - Basic concepts prerequisite
• data-extraction - Data collection principles
• heterogeneity-analysis - Understanding between-study variation
• bayesian-meta-analysis - Alternative modeling framework