# Real World Evidence

> Real-world evidence analysis in R, including target trial emulation, propensity scores, external controls, and bias analysis.

- **Type:** Skill
- **Install:** `agentstack add skill-choxos-biostatagent-real-world-evidence`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [choxos](https://agentstack.voostack.com/s/choxos)
- **Installs:** 0
- **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [choxos](https://github.com/choxos)
- **Source:** https://github.com/choxos/BiostatAgent/tree/main/plugins/r-tidy-modeling/skills/real-world-evidence

## Install

```sh
agentstack add skill-choxos-biostatagent-real-world-evidence
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# Real-World Evidence Analysis in R

## Overview

Methods for analyzing real-world data (RWD) to generate real-world evidence (RWE). Covers target trial emulation, comparative effectiveness research, propensity score methods for observational data, external control arms, bias quantification, and sensitivity analysis for unmeasured confounding.

## Target Trial Emulation

### Conceptual Framework

```r
# Target trial emulation framework
# Specify the target trial protocol, then emulate using observational data

# Key elements to specify:
# 1. Eligibility criteria
# 2. Treatment strategies
# 3. Assignment procedures
# 4. Follow-up period
# 5. Outcome
# 6. Causal contrast (ITT, per-protocol, etc.)
# 7. Analysis plan
```

### Using TrialEmulation Package

```r
library(TrialEmulation)

# Prepare data for target trial emulation
# Data should be in long format with time-varying covariates

# Example: Clone-censor-weight approach
trial_data 
  group_by(patient_id) |>
  mutate(
    # Check eligibility at each visit
    eligible = check_eligibility(age, lab_values, prior_treatment),
    # Time zero for each potential trial
    trial_start = if_else(eligible & treatment_initiated, visit_date, NA)
  ) |>
  ungroup()

# Step 2: Clone patients for each trial they're eligible for
# Step 3: Apply artificial censoring per assigned strategy
# Step 4: Weight for selection and artificial censoring
# Step 5: Pool and analyze
```

## Propensity Score Methods for RWE

### Propensity Score Estimation

```r
library(MatchIt)
library(WeightIt)
library(cobalt)

# Estimate propensity scores
ps_model  select(age, sex, bmi, smoking, diabetes, baseline_egfr),
  Q.SL.library = c("SL.glm", "SL.ranger", "SL.xgboost"),
  g.SL.library = c("SL.glm", "SL.ranger"),
  k_split = 5,
  verbose = FALSE
)

aipw_result$fit()$summary()

# Risk difference and risk ratio
aipw_result$summary()
```

## External Control Arms

### Historical Controls Integration

```r
# Combine trial data with external controls

# Step 1: Identify comparable external controls
external_controls 
  filter(
    # Match eligibility criteria
    age >= 18 & age = 30,
    no_prior_treatment == TRUE
  )

# Step 2: Propensity score matching/weighting
library(MatchIt)

combined  mutate(source = "trial", treatment = treatment),
  external_controls |> mutate(source = "external", treatment = 0)
)

# Match external controls to trial control arm characteristics
match_ext  filter(treatment == 0 | source == "external"),
  method = "nearest",
  caliper = 0.2
)

# Step 3: Analysis with matched external controls
matched_combined 
  group_by(patient_id) |>
  mutate(
    p_uncensored = 1 - predict(cens_model, type = "response"),
    ipcw = cumprod(p_uncensored)
  ) |>
  ungroup()

# Stabilized weights
rwd_long 
  mutate(
    sw = ps_weight * ipcw / mean(ps_weight * ipcw, na.rm = TRUE)
  )

# Weighted analysis with stabilized weights
library(survey)
design_ipcw 
  filter(adjusted_effect 
  summarise(across(everything(), ~mean(!is.na(.)) * 100)) |>
  pivot_longer(everything(), names_to = "variable", values_to = "pct_complete")

# 2. Plausibility
plausibility_checks 
  summarise(
    age_range_ok = all(age >= 0 & age = admission_date, na.rm = TRUE),
    lab_in_range = all(egfr >= 0 & egfr 
  count(year = year(index_date)) |>
  ggplot(aes(year, n)) +
  geom_col() +
  labs(title = "Patient Enrollment Over Time")

# 4. Treatment patterns
rwd |>
  group_by(treatment) |>
  summarise(
    n = n(),
    mean_age = mean(age),
    pct_male = mean(sex == "male") * 100,
    mean_followup = mean(followup_time)
  )
```

## Reporting RWE Studies

```r
# Create Table 1 for RWE study
library(tableone)

# Unweighted baseline characteristics
vars <- c("age", "sex", "bmi", "smoking", "diabetes", "egfr", "stage")
cat_vars <- c("sex", "smoking", "diabetes", "stage")

tab1_unweighted <- CreateTableOne(
  vars = vars,
  strata = "treatment",
  data = rwd,
  factorVars = cat_vars,
  test = FALSE
)

# Weighted baseline characteristics (after PS weighting)
tab1_weighted <- svyCreateTableOne(
  vars = vars,
  strata = "treatment",
  data = design_weighted,
  factorVars = cat_vars,
  test = FALSE
)

# Print with SMD
print(tab1_unweighted, smd = TRUE)
print(tab1_weighted, smd = TRUE)
```

## Key Packages Summary

| Package | Purpose |
|---------|---------|
| TrialEmulation | Target trial emulation |
| WeightIt | Propensity score weighting |
| MatchIt | Propensity score matching |
| cobalt | Balance diagnostics |
| AIPW | Doubly robust estimation |
| adjustedCurves | Adjusted survival curves |
| tipr | Tipping point analysis |
| EValue | E-values for unmeasured confounding |
| ipw | Inverse probability weighting |
| episensr | Quantitative bias analysis |

## Best Practices

1. **Protocol first**: Define target trial before touching data
2. **Eligibility**: Apply strict, transparent eligibility criteria
3. **Time zero**: Align treatment initiation with follow-up start
4. **Immortal time**: Avoid immortal time bias in study design
5. **Balance**: Check and report covariate balance (SMD < 0.1)
6. **Positivity**: Assess overlap and address violations
7. **Sensitivity**: Quantify impact of unmeasured confounding
8. **Reporting**: Follow RECORD/STROBE-RWE guidelines

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [choxos](https://github.com/choxos)
- **Source:** [choxos/BiostatAgent](https://github.com/choxos/BiostatAgent)
- **License:** MIT

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-choxos-biostatagent-real-world-evidence
- Seller: https://agentstack.voostack.com/s/choxos
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.