r-clinical

R Clinical Trials & Biostatistics

Biostatistics and clinical trial analysis in R: trial design, CDISC standards, TLF generation, survival analysis, biomarkers, and meta-analysis.

All code uses base pipe |>, <- for assignment, and tidyverse style.

Lazy references:

• Read references/cdisc-adam-guide.md for ADaM/SDTM dataset structure, naming conventions, and admiral package patterns
• Read references/tlf-templates.md for ready-to-use TLF code patterns (demographics, AE, KM curves, forest plots, waterfall plots)

Agent dispatch: For statistical methodology questions (mixed models, Bayesian approaches, multiplicity adjustments), dispatch to r-statistician.

---

Trial Design & Power Analysis

library(pwr)
library(gsDesign)

# Two-sample t-test: detect delta=0.5 with 80% power, alpha=0.05
pwr.t.test(d = 0.5, sig.level = 0.05, power = 0.80, type = "two.sample")

# Proportion test (e.g. ORR comparison)
pwr.2p.test(h = ES.h(p1 = 0.40, p2 = 0.25), sig.level = 0.05, power = 0.80)

# Group sequential design (interim analysis)
gs <- gsDesign(k = 2, test.type = 1, alpha = 0.025, beta = 0.20,
               sfu = sfO'Brien-Fleming)
gs$n.I   # sample sizes at each look
gs$upper$bound  # efficacy boundaries

Key packages: pwr, gsDesign, rpact (adaptive designs), PowerTOST (bioequivalence). Follow ICH E9(R1) for estimand framework.

---

Survival Endpoints (OS, PFS, DFS)

Boundary: Clinical trial survival endpoints with CDISC conventions. For general survival methodology outside clinical context, use r-stats instead.

library(survival)
library(survminer)

# Fit Kaplan-Meier
km_fit <- survfit(Surv(AVAL, CNSR == 0) ~ TRT01P, data = adtte)

# Log-rank test
survdiff(Surv(AVAL, CNSR == 0) ~ TRT01P, data = adtte)

# Cox proportional hazards
cox_fit <- coxph(Surv(AVAL, CNSR == 0) ~ TRT01P + AGE + SEX, data = adtte)
summary(cox_fit)   # HR, 95% CI, p-value

# Competing risks (e.g. death before progression)
library(cmprsk)
cr_fit <- cuminc(
  ftime = adtte$AVAL, fstatus = adtte$EVNTDESC,
  group = adtte$TRT01P
)

Censoring rules: follow CDISC ADTTE conventions — CNSR = 1 for censored, CNSR = 0 for event. Landmark analysis: survfit(..., start.time = t).

---

Biomarkers: ROC & Cutpoints

library(pROC)
library(cutpointr)

# ROC curve with 95% CI
roc_obj <- roc(response ~ biomarker, data = bm_df, ci = TRUE)
auc(roc_obj)       # AUC with 95% CI
plot(roc_obj, print.auc = TRUE)

# Optimal cutpoint (maximise Youden index)
cp <- cutpointr(data = bm_df, x = biomarker, class = response,
                method = maximize_metric, metric = youden)
summary(cp)

# Subgroup analysis scaffold
subgroups <- c("SEX", "AGEGR1", "RACE", "BMICAT")
subgroup_results <- purrr::map(subgroups, \(sg) {
  coxph(Surv(AVAL, CNSR == 0) ~ TRT01P,
        data = adtte |> dplyr::filter(.data[[sg]] == level))
})

---

Meta-Analysis

library(meta)
library(metafor)

# Fixed/random effects meta-analysis of log HRs
ma <- metagen(
  TE = log_hr, seTE = se_log_hr, studlab = study,
  data = meta_df, sm = "HR", method.tau = "REML",
  hakn = TRUE   # Hartung-Knapp correction
)
summary(ma)        # I², Q-test, pooled HR
forest(ma, sortvar = TE)

# Publication bias
funnel(ma)
metabias(ma, method.bias = "linreg")  # Egger's test

Heterogeneity: I² < 25% low, 25–75% moderate, > 75% high. Report I² and τ² alongside pooled estimate. Use metafor::rma() for meta-regression.

---

Genomics Bridge (Bioconductor)

# Differential expression
library(DESeq2)
dds <- DESeqDataSetFromMatrix(count_matrix, col_data, design = ~ condition)
dds <- DESeq(dds)
results(dds, contrast = c("condition", "treated", "control"))

# Limma/voom for microarray or normalised RNA-seq
library(limma)
fit <- lmFit(expr_matrix, design_matrix) |> eBayes()
topTable(fit, coef = 2, adjust.method = "BH", n = Inf)

# Enrichment
library(clusterProfiler)
enrichGO(gene = sig_genes, OrgDb = "org.Hs.eg.db", ont = "BP")

For full Bioconductor workflows, use BiocManager::install() for packages and consult Bioconductor workflows vignettes.

---

Regulatory Context

• ICH E9(R1): Define estimand (population, treatment, endpoint, intercurrent events, summary measure) before analysis
• ICH E6(R2): GCP — traceability and reproducibility requirements
• CONSORT: Report randomisation, allocation concealment, and flow diagram
• TLFs: Every regulatory table/listing/figure must be traceable to ADaM

CONSORT flow diagram scaffold:

library(ggplot2)
# Use DiagrammeR or a custom ggplot approach for CONSORT boxes
library(DiagrammeR)
grViz("
  digraph consort {
    Assessed [label='Assessed for eligibility\\n(n=500)']
    Excluded [label='Excluded (n=120)\\n- Not meeting criteria (n=80)\\n- Declined (n=40)']
    Randomised [label='Randomised (n=380)']
    ArmA [label='Arm A (n=190)']
    ArmB [label='Arm B (n=190)']
    Assessed -> Excluded; Assessed -> Randomised
    Randomised -> ArmA; Randomised -> ArmB
  }
")

---

MCP Integration (Optional)

All MCP features below are optional — the skill works fully without them. Use when connected MCP servers are available.

btw/mcptools (R session awareness)

• Before recommending admiral/pharmaverse functions: btw_tool_sessioninfo_is_package_installed to verify the package is available
• Before writing CDISC transformations: btw_tool_env_describe_data_frame to inspect actual dataset structure (variable names, types, expected CDISC domains)
• When uncertain about admiral function arguments: btw_tool_docs_help_page to read installed docs

Clinical Trials MCP

search_trials          — find trials by condition, phase, status
get_trial_details      — full protocol details for a specific NCT ID
analyze_endpoints      — summarise primary/secondary endpoint patterns
search_by_sponsor      — find trials by sponsor organisation
search_investigators   — look up principal investigators
search_by_eligibility  — filter trials by inclusion/exclusion criteria

Example: "Find Phase 3 trials for NSCLC comparing immunotherapy vs chemo" → search_trials(condition="NSCLC", phase="3") then get_trial_details(nct_id="NCT...") to pull endpoint definitions.

bioRxiv MCP

search_preprints           — keyword search across bioRxiv/medRxiv
get_preprint               — fetch full preprint metadata + abstract
search_published_preprints — find preprints that became journal articles
search_by_funder           — filter by funding source

---

Verification

After Cox model: check cox.zph() for PH assumption. After CDISC derivation: validate variable names against spec. Re-run if issues found.

Gotchas

Trap	Why It Fails	Fix
Using months instead of days for time-to-event	Surv() expects numeric time; months lose precision and break censoring intervals	Convert to days: as.numeric(difftime(end_date, start_date, units = "days"))
Forgetting to censor at analysis cutoff date	Patients without events who are still on study appear as events or get dropped	Set CNSR <- 1 and AVAL <- cutoff_date - start_date for patients without events before cutoff
Confusing OS, PFS, and DFS endpoint definitions	Each has different event/censoring rules; mixing them corrupts the analysis	Define endpoint in the ADTTE derivation: OS = death only; PFS = progression or death; DFS = recurrence or death
Not stratifying Kaplan-Meier by treatment arm	Single-arm KM hides the treatment comparison the user requested	Always use survfit(Surv(time, event) ~ TRT01P, ...) when comparing arms
Using survfit() without conf.type = "log-log"	Default "log" CI can exceed [0, 1]; "log-log" is regulatory preference	Pass conf.type = "log-log" to survfit() for regulatory submissions
Applying standard survival methods when competing risks apply	Death before progression biases PFS if treated as censored	Use cmprsk::cuminc() or tidycmprsk::cuminc() for competing risk endpoints
Generating full TLF suite when user asked for one KM plot	Scope creep wastes tokens and overwhelms the user	Deliver exactly what was asked; mention TLF templates in references/ if the user wants more

Examples

Happy Path: Kaplan-Meier survival curve with log-rank test

Prompt: "Plot overall survival by treatment arm with a risk table and log-rank p-value."

# Input — ADTTE dataset with CDISC conventions
library(survival)
library(survminer)

km_fit <- survfit(
  Surv(AVAL, CNSR == 0) ~ TRT01P,
  data = adtte,
  conf.type = "log-log"
)
survdiff(Surv(AVAL, CNSR == 0) ~ TRT01P, data = adtte)

# Output — KM plot with risk table
survminer::ggsurvplot(
  km_fit,
  data       = adtte,
  risk.table = TRUE,
  pval       = TRUE,
  xlab       = "Time (days)",
  ylab       = "Overall Survival Probability",
  legend.labs = c("Placebo", "Treatment")
)

Edge Case: Competing risks when death precedes progression

Prompt: "Estimate PFS but some patients died before progressing — standard KM is biased."

# WRONG — treating death-before-progression as censored biases KM upward
km_naive <- survfit(Surv(AVAL, CNSR == 0) ~ TRT01P, data = adtte_pfs)

# CORRECT — competing risks via cumulative incidence
library(tidycmprsk)

cuminc_fit <- cuminc(Surv(AVAL, factor(EVNTDESC)) ~ TRT01P, data = adtte_pfs)
#> EVNTDESC levels: "Progression", "Death", "Censored"

# Output — cumulative incidence curves per event type
ggcuminc(cuminc_fit, outcome = "Progression") +
  ggplot2::labs(x = "Time (days)", y = "Cumulative Incidence of Progression")

More example prompts:

• "How many patients do I need to detect an ORR improvement from 20% to 40%?"
• "Pool hazard ratios from 5 studies in a meta-analysis."
• "What Phase 3 NSCLC trials used PFS as primary endpoint?"