Bio Workflows Clinical Trial Pipeline

Version Compatibility

Reference examples tested with: statsmodels 0.14+, scipy 1.12+, tableone 0.9+, pyreadstat 1.2+, pandas 2.1+, numpy 1.26+, matplotlib 3.8+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Clinical Trial Analysis Pipeline

"Analyze my clinical trial data end to end" -> Load CDISC domain tables, prepare a subject-level analysis dataset, run primary statistical models, perform subgroup analyses, and generate regulatory-compliant tables and figures.

This is a workflow skill: it owns the chaining decisions and hand-offs, not the internals of any one step.

The governing principle

A clinical-trial analysis is a commit-then-execute pipeline: its trustworthiness is decided by whether the analysis was locked BEFORE the data was seen, and every seam failure is a wrong-but-silent handoff that answers a different question with no error thrown.

1. The ESTIMAND, locked in the SAP before unblinding, is the made-once commitment everything inherits. ICH E9(R1) defines it by 5 attributes: treatment condition, population, endpoint, intercurrent-event (ICE) handling strategy, and population-level summary. It is the precise definition of "what treatment effect the trial estimates", committed before analysis so the estimator is matched to it rather than chosen to flatter the data.
2. The estimator is a CONSEQUENCE of the estimand's ICE strategy, not a free modeling choice. A treatment-policy strategy needs all post-ICE data (MMRM/treatment-policy, or reference-based MI if truly missing); a hypothetical strategy censors/models the post-ICE data as if the ICE had not occurred (MMRM under MAR); a composite strategy folds the ICE into the endpoint. Choosing MMRM vs reference-based MI vs g-computation to flatter the data is an estimand-estimator mismatch.
3. The analysis population and all subgroups are defined a priori; the data never picks them. Randomization licenses causal interpretation for the ITT/FAS primary only — PP and subgroups do NOT inherit that protection. Pre-specify every subgroup and the multiplicity graph in the SAP; post-hoc data-driven subgroups are hypothesis-generating only. This is the clinical analog of "don't call hits before CN correction".
4. Baseline balance is not tested with p-values, and marginal is not conditional. In a randomized trial any imbalance is by definition chance, and baseline hypothesis testing is incoherent (Senn 1994) — report SMD (>0.1 notable, a convention Senn does not state) and adjust via pre-specified ANCOVA, don't test-then-decide. For a binary endpoint the primary is the MARGINAL risk difference via g-computation (FDA 2023); a conditional OR is a different parameter (non-collapsibility), reported as supportive.

Made-once commitments

Scientific Reasoning Framework

Before executing any analysis step, establish the causal framework. For an RCT, randomization justifies causal interpretation of the primary analysis, but subgroup analyses and observational comparisons within the trial (e.g., adherence effects) do not inherit this protection. Key decisions requiring scientific judgment at each step: (1) data preparation -- which aggregation strategy matches the estimand, (2) covariate selection -- include confounders and prognostic factors from the SAP, exclude mediators and colliders, (3) subgroup analysis -- test only biologically motivated interactions, (4) missing data -- link DS domain reasons to the assumed mechanism before choosing a method. The workflow below provides the technical steps; the scientific reasoning at each decision point determines whether the results are valid.

Workflow Overview

CDISC Domain Files (DM, AE, EX, LB)
    |
    v
[1. Data Preparation] ----> Subject-level dataset with outcomes and covariates
    |
    v
[2. Table 1] ------------> Baseline characteristics by treatment arm
    |
    v
[3. Primary Analysis] ---> Marginal RD via g-computation (conditional OR supportive)
    |
    v
[4. Categorical Tests] --> Chi-square / Fisher's exact for key associations
    |
    v
[5. Subgroup Analysis] --> Interaction terms, stratified ORs, forest plot
    |
    v
[6. Missing Data] -------> Multiple imputation sensitivity analysis
    |
    v
Results tables and figures

Step 1: Data Preparation

Goal: Create a single subject-level analysis dataset from CDISC domain tables.

Approach: Load domain files, aggregate event-level data to one row per subject, merge on USUBJID, and code the outcome variable.

import pandas as pd
import pyreadstat

dm, _ = pyreadstat.read_xport('dm.xpt')
ae, _ = pyreadstat.read_xport('ae.xpt')

# Aggregate: did each subject have the target adverse event?
target_ae = ae[ae['AEDECOD'] == 'COVID-19'].copy()
severity_map = {'MILD': 1, 'MODERATE': 2, 'SEVERE': 3, 'LIFE THREATENING': 4, 'FATAL': 5}
target_ae['AESEV_NUM'] = target_ae['AESEV'].map(severity_map)
had_event = target_ae.groupby('USUBJID')['AESEV_NUM'].max().reset_index()
had_event.columns = ['USUBJID', 'EVENT_SEVERITY']

analysis = dm[['USUBJID', 'ARM', 'ARMCD', 'AGE', 'SEX']].merge(had_event, on='USUBJID', how='left')
analysis['HAD_EVENT'] = analysis['EVENT_SEVERITY'].notna().astype(int)
analysis['TREATMENT'] = (analysis['ARMCD'] != 'PLACEBO').astype(int)

QC Checkpoint: Verify one row per USUBJID, no unexpected duplicates, treatment arms are present and reasonably balanced.

assert analysis['USUBJID'].is_unique, 'Duplicate subjects detected'
print(analysis['ARM'].value_counts())

Step 2: Table 1 Baseline Characteristics

Goal: Summarize demographics and baseline variables by treatment arm.

Approach: Use TableOne to generate a baseline table by arm with standardized mean differences and explicit missingness. Omit the baseline p-value column: in a randomized trial any imbalance is by definition due to chance, so a baseline p-value tests a null already known to be true (Senn 1994; CONSORT 2010/2025). Report SMD for balance instead. Table-construction and export mechanics (gtsummary/tableone, Word export, gene-symbol-safe supplements) live in reporting/publication-tables.

from tableone import TableOne

columns = ['AGE', 'SEX', 'RACE']
categorical = ['SEX', 'RACE']
table1 = TableOne(analysis, columns=columns, categorical=categorical,
                  groupby='ARM', pval=False, smd=True, missing=True)
print(table1.tabulate(tablefmt='github'))

Interpret SMD > 0.1 as meaningful imbalance. The response to a worrying imbalance on a prognostic covariate is to adjust for it (a pre-specified ANCOVA/model covariate), not to test it.

Step 3: Primary Analysis -- Logistic Regression

Goal: Estimate the treatment effect on the binary outcome as an adjusted odds ratio.

Approach: Fit a logistic regression with explicit reference category and clinically relevant covariates, then exponentiate coefficients to obtain ORs.

import statsmodels.formula.api as smf
import numpy as np

model = smf.logit(
    'HAD_EVENT ~ C(ARM, Treatment(reference="Placebo")) + AGE + C(SEX)',
    data=analysis
).fit()

or_table = pd.DataFrame({
    'OR': np.exp(model.params),
    'Lower_CI': np.exp(model.conf_int()[0]),
    'Upper_CI': np.exp(model.conf_int()[1]),
    'p_value': model.pvalues
})
print(or_table)
print(f'McFadden pseudo-R2: {model.prsquared:.4f}')

The logistic fit above yields the CONDITIONAL (adjusted) OR. When the estimand's summary measure is a marginal risk difference (the FDA 2023 primary for a binary endpoint), do NOT report this conditional OR as the primary effect. Compute the MARGINAL risk difference by g-computation: fit the covariate-adjusted model, predict each subject's outcome probability under both arms, average within arm, and difference; bootstrap the whole fit-and-predict for the CI. examples/clinical_trial_pipeline.py implements this; see clinical-biostatistics/logistic-regression for the estimator's assumptions and variance options. Report the conditional OR as supportive. The two differ by non-collapsibility and answer different questions.

QC Checkpoint: Verify model converged (no warnings), check for separation (coefficients > 10 or SE > 100), report pseudo-R-squared (McFadden > 0.2 is excellent; do not compare across pseudo-R2 types). Confirm the reported PRIMARY effect matches the estimand's summary measure (marginal RD via g-computation for a marginal estimand), not whichever the model emits by default.

Step 4: Categorical Tests

Goal: Test the crude association between treatment and outcome using contingency tables.

Approach: Build a 2x2 table, check expected cell counts, and choose chi-square or Fisher's exact accordingly.

from scipy.stats import chi2_contingency, fisher_exact

table = pd.crosstab(analysis['ARM'], analysis['HAD_EVENT'])
chi2, p, dof, expected = chi2_contingency(table, correction=False)

if (expected < 5).any():
    _, p = fisher_exact(table.values)
    print(f'Fisher exact p = {p:.4f}')
else:
    print(f'Chi-square p = {p:.4f} (chi2 = {chi2:.2f}, dof = {dof})')

Step 5: Subgroup Analysis

Goal: Test whether the treatment effect varies across pre-specified subgroups.

Approach: Fit a model with an interaction term and test it with a single interaction LR test. Subgroup-specific ORs are DESCRIPTIVE ONLY -- do not correct their individual p-values, do not interpret them; the alpha allocated to the subgroup family is handled by the pre-specified gMCP graph, not by a post-hoc correction. Visualize with a forest plot.

import matplotlib.pyplot as plt

# The HTE test is the treatment-by-subgroup INTERACTION, not per-subgroup significance.
# Fit the main-effects (restricted) and interaction (full) models, then LR-test the interaction.
main_model = smf.logit(
    'HAD_EVENT ~ C(ARM, Treatment(reference="Placebo")) + C(SUBGROUP)',
    data=analysis
).fit(disp=0)
interaction_model = smf.logit(
    'HAD_EVENT ~ C(ARM, Treatment(reference="Placebo")) * C(SUBGROUP)',
    data=analysis
).fit(disp=0)
# compare_lr_test exists only on linear-model results, NOT LogitResults -- compute the LR test by hand.
from scipy.stats import chi2
lr_stat = 2 * (interaction_model.llf - main_model.llf)
df_diff = int(interaction_model.df_model - main_model.df_model)
interaction_pval = chi2.sf(lr_stat, df_diff)
print(f'Treatment-by-subgroup interaction: LR chi2={lr_stat:.2f}, df={df_diff}, p={interaction_pval:.3f}')

# Subgroup-specific ORs are DESCRIPTIVE ONLY (to draw the forest plot). Do NOT interpret their
# individual p-values as evidence of a subgroup effect -- that is the invalid per-subgroup
# significance pattern; the interaction p-value above is the only valid HTE test.
labels, ors, lowers, uppers = [], [], [], []
for group in analysis['SUBGROUP'].unique():
    sub = analysis[analysis['SUBGROUP'] == group]
    sub_model = smf.logit(
        'HAD_EVENT ~ C(ARM, Treatment(reference="Placebo"))',
        data=sub
    ).fit(disp=0)
    or_val = np.exp(sub_model.params.iloc[1])
    ci = np.exp(sub_model.conf_int().iloc[1])
    labels.append(group)
    ors.append(or_val)
    lowers.append(ci[0])
    uppers.append(ci[1])

# Forest plot
fig, ax = plt.subplots(figsize=(8, 5))
y_pos = range(len(labels))
ax.errorbar(ors, y_pos,
            xerr=[np.array(ors) - np.array(lowers), np.array(uppers) - np.array(ors)],
            fmt='D', color='black', capsize=3, markersize=5)
ax.axvline(x=1.0, color='gray', linestyle='--', linewidth=0.8)
ax.set_yticks(y_pos)
ax.set_yticklabels(labels)
ax.set_xlabel('Odds Ratio (95% CI)')
ax.set_xscale('log')
plt.tight_layout()
plt.savefig('forest_plot.png', dpi=150)

QC Checkpoint: Interaction p-value reported from a single LR test. Alpha for the subgroup family allocated in the pre-specified gMCP graph (not a post-hoc p-value correction on the descriptive subgroup ORs). Forest plot shows overall estimate for context.