Clinical Research

Clinical Research

Operational support for running clinical studies: structuring a protocol so it survives review, choosing endpoints that are actually analysable, designing eligibility criteria that do not strangle accrual, planning sample size and power, and testing whether the site network can deliver the enrolment target.

Scope and limits. This skill supports study operations — planning, structuring, and auditing. It is not a substitute for a qualified biostatistician, and it is not regulatory advice. The sample-size calculator assumes a simple parallel design with no interim analyses, multiplicity adjustment, covariate adjustment, or clustering; any design departing from those assumptions requires a statistician. The protocol auditor checks structure and internal consistency, not regulatory acceptability. Every artifact produced here needs sign-off from qualified biostatistics, clinical, and regulatory affairs personnel before it enters a submission.

When to use this skill

• Planning a study and needing a defensible sample size before the budget and site count can be set
• Auditing a draft protocol for missing ICH E6 elements before it goes to an ethics committee or a sponsor review board
• Choosing between candidate endpoints where one is clinically meaningful and the other is achievable in the available sample
• Designing inclusion and exclusion criteria and needing to see the accrual cost of each additional restriction
• Assessing site feasibility — deciding how many sites, and which, are needed to hit an enrolment target inside the accrual window
• Diagnosing an under-accruing study and deciding between adding sites, extending the window, or amending eligibility

Inputs the skill expects

• Study phase, design (parallel, crossover, single-arm), and blinding
• The primary question in a form that names the comparison
• Candidate endpoints with their measurement instrument and timepoint
• Effect size assumptions and their source — prior study, pilot, or literature
• Expected dropout rate, from comparable studies where possible
• For feasibility: candidate sites with eligible population, prior accrual attainment, startup time, and competing studies

Clarify First

Before generating, confirm these inputs. If any is unknown or vague, ASK — do not assume:

• The primary endpoint and its measurement timepoint — everything downstream (sample size, visit schedule, site burden, cost) derives from it
• The effect size and where it came from — a literature effect and a pilot effect carry very different uncertainty, and a pilot-derived SD needs an inflation allowance
• Design features that break the simple formulas — interim analyses, co-primary endpoints, cluster randomisation, or crossover each require a different calculation and a statistician
• The enrolment window and site network available — a sample size that cannot be accrued is not a plan

Stop rule: ask only the 2-3 that most change the output. If the user says "just draft it," proceed and list your assumptions at the top of the artifact.

Workflows

Workflow 1 — Plan sample size and power

1. Fix the primary endpoint and its type: binary, continuous, or time-to-event. This selects the design, and the design selects the formula.
2. State the effect size you want to detect and where it came from. [PROVEN] Power for the smallest effect that would change clinical practice, not for the effect you hope to see — the latter systematically under-powers studies.
3. Set alpha (0.05 two-sided unless the protocol justifies otherwise), power (80% minimum, 90% where feasible), allocation ratio, and dropout rate.
4. Run the calculator. Read both the analysed n and the enrol n — the dropout inflation is the number that drives the budget.
5. If you have a fixed n constrained by budget or accrual, pass it as planned_n_per_group and read the achieved power instead.
6. Take the result to a biostatistician. This step is not optional.

python3 research-ops/clinical-research/scripts/sample_size_calculator.py \
  --input research-ops/clinical-research/assets/sample_power_spec.json \
  --format text

Workflow 2 — Audit a protocol outline

1. Map the draft protocol's sections onto the ICH E6 element keys.
2. Enter endpoints with role, measure, timepoint, and analysis population; enter inclusion and exclusion criteria verbatim.
3. Run the auditor. It reports missing sections, endpoints that cannot be analysed as written, eligibility criteria that conflict or duplicate, and gaps in the statistical and safety sections.
4. Clear every fail before circulating the draft. Missing sections and an unadjusted interim analysis are the two findings most likely to cost you a review cycle.

python3 research-ops/clinical-research/scripts/protocol_auditor.py \
  --input research-ops/clinical-research/assets/sample_protocol.json \
  --format text

Workflow 3 — Test the site network against the enrolment target

1. Collect per-site data: eligible annual population, self-reported accrual estimate, prior accrual attainment, startup time, coordinator status, and competing studies.
2. Run the feasibility scorer. It discounts self-reported estimates, blends in the site's historical attainment, caps against eligible population, and subtracts startup time from the accrual window.
3. Read the shortfall. If it is positive, the choice is more sites, a longer window, or looser eligibility — decide deliberately rather than discovering it at month 14.
4. Use --select to test whether a smaller, higher-quality site set beats a larger one. It usually does on cost, and often on total accrual.

python3 research-ops/clinical-research/scripts/site_feasibility_scorer.py \
  --input research-ops/clinical-research/assets/sample_sites.json \
  --select 5 --format text

Decision frameworks

Endpoint type drives everything

[PROVEN] Where the same clinical question can be posed as continuous or binary, the continuous version needs materially fewer participants. Dichotomising a continuous measure discards information and inflates n — do it only when the threshold itself is what is clinically meaningful.

Effect size sources and their reliability

That last row is a real failure mode. When the affordable sample size is back-solved into an effect size, the study is designed to fail and the failure is uninterpretable — you cannot distinguish "no effect" from "not enough people."