Course 3 · Week 1 — Designing studies

Cheatsheet — biostats_courses

Author

R. Heller

Observational designs

Design Strengths Weaknesses
Cohort temporality, incidence expensive, loss to follow-up
Case-control rare outcomes, cheap recall + selection bias
Cross-sectional prevalence, screening no temporality
Case-crossover within-person comparison for transient exposures only

STROBE checklist: https://www.strobe-statement.org/

Trial designs

Type When
Parallel-group most common; independent arms
Crossover stable chronic condition, carry-over washable
Cluster intervention at group level (schools, clinics)
Factorial two or more interventions, test interactions
Adaptive pre-specified modifications based on interim data
Non-inferiority new treatment not worse than control by margin \(\Delta\)
Equivalence two-sided non-inferiority

CONSORT for RCTs: https://www.consort-statement.org/

Bench / translational

  • Blocking reduces nuisance variation (plate, day, operator).
  • Factorial tests interactions efficiently.
  • Split-plot handles two levels of randomisation.
  • Pseudoreplication: technical replicates ≠ biological replicates.

Power — closed form

library(pwr)
pwr.t.test(d = 0.5, power = 0.80, sig.level = 0.05,
           type = "two.sample")
pwr.2p.test(h = ES.h(0.3, 0.2), power = 0.80)
pwr.r.test(r = 0.3, power = 0.80)
pwr.anova.test(k = 4, f = 0.25, power = 0.80)

Effect-size conventions (Cohen): small \(d\) = 0.2, medium 0.5, large 0.8.

Power — simulation

library(simr)
# Build a pilot model, increase N, or tweak fixef, then:
ps <- powerSim(model, nsim = 500, test = fixed("arm"))
ps

Simulation wins for any design the textbook skips: mixed models, adaptive rules, non-standard outcomes.

Decision rule for Week 1

  • Randomise if you can. If not, draw a DAG and name the biases.
  • Power calculation before the protocol freeze, not after data collection.
  • Cluster randomisation → design effect \(1 + (\bar m - 1)\rho\); inflate N.
  • Bench experiments → treat batch, plate, and operator as random effects.

Common pitfalls

  • Planning a cluster RCT without inflating N for design effect.
  • Borrowing a pilot effect size without acknowledging noise.
  • Running a non-inferiority trial as if it were superiority.
  • Treating technical replicates as if they were biological.

Further reading

  • Matthews, Introduction to Randomized Controlled Clinical Trials.
  • Chow & Liu, Design and Analysis of Clinical Trials.