Course 3 · Week 4 — Evidence synthesis, ID, pre-registration

Cheatsheet — biostats_courses

Author

R. Heller

Systematic review workflow

  1. PICO: Population, Intervention, Comparator, Outcome.
  2. Register protocol on PROSPERO.
  3. Build search across ≥ 2 databases; peer-review with a librarian.
  4. Two-screener title/abstract and full-text.
  5. Extract, appraise (RoB 2 / ROBINS-I), synthesise.
  6. Report with the PRISMA 2020 checklist and flow diagram.

Meta-analysis

library(metafor)
dat <- escalc(measure = "RR",
              ai = tpos, bi = tneg, ci = cpos, di = cneg,
              data = studies)
fit <- rma(yi, vi, data = dat, method = "REML")
fit
forest(fit, header = TRUE)
funnel(fit); regtest(fit)
Statistic Meaning
\(\tau^2\) between-study variance
\(I^2\) % of total variance due to heterogeneity
Egger’s test funnel-plot asymmetry (small-study effects)

Fixed-effect only when heterogeneity is effectively zero. Default to random effects.

Network meta-analysis

library(netmeta)
net <- netmeta(TE, seTE, treat1, treat2, studlab,
               data = dat, sm = "MD", random = TRUE)
netgraph(net)
netrank(net, small.values = "good")    # SUCRA ranking
decomp.design(net)                     # inconsistency

Transitivity + consistency are the core assumptions.

Infectious disease — SIR / SEIR

library(deSolve)
sir <- function(t, y, p) with(as.list(c(y, p)), {
  N <- S + I + R
  list(c(-beta * S * I / N,
          beta * S * I / N - gamma * I,
          gamma * I))
})
out <- ode(c(S = 999, I = 1, R = 0),
           times = 0:120,
           func = sir,
           parms = c(beta = 0.3, gamma = 0.1))

\(R_0 = \beta / \gamma\). Herd-immunity threshold \(\approx 1 - 1/R_0\).

Pre-registration and SAP

A pre-registration specifies, before data analysis:

  • Primary outcome + exact estimand.
  • Analysis model (fixed terms, adjustments, interactions).
  • Handling of missing data.
  • Multiplicity control.
  • Sample-size justification.

OSF and ClinicalTrials.gov are the two common registries.

Decision rule for Week 4

  • More than a handful of studies on a question → plan a systematic review.
  • Multiple comparators of interest → network meta-analysis.
  • Outbreak modelling → start with SIR, extend only as data justify.
  • Confirmatory analysis → pre-registered; exploratory analyses are labelled as such in the manuscript.

Common pitfalls

  • Declaring a review “systematic” without a pre-registered protocol.
  • Pooling studies with fundamentally different populations / outcomes.
  • Interpreting SUCRA rankings as if they were certain.
  • Calling a compartmental model a “forecast” without uncertainty.

Further reading

  • Higgins et al., Cochrane Handbook for Systematic Reviews.
  • Keeling & Rohani, Modeling Infectious Diseases in Humans and Animals.
  • Nosek et al. (2018), The preregistration revolution.