Course 3

ADVANCED · 4 WEEKS · 20 LABS

Study Design, Longitudinal Data & Causal Inference

Designing studies; handling missing, clustered, and time-to-event data; and making causal claims with the care they deserve.

What you’ll be able to do by the end

  • Pick a study design that actually answers your research question, and defend the choice in a pre-registration.
  • Distinguish MCAR, MAR, and MNAR missingness in a dataset you have, and apply multiple imputation correctly.
  • Fit linear and generalised linear mixed models to repeated-measures and clustered data, and interpret random-effect variances honestly.
  • Draw a DAG, derive the adjustment set from it, and recognise which causal claims a given design can support.
  • Apply propensity-score, IPTW, and g-methods analyses; know the difference between an average treatment effect, an ATT, and a heterogeneous treatment effect.
  • Perform meta-analysis of a small handful of studies, including network meta-analysis when several treatments are compared.

Who should take this course

Course 3 is aimed at researchers who are past the “is there an effect?” stage and into the “is the effect I estimated actually the effect I wanted?” stage. It assumes Courses 1 and 2, or equivalent comfort with regression, GLMs, and reading a residual plot. Clinicians running trials, epidemiologists with observational data, and bench scientists whose experiments are now clustered by batch or animal will all find this the most directly useful course on the site.

The shape of the four weeks

Week 1

Designing studies

Observational designs; RCTs; adaptive and non-inferiority trials; bench design; power by closed form and simulation.

Week 2

Missing data, longitudinal, time series

MCAR/MAR/MNAR; multiple imputation with mice; LMMs with lme4; GLMMs and GEE; time series basics.

Week 3

Survival II, causal inference, HTE

Time-varying covariates and landmarking; competing risks and multistate; DAGs; propensity scores and IPTW; g-methods and HTE.

Week 4

Evidence synthesis, ID, pre-registration

Systematic reviews, PRISMA; meta-analysis; network meta-analysis; SIR/SEIR with deSolve; pre-registration and SAPs.

Weekly summaries

Week 1 — designing studies. The first lab covers observational designs (cohort, case-control, cross-sectional, case-crossover) and STROBE. The RCT lab moves through parallel-group, crossover, cluster, and factorial trials; the adaptive lab adds non-inferiority and equivalence testing. Bench-science design (blocking, factorial, split-plot, pseudoreplication) gets its own lab because the typical omics lab ignores most of it. The week closes with a power lab that does closed-form calculations with pwr and WebPower and then redoes them by simulation with simr so you can handle any design analytically or otherwise. Key packages: pwr, WebPower, simr.

Week 2 — missing data, longitudinal, time series. The MCAR/MAR/MNAR lab clarifies what the acronyms mean using a small simulated study; the mice lab walks through multiple imputation, pooling, and convergence diagnostics. Linear mixed models with lme4 and lmerTest follow, then GLMMs and GEE for binary and count responses with lme4, glmmTMB, and geepack. A time-series primer on decomposition, ARIMA, and change-point detection closes the week. Key packages: mice, lme4, glmmTMB, geepack, forecast, changepoint.

Week 3 — survival II, causal inference, and heterogeneous effects. Time-varying covariates, landmarking, and the immortal-time bias open the week and set up the competing-risks and multistate-model lab that follows. The DAG lab introduces dagitty and ggdag and formalises adjustment-set selection. Propensity-score matching and IPTW follow, with balance diagnostics via cobalt. The final lab covers g-methods, instrumental variables, difference-in-differences, regression discontinuity, and heterogeneity of treatment effect. Key packages: survival, tidycmprsk, mstate, dagitty, ggdag, MatchIt, cobalt.

Week 4 — evidence synthesis, infectious disease, pre-registration. Systematic reviews and PRISMA open the week; meta-analysis with metafor follows; network meta-analysis with netmeta takes a full lab because graphical comparison methods are worth the investment. The penultimate lab uses deSolve to build SIR and SEIR compartmental models and to fit them to simulated epidemic data — an unusual but important skill in an era of recurrent outbreaks. The course closes with a lab on pre-registration and statistical analysis plans, including a reusable template. Key packages: metafor, netmeta, deSolve.

How to work through it

Course 3 is the densest on the site. A sustainable pace is one week of labs per fortnight. The DAG, propensity-score, and g-methods labs (W3 S3–S5) are the intellectual core of the course and repay a slow reading. The time-series lab (W2 S5) and the infectious-disease lab (W4 S4) are beautiful but skippable if your research domain has no use for them.

Further along