TruePositive

Home › Teaching

Teaching

Observational study design and causal inference

The organising idea is the target trial: state the randomised trial that would answer the question, then make explicit every respect in which the observational study departs from it. Most classic design errors are departures that were never noticed.

  • Estimands, and stating the quantity to be estimated precisely
  • Target trial emulation on a worked example
  • Immortal time, prevalent user and depletion of susceptibles bias
  • Propensity scores, weighting and overlap; the limits of matching
  • Time-varying exposure and confounding
  • Quantitative bias analysis and negative controls

Real-world evidence in regulatory and assessment settings

An account of how observational evidence is read by assessment bodies: where it is accepted without argument, where it is treated as supportive only, where it is rejected, and the reasons those distinctions fall where they do.

  • Uses of real-world evidence: burden of disease, comparative effectiveness, external controls, long-term outcomes
  • The EU Health Technology Assessment Regulation and the Joint Clinical Assessment
  • Data sources: registers, electronic health records, claims and cohorts
  • Recurring reasons observational evidence is rejected
  • Constructing an evidence generation plan

Analysing health data in secure research environments

Working within a controlled platform, where data cannot be exported, software cannot be installed freely, and every output is reviewed before release. Taught on a live environment, currently the NIH All of Us Researcher Workbench, and adaptable to a Danish register environment or a client platform.

  • Access tiers, approvals and data use agreements as constraints on analysis
  • Workspaces, compute environments and cost control
  • Controlled and referenced resources, and what follows for deletion and sharing
  • Reproducible analysis in R inside the environment, under version control
  • Cohort construction, and why rebuilt cohorts may not reproduce earlier counts
  • Preparing outputs for disclosure review

AI assistants in regulated research

A practical treatment of where large language model assistants save time in evidence work, the specific ways they fail in technical and regulated settings, and the handling rules that keep confidential and personal data out of them.

  • Tasks these tools perform reliably, and tasks they do not
  • Characteristic failures: fabricated citations, invented figures, outdated facts
  • Confidentiality and personal data in prompts
  • Verification practice and audit trail
  • Reusable prompts and instructions for recurring team tasks

Booking

A scoping conversation establishes who will attend and what they should be able to do afterwards. An agenda and set of learning outcomes are provided in writing before anything is committed, including which of your own cases will serve as examples. Materials, exercises and code are included and remain available for internal reuse.

Fees are charged per delivery rather than per participant. A quotation requires the dates and the approximate group size.

Course enquiries

State who will attend and what they should be able to do afterwards, and you will receive a proposed outline.

Discuss your analysis