W5 · Causal questions, trajectories and phenotypes

Workshop W5, teaching units 17–20: a collider created by selection on ICU admission, biomarker trajectories with informative dropout in a mixed model, clusters on a continuous gradient, and a 3Rs justification for the next experiment.
Author

R. Heller

Part III · Justifying complex models · Level: advanced · Teaching units: 17–20 (4 × 45 min) · Required reading: core readings of B02 and B13 (about 20 minutes) · Lab: Lab W5

Guiding question

Which explanation of a pattern survives a plausible counter-assumption: a cause, a selection effect, a dropout mechanism or an artefact of the method?

Learning objectives

By the end of this workshop you can

  1. place injury severity, ICU admission, a biomarker and an outcome in a directed acyclic graph (DAG), state the causal or prognostic target, and justify how each adjustment variable is handled;
  2. simulate how selection on a collider produces an association that does not exist in the full population;
  3. fit a linear mixed model for repeated measurements and report time unit, number of people, number of observations and the time-by-group interaction;
  4. examine late, informative dropout with a sensitivity analysis under a stated counter-assumption;
  5. show for one model family (clusters or flexible time functions) a stability or diagnostic check and name an alternative explanation;
  6. turn a preclinical finding into a justified next experiment using the 3Rs, with a criterion against doing it.

Preparation

  • Required reading (about 20 minutes): B02 · Describe, predict, explain – core reading and B13 · Longitudinal biomarkers – core reading.
  • Before the session: look at the trajectory figure in the B13 core reading. Mark the measurement days, the missing samples and the people who contribute to each time point.
  • Organisation: everyone computes Parts A and B. In Part C each group deepens one method and presents it to a group that chose the other one. A clinical–methodological tandem leads the session.

Teaching units

Unit Task Minutes Observable evidence
17 Place severity, ICU admission and outcome in a DAG. Simulate a collider effect by selecting a subgroup (Lab W5, Part A). 45 A causal estimand or prognostic target, and a reasoned treatment of every adjustment variable.
18 Draw a spaghetti plot and fit a mixed model for repeated measurements. Examine the influence of late dropout under a plausible counter-assumption (Part B). 45 A result with time unit, number of people, interaction and sensitivity finding.
19 Elective station clusters: a continuous gradient is grouped artificially. Elective station longitudinal models: a flexible time function against a simple slope (Part C). 45 An alternative explanation plus a stability or diagnostic figure.
20 Turn a hypothesis from a preclinical finding into a next experiment, weighing the 3Rs (Replacement, Reduction, Refinement), sample size, species and the values of people affected. 45 A justification for the experiment with a testable alternative and a criterion against a further experiment.

The setting

A fictitious cohort of people with traumatic spinal cord injury has repeated measurements of a trace-element biomarker during the first four weeks. Severely injured people are admitted to intensive care more often, and people with complete lesions miss the late measurement more often. Each of these facts can create a pattern that looks like biology. All values are synthetic.

Materials

Four perspectives in this workshop

  • Ethics. A subtype label can change how people are treated. A cluster that is only a cut through a gradient should not become a diagnosis. In unit 20, the people who would benefit and the animals that would be used are weighed explicitly (NC3Rs: the 3Rs).
  • Regulation. Indirect: a biomarker that later guides treatment would need evidence for that intended purpose, not only an association in a selected cohort.
  • Data protection. Rare injury patterns and exact sampling dates can identify people; the shared tables carry pseudonyms and study days only.
  • Security. Indirect: a result that cannot be reproduced from the script and the stored seed is not trustworthy, whatever its cause.

Transition to W6

You have a report with a baseline, a counter-check and a proposal for the next experiment. W6 moves from explanation to forecasting and simulation: which assumptions make a number good enough for planning?

Further reading

Sources