Post on 22-Dec-2015
George TH Ellison PhD DSc
Division of Epidemiology and Biostatistics
Leeds Institute of Genetics, Health and Therapeutics
g.t.h.ellison@leeds.ac.uk
Wendy Harrison (Leeds) and Graham Law (Leeds)
Johannes Textor (Utrecht)
Teaching DAGs to support MBChB students design, analyze and critically appraise
clinical research
DAGs help us distinguish between:
- nonparametric theoretical models of causality; and
- optimal parametric statistical models for testing these
DAGs can be used at every stage of quantitative research:
- optimising the number of variables measured (design)
- optimising adjustment for confounding (analysis)
- evaluating published statistical models (critical appraisal)
Why DAGs?
Why teach statistical modeling in MBChB? Most clinical research/audit uses an observational design Most observational research is poorly/implicitly modelled
What is a DAG (Directed Acyclic Graph)? A type of ‘causal path diagram’ with: unidirectional
(‘causal’) arrows linking variables; and no circular paths
DAGitty.net applies algorithms automatically
Challenges facing the application of DAGs Algorithms are tedious and time-consuming to apply DAGs with more than a handful of variables are complex
Cross-tabulation might help as variables
causes one above
caused by one above
no causal relationship
Comparing three ways of drawing DAGs Three one-hour tutorials using three approaches:
(i) ‘graphical’; (ii) ‘cross-tabulation’; and (iii) ‘relational’
Each approach evaluated based on:
- how many variables were included in the DAG
- mediators/confounders correctly identified*
- student feedback on ease of use and interpretation
All participants were third year MBChB students who had
completed a year-long critical appraisal course
The context was a published paper on an accessible topic:
‘determinants of pregnancy-associated weight gain’
Handouts contained 10, 20 and 30 variables:
Group using ‘graphical’ approach:
Group using ‘cross-tabulation’ approach:
Group using ‘cross-tabulation’ approach:
Group using ‘relational’ approach:
Group using ‘relational’ approach:
What do I (now) think the DAG should be?
Focusing on the ‘relational’
None of the students were able to attempt including more
than 10 variables in their cross-tabulation 86% correctly identified covariates that should have been
classified as ‘mediators’ by Harris et al. 1999... Fewer than 5% correctly identified the only covariate that
is likely to have acted as ‘confounder’ (maternal age) A disproportionate use of ‘competing exposure’ as a
classification for covariates that are likely to have been
‘mediators’ suggests students were reluctant to identify
‘exposure’ as a potential/likely/theoretical cause
Focusing on the ‘relational’ Most students found it ‘Difficult’...
Why?
- understanding DAGs and DAG-related terminology
- “Time consuming” debating/agreeing links and directions
- “...so much depends on variation and opinion”
Summary It is feasible to teach DAGs to MBChB students
Most students are capable of distinguishing between
‘confounders’, ‘mediators’ and ‘competing exposures’
‘Cross-tabulation’ and ‘relational’ were slower to apply
but less likely to result in errors
Suggestions for future development:
- include a quiz to strengthen initial knowledge
- (perhaps) avoid group work (at least initially)
- reward recognition of ‘subjective causality’
- explore an approach that involves removing rather than
including causal paths (‘arcs’)