www.dagitty.net Dealing with confounders just got easier! George TH - PowerPoint PPT Presentation

School of Medicine FACULTY OF MEDICINE AND HEALTH www.dagitty.net Dealing with confounders just got easier! George TH Ellison PhD DSc TIME Research Group - Division of Biostatistics Leeds Institute of Genetics, Health and Therapeutics University of Leeds - School of Medicine g.t.h.ellison@leeds.ac.uk

What are confounders? A confounder … - causes the ‘outcome’ (i.e. the dependent variable) - causes the ‘exposure’ (i.e. the independent variable) A mediator … - causes the ‘outcome’ (i.e. the dependent variable) - is caused by the ‘exposure’ (i.e. the independent variable) A competing exposure … - causes the ‘outcome’ (i.e. the dependent variable) - is unrelated to the ‘exposure’ (i.e. the independent variable)

Mediator Competing exposure Confounder Outcome Exposure Confounder

Why are confounders important? Confounders alter the relation between exposure and outcome - can create or strengthen a relationship - can weaken or (on rare occasions) remove a relationship What do we mean by ‘causes’? A ‘causal’ relationship can be… - functional (e.g. no contraception  teenage mother) - empirical (e.g. teenage grandmother  teenage mother) - speculative (e.g. unemployment  teenage mother) - hypothetical (e.g. teenage grandfather  teenage mother)

How do you identify confounders? Draw a ‘Directed Acyclic Graph’ to summarise all causal relationships between your known/available variables What is a ‘DAG’? A type of ‘causal path diagram’ with unidirectional (‘causal’) arrows linking variables, and no circular paths What does a ‘DAG’ look like? You’ve already seen one…

Mediator Competing exposure Confounder Outcome Exposure Confounder

How do I draw a DAG? Using a simple spreadsheet… I’ll show you now… What if I get it wrong? No worries… if you can’t figure out which variables have a causal link to other variables then draw every possible DAG: it’s a really explicit approach to conducting sensitivity analyses What next? Using www.dagitty.net you can identify which variables need to be adjusted for in your multivariable statistical analyses… this is often fewer than you think…

Voting for UKIP An example: Pipe Smoking

An example: Pipe Smoking and UKIP  causes one above  caused by one above  no causal relationship

Student LikePurpleYellow Age VoteUKIP SmokePipe Gender

This is where you find which variables need to be adjusted This is where the ‘Model Text Data’ can be copied and saved in a text file

Summary - only confounders cause both exposure and outcome - confounders alter relation between exposure and outcome - DAGs help to summarise/visualise causal relationships - www.dagitty.net can help identify confounders Reference Law GR, Green R, Ellison GTH. Confounding and causal path diagrams in: Tu Y-K, Greenwood DC (eds). Modern Methods for Epidemiology . Springer, London: 2012. (available from www.leeds.ac.uk/light/research/time/)