o This also means that our research conclusions rest, in part, on our assumption that no causal relationship exists between them
Software
A possibly neglected issue in the promotion of causal diagrams has been the availability of software and published guidance on the choices that are available. A number of software packages have been developed over the years to facilitate the drawing and analysing of causal diagrams. One of the first was TETRAD in 1986,265 becoming the TETRAD Project in
1998,266 but it was aimed primarily at structural equation modelling. It has since been
expanded and is available at www.phil.cmu.edu/projects/tetrad/, however, it is still not really aimed at most types of health research.
The only software package specifically designed to create DAGs that has been made known to health research through publications in epidemiology journals is DAGitty,267 available at www.dagitty.net and also as the R package ‘dagitty’.268 As such, to our knowledge, it is the
only package that has been mentioned whenever the software used to create a DAG is listed in an article. And while it is being improved from time to time, it is non-commercial software with very few programmers, so progress is slow, and its limited features and interface full of what to many, is technical jargon, may act to discourage some researchers from getting started with causal diagrams.
Alternatives to DAGitty are mostly diagramming software packages like Microsoft Visio (visio.microsoft.com), LucidChart (www.lucidchart.com) and Gliffy (www.gliffy.com). However, while easy to use, they do not offer features that are specific to DAGs.
3.5 Uses of causal diagrams
The widespread use of diagrams to convey abstract information shows it is generally accepted that diagrams can assist in the understanding of abstract concepts, at least
sometimes.269 Research in cognitive science has suggested that diagrams can make it easier
to find the information relevant to a concept,270 such as the causal paths between variables
alternative possibilities by making all the possibilities explicit,271
,272 such as when a researcher
is forming conclusions at the end of a study, based partly on alternative explanations for the results.
Causal diagrams, which in most cases are DAGs, provide an intuitive framework that can help researchers conceive of and understand the biases that might influence a study, and can make communicating more difficult concepts easier than explaining solely with words.59 This
makes DAGs a useful tool to enhance the communicating of concepts relating to bias, whether teaching basic concepts59
,150
,253
,273 or publishing the results of methodological
research.274
–277 This is especially the case with the structural classification of bias, covered in
the previous section, but DAGs have also been used to explain more specific types of bias, such as different types of time-dependent confounding,278 missing data biases,244
,279
–281 and
possible explanations for apparent paradoxes such as Simpson’s paradox,282 the birth weight
paradox,283 and the obesity paradox.284
–286
It is now well established that an analysis of observational data should take into
consideration not only the study design, but also substantial background subject-matter knowledge if the goal is to obtain evidence regarding a causal association.255
,287 Otherwise,
important uncontrolled confounding might not be considered when making inferences, or variables might be included in a model that instead of reducing bias, increases it via collider bias. Also, by constructing a causal DAG that aims to adequately represent background causal knowledge, a researcher or statistician might be prompted to include variables that otherwise would not have been considered.
This means that if a DAG is constructed during the planning stage of a study, potential confounders that otherwise might not have been considered, can instead be either
controlled by modifying the design, or else have data collected on that variable so it can be used to adjust the analysis.224 The DAG can also be used to communicate this understanding
to fellow investigators or study staff, or to ask for feedback from subject matter experts.59
Once a study’s data has been collected, a DAG can be useful in identifying previously unconsidered sources of bias, such as from missing data,244 loss to follow-up279 or time-
3.5 Uses of causal diagrams
dependent confounding.288 And this can help plan the analysis with the most appropriate
methodology.47
It is also possible to use a DAG to identify a minimally sufficient set of variables that is needed to control for confounding in the analysis.224 This would exclude variables such as
intermediates on the causal pathway between the exposure and the outcome. The program DAGitty was recently criticised, however, because it can calculate such a set automatically. This may potentially mislead a researcher into thinking they could successfully control for confounding by adjusting for the variables DAGitty chose, even though important
confounders were not included in the DAG.76
Finally, a DAG can help with the interpretation and communication of the results. By making the assumptions on which causal inferences rest more explicit, such as the possibility of confounding from sources that were not controlled, conclusions by researchers might be more likely to be adequately cautious. The DAG can, and should, also be included with any published report, to help communicate the sources of bias identified, how they were controlled in the design and the analysis, and the assumptions and associated uncertainty that remains following the analysis. Unfortunately, it is still not uncommon to find articles that merely mention that a DAG was used to help select the model covariates, without providing the DAG itself.