Why We Should Teach Causal Inference: Examples in Linear Regression With Simulated Data

Abstract

Basic knowledge of ideas of causal inference can help students to think beyond data, that is, to think more clearly about the data generating process. Especially for (maybe big) observational data, qualitative assumptions are important for the conclusions drawn and interpretation of the quantitative results. Concepts of causal inference can also help to overcome the mantra "Correlation does not imply Causation. " To motivate and introduce causal inference in introductory statistics or data science courses, we use simulated data and simple linear regression to show the effects of confounding and when one should or should not adjust for covariables.

Might Temporal Logic Improve the Specification of Directed Acyclic Graphs (DAGs)?

“…Notwithstanding George Smiley's caution and the fallibility of method, statistical expertise plays an increasingly important role in generating and interpreting quantitative evidence to inform policy and practice (Teater et al 2017;Flyberg et al 2019); and training in statistical skills has always needed to keep pace with ongoing developments in analytical practice (Tu and Greenwood 2012;Porta et al 2015;Efron and Hastie 2016;Hokimoto 2017;Lübke et al 2020). Indeed, perhaps the most important contribution that statistics can make to evidence-based decision-making (and, by extension, the contribution that statistical skills training can make to professionals and the lay public, alike) is in revealing and dealing with the many different sources of bias that can occur when analysing and interpreting obvious, observable differences between ostensibly comparable phenomena (Flyberg et al 2019).…”

“…Many of the same sources of bias that led to the earlier focus on experimentation and randomisation particularly those relating to confounding and sampling biascontinue to threaten the validity of observational analyses, not least because design constraints mean that such studies exert little control over the allocation of naturally occurring 'exposures' (be these physical, biological or social phenomena). Nonetheless, recent efforts to address confounding and sampling bias have made substantial progress; and the emergence of 'causal inference' as a novel interdisciplinary field spanning statistics, mathematics and computing (as well as the applied social and biomedical sciences), has helped in the translation of abstract theoretical techniques into accessible and practical applications (Porta et al 2015;Lübke et al 2020).…”

Statistics education from a data‐centric perspective

“…Despite its limitations, the general linear model remains a pivotal tool for data analysis and continues to provide avenues for progress [18]. Moreover, causal modeling too can be included [24].…”

Opportunities to improve learning analytics for student support when using online assessment tools

“…There is also possible confounding between the variables of the model, as is often the case in the complex world of learning (Davis & Sumara, 2006), and interaction effects were not considered for this initial study. For instance, you could posit that the (currently unseen) positive effect of starting early on an assessment is mediated through the LMS activity which would then be masked in the regression model used here; Lübke et al (2020) has explanatory examples that highlight this. Teasing out the genuine direct effects of the predictor variables requires more sophistication than what is presented here.…”