Considerations for Causal Inference Studies

Epidemiologists and researchers design studies with one or more of four primary objectives: description, prognostic modelling, association, or causal inference. Association studies often serve as a precursor to causal inference but are not necessarily equivalent. The concept of cause and effect is innate to human understanding, but its definition and application can vary across science disciplines. Philosophically, one can identify flaws or practical limitations in attempting to prove causal propositions, and the debate about the interplay of logic, belief, and observation in assessing causal claims is not completely resolved. In the setting of epidemiology, it is practical to view causal inference as measuring an effect rather than applying criteria to determine the existence or not of an effect [1, 2]. Causal inference is the methodological approach to provide evidence for the estimation of the isolated effect of an exposure or intervention on an outcome. It is used to study an exposure or intervention that can be controlled or changed and how an outcome may correspondingly change [1]. The establishment of causality underpins randomised controlled trials (RCTs) [3]. Their primary strength in determining causation is their ability to account for confounders, both known and unknown. Randomisation, when done well, reduces bias by evenly distributing characteristic factors across comparison groups such that confounders are similarly distributed, ensuring that changes in outcomes result solely from the intervention. RCTs can have limitations to do with limited generalisability and external validity due to strict inclusion criteria and may themselves suffer from low methodological quality [4, 5]. However, RCTs may be infeasible in real-world settings due to unethical interventions, for example, the effect of smoking on pregnancy outcomes, or impracticality, for example, rare outcomes or long-term exposures [4]. Therefore, there is a need to draw causal effects from observational and longitudinal data. Causal inference methodologies use observational and longitudinal data to emulate a target trial that would estimate the causal effect of interest. Counterfactual or potential outcomes are central for this as they are contrasted under treatment or non-treatment, forming the basis for causal estimation. The treated group's data are used to model expected outcomes for the treated and the counterfactual treatment outcomes for the untreated and vice versa using the untreated group's data. The average difference between the groups’ expected potential outcomes defines the average treatment effect, a key causal estimand [1] but there are many others [6]. Implementation science suggests that population attributable effects perhaps should be prioritised [7]. Due to the lack of treatment randomisation, controlling for confounding is critical, necessitating a hypothesised causal structure between exposure, confounders, and outcome variables [8]. Directed acyclic graphs are a vital tool for this, making causal assumptions explicit and incorporating domain knowledge to guide analyses [9]. This is important as causality is a multifaceted and intricate phenomenon that relies not only on the information available in the data but also on contextual external and pre-existing knowledge. This is crucial in the setting of effect mediation or modification [10]. Specific assumptions collectively support causal inference. The causal question must align with the analysis, implying the exposure is well-defined and consistent for all individuals (i.e., no multiple versions of exposure). Additionally, it is assumed that one individual's exposure does not influence another's outcome, a principle known as no interference, or the stable-unit-treatment-value assumption. Further, positivity is assumed, meaning all individuals have a non-zero probability of experiencing each level of exposure. Finally, exchangeability requires that potential outcomes be independent of the assigned exposure level, conditional on covariates adjusted for in the analysis [1]. This means the exposure influences the outcome, but the assignment of specific exposure levels is independent of the outcome [8]. Many well-established statistical methods facilitate causal inference, including regression techniques such as logistic, linear, non-linear, quantile, multi-level, and structural equation modelling. Over the past 40–50 years, innovative methods such as mediation analysis, g-computation, inverse probability weighting, propensity scores, instrumental variables, doubly robust estimators, and techniques for longitudinal and optimal dynamic treatment regimens have been developed and widely utilized by applied researchers to address significant causal questions [11]. Bayesian causal inference can be used to address the common issue of missing data in observational studies [12]. Additionally, the growing use of genetic data in birth cohort studies has popularised Mendelian randomisation as a method for investigating causality [13]. In recent decades, causal inference applications have grown exponentially and include methods for high-dimensional data, precision medicine, causal machine learning, and causal discovery [14]. It is important to stress that causal inference is more than a purely statistical or modelling methodology. It requires clarity about definitions of the exposure, outcome, population of interest, choice of appropriate estimand and analytical method and, crucially, incorporates contextual and domain knowledge [1, 8, 10] (Figure 1). As should be expected, not all observational data are suitable for causal inference. Limitations due to unknown confounders need to be considered. Multicausality may necessitate the identification of a minimal set of conditions rather than a single cause. Furthermore, one observational study is typically insufficient to determine causality and should be replicated in different settings. Within the broader field of causal inference, the modern epidemiological counterfactual emphasis briefly described here represents one perspective. While it has strengths, it may overly constrain what is considered a cause, narrowing the field's scope and increasing the risk of erroneous conclusions, potentially causing harm, particularly if key biological and social assumptions are poorly formulated. This modern approach diverges somewhat from the recent surge of literature in philosophy, history of science, and other natural and social sciences, where lively debates continue about the types and processes of causal inference and their explanation. It has been proposed that, to enhance causal inference and explanation, we should emphasise the importance of causal triangulation, where epidemiologists utilise diverse study designs with distinct and unrelated biases. Hypotheses should be tested across various populations and historical periods to assess robustness against confounding structures and the analytical methods employed [15]. Despite these challenges, with careful consideration of the available data, appropriate methodologies outlined here can yield robust causal inference from observational studies. Open access publishing facilitated by Curtin University, as part of the Wiley - Curtin University agreement via the Council of Australian University Librarians. The authors declare no conflicts of interest.