Causal inference and observational study design

No. Adjusting for confounders is a first step that is often mistaken for the last, but an adjusted regression answers a question of conditional association, not a causal one. This matters when the real question is "what effect does X have on Y?" and not just "are X and Y associated after adjustment?". The difference is fundamental. A multivariable regression adjusted for age and sex estimates a conditional association. Estimating a causal effect requires a design with explicit causal assumptions, an adjustment set derived from the causal structure of the problem rather than from statistical significance, and a sensitivity analysis that assesses how far the conclusion would depend on possible unmeasured confounding. Choosing which variables to adjust for by looking at p-values, without a causal framework, can introduce bias instead of removing it. What sets causal inference apart is not a more sophisticated model, but making explicit and defensible the assumptions under which an association can be read as an effect.

A DAG (directed acyclic graph) is the formal representation of the causal structure of the problem: what causes what, according to domain knowledge, drawn before any model is run. It applies to any study with a causal aim, because the causal structure determines which variables to adjust for, not the other way round. The DAG identifies the minimal adjustment set, that is, which variables to control for and, above all, which to leave alone: adjusting for a mediator removes part of the effect you want to estimate, and adjusting for a collider introduces bias where there was none. It also shows whether the causal question is identifiable at all with the available data, so that you can be transparent when it is not. Building the DAG with the research team, who bring the domain knowledge, is what anchors the whole analysis that follows. Deciding the adjustment without a DAG means choosing blindly which biases are corrected and which are created.

Without randomisation, the groups differ in characteristics that affect the outcome, and propensity score methods aim to balance those differences so that the comparison is more defensible. They apply when you want to estimate the effect of an intervention or exposure with observational data from registries, cohorts or administrative databases. The propensity score, the probability of receiving the treatment given the covariates, is used to match, to weight (IPTW with stabilisation and truncation) or to stratify. It can be combined with doubly robust estimation, which improves protection against misspecification if at least one of the models involved is correctly specified. The key lies in the balance assessment: a propensity score is validated by checking that the covariates end up balanced between groups, using standardised differences and comparison of distributions, not a p-value. If matching or weighting does not achieve acceptable balance, this is reported and alternatives are assessed; a result is not forced. Balancing the observed groups reduces bias, but it does not replace randomisation or remove unmeasured confounding.

Every causal estimate from observational data rests on assumptions that cannot be verified with the data themselves, so the job is not to "prove causality" but to make those assumptions explicit and quantify how sensitive the conclusion is to their violation. This applies to any observational study with a causal question, where honesty about the assumptions is part of the rigour. Instead of taking the assumptions for granted and moving on, we run a quantitative sensitivity analysis, such as the E-value or quantitative bias analysis. It estimates, under explicit assumptions, what magnitude of unmeasured confounding would be compatible with explaining away or weakening the estimated effect. That turns an implicit assumption into a quantifiable argument: an effect that only weakens under implausible confounding scenarios is more defensible than one that vanishes with a small bias. The conclusion is not presented as definitive proof equivalent to a trial, but as a causal estimate with its identification assumptions stated and its robustness quantified. That transparency is exactly what a demanding reviewer asks for.

Yes, if the design allows a defensible comparison; observing a before and an after is not enough. It applies when there is no randomised control group, but there are pre- and post-implementation data and a group or discontinuity from which a reasonable counterfactual can be built, and you want to separate the effect of the change from the trends that were already under way. Difference-in-differences compares the trajectory of the affected group with that of a comparison group, under the parallel trends assumption, which must be examined, not assumed. Depending on the case, regression discontinuity designs may also fit. The underlying premise is that, without randomisation, the study design becomes the analysis: designing the comparison well matters more than applying a sophisticated model to a poorly framed comparison. What you receive is an impact estimate with its design and assumptions made explicit, not an automatic attribution of the observed change to the policy.

Causal inference and observational study design is the broader umbrella: estimating the effect of an intervention or exposure on an outcome when there is no randomisation, with DAGs, propensity scores, difference-in-differences, g-computation and sensitivity analysis. Causal mediation and moderation analysis is one part of that umbrella, with its own audience and its own point of entry. It fits when the question is why or for whom an effect occurs: decomposing it into direct and indirect effects through a mediator, or seeing whether it varies across subgroups. They do not answer quite the same question. Mediation decomposes an effect into direct and indirect pathways under additional causal assumptions; general causal inference estimates the total or marginal effect of an exposure, intervention or policy. If you are looking specifically for "mediation analysis", there is a dedicated service for it; if your causal question is broader, this is where you start. The two are handled in coordination: mediation is a special case within this framework.

The methodological toolkit overlaps, with propensity scores, observational design and confounding control, but the framing and the context differ. Real-world evidence fits when the frame is generating clinical evidence from real-world data for a biopharmaceutical or medical affairs context, with its own language and its own audiences. Causal inference and observational study design is the frame for causal questions in health research and public health: the effects of interventions, exposures or policies in registries, cohorts or administrative data. Neither is more rigorous than the other: they share methods, but they address the same family of questions from different framings. If the need is framed as generating real-world evidence for biopharma, the natural frame is RWE; if it is framed as estimating a causal effect in a research or public health study, it is this service. The choice is one of context, not of method.

At the end of a causal inference analysis you receive the effect estimate together with everything needed to defend how it was obtained: the DAG that formalises the assumptions, a justified adjustment set and the estimation strategy chosen for the question (matching, weighting, g-computation, difference-in-differences or mediation). We include balance checks and sensitivity analyses for unmeasured confounding. The R code makes it possible to trace each result back to the model, variables and assumptions behind it. For peer review, we set out the identification assumptions and show what changes when they are challenged. The research team contributes the subject knowledge, decides the substantive interpretation and signs the work; we provide the causal framework and estimation. The result is not an unqualified causal claim, but an estimate whose identification, uncertainty and robustness can be examined.