# Does RCTs imply causation?

Can statistics help to definitively deduce causality? RCTs require that you state the set of variables that you believe influence a given outcome of interest, a priori, therefore if one wants to understand the effect of a treatment/intervention, the typical procedure would be to model the intervention distribution based on the variables which act on the intervention, however, if the variables which you have listed a priori is not definitive, then the effect would be biased.

There are a few other well-known methods for establishing causality, for example, the Bradford Hill guideline in Epidemiology is not entirely empirical since it is based on assumptions that can vary depending on the dataset that is gathered, for example, an assumption is that the effect size is greater than that of its confounders but since confounders are exogenous, then its effect can never be known.

• Does this answer your question? Is there any theory or field of study that concerns itself with modeling causation rather than correlation? Jun 16 at 15:57
• Well, I would say that the answer to the question in your post is "Yes," and the post to which I linked shows how that works, even in some situations where RCTs are not available (not ethical, e.g.). There are algorithms in the New Causal Revolution that can deduce causal graph structures, with a few limitations (can't distinguish between two graphs with the same $v$ structures). Jun 16 at 16:06
• Well, sure, but that's not technically the question you asked. You asked if statistics can help to definitively deduce causality. I quite frankly don't see how, in most cases, you could deduce causality without statistics. Jun 16 at 16:20
• From another thread: even with an intervention you have to assume that all the other variables that could be affecting the dependent variable are fixed while you are intervening. Otherwise, something else might have happened simultaneously with your intervention that could mask the effect of the intervention. Thus causality is always coming from an assumption. Of course, in an experimental setting you can make the assumption more plausible than otherwise, but strictly speaking the assumption is unavoidable. (What if there is an evil spirit that affects Y whenever you are intervening onto X?) Jun 16 at 20:39
• "RCTs require that you state the set of variables that you believe influence a given outcome of interest" No they don't. In fact, that's the whole point of them; you don't need to know anything about the causal process or system to interpret the estimate from a successful RCT as causal. "RCT is not entirely empirical. It does not balance confounders." Again, this exactly what RCTs do. They balance the expected (unobserved) potential outcomes so no model for the outcome is required. The difference in outcome means is the causal effect (assuming the RCT was successful).
– Noah
Jun 17 at 0:20

The RCT provides the most compelling evidence of causation, even more so when blinding is performed. It meets the most stringent criteria for declaring an effect to be causative. As always, an answer in list form is helpful but not exhaustive.

1. Randomization: balances the possible confounding factors
2. Administration of a placebo is possible only in a controlled clinical study setting, at least in such a way that comparison is possible.
3. Prospective/longitudinal design: there is no retrospective or cross-sectional analysis.
4. Oversight by independent ethics, safety, and regulatory committees: The number of independent people with the power to veto a randomized study further justifies its conduct. A study without equipoise or adequate monitoring/conduct is at risk of being halted, stopped, and the PIs can actually face disciplinary action personally and professionally. For instance, investigators guilty of misconduct can be blacklisted and, in doing so, be prevented from participating in
5. Cost barrier: of all the study designs, RCTs may be the most cost-prohibitive to do. As such, the funders are yet another independent group weighing in on the rationale and probable success of the study. Refining the scope and reducing the overall number of studies performed ameliorates the issue of publication bias.

There are other factors and considerations however:

1. Participation bias: subjects may enroll to studies because they want to gain access to an exclusive treatment, and their ongoing participation may be subject to whether they think they're getting it; i.e. will not consent to a blinded study, or will unenroll due to "lack of benefit" when they suspect they're getting a control (standard of care +/-
2. Blinding may not be possible: the novel treatment may be an oral form of a previous subQ (subcutaneous) injection, to investigate the tolerability, preference, and QoL, it doesn't make sense to require the active arm to inject saline.
3. Critical analyses may themselves not be randomized: so called per-protocol analyses for instance may only analyze subjects who are on treatment, or who received the planned treatment, excluding subjects who were not compliant or who disenrolled before even having the chance to receive treatment
4. Contamination and crossover effects still affect randomized analyses. Subjects will talk to each other in infusion centers, or they come from the same household. Randomization will not remove the residual correlation, nor will it account for variations in actual treatment in the intent-to-treat analysis. To point 1, subjects who go off study may be allowed to receive the investigational treatment which would attenuate the treatment effect if the subject was previously on control and the treatment was effective.