Observation, Explanation and Intervention

Suppose researchers observe that people who sleep less tend to perform worse on a cognitive task.

That finding can be real and scientifically important.

But it does not yet tell us why the pattern exists.

It does not by itself establish that reduced sleep causes the poorer performance.

And even if stronger evidence later supports a causal role for sleep, another question remains: would deliberately increasing sleep improve performance under the conditions we care about?

These are related questions, but they are not the same.

A science of human change has to distinguish among what happens, why or how it happens, and what changes when we deliberately alter something.

Observation, explanation and intervention are not mutually exclusive kinds of study. They are different questions evidence can help answer and different inferential achievements research can make.

The danger is not that science asks these questions separately.

The danger is that evidence answering one is too quickly promoted into an answer to another. This is where the prior question of what counts as evidence for a given claim becomes unavoidable.

Evidence should be interpreted according to the question it actually answers, not the stronger question we wish it had answered.

Observation tells us what occurs

Observation is not a weak form of science.

Before a phenomenon can be explained, it has to be characterised.

Researchers may observe how common a behaviour is; how an outcome changes over time; whether two variables tend to occur together; whether behaviour differs across contexts; whether one pattern tends to precede another; and whether naturally occurring differences are associated with different outcomes.

These observations can reveal stable and important regularities.

Suppose shorter sleep is repeatedly associated with poorer performance.

That tells us something real about the pattern.

But several causal explanations may still be compatible with the same observation.

Perhaps reduced sleep contributes to poorer performance. Perhaps demanding work contributes both to shorter sleep and poorer performance. Perhaps stress affects both. Perhaps poorer functioning itself disrupts sleep. Perhaps several processes interact.

The observed pattern does not uniquely select among them.

That does not mean observational evidence is irrelevant to causation. Observational evidence can contribute to causal inference when the design, assumptions and analytical strategy support that inference.

The important point is narrower: an observed pattern does not make its causal explanation self-evident.

Researchers also observe indicators through particular measurements and operational definitions rather than entire constructs directly.

So even before causal interpretation begins, another question matters: what exactly was measured?

That is why observation remains connected to Measurement and Operational Definition.

A pattern can be real. Its interpretation can still be open.

A pattern can be real without its cause being known.

Prediction and explanation answer different questions

Observation often leads naturally to prediction.

If a pattern is stable enough, we may become good at forecasting where an outcome will appear.

A model might predict which people are most likely to disengage from a behaviour-change programme using their previous attendance, recent engagement and behavioural history.

The prediction might be highly accurate.

But predictive success does not necessarily tell us what produces disengagement.

A variable can contain useful information about an outcome without being the process that causes it.

A model can therefore answer who is likely to disengage? while remaining incomplete as an answer to why do they disengage?

Prediction and explanation are both scientifically useful. They simply serve different purposes.

This is one reason the distinction among Correlation, Prediction, Causation and Mechanism matters so much.

Prediction can tell us where an outcome is likely to appear without telling us what produces it.

Explanation asks why or how the pattern occurs

Explanation goes beyond establishing that a pattern exists.

In this essay, explanation means an evidence-supported account of why or how a phenomenon is produced, sustained or constrained.

Depending on the question, that may involve causal influences; mechanisms; enabling conditions; interactions; triggers; and constraints.

An explanation therefore need not reduce a complex outcome to one isolated causal variable. The Library's distinction among Causes, Conditions, Triggers and Constraints is important precisely because different features of a causal system can play different explanatory roles.

Return to sleep and performance.

Suppose stronger evidence supports the claim that reducing sleep causally affects performance on the task.

We now know more than we did from the original association.

But establishing a causal contribution can advance explanation without exhausting it.

Questions may remain about which processes mediate the effect; which components of performance are affected; the conditions under which the effect changes; and how much of the everyday association is attributable to sleep rather than other influences.

Causal contribution and complete explanation are therefore not the same achievement.

And even a strong causal explanation still does not automatically identify the best intervention.

This is where one of the most important distinctions in a science of change appears: knowing what contributes to an outcome does not automatically tell us how best to change that outcome.

A causal factor is not automatically an intervention target

Suppose a past exposure or developmental event contributes causally to a current behavioural pattern.

That historical influence can matter to the explanation even though the past itself can no longer be changed. Its causal importance does not disappear because it is no longer directly manipulable.

An intervention must operate somewhere in the current or future causal system.

That already separates two questions: what caused or contributed to this outcome, and what can we feasibly alter now to influence what happens next?

But manipulability alone is not enough.

A causally relevant factor might technically be changeable while still being a poor intervention target. Changing it might be difficult; ethically inappropriate; poorly timed; weak relative to another available target; or limited by surrounding conditions.

What is most important to explaining an outcome is therefore not necessarily what is most useful to change.

Nor does a science of causation need to define causes only as things human beings can directly manipulate.

A factor can matter causally without being a feasible or useful point of intervention.

Good causal explanation can substantially improve intervention design. It can identify what matters, eliminate implausible targets and reveal where leverage may exist.

But it does not mechanically determine the intervention.

Knowing what contributed to an outcome narrows the intervention problem; it does not necessarily solve it.

Intervention asks what changes when the system is altered

Intervention asks: what happens when some feature of the system or its conditions is deliberately altered?

Under an appropriate design and assumptions, intervention evidence can support a causal claim about the effect of that alteration on an outcome.

The relevant causal contrast can sometimes be studied through an actual intervention and sometimes through research designs intended to estimate what would happen under a specified intervention.

Return to the sleep example.

Suppose researchers deliberately increase sleep opportunity under controlled conditions and later observe improved task performance.

If the design justifies the causal inference, the result provides evidence that changing sleep opportunity affects performance under the tested conditions.

That is a different achievement from merely observing that sleep and performance are associated.

This does not mean such findings can never generalise. It means generalisation is an additional inference that requires its own support.

Intervention should also not be equated with one particular research design.

Randomisation is one powerful strategy for supporting causal inference about interventions. But it is not the definition of intervention evidence, nor does randomisation by itself establish the mechanism or show that the same effect will occur in different populations, contexts or implementations.

Intervention evidence is also claim-specific.

A strong study may support the claim that changing A changes Y under these conditions. That does not automatically establish that A changes Y through mechanism M. And it certainly does not automatically establish that theory T is the complete explanation of why A changes Y.

This is where Replication, Robustness and Scientific Confidence becomes essential. The same intervention result may justify high confidence in one claim and much lower confidence in another.

Intervention evidence can show that changing something changes an outcome without explaining everything that happened in between.

Intervention effect is not mechanism

Suppose a theory proposes that intervention A works through mechanism M to produce outcome Y.

Researchers test A and find that Y changes.

Under an appropriate design, that result may support the claim that A affects Y. But that is not yet the same as establishing that A affects Y through M.

Intervention A may change several features of the system at once. It could affect attention; expectations; opportunities; social interaction; behaviour; reinforcement; or environmental conditions.

Mechanism M may be important. Another process may matter more. Several pathways may contribute.

This is why intervention studies often require separate evidence about mediators, moderators and mechanisms.

Mechanistic understanding can also be partial and still scientifically valuable. The point is not that science must either know the complete mechanism or know nothing about it.

The point is that the intervention effect alone does not tell us how much of the mechanism has actually been established.

An intervention can work without revealing the complete mechanism through which it works.

This does not make intervention evidence explanatorily irrelevant. Quite the opposite.

Interventions can rule against some causal accounts; strengthen others; establish direction of influence under appropriate conditions; and reveal unexpected dependencies.

Intervention evidence can contribute to explanation without completing it.

Intervention success does not uniquely confirm its theory

Scientific theories often motivate interventions.

Suppose theory T predicts that A changes M and that M changes Y. An intervention using A changes Y.

That result is relevant to theory T. It may support one of the theory's predictions.

But it does not uniquely confirm the theory.

Another theory may also predict that A changes Y. A may affect several processes besides M. The intervention may work for reasons the original theory did not anticipate.

The more specifically a theory predicts an intervention result that plausible alternatives do not, the more informative that result can become for comparing explanations.

But the general rule remains: intervention success can provide evidence relevant to a theory without uniquely confirming that theory.

This reveals another important distinction: intervention effectiveness and explanatory adequacy are different properties.

An intervention can be effective while its mechanism remains uncertain. An explanation can be scientifically informative while offering no immediately useful lever for change.

Neither achievement should be mistaken for the other.

Failure also needs interpretation

Failed interventions are easy to overread in the opposite direction.

Suppose a theory predicts that changing X should change Y. Researchers intervene on X, and Y does not change.

Does that prove X was never causally relevant?

Not necessarily.

First we have to ask whether the intervention actually changed X in the way the theory required.

An intervention intended to alter a causal process may fail to engage that process. Implementation may fail. The manipulation may be too weak. Compensating processes may preserve the outcome.

But these possibilities cannot become an all-purpose excuse.

If a well-designed intervention successfully changes the intended target and the predicted outcome still does not occur, confidence in the relevant prediction should decrease.

What an intervention failure tells us depends on what the intervention actually changed and what the theory predicted would follow.

Failure is evidence. The question is which claim it bears on.

The same result could be strong evidence against the claim that changing X under these conditions changes Y, while being much weaker evidence against the claim that X has ever contributed causally to Y in any context.

Those are not the same claim.

Observation, explanation and intervention work together

The distinctions in this essay should not leave the impression that science is divided into three sealed compartments.

Observation, explanation and intervention interact iteratively.

Observation identifies a pattern. That pattern motivates competing explanations. Explanatory work generates causal predictions. Intervention can test some of those predictions. The intervention produces new observations. Unexpected effects, failures and boundary conditions force explanations to change. Revised explanations motivate new observations and interventions.

The relationship is therefore not a universal ladder climbing from weaker to stronger knowledge.

Some designs provide stronger evidence for particular claims than others. But there is no universal rule that makes intervention evidence superior for every scientific question.

A descriptive study may be exactly what is needed to establish the distribution of a phenomenon. An explanatory study may clarify an important causal structure that no available intervention can directly manipulate. An intervention study may establish an effect while remaining mechanistically incomplete.

Observation, explanation and intervention interact iteratively rather than forming a universal ladder of evidential superiority.

They are questions and inferential achievements, not mutually exclusive study types.

Three questions, three achievements

The distinction can be compressed.

That X and Y occur together supports an association; it does not establish that X causes Y.

That X predicts Y supports predictive information about Y; it does not establish why Y occurs.

That evidence supports X as a causal contributor to Y supports causal understanding at the level tested; it does not establish that X is the best practical intervention target.

That altering X changes Y under an appropriate design supports a causal effect under the tested conditions; it does not establish the mechanism through which the effect occurred.

That intervention A changes Y supports the claim that A can affect Y under the tested conditions; it does not establish that the theory motivating A is uniquely correct.

That evidence supports M as part of the causal pathway supports the claim that M contributes to the process under the tested conditions; it does not establish that manipulating M alone will reliably reproduce the outcome.

The essay can therefore return to three simple questions: what happened, why or how did it happen, and what changed when we intervened?

A mature science of human change needs all three.

Without observation, we do not know what requires explanation. Without explanatory inquiry, patterns can be mistaken for their own causes. Without intervention evidence, causal understanding does not automatically tell us what deliberate changes will work.

But even when all three are available, the evidence still has to be interpreted according to what was measured; what the design can establish; what competing explanations remain; how robust and replicated the result is; and which claim is actually under evaluation.

That is the deeper discipline this essay establishes.

Evidence should be interpreted according to the question it actually answers, not the stronger question we wish it had answered.