Group Averages and Individual Lives

Imagine a study evaluating a programme designed to improve a measured outcome.

Across the study group, the measured outcome improves by five points on average.

That finding may be scientifically meaningful. If the study design supports a causal interpretation, the average difference may estimate an average causal effect.

But it does not mean that every participant improved by exactly five points.

Some observed changes may be larger. Some may be smaller. Some may show little change at all. The amount and pattern of variation depend on the data.

Now reverse the mistake.

Suppose one participant shows almost no observed improvement.

That does not automatically show that the group result was false.

The average and the individual answer related but different questions.

An average can be informative without being a description of every individual.

That distinction matters because much of science learns about people by studying groups, while the questions we care about often concern a particular person.

The relevant problem is therefore not whether group evidence can apply to individuals.

It can.

The problem is how far the inference legitimately travels.

What does this group-level evidence make reasonable to expect about this individual, and how much uncertainty remains?

An Average Is a Summary of a Distribution

An average compresses information.

Suppose several people complete the same task and their scores are combined into a mean. That mean tells us something about the centre of the observations.

It does not tell us how those observations are distributed around it.

Two datasets can have the same mean while looking very different.

In one, most values may cluster closely around the average.

In another, they may be widely dispersed.

Two groups can have different means while their distributions overlap substantially.

Averages summarize distributions; they do not replace them.

This matters because hearing a group average can invite us to imagine an “average person” who somehow embodies it.

But the mean belongs to the distribution.

An average is a property of a distribution, not a description of every person within it.

A mean alone does not tell us:

how much variation surrounds it;

how much two groups overlap;

whether values cluster tightly;

whether there are important tails or subgroups.

And, as established in Measurement and Operational Definition, interpretation depends first on what was measured and how well the measure represents the construct of interest.

A precisely calculated average of a poorly measured outcome remains limited by the measurement itself.

D2.6 adds another principle:

Even a well-measured average leaves individual variation to be interpreted.

Average Effect Does Not Mean Uniform Effect

The distinction becomes more demanding when we move from average scores to average effects.

Suppose a well-designed study supports the conclusion that an intervention improves an outcome on average.

An average causal effect describes the average causal contrast supported under the design and assumptions of the study.

It does not establish the same causal effect for every individual.

An average effect describes an average, not a rule that every person follows.

Three different quantities can easily be confused here.

One is observed change: what happened to the measured outcome for a person.

Another is the average causal effect estimated across the relevant population.

A third is the causal effect for one particular individual.

Those are not interchangeable.

If someone’s score rises seven points after an intervention, seven points is their observed change.

Observed change is an outcome, not automatically a measurement of that person’s causal treatment effect.

Other influences may also have contributed to what was observed. Likewise, little observed change does not by itself establish that the intervention caused no change for that person.

This is a direct application of the distinction developed in Correlation, Prediction, Causation and Mechanism: the form of the evidence constrains the form of the conclusion.

Statistical reliability does not remove this distinction.

A group effect may be estimated reliably while still being:

modest in magnitude;

variable in observed expression;

uncertain when translated to one person.

Statistical reliability, effect magnitude and individual uniformity are different properties.

Group Differences Are About Distributions

The same discipline applies when groups are compared.

Suppose Group A scores higher than Group B on average.

That statement can be correct even if their distributions overlap substantially.

Some people in Group B may score above many people in Group A.

Some people in Group A may score below many people in Group B.

The difference between the means remains real.

So does the overlap.

A difference between group averages is not a boundary separating every member of one group from every member of another.

The inferential error happens when:

Group A has a higher mean

quietly becomes:

members of Group A are high, while members of Group B are low.

The second statement is stronger.

It turns a distributional difference into a deterministic classification of individuals.

Group-level findings can still be scientifically important. They can identify systematic differences that matter for explanation, prediction and later research.

But the interpretation must stay at the level the evidence actually supports.

This distinction becomes especially important in downstream discussions of personality, genetic influence and cognitive ability.

Group differences can matter without becoming individual destinies.

Population Evidence Changes Probabilities, Not Destinies

Relevant population evidence can change what is reasonable to expect about a person.

Suppose a predictor is associated with a much higher probability of an outcome under relevant conditions.

That matters.

A probability of 80 percent and a probability of 10 percent are both uncertain, but they do not support the same expectation.

Population evidence changes what is reasonable to expect without uniquely fixing one person’s future.

This is why saying:

statistics do not apply to individuals

is too strong.

Population evidence can inform individual inference by making some outcomes more plausible and others less plausible.

What it generally does not provide is certainty.

Uncertainty limits inference; it does not erase information.

How directly the evidence applies depends on what kind of population generated it and how closely the individual and their conditions match that evidence.

Relevant questions include:

Was this kind of person represented in the study?

Are the circumstances comparable?

Is the predictor measured in the same way?

Does the model remain appropriate under the current conditions?

What other information changes the expectation?

Population evidence creates a starting point.

Individual information can sometimes refine it.

Conditional Evidence Can Improve Individual Inference

An overall average is not always the most informative estimate available.

Suppose an intervention has one average effect across a broad population, but strong evidence shows that the effect differs systematically according to a particular baseline characteristic or condition.

That information can refine the expectation.

Instead of relying only on:

the average effect across everyone studied,

we may be able to estimate:

the average effect under a more specific set of conditions.

Reliably supported conditional information can improve individual inference, but greater specificity does not guarantee greater accuracy.

Where a person starts can matter.

Context can matter.

Validated moderators can matter.

Similarity to the studied population can matter.

Measurement quality can matter.

But collecting more personal characteristics does not automatically make a model better.

A variable may predict outcome without modifying the effect of an intervention.

A relationship found in one dataset may fail elsewhere.

A complicated model may look highly personalised while mostly fitting noise.

This is why Mediators, Moderators and Mechanisms matters here.

Reliable moderators can make inference more conditional.

They do not turn conditional averages into certainty.

A Subgroup Is Still a Group

One way to refine an overall average is to examine a narrower subgroup.

That can be scientifically valuable.

A subgroup defined by a relevant baseline characteristic may provide an estimate more applicable to a particular person than the overall population average.

But the subgroup still contains multiple people.

A narrower group can improve relevance without becoming an individual.

Variation can remain within that subgroup.

And uncertainty exists at two levels.

First, the subgroup estimate itself may be uncertain.

Second, individual outcomes can still vary around that subgroup estimate.

Uncertainty can remain both in the subgroup estimate itself and in how individuals vary within that subgroup.

Subgroup analysis also introduces another danger: false precision.

When groups are small, or many possible distinctions are explored, seemingly specific findings can become unstable.

More specific evidence is not automatically more reliable evidence.

The goal is not maximal subdivision.

It is well-supported conditional inference.

Observed Variation Is Not Automatically Causal Heterogeneity

Suppose participants in an intervention show very different observed changes.

One improves substantially.

Another changes a little.

Another hardly changes.

It is tempting to conclude:

the intervention works differently for different people.

That may be true.

But the observed pattern alone does not establish it.

Two kinds of heterogeneity must be separated.

Outcome heterogeneity means people show different observed outcomes or observed changes.

Causal-effect heterogeneity means the effect of the exposure or intervention itself differs across people or conditions.

Heterogeneity of outcomes and heterogeneity of causal effects are different claims.

Observed outcomes can vary because of:

different baselines;

measurement error;

ordinary within-person fluctuation;

events unrelated to the intervention;

other uncontrolled influences;

genuine heterogeneity in causal effect.

The existence of variation therefore creates an interpretive problem.

It does not solve it.

Different observed outcomes do not, by themselves, identify different causal effects.

To claim heterogeneous causal effects, evidence must support differences in the effect of the intervention or exposure itself—not merely differences in outcomes observed after it.

This is where D2.2 and D2.3 work together.

Measurement discipline asks how much variation may arise from the measurement process.

Causal discipline asks what part of the observed difference can legitimately be attributed to the intervention or exposure.

Variation is evidence to interpret, not automatic proof of causal heterogeneity.

True effect heterogeneity can exist.

But it must be demonstrated rather than presumed.

Between People Is Not the Same as Within a Person

Another inferential boundary is especially important in a Library concerned with change.

Suppose that across a population, people who use a particular study strategy more often tend to perform better.

That is a between-person association.

People differ in strategy use, and those differences are associated with differences in performance.

Now ask:

What happens if one particular learner increases their use of that strategy?

That is a within-person change question.

The first finding does not automatically answer the second.

Differences between people and changes within a person are different evidential questions.

A between-person relation may reflect differences in:

prior skill;

available time;

motivation;

environment;

other characteristics.

Even if the association itself is robust, a separate question remains about what happens when one person changes.

This does not make between-person evidence weak or invalid.

A between-person association can be valid and important while remaining insufficient evidence for the corresponding within-person change claim.

Nor should the distinction be exaggerated into:

between-person findings never apply within individuals.

Sometimes the two relations align.

The principle is narrower:

A between-person association does not automatically establish the corresponding within-person relation.

For a science of human change, this distinction is foundational.

Evidence about how people differ is not automatically evidence about what happens when a person changes.

An Average Pattern Does Not Reveal Every Individual Mechanism

Group-level effects can also tempt us to imagine one underlying mechanism operating identically in everyone.

Suppose an intervention produces a reliable average effect, and evidence suggests that a particular mediator contributes to that effect.

That does not establish that the same pathway operated to the same degree in every participant.

But individual variation does not justify the opposite assumption either.

A group-level mechanism does not establish identical person-level mechanisms, and observed individual variation does not establish that every person had a different mechanism.

The aggregate pattern leaves individual mechanism partly open unless person-level evidence resolves it.

This is where later nodes become relevant.

The Same Outcome Can Have Different Causes develops how similar outcomes may arise through different pathways.

The Same Cause Can Produce Different Outcomes develops how a common influence can have different consequences.

D2.6 needs only the inferential rule:

Aggregate similarity does not establish individual causal similarity.

Exceptions and Individual Cases

The population-to-individual distinction also works in reverse.

Suppose research finds that an outcome is more common under one condition than another.

Someone responds:

I know a person with that condition who had the opposite outcome.

That case may be entirely real.

It does not automatically refute the group result.

Probabilistic claims permit exceptions.

A claim that:

an outcome is more common

is different from:

an outcome always occurs.

One counterexample contradicts the second.

It does not logically contradict the first.

Individual cases can still be scientifically valuable.

They can reveal:

that an outcome is possible;

an unexpected harm;

a mechanism worth investigating;

a condition missing from an existing explanation;

that a purportedly universal claim is false.

What one case usually cannot establish by itself is:

population prevalence;

the average effect;

the full distribution;

the probability for other people.

An individual case can be informative without substituting for evidence about a population.

Population evidence and individual cases therefore have different evidential roles.

Neither becomes more scientific by pretending to answer the other’s question.

Knowing the Average Precisely Is Not Knowing the Individual Precisely

Imagine a very large study.

Because the sample is large, the population mean or average effect may be estimated with considerable precision.

It is tempting to infer that individual prediction must therefore also be precise.

That does not follow.

Individual outcomes can remain widely dispersed around a group quantity that is itself estimated extremely well.

Knowing the average precisely does not mean knowing each individual precisely.

The two forms of uncertainty concern different targets.

One concerns uncertainty about the population quantity being estimated.

The other concerns how uncertain the outcome remains for a particular person even if the population quantity were known very well.

A large sample can substantially reduce the first while leaving the second much larger.

This distinction is one reason estimates of population parameters and predictions for individual observations require different inferential tools.

D2.6 does not need the technical machinery.

The conceptual point is enough:

Precision at one level does not automatically transfer to another.

Group Evidence Still Matters

After enough cautions about averages, a different mistake becomes possible.

Perhaps group research is simply too crude to matter for individuals.

That conclusion would be wrong.

Population evidence can reveal:

systematic associations;

reliable group differences;

average intervention effects;

baseline probabilities;

credible moderators;

candidate mechanisms.

These findings constrain what is reasonable to expect about individual cases.

They can make some explanations less plausible and others more plausible.

They can show when an anecdote is common or unusual.

And population studies often provide part of the evidence from which more conditional individual predictions are developed.

The alternative to deterministic use of group evidence is probabilistic use—not abandonment of group evidence.

Population evidence is generated from individuals, but applying it to another individual still requires assumptions about similarity, conditions and uncertainty.

How much uncertainty remains is itself an empirical question.

Individual prediction may sometimes be weak.

In other settings, the evidence can narrow possible outcomes substantially.

D2.6 does not assume one universal level of uncertainty.

It requires the level of confidence to match the evidence.

From Population Evidence to One Person

Return to the hypothetical programme with a five-point average improvement.

The group finding remains meaningful.

But when one person asks:

Will this work for me?

a second inferential task begins.

What population generated the evidence?

How similar is this person to that population?

Where do they begin?

Are there reliable moderators?

How variable were outcomes?

What did the measure capture?

Does the evidence support prediction, causation or mechanism?

How much uncertainty remains around the person-level inference?

The answer may become more specific when reliable conditional information is available.

Or it may remain broad when the relevant uncertainty cannot be reduced.

Either can be scientifically appropriate.

What matters is that the translation from population evidence to the person is made explicitly rather than assumed.

Group evidence can constrain what is reasonable to expect about one person without determining that person’s trajectory, mechanism or outcome.

That is the relation D2.6 establishes.

A group-level result need not describe every individual to be useful.

An individual need not match the average for the average to remain scientifically meaningful.

The translation requires distributions, conditions, measurement, relevant moderators and uncertainty.

From population evidence to one life, the inference is probabilistic and conditional—not arbitrary, and not predetermined.