Module 4 — Variation & Distributions

Module Learning Objective

This module explains why healthcare systems naturally exhibit variation and rarely behave in perfectly predictable ways — and why understanding variation is often more important than understanding averages.

Why Looking at the Data Matters

Healthcare analytics often relies on summary statistics such as:

averages (means)
medians
percentages
rates
standard deviations

These measures are useful because they simplify complex data.

However, summary statistics can sometimes be misleading.

Very different underlying patterns may produce:

the same mean
similar variation
similar summary statistics

while representing very different operational realities.

This creates an important lesson:

Always look at the shape of the data — not just the summary numbers.

A Statistical Cautionary Tale

In statistics, there are famous examples showing how datasets can appear almost identical numerically while looking completely different when plotted.

Examples include:

Anscombe’s quartet https://en.wikipedia.org/wiki/Anscombe%27s_quartet
Datasaurus Dozen https://en.wikipedia.org/wiki/Datasaurus_dozen

These examples demonstrate a simple but powerful message:

Datasets with similar summary statistics may behave in completely different ways when visualised.

For example:

Two datasets may have:

the same average
the same variance
the same correlation

yet reveal very different patterns when visualised.

One may show:

a clear trend

while another may contain:

clusters
outliers
hidden non-linearity
very different distributions

Key Lesson

Numbers summarise data — but plots help us understand it.

A Healthcare Example: Length of Stay

Healthcare data often behaves this way.

Length of Stay provides one of the clearest examples.

Healthcare data often behaves this way.

Length of Stay provides one of the clearest examples.

Two hospitals may report the same average Length of Stay while experiencing very different operational realities.

We return to this example shortly.

Trust A

Most patients stay: 4–8 days

Variation is relatively predictable.

Trust B

Most patients stay: 1–3 days

But a relatively small cohort stays much longer: - 30–60+ days

due to:

frailty
multimorbidity
delayed discharge
social care complexity

Operationally:

the systems may feel completely different despite identical averages.

This highlights an important lesson:

identical summary statistics do not imply identical operational systems.

Common Patterns of Variation in Healthcare

Before returning to Length of Stay in more detail, it is useful to understand some common ways healthcare activity behaves statistically. These patterns help explain why healthcare systems rarely behave in smooth or predictable ways.

Normal Distribution

Characteristics:

symmetrical
clustered around a central value
predictable variation

Healthcare examples:

adult height
blood pressure (in selected populations)
laboratory measurements

Interpretation

Variation is relatively predictable.

Right-Skewed Distribution (Most Common in Healthcare)

Characteristics:

many small values
fewer large values
long tail

Healthcare examples:

Length of Stay
ED attendances per patient
outpatient waiting times
frequent attenders
delayed discharge

Interpretation

A small number of patients often drive disproportionate operational demand.

Counts & Proportions

Healthcare systems often monitor:

counts (falls, incidents, infections)
proportions (% compliance, mortality, readmissions)

These are useful because healthcare activity is often uneven and noisy.

Bimodal Distributions

Characteristics:

two peaks

Healthcare examples:

ED waiting times split between minors and majors
mental health LoS groups
elective vs emergency pathways

Interpretation

May suggest:

different patient groups
hidden segmentation
pathway effects

Length of Stay (LoS): Looking Beyond the Average

Length of Stay (LoS) is one of the most common healthcare measures used to monitor hospital performance, patient flow and operational pressure.

At first glance, LoS may appear simple:

How long do patients stay in hospital on average?

However, Length of Stay data rarely behaves in a simple or predictable way.

In reality, LoS is often highly skewed.

Most patients may stay for only a short period of time, while a much smaller group experience considerably longer stays.

This creates what is known as a long tail distribution.

For example:

many patients may stay 1–3 days
fewer patients stay 7–14 days
a relatively small number may remain in hospital for 30, 60 or even 90+ days

This means the “average” Length of Stay can sometimes hide what services are actually experiencing operationally.

Understanding Skew and Tails

In a symmetrical distribution, most values cluster evenly around the middle.

Healthcare data often behaves differently.

Length of Stay commonly demonstrates a right-skewed distribution, where:

many short stays occur
progressively fewer longer stays occur
a small number of patients create a long operational tail

The important insight is:

not all patients contribute equally to hospital bed occupancy.

A relatively small number of patients can consume a disproportionate number of bed days.

This matters enormously operationally.

Why Longer stays may occur

Longer stays are often not random.

They may be disproportionately associated with patients who experience:

moderate or severe frailty
multiple long-term conditions (multimorbidity)
cognitive impairment
social care needs
housing instability
complex discharge requirements

For example:

A relatively fit patient admitted with pneumonia may recover and return home quickly.

In contrast, an older frail patient with:

heart failure
diabetes
dementia
reduced mobility

may require:

therapy input
social care assessment
discharge planning
community support coordination

Even if medically fit, discharge may remain delayed due to wider system complexity.

This means Length of Stay is often influenced by much more than clinical recovery alone.

Delayed Discharge and Operational Pressure

Delayed discharge provides an important example of how distributions matter.

Imagine two hospitals with a similar average Length of Stay.

On paper:

Imagine two hospitals:

Trust	Mean Length of Stay
Trust A	6 days
Trust B	6 days

The averages look identical.

The averages appear identical.

However:

Trust A

Most patients stay: 4–8 days

Trust B

Most patients stay:

1–3 days

But a smaller group experience:

30–60+ day stays due to delayed discharge and frailty-related complexity

Operationally, these hospitals may feel very different.

Trust B may experience:

bed occupancy pressure
delayed patient flow
discharge bottlenecks
corridor care risk
escalation pressures

even though the average Length of Stay is the same.

This highlights an important lesson:

identical averages do not imply identical operational systems.

Poor Interpretation vs Better Interpretation

Observed:

Mean LoS = 6 days in both trusts

Poor interpretation:

“The hospitals perform similarly.”

Better interpretation:

“The averages are similar, but the underlying distribution and operational pressures may differ substantially.”

Why This Matters for Population Health Management (PHM)

Understanding who sits in the long tail matters.

If a relatively small cohort consumes a large proportion of bed days, important questions emerge:

Who are these patients?
What risks or characteristics do they share?
Are there opportunities for earlier intervention?
Could support be delivered upstream?

This moves analytics beyond description and into action.

For example:

Healthcare systems may identify cohorts with:

moderate/severe frailty
repeated non-elective admissions
multimorbidity
social care dependency
discharge delays

These cohorts may benefit from:

anticipatory care
frailty pathways
integrated neighbourhood working
enhanced discharge planning
community-based intervention

Why This Matters for Pathway Redesign and Demand Management

Variation in Length of Stay can reveal where systems are struggling.

The goal is often not simply:

“reduce average Length of Stay”

Instead, the more useful question may be:

“What is driving long stays, and which parts of the pathway can realistically be redesigned?”

Examples might include:

earlier frailty identification
improved discharge coordination
hospital-at-home models
better rehabilitation pathways
integrated community support

This is where analytics becomes operationally useful.

Rather than reporting averages alone, healthcare analytics can help answer:

Which patients are driving demand, why, and what can realistically be done differently?

Key Insight

A relatively small cohort may consume a disproportionate share of bed days — and understanding this cohort may unlock some of the greatest opportunities for pathway redesign, prevention and demand management.

Statistical Process Control (SPC) Charts

So far we have explored variation across patients.

What Are SPC Charts?

SPC charts help distinguish:

But healthcare systems also vary over time.

This raises another important question:

is change meaningful — or simply normal fluctuation?

This is where SPC charts help.

common cause variation

(normal system variation)

special cause variation

(unusual change requiring investigation)

This is extremely NHS relevant.

When SPC Charts Work Well

When monitoring:

Counts (c-chart/u-chart)

Example:

falls per ward
complaints
infections

Proportions (p-chart)

Example:

% stroke patients thrombolysed
% patients seen within target

Example: Counts

Monthly falls on Ward A.

Question:

Is variation normal — or has something changed?

Example: Proportion

% stroke patients receiving assessment within target timeframe.

Question:

Are we seeing sustained improvement or random fluctuation?

The examples below illustrate how SPC can help distinguish expected fluctuation from meaningful change in healthcare systems.

When NOT To Use SPC

Important section.

Avoid when:

tiny numbers
unstable denominators
changing definitions
poor data quality
major pathway redesign underway
inappropriate comparisons

And importantly:

SPC does not explain why variation occurred.

It only flags:

something unusual may have happened.

That distinction matters.

Looking Ahead: When Understanding Variation Is Not Enough

Understanding variation helps us interpret healthcare systems.

But sometimes decision-makers need to move beyond:

“What happened?”

and ask:

“What might happen if the system changes?”

Examples include:

adding ED capacity
reducing discharge delays
workforce shortages
winter demand pressures
elective backlog recovery

This is where simulation models can help.

Rather than relying only on averages or historical trends, simulation allows healthcare systems to explore:

What might happen under different scenarios?

Simulation will be explored in more detail in a later module in the series.

Key Takeaways

Healthcare systems naturally exhibit variation — averages rarely tell the whole story
Summary statistics can look similar while masking very different operational realities
Understanding distributions helps explain demand, flow and system pressure
A relatively small cohort of patients may consume a disproportionate share of bed days
Statistical Process Control (SPC) helps distinguish expected variation from meaningful change
SPC highlights signals — but does not explain causes
When systems become highly variable or operationally complex, simulation models may help explore future scenarios

Questions Decision-Makers Should Ask

Are we looking only at an average — or the full distribution of activity?
Could a relatively small cohort be driving disproportionate operational pressure or demand?
Are we seeing expected variation — or evidence of something meaningfully different?
Is this signal likely to reflect real change or random fluctuation? Would SPC help us distinguish?
Does this metric hide operational complexity (for example, frailty, delayed discharge or multimorbidity)?
Are there different patient groups or pathways hidden within the data?
What operational actions would realistically follow from this analysis?
Has the system become too complex for averages and trend lines alone?
Would scenario modelling or simulation help us test future options?