Module 11: Common Analytical Traps & Data Fallacies
Learning Objectives
By the end of this module you should be able to:
- Recognise common analytical traps that can lead to misleading conclusions.
- Understand why observed improvement does not always imply intervention success.
- Identify common sources of bias in healthcare analysis and evaluation.
- Recognise the influence of incentives, observation and selection effects.
- Apply a more critical approach to interpreting healthcare data and evidence.
Why This Matters
Healthcare organisations are increasingly data-driven.
Yet many analytical mistakes occur not because the data is wrong, but because the interpretation is incomplete.
Decision-makers are often required to make judgements under uncertainty using information that is influenced by:
- natural variation
- human behaviour
- selection effects
- incentives
- measurement limitations
Understanding common analytical traps helps organisations avoid drawing overly confident conclusions from incomplete evidence.
Many of these concepts build directly on ideas introduced in previous modules, particularly:
- Correlation vs Causation
- Variation & Distributions
- Standardisation & Fair Comparison
- Evaluating Interventions & Schemes
- Evaluating Methods in Practice
Correlation Does Not Equal Causation
Observed
Higher deprivation is associated with higher ED attendance.
Tempting Conclusion
Deprivation causes ED attendance.
What Else Might Be Happening?
The relationship may be real.
However, multiple interacting factors may contribute to the observed pattern, including:
- underlying health need
- frailty
- multimorbidity
- access to primary care
- housing
- transport
- social circumstances
Why This Matters
An observed relationship does not automatically explain why it exists.
Correlation should start questions, not end them.
Regression to the Mean
Observed
A practice with unusually high admission rates receives support.
Six months later:
Admission rates improve.
Tempting Conclusion
The intervention caused the improvement.
What Else Might Be Happening?
Extremely high or low observations often move closer to average over time even when no intervention occurs.
This phenomenon is known as:
Regression to the Mean
Why This Matters
Without an appropriate comparison group, natural variation can easily be mistaken for intervention impact.
The Hawthorne Effect
Observed
Hand hygiene compliance improves substantially during a pilot.
Tempting Conclusion
The intervention was successful.
What Else Might Be Happening?
People may change their behaviour simply because they know they are being observed.
Examples include:
- monitored improvement programmes
- audit periods
- pilot projects
- temporary performance initiatives
Why This Matters
Some improvement may be attributable to increased attention and observation rather than the intervention itself.
A key question is:
Would the improvement still occur if nobody knew they were being measured?
Selection Bias
Observed
Patients enrolled in a programme experience better outcomes.
Tempting Conclusion
The programme caused the improvement.
What Else Might Be Happening?
Participants may differ systematically from non-participants.
For example:
- more motivated patients may enrol
- lower-risk patients may participate
- clinicians may selectively refer patients
Why This Matters
The observed outcomes may reflect differences between groups rather than the intervention itself.
Survivorship Bias
Observed
Several successful pilots have been presented nationally.
Tempting Conclusion
Most pilots are successful.
What Else Might Be Happening?
Successful projects are often more visible than unsuccessful ones.
Failed pilots may:
- never be published
- receive less attention
- be quietly discontinued
Why This Matters
Looking only at successful examples can create an unrealistic view of effectiveness.
Financial Incentive Effects
Observed
Admissions reduce after a financial incentive scheme is introduced.
Tempting Conclusion
The pathway redesign caused the reduction.
What Else Might Be Happening?
Behaviour may have changed because incentives changed.
Examples include:
- payment mechanisms
- commissioning incentives
- performance targets
- contractual arrangements
Why This Matters
A key question is:
Are we observing the effect of the intervention, the effect of the incentive, or both?
The League Table Fallacy
Observed
Trust A ranks 10th and Trust B ranks 30th.
Tempting Conclusion
Trust A is performing better.
What Else Might Be Happening?
Rankings may be influenced by:
- population characteristics
- deprivation
- case-mix
- random variation
- measurement choices
Why This Matters
League tables often simplify complex realities and should rarely be interpreted in isolation.
The Precision Fallacy
Observed
Predicted savings are estimated at:
£2,483,726
Tempting Conclusion
A precise number must be highly accurate.
What Else Might Be Happening?
Forecasts remain estimates.
Precision and certainty are not the same thing.
Why This Matters
Decision-makers should focus on:
- assumptions
- uncertainty
- confidence intervals
- plausible ranges
rather than precision alone.
Confirmation Bias
Observed
A dashboard appears to support an existing belief.
Tempting Conclusion
The evidence confirms the belief.
What Else Might Be Happening?
People naturally seek evidence that supports existing views and may overlook contradictory information.
Why This Matters
Good analytical practice requires asking:
- What evidence challenges this conclusion?
- What alternative explanations exist?
- What information might we be missing?
Key Takeaways
- Correlation does not automatically imply causation.
- Extreme observations often move closer to average over time.
- People may change behaviour simply because they know they are being observed.
- Intervention groups may differ from comparison groups.
- Successful examples are often more visible than unsuccessful examples.
- Financial incentives can influence behaviour and outcomes.
- Rankings can oversimplify complex realities.
- Precision does not guarantee accuracy.
- Confirmation bias affects everyone.
- Better decisions come from questioning interpretations, not simply accepting results.
Questions Decision-Makers Should Ask
- Could there be alternative explanations for this result?
- Are we confusing association with causation?
- Could regression to the mean be contributing to the observed change?
- Are people behaving differently because they know they are being measured?
- Are the intervention and comparison groups genuinely comparable?
- Could selection bias be influencing outcomes?
- Are financial incentives affecting behaviour?
- Does the apparent ranking tell the whole story?
- Are we mistaking precision for certainty?
- What evidence might challenge our current interpretation?
Part 1 Knowledge Check
Question 1
A practice with unusually high admission rates receives support. Admissions subsequently fall.
What is a possible alternative explanation?
Question 2
Staff compliance improves during a monitored pilot.
What effect may be present?
Question 3
A programme appears successful because highly motivated patients chose to participate.
What bias may exist?
Question 4
Two ICBs have similar age-standardised rates but very different frailty burdens.
Which statement is most accurate?
Question 5
A dashboard shows a forecast to one decimal place.
What should you remember?
Question 6
A hospital introduces a pilot to improve discharge documentation.
During the pilot, compliance increases substantially.
Which explanation should also be considered?
Final Reflection
Across the first ten modules, a common theme emerges:
Data rarely provides definitive answers on its own.
Good decision-making requires understanding:
- context
- uncertainty
- variation
- evidence
- bias
- incentives
- interpretation
The role of analytics is not to eliminate uncertainty.
The role of analytics is to help organisations make better decisions despite uncertainty.
A good analyst asks:
What does the data show?
A great analyst also asks:
What else might explain what we are seeing?
That question sits at the heart of health data literacy.