Statistics for Healthcare Professionals

Executive premise

Every number is a claim about a process.

Statistical intelligence is the leadership capacity to determine whether a number is valid, meaningful, actionable, equitable, and sufficiently certain for the decision at hand.

01

Measurement precedes analysis

Definitions, provenance, missingness, denominators, and representativeness determine whether a statistic is fit for use.

02

Variation is information

Leaders must distinguish common-cause movement from special-cause change before rewarding, blaming, scaling, or abandoning a process.

03

Magnitude and uncertainty lead

Effect size, interval estimates, operational thresholds, and consequences are more useful than a binary significance label.

04

Prediction is not causation

Association, prediction, and causal effect answer different questions and require different evidence.

05

Models require lifecycle governance

Discrimination alone is insufficient. Calibration, external validation, fairness, workflow fit, drift, and net benefit matter.

06

Dashboards need response architecture

Measures become management systems only when linked to owners, thresholds, protocols, balancing measures, and learning cycles.

Decision architecture

Start with the decision, not the dataset.

A statistic becomes useful only when it is connected to a defined decision and a learning cycle.

1Define the decisionOwner, options, threshold, time horizon

2Validate the measureDefinition, denominator, provenance, missingness

3Describe variationShape, center, spread, subgroup mix

4Estimate uncertaintyIntervals, sample size, sensitivity

5Test or modelDesign, assumptions, validation, utility

6Decide, act, and learnOwner, cadence, stop or scale criteria

Four recurrent interpretive failures

Where executive decisions most often go wrong

The dashboard is organized to correct four high-cost habits identified in the research.

Mean as typicalSkewed distributions hide the actual patient experience and tail burden.

Point estimate as factUncertainty is omitted, so precision is inferred rather than measured.

Noise as signalTwo-point comparisons and weekly fluctuations trigger tampering.

Accuracy without prevalenceVendor performance claims collapse when the deployment population changes.

Measurement and data fitness

Do not calculate before you validate what the number represents.

Use the scale selector, data-fitness review, and distribution laboratory to test whether a metric preserves the operational story.

01

Measurement scale selector

Select the type of variable to see legitimate summaries and common analytic errors.

Healthcare examplePayer, diagnosis group, modality, site

Legitimate summariesCounts, proportions, mode, chi-square tests

Leadership cautionNumeric codes remain labels and should not be averaged.

02

Seven-question data fitness review

Rate each domain from 0 to 4. The score is a governance screen, not a certification.

Definition2 Provenance2 Completeness2 Correctness2 Concordance2 Currency2 Bias and representativeness2

50/ 100

Usable only with documented limitations and targeted remediation.

03

Distribution laboratory: mean, median, tail, and coefficient of variation

Generate a synthetic right-skewed wait-time distribution. Watch how the mean and upper percentiles respond when the tail grows.

Encounters500 Typical wait, minutes18 Tail intensity6

Analyze de-identified local values

Paste comma-, space-, or line-separated numeric wait times Do not enter names, identifiers, or protected health information.

Mean0min

Median0min

90th percentile0min

Coefficient of variation0%

MCARMissingness is unrelated to observed or unobserved values.Complete-case analysis may remain unbiased, but precision is lost.

MARMissingness can be explained by observed variables.Multiple imputation or likelihood-based methods may be defensible.

MNARMissingness depends on unobserved values after adjustment.Sensitivity analysis and process investigation are required.

03A

EHR data-quality evidence

Share of 103 reviewed publications assessing each dimension. These are literature frequencies, not hospital performance targets.

Completeness74%

Correctness51%

Concordance45%

Currency34%

Plausibility28%

Conformance17%

Bias11%

Completeness dominates the literature, but a populated field may still be wrong, outdated, nonconforming, or systematically biased.

03B

Preserve the decision-relevant structure

Select the display that matches the question rather than decorating a point estimate.

Process over timeRun chart, control chart, or interrupted time seriesAvoid two selected months.

DistributionHistogram, box plot, density, or percentilesAvoid the mean alone.

Estimate precisionPoint estimate with confidence intervalAvoid a p-value without magnitude.

Group comparisonDot or interval plot on a common scaleAvoid truncated axes and 3D bars.

Prediction modelCalibration, ROC, decision curve, and subgroup reviewAvoid accuracy or AUC alone.

Patient flowProcess map linked to time and defect dataAvoid a dashboard without workflow context.

Probability, sampling, and inference

Translate uncertainty into operational decisions.

These calculators connect probability distributions, sampling precision, confidence intervals, absolute effects, and method selection to healthcare management.

04

Normal duration model

Estimate the probability that a standardized examination exceeds the scheduled slot.

Mean minutes Standard deviation Scheduled slot

0%estimated slot-overrun probability

z = (slot – mean) / standard deviation

05

Poisson surge model

Estimate the chance that daily volume reaches or exceeds the capacity threshold.

Average daily volume Capacity threshold

0%probability of a surge day

P(X = k) = e^-λ λ^k / k!

06

Sampling precision and the CLT

See how sample size changes standard error and the approximate margin of error.

Population standard deviation Sample size Target margin of error

Standard error0

95% margin0

Required n0

SE(x̄) = s / √n

07

Effect and confidence interval laboratory

Interpret the estimate relative to both the null and the minimum operationally important effect.

Estimated change Standard error Meaningful threshold

Lower 95%0

Point estimate0

Upper 95%0

08

Absolute effect and NNT calculator

Convert relative claims into absolute operational consequences.

Control risk, % Intervention risk, % Eligible population

ARR0pp

Relative reduction0%

NNT0

Events avoided0

ARR = risk_control – risk_intervention | NNT = 1 / ARR

09

Method selector

Choose the analytic structure. The recommendation is a starting point, not an automated methodological opinion.

Outcome structure Comparison Dependence

Recommended starting method

Two-sample t test

Check independence and distribution. Welch’s test addresses unequal variances; consider robust or permutation methods when assumptions are poor.

10

Decision-risk matrix

Translate Type I and Type II errors into organizational consequences.

False-positive consequence Missed-signal consequence Evaluation power, %

No real effect

Real effect

Act

Type I error
false-positive action

Correct detection

Do not act

Correct restraint

Type II error
missed signal

Causal reasoning and improvement science

Design determines what a result is allowed to mean.

Use the question architecture, causal map, and process-control laboratory to separate description, prediction, causal effect, and real process change.

Descriptive question

What is happening, for whom, and with what variation?

Use valid measures, distributions, denominators, time order, and uncertainty. Avoid causal language.

11

Simplified causal structure

A staffing and wait-time analysis must make shared causes visible before adjustment choices are made.

A directed acyclic graph does not prove causality. It clarifies the assumptions required for a causal interpretation.

12

Operational design matcher

Match the evaluation design to decision risk, reversibility, and available comparison structure.

Randomization feasible? Comparison group available? Clear threshold or rollout timing? Decision risk

Preferred design family

Difference-in-differences or controlled interrupted time series

Compare change over time with a credible comparison group and examine whether pre-intervention trends are sufficiently parallel.

13

Statistical process-control laboratory

Control limits describe expected process variation under a stable baseline. They are not targets, specifications, or confidence intervals.

Baseline periods14 Intervention effect, minutes10 Noise level3

Center line0min

Upper limit0min

Lower limit0min

Signal statusReviewing

Outcome

Patient, workforce, financial, or operational result that matters.

Process

Whether the mechanism expected to produce the result is occurring.

Balancing

Whether harm, burden, delay, or cost has shifted elsewhere.

Equity

Whether benefits and burdens differ across meaningful groups.

Prediction, diagnostic accuracy, and AI governance

A model can rank well and still be unsafe to deploy.

Performance must be interpreted through prevalence, calibration, transportability, fairness, workflow capacity, and lifecycle controls.

14

Prevalence-adjusted diagnostic calculator

Estimate what a positive result means in the deployment population.

Sensitivity, %90 Specificity, %90 Prevalence, %2 Population1000

PPV0%

NPV0%

False alarms0

Missed cases0

Outcome presentOutcome absent Positive

TP

FP

Negative

FN

TN

15

Workforce-throughput regression

Estimate expected MRI throughput from staffing within the observed range. The relationship is predictive, not automatically causal.

Technologists on shift6 Expected gain per FTE6.9 Residual uncertainty6

Expected exams0

Approximate low0

Approximate high0

Do not extrapolate beyond the observed staffing range. Demand, case mix, equipment uptime, and protocol complexity may confound the observed association.

16

Minimum governance architecture for predictive models and AI

Check each domain only when the required evidence exists and an accountable owner is named.

Purpose and targetDefined decision, users, outcome, prohibited uses, and review against intended purpose. Data and representativenessProvenance, missingness, site coverage, subgroup coverage, and input-drift monitoring. PerformanceDiscrimination, calibration, uncertainty, utility, benchmark comparison, and threshold review. EquitySubgroup performance, label and proxy analysis, access effects, and harm review. WorkflowHuman factors, alert burden, capacity, escalation pathways, adoption, and downstream delay. AccountabilityClinical, operational, analytic, and executive owners, incident response, audit trail, and retirement criteria.

0of 6 governance domains documented

Deployment readiness has not been established.

VITALS and executive governance

Measure the interpretive health of the leadership system.

VITALS organizes six domains of statistical competency. Statistical Decision Yield measures whether uncertainty is actually present in consequential decisions.

17

VITALS self-assessment

Rate current organizational capability from 0 to 4. Use the profile to target governance improvement.

VVariation literacyIs this movement signal or noise, and how do we know?2 IInference disciplineWhat sample is this based on, and how precise is it?2 TTesting rigorCould the evaluation detect the effect that matters?2 AAccuracy calculusAt our prevalence, what does a positive result mean?2 LLinkage modelingWhat does the model predict here, and where does it stop applying?2 SStewardship of decisionsWhich decisions carried an explicit uncertainty estimate?2

50/ 100

Developing capability with inconsistent executive use.

20

Executive decision charter builder

Document the decision architecture before analysis begins. Entries remain in the browser and are not submitted.

Decision and accountable owner Target population Intervention or exposure Comparator Outcome and time horizon Minimum meaningful effect Data quality and design Equity, balancing, and action rules

Capability, accountability, and implementation

Build statistical intelligence as an operating capability.

Technology does not create interpretive maturity. Governance, shared definitions, analytic competence, leadership behavior, and repeated evidence use do.

Level 1

Reporting without an interpretive operating model

Metrics are produced, but definitions, denominators, uncertainty, and response rules are inconsistent. The immediate priority is a metric inventory and ownership structure.

0-90 days

Define and stabilize

Inventory executive metrics and assign owners.
Document numerator, denominator, eligibility, and data source.
Identify high-risk dashboards and algorithms.
Train leaders on variation, intervals, and effect size.

Evidence: critical definitions and data defects have named owners.

3-6 months

Standardize interpretation

Convert key measures to time-ordered displays.
Add balancing and equity measures.
Establish analytic review for high-impact decisions.
Require effect magnitude and uncertainty reporting.

Evidence: reviews distinguish common causes, special causes, and practical significance.

6-12 months

Govern the lifecycle

Create a model registry and monitoring controls.
Conduct temporal or external validation.
Embed prospective evaluation in major initiatives.
Publish enterprise statistical governance policy.

Evidence: models have owners, drift monitoring, incident response, and retirement rules.

21

Leadership-domain applications

Select a domain to review the statistical questions that belong in its governance routine.

Quality and patient safety

Prioritize rates with valid exposure denominators, time-ordered displays, balancing measures, and reliability-aware comparisons. Use risk adjustment carefully and do not normalize preventable disparities.

Are rare events aggregated over a defensible window?
Does the control chart show a real process change?
What balancing measure could reveal shifted harm?

22

Distributed accountability

Statistical intelligence is a shared operating system, not an analytics department task.

Board and executive teamSet decision thresholds, risk tolerance, and governance expectations.

Operational and clinical leadersDefine measures, explain workflow, and act on signals.

Analysts and data scientistsDesign analysis, quantify uncertainty, and validate models.

Data stewards and informaticsMaintain definitions, provenance, quality, and access.

Quality-improvement teamsTest change with time-ordered data and context.

Compliance, ethics, and equity leadersAssess legal, ethical, distributional, and patient impacts.

23

Executive formula reference

Formulas clarify the claim. They do not replace design review or methodological expertise.

Search formulas

Mean

x̄ = Σxᵢ / n

Average level; sensitive to extreme values.

Weighted mean

x̄w = Σwᵢxᵢ / Σwᵢ

Combines observations or groups with explicit weights.

Sample variance

s² = Σ(xᵢ - x̄)² / (n - 1)

Average squared dispersion with sample correction.

Standard error

SE(x̄) = s / √n

Uncertainty of an estimated mean under the sampling model.

Confidence interval

estimate ± critical value × SE

Compatibility range under stated model assumptions.

Z score

z = (x - μ) / σ

Distance from the mean in standard-deviation units.

Relative risk

RR = risk₁ / risk₀

Ratio of outcome risks.

Odds ratio

OR = ad / bc

Ratio of odds; generally not equal to a risk ratio.

Absolute risk reduction

ARR = risk₀ - risk₁

Difference in event probability.

Number needed to treat

NNT = 1 / ARR

Patients treated per additional favorable outcome.

Sensitivity

TP / (TP + FN)

Detection among people with the outcome.

Specificity

TN / (TN + FP)

Correct negatives among people without the outcome.

Positive predictive value

TP / (TP + FP)

Outcome frequency among positive results; depends on prevalence.

Linear regression

E(Y|X) = β₀ + ΣβⱼXⱼ

Conditional mean model. Causal interpretation requires design assumptions.

Logistic regression

log[p / (1 - p)] = β₀ + ΣβⱼXⱼ

Model for log odds of a binary outcome.

Three-sigma control limits

center line ± 3σ

Expected range of stable process variation.

24

Executive glossary

Search the language leaders need to interrogate statistical claims.

Search terms

AssociationA statistical relationship between variables; it does not establish causation.

CalibrationAgreement between predicted probabilities and observed outcome frequencies.

Common-cause variationVariation generated by the stable underlying process.

Confidence intervalAn interval procedure designed for stated long-run coverage under assumptions.

ConfoundingMixing of an exposure-outcome relationship with the influence of a shared cause.

DiscriminationA model’s ability to rank individuals with and without an outcome.

Effect sizeMagnitude and direction of a difference or association.

EstimandThe precise quantity an analysis seeks to estimate.

External validationEvaluation of a model on meaningfully separate data.

P-valueA measure of incompatibility between data and a specified statistical model.

Risk adjustmentAdjustment for baseline risk to improve comparability, subject to model and fairness concerns.

Special-cause variationVariation suggesting a source outside the stable process.

Statistical powerThe probability that a test rejects the null under a specified alternative.

25

Evidence anchors used in the research synthesis

The dashboard translates the report’s methodological evidence domains into executive questions and operating controls.

Data qualityWeiskopf and Weng; Kahn et al.; Lewis et al.Are the data fit for this use, population, and time period?

Statistical interpretationGreenland et al.; Wasserstein and Lazar; Wasserstein et al.What is the effect, how uncertain is it, and which decisions remain compatible?

Improvement scienceThor et al.; Perla et al.; Fretheim and Tomic; Ogrinc et al.Does the time series show a stable change, and what mechanism could explain it?

Causal inferenceHernánWhat intervention effect is being estimated, compared with what alternative?

Prediction and AIVan Calster et al.; Collins et al.; Riley et al.; Moons et al.Will the model remain accurate, calibrated, useful, and equitable here?

Dashboard effectivenessXie et al.Is visualization integrated with workflow, feedback, accountability, and action?

Final leadership imperative

Require every consequential number to answer five questions.

What was measured?Compared with what?How uncertain is it?What alternatives remain?What action and learning cycle follows?