Quantitative Data and Health Care Analytics

Introduction

Health care organizations generate large quantities of quantitative data through clinical encounters, laboratory systems, electronic health records (EHRs), claims, finance, staffing, patient surveys, registries and remote-monitoring devices. These data can support evaluation of quality, safety, effectiveness, equity, efficiency and value. However, data volume does not guarantee valid evidence. A technically sophisticated dashboard can still mislead if the sample is biased, definitions are inconsistent, missingness is ignored, or comparisons fail to account for differences in patient risk.

Health care analytics is therefore more than calculation. It is a disciplined process of converting a decision problem into a measurable question, obtaining fit-for-purpose data, analyzing uncertainty, and communicating results in a way that supports responsible action. This chapter focuses on the practical and methodological foundations required by health care managers, evaluators and policy professionals.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Distinguish major sources and types of quantitative health care data.
2. Plan a sampling strategy and identify common sources of sampling and measurement error.
3. Design a defensible data-collection process for clinical, administrative, claims, survey and patient-reported data.
4. Select and interpret descriptive, comparative and multivariable analyses.
5. Interpret effect sizes, confidence intervals, statistical significance and practical importance.
6. Explain risk adjustment, stratification and the dangers of inappropriate comparisons.
7. Design useful dashboards and communicate uncertainty to decision-makers.
8. Apply principles of reproducible, ethical and governed analytics.

1. From Management Question to Analytic Question

A useful analysis begins with a decision rather than with an available dataset. A management question such as “Is our diabetes service performing well?” should be translated into a specific evaluative question. A strong formulation identifies:

• Population: Which patients or service users are included?
• Intervention or exposure: What service, policy or characteristic is being examined?
• Comparator: What is the alternative, baseline or benchmark?
• Outcome: What measurable result matters?
• Time horizon: When is the outcome assessed?
• Purpose: Is the analysis descriptive, predictive, causal or evaluative?

For example: “Among adults discharged after heart failure, does pharmacist-led follow-up reduce 30-day readmissions compared with usual care?” This question specifies a population, intervention, comparator, outcome and time period. It also signals that confounding, follow-up completeness and competing events require attention.

A data dictionary should be created before analysis. It defines each variable, its operational meaning, coding, unit, permissible values, source, time reference and quality checks. For example, “readmission within 30 days” must specify whether planned readmissions, transfers and admissions to outside hospitals are included.

2. Types and Sources of Quantitative Health Data

Clinical data

Clinical data include diagnoses, observations, medications, procedures, laboratory results, imaging findings and clinical outcomes. EHR data are often longitudinal and detailed, but they may contain copied-forward notes, coding variation, delayed entry and differences in documentation practices. A recorded absence of a condition is not always evidence that the condition was assessed and absent.

Administrative and operational data

Administrative systems record appointments, length of stay, staffing, bed occupancy, referrals, waiting times, costs and service utilization. They are valuable for monitoring access and efficiency, but operational definitions may change when workflows or information systems change.

Claims data

Claims data are generated for reimbursement. They provide broad coverage of diagnoses, procedures, utilization and costs, especially across providers. Their limitations include delayed submission, upcoding incentives, limited clinical detail and imperfect identification of disease severity. A diagnosis code may represent a suspected condition, a historical condition or a billing justification rather than a confirmed clinical finding.

Surveys and patient-reported data

Surveys measure experiences, satisfaction, symptoms, functioning, quality of life and self-reported behavior. Patient-reported outcome measures (PROMs) assess health status from the patient’s perspective, while patient-reported experience measures (PREMs) assess experiences of care. Valid interpretation requires attention to instrument validity, reliability, language, mode of administration, response rates and changes in the respondent population.

Registries, sensors and linked data

Registries use predefined clinical variables and may provide high-quality information for particular conditions. Wearables and remote-monitoring devices generate frequent observations but introduce issues of calibration, adherence, device access and algorithmic processing. Linkage of clinical, claims and social data can improve analysis but increases risks of identity disclosure, linkage error and unjustified inference.

3. Sampling and Study Populations

Sampling determines whose data are observed and therefore to whom conclusions can reasonably apply. A target population is the full group of interest; a sampling frame is the list or mechanism from which observations are selected; the analytic sample is the set ultimately analyzed.

Probability sampling

In simple random sampling, every eligible unit has a known, usually equal, probability of selection. Systematic sampling selects every kth unit after a random start. Stratified sampling divides the population into important subgroups and samples within each; it can improve precision and ensure representation of small groups. Cluster sampling selects naturally occurring groups, such as clinics or hospitals, and then observes units within them. Cluster sampling is efficient but observations within a cluster may be correlated.

In a stratified design, sampling fractions may intentionally differ. Weighted analysis is then needed if estimates are intended to represent the population. A basic sampling weight is the inverse of the selection probability:

where π_i is the probability that unit i is selected. Nonresponse and post-stratification adjustments may also be incorporated. Weights should be documented and used with methods appropriate to the survey design.

Non-probability sampling

Convenience, volunteer, purposive and consecutive samples are common in service evaluation. They may be practical, but selection probabilities are unknown and generalizability is limited. A consecutive sample of all patients attending one clinic is not equivalent to a random sample of all eligible patients in a region.

Sample size and precision

For estimating a proportion in a simple random sample, a commonly used approximate sample-size formula is:

Here, n is the required sample size, Z is the standard-normal critical value for the desired confidence level, p is the anticipated proportion, and e is the desired absolute margin of error. The formula assumes independent observations, a probability sample and a sufficiently large population. If p is unknown, p = 0.5 gives the largest required sample size for a fixed e.

For a 95% confidence level, Z is approximately 1.96. If p = 0.50 and e = 0.05:

This is not automatically the correct sample size for an evaluation. It does not account for clustering, nonresponse, design effects, finite populations, multiple outcomes or statistical power for a comparison. For a cluster design, a simple approximation is:

where DEFF is the design effect, m is the average cluster size and ρ is the intracluster correlation coefficient. The effective sample size is approximately the actual sample size divided by DEFF. Even a small ρ can materially reduce precision when clusters are large.

Power calculations for comparisons should be based on the minimum clinically important difference, outcome variability or baseline event rate, significance level, desired power and expected attrition. Statistical consultation is advisable for complex longitudinal, clustered or time-to-event designs.

4. Data Collection and Data Quality

A defensible collection plan specifies who collects data, when, from where, using which instrument, with what training and under what quality controls. Collection procedures should be standardized without eliminating clinically meaningful judgment.

Important quality dimensions include:

• Completeness: Are expected records and fields present?
• Validity: Does the variable measure what it claims to measure?
• Reliability: Would repeated measurement produce similar results under stable conditions?
• Timeliness: Are records available when decisions require them?
• Consistency: Are definitions and codes stable across sites and periods?
• Uniqueness: Are duplicate patients or encounters identified?
• Plausibility: Are values within clinically credible ranges?

A data-quality audit may compare a sample of electronic records with source documents. Automated checks can identify impossible dates, negative lengths of stay, implausible vital signs, duplicate identifiers and mismatched units. Changes in coding systems or EHR configuration should be recorded because apparent trends may reflect measurement change rather than changes in care.

Missing data

Missingness should be described by variable, time, site and patient subgroup. Three conceptual mechanisms are commonly distinguished:

• MCAR: Missing completely at random; missingness is unrelated to observed or unobserved values.
• MAR: Missing at random conditional on observed variables.
• MNAR: Missing not at random; missingness depends on unobserved values even after conditioning on observed information.

Complete-case analysis is simple but can reduce precision and introduce bias, particularly when missingness is substantial or differential. Multiple imputation can be appropriate under a plausible MAR assumption when the imputation model includes variables related to missingness and the outcome. Sensitivity analyses should examine how conclusions change under plausible MNAR scenarios. Missingness itself may be a quality or equity indicator, not merely a nuisance.

5. Descriptive and Comparative Analysis

Descriptive analysis should precede modeling. For continuous variables, report mean and standard deviation when distributions are reasonably symmetric; report median and interquartile range when skewness or outliers make these more informative. For categorical variables, report counts and percentages, including denominators.

The arithmetic mean is:

where x_i is observation i and n is the number of observations. The sample standard deviation is:

The denominator n − 1 provides the usual unbiased estimate of population variance under independent random sampling. These summaries are not automatically appropriate for clustered or weighted data.

Common health care measures include:

• Risk: number experiencing an event divided by the number at risk.
• Rate: events divided by person-time, such as admissions per 1,000 patient-months.
• Prevalence: existing cases divided by the population at a specified time.
• Incidence: new cases arising over a specified period among those at risk.
• Length of stay: total inpatient days divided by eligible discharges, with a clearly defined handling of same-day stays and outliers.

For two groups, the risk difference is:

and the risk ratio is:

where p₁ is risk in the intervention or exposed group and p₀ is risk in the comparator group. The odds ratio is:

Risk difference is often easiest for managers to interpret because it expresses absolute change. Relative measures can appear large when baseline risk is small. Odds ratios should not be casually described as risk ratios, especially for common outcomes.

For a favorable treatment that reduces risk, the number needed to treat is:

where RD is expressed as a proportion, not a percentage. NNT should be accompanied by the time horizon and confidence interval. It is meaningful only when the populations, follow-up periods and outcome definitions are comparable.

6. Effect Sizes and Uncertainty

A p-value addresses how compatible the observed data are with a specified null model; it does not measure clinical importance, probability that the hypothesis is true, or probability that results will replicate. A confidence interval communicates estimate precision and a range of values compatible with the data and model under repeated-sampling reasoning.

For an estimated mean with standard error SE, a large-sample 95% confidence interval is approximately:

The value 1.96 applies to an approximately normal sampling distribution and a 95% interval. Small samples or non-normal estimators may require a t critical value, bootstrap methods or other approaches. Confidence intervals should reflect clustering, weighting and model uncertainty where applicable.

Consider readmission risks of 12% under usual care and 9% under a new program:

The absolute reduction is 3 percentage points. The relative risk is:

Thus, the program is associated with a 25% relative reduction and a 3-percentage-point absolute reduction. If the eligible population contains 2,000 patients, the observed difference corresponds to approximately 60 fewer readmissions, assuming comparable follow-up and no major selection differences. Statistical and clinical interpretation still requires a confidence interval, harms, costs, implementation burden and the credibility of the design.

7. Stratification, Risk Adjustment and Causal Caution

A crude comparison can be misleading when groups differ in age, comorbidity, disease severity, socioeconomic circumstances or referral patterns. Stratification examines results within categories, such as age bands or severity levels. It is transparent and useful for detecting effect heterogeneity, but too many strata can produce sparse data.

Risk adjustment uses measured patient characteristics to estimate expected outcomes and support fairer comparisons among providers or populations. A logistic regression model for a binary outcome can be expressed as:

where p is the modeled probability of the outcome, β₀ is the intercept, β_j are coefficients, and X_j are predictors. The model assumes a suitable link function, correctly specified predictors, appropriate independence or correlation structure, and adequate data quality. The exponentiated coefficient e^β_j is an adjusted odds ratio for a one-unit change in X_j, holding other modeled variables constant.

Risk adjustment is not a universal correction. It can perpetuate inequity if social disadvantage is treated only as a patient risk factor, conceal poor care if inappropriate mediators are adjusted for, or produce unstable estimates in small providers. Variables occurring after treatment should generally not be adjusted for when estimating the total treatment effect. Clinical judgment, causal diagrams and sensitivity analyses are valuable safeguards.

Standardized outcomes can be summarized using observed-to-expected ratios:

An O/E ratio below 1 indicates fewer observed events than expected under the model; above 1 indicates more. The interpretation depends entirely on the reference population, model calibration and uncertainty. It is not proof of superior or inferior care.

Risk adjustment should be complemented by stratified reporting. Overall averages can hide disparities: a hospital may improve overall while outcomes worsen for rural patients, people with disabilities or a particular ethnic group. Equity dashboards should show denominators, uncertainty and data-completeness differences across groups.

8. Regression, Time-to-Event and Longitudinal Analysis

Regression allows analysts to estimate associations while controlling for measured variables and, when appropriate, interactions. Linear regression is commonly used for continuous outcomes, logistic regression for binary outcomes, and count models such as Poisson or negative binomial regression for events. Model choice should follow the outcome-generating process rather than convenience.

Time-to-event methods are useful when follow-up differs among patients or censoring occurs. The Kaplan–Meier method estimates the probability of remaining event-free over time under non-informative censoring assumptions. A Cox proportional hazards model estimates a hazard ratio, which compares instantaneous event rates under a proportional-hazards assumption. A hazard ratio is not generally the same as a risk ratio and should not be interpreted as a probability ratio without care.

Repeated measurements from the same patient are correlated. Standard methods that assume independent observations may understate uncertainty. Mixed-effects models, generalized estimating equations or cluster-robust methods can account for within-patient or within-site correlation, provided their assumptions and implementation are appropriate.

Prediction and causal inference are distinct. A predictive model may accurately identify high-risk patients without showing that an intervention will improve their outcomes. A variable can be predictive but not causally actionable. Conversely, a causal factor may be difficult to predict. Managers should ask whether the purpose is forecasting, resource allocation, explanation, or estimating the effect of a change.

9. Dashboards and Decision-Focused Communication

A dashboard is an information system for repeated monitoring, not a substitute for analysis. Effective dashboards connect indicators to decisions and actions. Each measure should have a documented numerator, denominator, inclusion rule, time window, owner, refresh frequency, target and escalation rule.

Good dashboard design includes:

1. A small number of decision-relevant measures.
2. Clear labels and units.
3. Visible denominators and data dates.
4. Trends rather than isolated values.
5. Stratification by relevant populations.
6. Display of confidence intervals or control limits when appropriate.
7. Explicit distinction between target, benchmark and statistical control limit.
8. Accessible colors and alternatives to color-only meaning.
9. Drill-down to source definitions and records.
10. Warnings when data are incomplete or definitions changed.

Control charts can distinguish common-cause variation from potentially special-cause variation. However, limits must be calculated using an appropriate baseline and method; arbitrary red-amber-green thresholds can create false alarms. A dashboard should not rank clinicians or sites without considering reliability, case mix, sample size and unintended consequences.

10. Reproducible Analytics and Governance

Reproducibility means that another authorized analyst can understand and, within the limits of the data environment, reproduce the analysis. A reproducible workflow generally includes version-controlled code, immutable or documented source extracts, a data dictionary, prespecified analytic decisions, automated tests, parameter logs, model specifications, and a record of exclusions and missing-data handling.

A useful analytic record answers: What question was asked? Which population was eligible? Which records were excluded and why? What transformations were applied? Which software and package versions were used? What assumptions were made? Which results were exploratory? How were results validated?

Health data governance must address confidentiality, lawful data use, access controls, retention, secondary use, consent where applicable, and disclosure risk. De-identification reduces but does not eliminate re-identification risk, especially when datasets are linked. Role-based access, audit trails, encryption, minimum-necessary access and secure computing environments are practical controls.

Ethical analytics also requires attention to fairness. An algorithm may perform differently across groups because of biased labels, unequal access to care, missing data or historical inequity. Validation should report performance by relevant subgroup. For binary prediction, discrimination may be summarized by the area under the receiver operating characteristic curve, while calibration examines whether predicted probabilities correspond to observed frequencies. Neither metric alone establishes fairness or clinical usefulness.

Artificial intelligence can assist with coding, anomaly detection, forecasting and natural-language extraction, but generated outputs require validation and human accountability. Automation should not conceal uncertainty, replace professional judgment, or make unreviewable decisions about access to care. Model drift, changes in clinical practice and shifts in population composition require ongoing monitoring.

11. Illustrative Case: Evaluating a Hospital Readmission Program

Background

A regional hospital introduced a transitional-care program for adults discharged after heart failure. The program included medication reconciliation, a telephone call within 72 hours and a follow-up appointment within seven days. Managers wished to know whether the program reduced 30-day readmissions and whether benefits were equitable.

Problem and analytic design

The evaluation team used EHR and claims data for 12 months before and 12 months after implementation. The eligible population was defined as adults discharged alive after an index heart-failure admission. Planned readmissions were excluded according to a prespecified algorithm. The primary outcome was any unplanned readmission within 30 days.

Because participation was not randomized, patients receiving the program could differ from usual-care patients. The team therefore reported crude results, stratified results by age and comorbidity burden, and an adjusted model containing baseline characteristics measured before discharge. They did not adjust for post-discharge follow-up, because follow-up may be part of the intervention pathway.

Findings

There were 1,000 eligible discharges in usual care and 1,000 after implementation. Readmission declined from 120 to 90 cases. The crude risks were 12% and 9%, producing an absolute reduction of 3 percentage points and a relative risk of 0.75. The team calculated a confidence interval and conducted sensitivity analyses for missing outside-hospital admissions.

The overall improvement was larger among patients with low-to-moderate comorbidity than among those with very high comorbidity. However, participation was lower among patients living far from the hospital and among those with incomplete telephone contact information. The apparent overall benefit could therefore partly reflect differential enrollment and incomplete outcome capture.

Approach and solution

Managers introduced three safeguards: active outreach using alternative contact routes, monitoring of enrollment and outcome completeness by subgroup, and a quarterly review of risk-adjusted and stratified results. The data dictionary was revised to specify transfer handling and outside-hospital linkage. The dashboard displayed absolute risk, relative risk, denominators, confidence intervals and a data-completeness indicator rather than a single traffic-light score.

Outcome and learning points

The program remained promising, but the evaluation did not establish that it alone caused the reduction. The strongest conclusions concerned the observed association, its size and the conditions under which it occurred. The case demonstrates that effect size, risk adjustment, equity analysis, data completeness and implementation monitoring must be interpreted together.

Challenges and Limitations

Quantitative health care evaluation faces recurring limitations. Administrative data may not represent clinical reality; observational comparisons remain vulnerable to confounding; measurement definitions may change; missing data can be informative; small samples create unstable estimates; and multiple analyses increase the chance of apparently positive findings. Risk adjustment can improve comparability but cannot correct for unmeasured severity or biased measurement. Dashboards may encourage target chasing, gaming or neglect of unmeasured outcomes. Finally, statistical significance can be absent for a clinically important effect in a small sample, while a trivial effect can be statistically significant in a very large dataset.

Best practice is therefore to combine quantitative results with clinical review, patient and staff perspectives, process knowledge and sensitivity analysis. The appropriate question is not simply whether a number changed, but whether the change is credible, important, equitable, actionable and sustainable.

Chapter Summary

Health care analytics converts data into evidence for decisions about quality, safety, effectiveness, efficiency and equity. Sound evaluation begins with a precise question, an explicit population and a fit-for-purpose data source. Sampling design affects representativeness and precision; data collection requires valid definitions, quality checks and transparent treatment of missingness.

Descriptive statistics should precede modeling. Absolute and relative effect measures answer different questions, while confidence intervals communicate uncertainty. Risk adjustment and stratification can improve interpretation but do not remove all bias and may conceal inequities if used uncritically. Regression, time-to-event methods and longitudinal models must match the outcome and dependence structure.

Dashboards should be decision-focused, transparent and equity-sensitive. Reproducible code, data dictionaries, version control and analytic documentation support trustworthy results. Governance, privacy, fairness and human accountability are essential when clinical data and artificial intelligence are used. Ultimately, quantitative evidence is strongest when statistical reasoning is integrated with clinical expertise, operational knowledge and the lived experience of patients.

Key Terms

Sampling frame: The list or mechanism from which a sample is selected.

Target population: The full population to which an evaluation seeks to apply conclusions.

Stratified sampling: Sampling separately within defined population subgroups.

Cluster sampling: Sampling naturally occurring groups rather than individual units independently.

Sampling weight: A value, often the inverse of selection probability, used to produce population-representative estimates.

Patient-reported outcome measure: A standardized measure of health status reported directly by a patient.

Data dictionary: A structured description of variables, definitions, coding and units.

Missing completely at random: Missingness unrelated to observed or unobserved data.

Missing at random: Missingness explainable by observed variables, conditional on those variables.

Risk difference: The difference in event risks between two groups.

Risk ratio: The ratio of event risks between two groups.

Odds ratio: The ratio of odds of an outcome between two groups.

Confidence interval: An interval procedure expressing sampling uncertainty around an estimate.

Risk adjustment: Statistical adjustment for measured differences in patient risk or case mix.

Stratification: Analysis within defined subgroups.

Confounding: Distortion of an exposure–outcome association by a related third variable.

Design effect: Inflation of variance caused by features such as cluster sampling.

Reproducible analytics: An analysis documented and implemented so authorized others can inspect and reproduce it.

Calibration: Agreement between predicted probabilities and observed outcome frequencies.

Data governance: Policies and controls governing data quality, access, security, use and accountability.

Review Questions

Short Answer Questions

1. Distinguish a target population, sampling frame and analytic sample.
2. Why can a risk difference be more useful to a health care manager than a relative risk?
3. State two limitations of claims data for clinical evaluation.
4. What is the purpose of a data dictionary?
5. Explain why missingness should be examined by subgroup.

Long Answer Questions

1. Compare probability and non-probability sampling in a hospital service evaluation. Discuss representativeness, feasibility and sources of bias.
2. Explain the roles of stratification and risk adjustment, including circumstances in which adjustment may create misleading conclusions.
3. Describe how a health care organization should design a reproducible analytics workflow.
4. Discuss how confidence intervals, effect sizes and p-values should be used together in managerial decision-making.
5. Design a dashboard for monitoring a transitional-care program, specifying measures, denominators, stratifications and governance safeguards.

Analytical / Application Questions

1. A program reduces 30-day readmissions from 15% to 12%. Calculate the risk difference, relative risk and approximate NNT, and explain which measure should be emphasized to a hospital board.
2. A hospital’s overall mortality improves, but mortality among rural patients increases. Propose an analytic investigation using stratification, data-quality checks and risk adjustment.
3. A predictive model identifies patients at high risk of emergency admission but performs poorly for one ethnic group. Analyze possible causes and recommend technical, governance and clinical responses.

Multiple Choice Questions with Answers

Question 1

Which feature most directly improves the representativeness of a probability sample?

A. Every eligible unit has a known selection probability

B. Only patients who volunteer are included

C. The largest available clinic is selected

D. Analysts exclude difficult-to-contact patients

Correct Answer: A. Every eligible unit has a known selection probability

Question 2

If usual-care risk is 12% and intervention risk is 9%, what is the absolute risk reduction?

A. 3 percentage points

B. 25 percentage points

C. 0.75 percentage points

D. 21 percentage points

Correct Answer: A. 3 percentage points

Question 3

What does an odds ratio of 1.0 indicate, under the specified model?

A. Equal modeled odds between groups

B. Equal sample sizes between groups

C. A 100% increase in risk

D. Perfect prediction

Correct Answer: A. Equal modeled odds between groups

Question 4

The sample-size formula for estimating a proportion is most directly based on which desired quantity?

A. Absolute margin of error

B. Number of regression predictors only

C. Hospital bed capacity

D. Average patient age only

Correct Answer: A. Absolute margin of error

Question 5

Which statement best describes risk adjustment?

A. It accounts for measured differences in patient risk or case mix

B. It guarantees causal inference

C. It removes all effects of socioeconomic disadvantage

D. It replaces the need for data validation

Correct Answer: A. It accounts for measured differences in patient risk or case mix

Question 6

Why may cluster sampling produce less precision than a simple random sample of the same size?

A. Observations within clusters may be correlated

B. It always excludes small populations

C. It cannot use sampling weights

D. It measures only categorical variables

Correct Answer: A. Observations within clusters may be correlated

Question 7

A confidence interval primarily communicates what?

A. Sampling uncertainty around an estimate

B. The probability that the null hypothesis is true

C. The clinical importance of an intervention by itself

D. The percentage of missing records

Correct Answer: A. Sampling uncertainty around an estimate

Question 8

Which dashboard practice is most appropriate?

A. Show definitions, denominators, dates and relevant uncertainty

B. Use a single color without numerical values

C. Rank providers without considering sample size

D. Hide missing-data information from users

Correct Answer: A. Show definitions, denominators, dates and relevant uncertainty

Question 9

Which situation is most consistent with MNAR missingness?

A. Patients with worse unobserved symptoms are less likely to complete the survey

B. A random computer failure removes records independently of all variables

C. Missingness is fully explained by recorded age and sex

D. Every patient has complete data

Correct Answer: A. Patients with worse unobserved symptoms are less likely to complete the survey

Question 10

What is a central requirement of reproducible analytics?

A. Documented data definitions, code and analytic decisions

B. Using the most complex model available

C. Removing all exploratory analyses from the record

D. Reporting only statistically significant findings

Correct Answer: A. Documented data definitions, code and analytic decisions