People Analytics and Evidence-Based HR Decisions

Introduction

People analytics is the disciplined use of workforce data, statistical reasoning, organisational theory and professional judgement to improve decisions affecting people at work. It is more than producing HR reports. Reporting describes what has happened; analytics investigates why it happened, estimates what may happen next, evaluates alternative actions and supports decisions under uncertainty.

For HR professionals, the central challenge is not the absence of data but the conversion of data into valid evidence. Organisations commonly possess information from human-resource information systems, recruitment platforms, learning systems, performance processes, surveys, payroll, absence records, collaboration tools and operational systems. These sources may be incomplete, inconsistent, biased or collected for administrative rather than analytical purposes. Consequently, analytical maturity depends as much on question formulation, measurement quality and causal reasoning as on software or algorithms.

Evidence-based HR combines four forms of evidence: the best available research, organisational data, professional expertise and the values and experiences of affected stakeholders. A statistically significant result is not automatically an important managerial result, and a predictive model is not evidence that a proposed intervention will cause improvement. Responsible people analytics therefore requires attention to validity, fairness, privacy, transparency and implementation.

Learning Objectives

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

1. Distinguish descriptive, diagnostic, predictive and prescriptive analytics.
2. Translate an HR concern into a precise workforce question and an estimand.
3. Select valid metrics, define their denominators and assess data quality.
4. Design useful dashboards for different managerial audiences.
5. Interpret regression, classification, survival and experimental results cautiously.
6. Distinguish association from causation and identify confounding, selection bias and measurement error.
7. Evaluate interventions using experiments and quasi-experimental methods.
8. Communicate uncertainty, practical significance and limitations responsibly.
9. Apply ethical, legal and governance principles to people analytics and AI.

1. From HR Reporting to Evidence-Based Decision-Making

A useful analytical chain is:

Business problem → workforce question → construct and measure → data → analysis → decision → intervention → evaluation.

Failure at an early stage cannot usually be repaired by sophisticated modelling. For example, “Which employees are likely to leave?” is incomplete unless the organisation specifies the population, time horizon, event definition and purpose. A more precise question is: “Among permanent employees active on 1 January, which observable factors are associated with voluntary resignation within the next six months, and which retention actions are effective and acceptable?” This wording separates prediction from intervention evaluation.

An estimand is the quantity a decision-maker wants to learn. Examples include:

• the proportion of new hires still employed after twelve months;
• the average difference in absence days between two teams;
• the expected effect of a manager-training programme on engagement;
• the probability that an employee will resign within six months;
• the cost per successful hire.

The estimand must be distinguished from the estimator, which is the method used to calculate it, and from the estimate, which is the numerical result obtained from a sample.

Four analytical purposes

The categories are connected but not interchangeable. A dashboard showing high turnover is descriptive. A comparison indicating that turnover is concentrated among employees with short tenure is diagnostic, although it does not establish that short tenure causes turnover. A model estimating resignation risk is predictive. A recommendation to improve scheduling, compensation or supervisory practice is prescriptive and requires evidence about likely consequences, costs and fairness.

2. Formulating Strong Workforce Questions

A good HR question identifies five elements:

1. Population: Who is being studied?
2. Outcome: What constitutes success, failure or change?
3. Exposure or intervention: What factor or action is being examined?
4. Time: Over what period is the outcome measured?
5. Decision: What will be done with the answer?

A practical structure is: For [population], how does [exposure or intervention] affect [outcome] over [time period], compared with [counterfactual or comparison group], and for which groups?

The comparison group is essential for causal questions. “Did engagement rise after the leadership programme?” is weaker than “Did engagement rise more among participating teams than among comparable non-participating teams during the same period?”

Questions should also distinguish descriptive, predictive and causal aims. A descriptive question asks how many employees left. A predictive question asks which current employees have elevated predicted risk. A causal question asks whether a change in pay, work design or management practice would reduce regretted turnover. The data and methods required differ substantially.

Constructs and operational definitions

HR concepts such as engagement, inclusion, burnout, performance and potential are constructs rather than directly observable objects. They require operational definitions. A construct may be represented by a survey scale, behavioural indicator, administrative event or combination of measures. Analysts should document:

• the conceptual definition;
• the operational measure;
• the unit of analysis;
• the observation period;
• inclusion and exclusion rules;
• missing-data treatment;
• known sources of measurement error.

A survey item asking whether an employee is satisfied with pay is not equivalent to total job satisfaction. Similarly, absence records may omit presenteeism, informal leave or unrecorded sickness. Measurement validity is therefore a substantive HR issue, not merely a technical issue.

3. Metrics and Measurement Discipline

A metric is useful only when its numerator, denominator, population and time period are explicit. Consider turnover:

Here, separations must be defined, average headcount may be calculated as the mean of beginning and ending headcount or by a more frequent average, and the period must be stated. If beginning headcount is 1,000 and ending headcount is 1,100, with 84 separations, the simple average headcount is 1,050 and turnover is:

This rate is not directly comparable with a rate calculated using only voluntary resignations or monthly average headcount.

Common HR measures

A leading indicator changes before an outcome, such as manager feedback frequency or workload volatility. A lagging indicator records an outcome, such as regretted turnover. Leading indicators are useful for intervention but may be weak proxies. Analysts should not optimise a proxy at the expense of the underlying objective, a problem known as Goodhart’s law.

Data quality

Data quality should be evaluated across accuracy, completeness, consistency, timeliness, uniqueness and validity. A dashboard can be visually polished while relying on duplicate employee records, changed definitions or missing contingent workers. A data dictionary and metric catalogue should specify ownership, source system, refresh frequency, calculation logic and acceptable limitations.

4. Dashboards and Analytical Communication

A dashboard is a decision interface, not a warehouse of charts. It should begin with the decision it supports. An executive dashboard may show workforce size, critical-skill capacity, labour cost and regretted turnover. A business-partner dashboard may show team-level trends, hiring funnel conversion and absence patterns. A manager dashboard may focus on actionable workload, development and staffing information, subject to privacy controls.

Effective dashboards generally include:

• a small number of decision-relevant metrics;
• clearly labelled denominators and time periods;
• trend and comparison views;
• segmentation by relevant groups;
• targets with explicit rationale;
• explanations of uncertainty and data limitations;
• drill-down access governed by role and purpose.

Visualisation should match the analytical task. Lines show trends, bars compare categories, distributions reveal variation, and scatterplots help examine relationships. Colour should not imply causation or moral judgement. Ranking individual employees by “risk” can encourage stigma, gaming and self-fulfilling decisions; aggregated or team-level views are often safer.

5. Statistical Models for People Analytics

Descriptive statistics

The arithmetic mean is:

where xᵢ is an observation and n is the number of observations. The median is the middle ordered value and is often more appropriate for skewed variables such as salary or time to hire. The sample standard deviation is:

The denominator n − 1 provides the usual unbiased estimator of population variance under independent random sampling. Organisational data may not be a random sample, so inferential interpretations require care.

Regression

For a continuous outcome, a linear regression can be written as:

where yᵢ is the outcome for person or unit i, xᵢⱼ are predictors, β coefficients, β₀ the intercept and εᵢ the error term. A coefficient estimates the expected change in y associated with a one-unit increase in a predictor, holding included predictors constant. This interpretation does not prove causation.

Important assumptions for conventional inference include an appropriate functional form, independent or correctly modelled errors, adequate measurement, and, for some tests, homoscedasticity and approximately suitable residual behaviour. Clustered employees within teams violate simple independence; standard errors or multilevel models may need to account for team structure.

For a binary event such as resignation, logistic regression models log odds:

where pᵢ is the predicted probability of the event. The quantity exp(βⱼ) is an odds ratio, not a probability difference. An odds ratio of 1.50 means the odds are estimated to be 50% higher for a one-unit predictor increase, conditional on the model; it does not mean that probability rises by 50 percentage points.

Prediction and validation

A useful predictive model must be evaluated on data not used to fit it. Key concepts include:

• Discrimination: whether higher-risk cases tend to rank above lower-risk cases, often assessed with area under the ROC curve.
• Calibration: whether predicted probabilities correspond to observed event frequencies.
• Precision and recall: especially relevant when the event is rare.
• Stability: whether performance persists across time, departments and demographic groups.
• Utility: whether using the model improves decisions after costs and harms are considered.

A model with excellent discrimination may still be poorly calibrated or operationally useless. Data leakage occurs when information unavailable at decision time enters training data, producing unrealistic performance.

Time-to-event analysis

Turnover is a time-to-event outcome, and some employees have not yet left by the end of observation. Treating all such cases as “stayers” can bias analysis. Survival methods handle censoring. The hazard function represents the instantaneous event rate conditional on surviving up to time t. A Cox proportional hazards model is commonly expressed as:

where h₀(t) is the baseline hazard. The proportional-hazards assumption means hazard ratios are constant over time, an assumption that should be assessed rather than presumed.

6. Association, Causation and Evaluation

Correlation or regression association cannot by itself establish that an HR practice caused an outcome. A relationship may reflect confounding, reverse causality, selection, common causes or measurement artefacts.

A causal effect compares outcomes under two potential conditions for the same unit: treatment and no treatment. Because the same employee cannot simultaneously experience both conditions, the individual causal effect is unobservable; research designs estimate average effects using credible comparisons.

Randomised experiments

In an A/B test, eligible units are randomly assigned to intervention or control. Randomisation supports exchangeability in expectation, while pre-specified outcomes, adequate sample size, implementation fidelity and ethical safeguards remain necessary. The average treatment effect can be represented as:

where Y(1) is the potential outcome under intervention, Y(0) under control, and E denotes the population average. Randomisation does not guarantee balance in a small sample and does not solve non-compliance, attrition or spillovers automatically.

Quasi-experimental designs

When randomisation is impractical, useful designs include:

• Difference-in-differences: compares before–after change in a treated group with change in a comparison group. Its key assumption is parallel trends: absent the intervention, both groups would have followed similar trends.
• Interrupted time series: examines whether an intervention changes the level or slope of a sufficiently long outcome series, while accounting for seasonality and concurrent events.
• Regression discontinuity: exploits a treatment rule based on a threshold; validity depends on no precise manipulation around the cutoff and continuity of potential outcomes.
• Matching or weighting: improves comparability on observed covariates but cannot remove unmeasured confounding.

The appropriate design depends on the decision, intervention, timing, unit of assignment and ethical constraints. Statistical adjustment is not a substitute for thoughtful research design.

7. Prescriptive Analytics and Decision Quality

Prescriptive analytics links predicted outcomes to available actions, constraints and objectives. A decision model should specify:

• possible actions;
• expected benefits and costs;
• resource and policy constraints;
• distributional effects;
• uncertainty;
• reversibility and monitoring requirements.

For example, a retention intervention may be targeted to employees with high predicted risk. Yet high risk does not imply high treatment benefit. Prescriptive targeting should ideally estimate uplift or heterogeneous treatment effects: who is likely to benefit from the intervention compared with not receiving it. Otherwise, resources may be spent on people who would stay anyway, while employees who could benefit are missed.

A simple expected-value formulation is:

The expression is appropriate only when probabilities, values and costs are defined on a comparable basis and important outcomes are not omitted. In HR, values may include employee wellbeing, fairness, trust and legal risk, not merely financial return.

8. AI in People Analytics

Machine learning can process high-dimensional data and identify nonlinear relationships, interactions or complex patterns. Common applications include forecasting workforce demand, matching skills to roles, analysing open-text survey responses and supporting recruitment operations. However, AI does not eliminate the need for human judgement or causal reasoning.

Key risks include:

• historical bias encoded in labels or decisions;
• proxy discrimination through variables correlated with protected characteristics;
• automation bias and excessive trust in system outputs;
• privacy intrusion and function creep;
• opaque explanations and weak contestability;
• model drift as jobs, policies and labour markets change;
• generative-AI hallucination, leakage of confidential data and inconsistent classification.

A responsible AI lifecycle includes purpose limitation, data minimisation, documentation, representative validation, subgroup performance testing, human review, appeal mechanisms, access control, audit logs, monitoring and retirement criteria. High-impact employment decisions should not be delegated solely to an unreviewable score. Legal requirements vary by jurisdiction, but general principles include non-discrimination, transparency, proportionality, data protection and the right to challenge consequential decisions.

Fairness is multidimensional. Possible assessments include comparing selection rates, error rates, calibration or treatment effects across groups. No single fairness metric is universally correct, and some metrics cannot all be satisfied simultaneously when base rates differ. Fairness analysis must therefore be connected to the decision context, legal duties and stakeholder values.

9. Communicating Findings Responsibly

An analytical report should answer five questions:

1. What decision was examined?
2. What data and method were used?
3. What was found, including uncertainty and magnitude?
4. What alternative explanations and limitations remain?
5. What action is recommended, for whom, and how will it be evaluated?

Analysts should distinguish statistical significance from practical significance. A small effect may be statistically detectable in a very large dataset but irrelevant to operations. Conversely, an important effect may be estimated imprecisely in a small population. Confidence intervals communicate a range of values compatible with the data under stated assumptions; they are not guarantees that the true value lies in the interval.

Communication should avoid causal verbs when the design supports only association. “Employees with irregular schedules had higher turnover” is different from “irregular schedules caused turnover.” The latter requires a credible causal design and careful consideration of alternative mechanisms.

10. Illustrative Case: Evaluating a Flexible-Scheduling Intervention

Background

A fictional customer-service organisation with 2,400 employees observes rising voluntary turnover among contact-centre agents. Exit interviews suggest schedule inflexibility, but the organisation also suspects workload, supervisory climate and external labour-market competition.

Problem

Leaders propose allowing eligible agents to exchange shifts through a digital scheduling system. The decision is whether to expand the programme, not merely whether users of the system differ from non-users.

Analysis

The analytics team defines the primary outcome as voluntary resignation within six months, with secondary outcomes of absence, service quality and employee-reported schedule control. The eligible population excludes employees on extended leave and those in roles where shift exchange is operationally impossible. Baseline data include tenure, team, prior absence, schedule variability, supervisor, pay band and survey measures.

A pilot is conducted in eight comparable teams. Four teams are randomly assigned to receive the system immediately and four to a delayed-access control condition. Randomisation occurs at team level to reduce spillovers. The analysis is intention-to-treat: employees are analysed according to their team assignment, regardless of whether they use the tool.

Suppose voluntary resignation is 9% in the intervention group and 12% in the control group during six months. The unadjusted risk difference is:

The relative risk is:

Thus, the intervention group experienced 25% lower relative risk in this illustrative result. The absolute reduction is three percentage points, which is usually more directly useful for planning. If 300 eligible employees receive the intervention, the illustrative number needed to treat is:

Approximately 34 employees would need the intervention to prevent one additional resignation under the estimated average effect. This interpretation assumes the estimated effect is causal, stable and applicable to the target population; uncertainty should be reported with a confidence interval.

Approach and outcome

The team checks baseline balance, implementation, spillovers, subgroup effects and service quality. Survey results suggest improved schedule control, while service quality is unchanged. Usage is lower among new employees and employees with limited digital access, so the organisation provides training and a non-digital exchange process. Expansion is approved with quarterly monitoring rather than unconditional rollout.

Learning points

The case illustrates the difference between user-versus-non-user comparisons and randomised evaluation, the importance of absolute effects, the need to monitor unintended consequences, and the value of analysing access and effects across groups. It also shows that a statistically favourable result does not settle questions of cost, fairness, scalability or employee consent.

11. Governance and Operating Model

A mature people-analytics function normally establishes:

• a data governance council with HR, legal, privacy, security, technology and employee representation;
• a catalogue of approved metrics and definitions;
• role-based access and retention rules;
• model documentation and version control;
• impact assessments for high-risk analytical systems;
• procedures for correction, appeal and incident response;
• independent review of high-impact decisions;
• training for managers on interpretation and appropriate use.

The three-lines perspective is useful: analysts and HR owners manage risks in daily work; specialist risk, privacy and legal functions provide oversight; internal audit or an equivalent independent function evaluates governance effectiveness. Governance should be proportionate to impact: aggregate workforce planning generally presents less individual risk than automated promotion or termination recommendations.

12. Best-Practice Checklist

Before approving an HR analytical product, ask:

1. Is the decision and target population clearly defined?
2. Is the outcome meaningful, measurable and not merely a convenient proxy?
3. Are data provenance, quality and missingness documented?
4. Does the method fit the question and data structure?
5. Is the comparison or counterfactual credible?
6. Have practical significance and uncertainty been reported?
7. Has performance been validated over time and across relevant groups?
8. Could the system create discrimination, privacy harm or stigma?
9. Can affected people understand, challenge or correct the decision?
10. Will the intervention and its consequences be monitored after implementation?

Chapter Summary

People analytics converts workforce data into evidence for decisions, but analytical sophistication cannot compensate for a poorly defined question or weak measurement. Descriptive analytics identifies patterns; diagnostic analytics investigates explanations; predictive analytics estimates future outcomes; and prescriptive analytics compares actions under constraints. Strong practice begins with a defined population, outcome, exposure, time horizon and decision.

Metrics require explicit numerators, denominators, units and inclusion rules. Dashboards should serve decisions rather than display data indiscriminately. Regression, classification and survival models can support explanation and prediction, but their assumptions, validation and interpretation must be made explicit. Association is not causation. Randomised experiments and well-designed quasi-experiments provide stronger evidence about interventions than uncontrolled comparisons.

Responsible people analytics integrates data quality, statistical reasoning, professional judgement, employee voice, privacy, fairness and governance. AI can extend analytical capacity, but it can also reproduce historical bias and intensify opacity. The professional standard is therefore not simply accurate prediction: it is valid, useful, fair, contestable and continuously evaluated decision support.

Key Terms

People analytics: Systematic use of workforce data, analysis and judgement to inform people-related decisions.

Evidence-based HR: HR practice integrating research evidence, organisational data, professional expertise and stakeholder values.

Estimand: The population quantity a study seeks to learn.

Descriptive analytics: Analysis describing what has occurred or is occurring.

Diagnostic analytics: Analysis investigating patterns and possible explanations.

Predictive analytics: Estimation of likely future outcomes or events.

Prescriptive analytics: Analysis recommending actions under objectives and constraints.

Construct: An abstract concept such as engagement or potential that requires measurement.

Leading indicator: A measure that may change before the outcome of interest.

Lagging indicator: A measure recording an outcome after it occurs.

Data leakage: Use of information in model development that would not be available at decision time.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Discrimination: A model’s ability to rank higher-risk cases above lower-risk cases.

Confounding: Distortion of an exposure–outcome relationship by a common cause.

Average treatment effect: Average causal difference between intervention and control potential outcomes.

Difference-in-differences: Quasi-experimental comparison of changes over time between treated and comparison groups.

Censoring: Incomplete observation of the time until an event occurs.

Uplift: Difference in an individual or subgroup outcome caused by receiving an intervention rather than not receiving it.

Function creep: Use of data for purposes beyond the original justified purpose.

Model drift: Decline or change in model performance as contexts and data evolve.

Review Questions

Short Answer Questions

1. Distinguish descriptive, diagnostic, predictive and prescriptive analytics, giving one HR example of each.
2. Why must an HR metric specify its numerator, denominator, population and time period?
3. Explain why a predictive model does not necessarily identify an effective intervention.
4. What is the difference between statistical significance and practical significance?
5. State two ethical risks associated with AI-based people analytics.

Long Answer Questions

1. Explain how an HR professional should move from a broad organisational concern to a well-defined workforce question and estimand.
2. Compare regression, classification and survival analysis as tools for people analytics, including their assumptions and interpretation.
3. Discuss how randomised experiments and quasi-experimental designs support causal HR decisions.
4. Design a dashboard for a senior HR leadership team and justify the metrics, comparisons and safeguards you would include.
5. Evaluate the claim that more employee data necessarily produces better HR decisions.

Analytical / Application Questions

1. An organisation reports 120 separations, beginning headcount of 2,000 and ending headcount of 2,200. Calculate the turnover rate using the simple average-headcount definition, and identify two limitations of the measure.
2. A resignation model has strong discrimination but is poorly calibrated for part-time employees. Should it be used for individual retention decisions? Develop a validation and governance response.
3. A company introduces manager training in three departments and observes lower turnover than in other departments. Propose a credible evaluation design and identify at least three threats to causal interpretation.

Multiple Choice Questions with Answers

Question 1

Which analytical question is primarily predictive?

A. How many employees resigned last quarter?

B. Which current employees are likely to resign within six months?

C. Why did absence increase in the manufacturing unit?

D. Which intervention should receive the largest budget?

Correct Answer: B. Which current employees are likely to resign within six months?

Question 2

An organisation has 84 separations, beginning headcount of 1,000 and ending headcount of 1,100. Using simple average headcount, what is the turnover rate?

A. 7.6%

B. 8.0%

C. 8.4%

D. 9.1%

Correct Answer: B. 8.0%

Question 3

What does an odds ratio of 1.50 in a logistic regression most directly indicate?

A. A 50-percentage-point increase in probability

B. A 50% increase in odds for a one-unit predictor increase, conditional on the model

C. A 1.50-unit increase in the outcome

D. A 50% causal increase in resignations

Correct Answer: B. A 50% increase in odds for a one-unit predictor increase, conditional on the model

Question 4

Which feature is essential to a credible difference-in-differences design?

A. The treatment group must be larger than the comparison group

B. The outcome must be normally distributed

C. The groups would have followed parallel trends absent the intervention

D. Every employee must receive the intervention

Correct Answer: C. The groups would have followed parallel trends absent the intervention

Question 5

What is data leakage in predictive people analytics?

A. Loss of data through a cybersecurity incident

B. Use of information unavailable at decision time during model development

C. A model with too few predictors

D. Random assignment to a control group

Correct Answer: B. Use of information unavailable at decision time during model development

Question 6

Which statement best distinguishes calibration from discrimination?

A. Calibration concerns ranking; discrimination concerns privacy

B. Calibration concerns probability accuracy; discrimination concerns ranking cases by risk

C. Calibration concerns causality; discrimination concerns sample size

D. Calibration and discrimination are identical measures

Correct Answer: B. Calibration concerns probability accuracy; discrimination concerns ranking cases by risk

Question 7

If an intervention reduces resignation from 12% to 9%, what is the absolute risk reduction?

A. 3 percentage points

B. 25 percentage points

C. 25%

D. 0.75 percentage points

Correct Answer: A. 3 percentage points

Question 8

Why can a user-versus-non-user comparison of a flexible scheduling tool be misleading?

A. Users may differ systematically from non-users before using the tool

B. Percentages cannot be used in HR research

C. Randomisation always creates bias

D. Turnover cannot be measured over six months

Correct Answer: A. Users may differ systematically from non-users before using the tool

Question 9

Which practice best reflects responsible use of a high-impact HR algorithm?

A. Allowing the score to determine termination automatically

B. Removing all documentation to protect intellectual property

C. Providing validation, human review, appeal and subgroup monitoring

D. Using historical managerial decisions as unquestioned labels

Correct Answer: C. Providing validation, human review, appeal and subgroup monitoring

Question 10

Why is uplift potentially more useful than predicted turnover risk for targeting a retention intervention?

A. It identifies who is likely to benefit from the intervention relative to not receiving it

B. It guarantees that every targeted employee will remain

C. It eliminates the need for an outcome measure

D. It converts an association into a legal conclusion

Correct Answer: A. It identifies who is likely to benefit from the intervention relative to not receiving it