People Analytics and Evidence-Based HR Decision Making

Introduction

Human resource management increasingly depends on the ability to convert workforce information into sound organizational decisions. Organizations collect data about recruitment, performance, compensation, absence, engagement, learning, mobility, diversity, and turnover. Yet data collection alone does not create strategic value. Value arises when managers formulate important questions, select valid measures, analyze evidence appropriately, recognize uncertainty, and act responsibly.

People analytics is the systematic use of workforce data, statistical reasoning, research design, and organizational knowledge to improve decisions affecting employees and organizational performance. Evidence-based HR decision making is broader than analytics. It integrates the best available research evidence, internal organizational data, professional expertise, stakeholder values, and contextual constraints. A dashboard can display a pattern; it cannot by itself explain why the pattern exists or establish that an intervention caused improvement.

For graduate managers, the central challenge is not learning to produce more metrics. It is learning to distinguish useful evidence from attractive but misleading information. This requires attention to measurement validity, causality, data quality, privacy, algorithmic bias, return on investment, and managerial communication.

Learning Objectives

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

1. Formulate decision-relevant HR questions and translate them into measurable constructs.
2. Design HR metrics, key performance indicators, and dashboards that support action.
3. Evaluate workforce data quality and select suitable analytical methods.
4. Distinguish association from causation and assess evidence from experiments and quasi-experiments.
5. Estimate HR program costs, benefits, and return on investment.
6. Communicate analytical findings to technical and non-technical audiences.
7. Identify privacy, fairness, governance, and responsible-AI risks in people analytics.
8. Develop an evidence-based HR decision process for practical organizational use.

From HR Reporting to Evidence-Based Decision Making

Traditional HR reporting answers descriptive questions: How many employees do we have? What was the turnover rate? How many training hours were delivered? People analytics extends this work through four progressively more demanding questions:

These categories are not a hierarchy in which predictive analytics is always superior. A well-designed descriptive report may be more useful than an inaccurate prediction model. The appropriate level of analysis depends on the decision, available evidence, risks, and costs of error.

Evidence-based management commonly draws on four sources: organizational data, external scientific research, professional expertise, and the values and experiences of affected stakeholders. These sources may conflict. For example, internal data may suggest that a flexible-work policy is associated with better retention, while research indicates that its effects differ by occupation. Managers must therefore interpret evidence in context rather than apply universal prescriptions mechanically.

A disciplined decision cycle is:

1. Define the decision and its consequences.
2. Formulate a specific, answerable question.
3. Identify relevant constructs and measures.
4. Gather and assess evidence.
5. Analyze patterns and competing explanations.
6. Select and implement an intervention.
7. Evaluate outcomes and unintended effects.
8. Document learning and revise the decision process.

Formulating Strong HR Questions

Weak questions produce weak analytics. “How can we improve engagement?” is a broad aspiration, not an analytical question. A stronger formulation identifies the population, outcome, intervention or exposure, comparison, and time horizon. For example: “Among customer-service employees with at least six months of tenure, does supervisor coaching delivered monthly reduce voluntary turnover during the following year compared with current practice?”

A useful question should specify:

• the decision owner;
• the population affected;
• the outcome of interest;
• the relevant time period;
• the intervention or explanatory factor;
• the comparison condition; and
• the action that will follow from the answer.

Managers should also distinguish constructs from measures. “Engagement” is a construct, while a survey score is one operational measure of it. “Performance” may be represented by sales, customer outcomes, error rates, peer evaluations, or supervisor ratings. Each measure captures only part of the construct and may contain systematic bias.

A measurement plan should state the definition, numerator, denominator, data source, frequency, owner, and limitations of each metric. It should also specify whether higher values are desirable, undesirable, or context-dependent.

Metrics, Rates, and Dashboards

A metric is a quantified representation of an organizational phenomenon. A key performance indicator (KPI) is a metric selected because it is closely connected to an important objective and supports managerial action. A dashboard is a structured visual display of selected metrics, trends, comparisons, targets, and alerts.

Common HR measures

Headcount is the number of employees at a specified point in time. Full-time-equivalent (FTE) headcount adjusts for work effort and is often more useful when employees work different schedules. Turnover measures employee separations, but its interpretation depends on whether the numerator includes voluntary, involuntary, regrettable, retirement, and internal-movement events.

A basic period turnover rate is:

The average headcount is often approximated as:

The formula assumes that beginning and ending headcount reasonably represent the workforce throughout the period. If headcount changes sharply, monthly or daily average headcount is preferable. The result is a percentage, not a probability that a particular employee will leave.

A selection ratio is:

A training completion rate may be calculated as:

These measures are useful but incomplete. High completion does not prove learning, and low turnover does not necessarily indicate a healthy workplace if employees are unable to leave or if vacancies are frozen.

A dashboard should connect three layers:

1. Outcome indicators: business and workforce results, such as regrettable turnover, safety incidents, or productivity.
2. Driver indicators: factors plausibly related to outcomes, such as workload, manager behavior, internal mobility, or schedule predictability.
3. Action indicators: activities that management can change, such as coaching frequency, staffing levels, or onboarding quality.

Dashboards should display definitions, data dates, denominators, targets, comparison groups, and confidence or uncertainty where appropriate. Color coding must not conceal small sample sizes or unstable estimates. A dashboard should answer “What decision does this support?” rather than merely “What information can we display?”

Worked example: interpreting turnover

A business unit begins a quarter with 240 employees and ends with 260. During the quarter, 18 employees leave. Average headcount is (240 + 260) ÷ 2 = 250. The quarterly turnover rate is (18 ÷ 250) × 100 = 7.2%. If annualized by simply multiplying by four, the result is 28.8%, but this is only a rough extrapolation and assumes comparable turnover in each quarter. It is not equivalent to calculating the proportion of the original workforce that would remain after four periods, because employees can leave and be replaced during the year.

Data Quality and Analytical Design

People data are generated for operational purposes, not always for research. Payroll systems, applicant-tracking systems, surveys, learning platforms, scheduling systems, and performance tools may use different identifiers, time periods, definitions, and missing-value codes. Combining them without governance can produce false precision.

Important dimensions of data quality include:

• Completeness: whether required records or fields are present.
• Accuracy: whether values represent the underlying reality.
• Consistency: whether definitions agree across systems and periods.
• Timeliness: whether data are available when decisions are made.
• Validity: whether a measure captures the intended construct.
• Uniqueness: whether duplicate records have been removed.
• Representativeness: whether observed cases reflect the population of interest.

A data-quality audit should document source systems, extraction dates, inclusion rules, transformations, missingness, duplicate handling, and known limitations. Missing data should not automatically be replaced with zero. For example, a missing performance rating may mean that no rating was required, that the manager failed to submit it, or that the employee was absent during the cycle. Each interpretation has different implications.

Analytical design begins before statistical modeling. A cross-sectional survey can identify associations at one point in time but is weak evidence for temporal ordering. Longitudinal data can establish whether a possible cause precedes an outcome, although unmeasured confounding may remain. A comparison group improves interpretation of change. Random assignment, when feasible and ethical, provides the strongest basis for estimating the causal effect of an intervention.

Descriptive and Inferential Analysis

Descriptive statistics summarize observed data. The mean is:

where xᵢ is observation i and n is the number of observations. The mean is sensitive to extreme values. The median, the middle ordered observation, may better represent skewed pay, tenure, or absence distributions.

The sample variance is:

The sample standard deviation is s = √s². The denominator n − 1 is used when estimating population variability from a sample under standard sampling assumptions. These statistics describe dispersion in the same units as the underlying measure for the standard deviation.

A correlation coefficient describes the strength and direction of a linear association between two variables. It does not establish that one variable causes the other. For instance, engagement and retention may be positively correlated because supportive managers improve both, because high-performing teams receive more resources, or because employees who intend to stay report greater engagement.

Regression models can adjust for observed differences and estimate conditional associations. A simple linear model is:

where yᵢ is the outcome, xᵢ is a predictor, β₀ is the intercept, β₁ is the expected change in the outcome for a one-unit change in the predictor, and εᵢ is the unexplained error. Interpretation requires attention to scale, omitted variables, nonlinearity, outliers, dependence among observations, and measurement error. For binary outcomes such as leaving versus staying, logistic regression is often more appropriate than ordinary linear regression, but its predictions still depend on data quality and assumptions.

Statistical significance should not be confused with managerial importance. A very small effect may be statistically detectable in a large workforce but not worth acting upon. Conversely, a practically important effect may be uncertain in a small sample. Managers should consider effect size, uncertainty, costs, benefits, and consequences of error.

Causality and Evaluation

The key causal question is counterfactual: What would have happened to the same employees at the same time if they had not received the intervention? The unobserved alternative is the counterfactual outcome. Because it cannot normally be observed directly, researchers use designs that approximate it.

A randomized controlled trial assigns eligible units to intervention and control conditions by chance. Randomization tends, in expectation, to balance both observed and unobserved characteristics. In HR, randomization may involve employees, teams, locations, or time periods. Ethical and operational constraints must be addressed, including informed participation, fairness, spillover between groups, and the duty not to withhold essential services.

When randomization is not feasible, quasi-experimental designs can be useful. Examples include:

• Difference-in-differences: compares change over time in an intervention group with change in a comparison group. The crucial parallel-trends assumption requires that, absent the intervention, both groups would have followed similar trends.
• Interrupted time series: examines whether an outcome changes beyond its prior pattern after an intervention. It requires sufficient observations before and after implementation and attention to other events occurring at the same time.
• Matching or weighting: constructs a comparison group with similar observed characteristics. It cannot remove bias from unmeasured differences.
• Regression discontinuity: uses a known threshold for eligibility and compares cases just above and below the cutoff. Its effect generally applies near that threshold.

Pre–post comparisons are easy to conduct but often misleading. An improvement may result from seasonality, economic change, regression to the mean, concurrent initiatives, or natural recovery rather than the HR intervention.

Predictive analytics requires a different standard from causal analysis. A model can predict resignation risk without identifying what would prevent resignation. Using a prediction as if it were a causal explanation may lead to ineffective or discriminatory interventions. Prediction systems should therefore be evaluated for calibration, discrimination, stability across groups, and practical utility, not merely overall accuracy.

Return on Investment and Business Cases

HR initiatives should be evaluated using both financial and nonfinancial outcomes. A conventional return on investment is:

Program costs may include vendor fees, employee time, manager time, technology, implementation, communication, and opportunity costs. Financial benefits may include avoided turnover costs, increased contribution margin, reduced absence costs, or measurable productivity gains.

Suppose a retention program costs $120,000. The organization estimates that it prevents 10 avoidable departures, with an estimated replacement cost of $18,000 per departure. Gross benefit is 10 × $18,000 = $180,000. Net benefit is $180,000 − $120,000 = $60,000. ROI is ($60,000 ÷ $120,000) × 100 = 50%. The estimate is only as credible as the assumptions about avoidable departures and replacement cost. Sensitivity analysis should test alternative assumptions. If only five departures were prevented, benefit would be $90,000 and ROI would be −25%.

Cost-benefit analysis should include distributional effects. An intervention may produce a positive average return while disadvantaging a particular occupational group. Nonfinancial effects—trust, safety, employee voice, legal risk, and psychological well-being—should not be ignored merely because they are difficult to monetize.

Communicating Insights and Making Decisions

Analytical communication is an exercise in translation. Senior leaders usually need the decision, magnitude, uncertainty, business relevance, and recommended action. Technical audiences need definitions, methods, assumptions, diagnostics, and limitations. Employees and representatives need understandable explanations of purpose, data use, rights, and safeguards.

A concise evidence narrative can follow this sequence:

1. Decision: What choice must be made?
2. Finding: What does the evidence show?
3. Magnitude: How large and practically important is the effect?
4. Confidence: How certain is the conclusion?
5. Alternatives: What other explanations remain plausible?
6. Recommendation: What should be done, by whom, and by when?
7. Evaluation: How will the decision be tested?

Visualizations should use comparable scales, clear denominators, meaningful time periods, and annotations for policy or system changes. Avoid truncated axes when they exaggerate differences. Avoid presenting forecasts as facts. Every chart should identify the population, unit of analysis, data period, and source.

Privacy, Fairness, and Governance

People analytics involves sensitive information and unequal power relationships. Responsible practice requires purpose limitation, data minimization, access control, retention limits, security, transparency, and lawful processing. Organizations must comply with applicable employment, privacy, anti-discrimination, and data-protection requirements; legal obligations vary by jurisdiction and should be reviewed with qualified counsel.

Important governance questions include:

• Was the data collected for this purpose, or is a new use justified and disclosed?
• Can individuals be reidentified from supposedly anonymous data?
• Who can access individual-level information?
• Are employees informed and able to raise concerns?
• Is the intervention proportionate to the risk being addressed?
• How will errors be corrected and decisions challenged?
• Does the system create disparate effects across protected or vulnerable groups?

Fairness is not a single universal metric. Depending on the decision and legal context, analysts may examine selection rates, false-positive and false-negative rates, calibration, equal opportunity, or group-specific error costs. Some fairness criteria cannot generally be satisfied simultaneously when groups have different base rates. Therefore, fairness requires normative judgment, stakeholder involvement, and attention to the decision process—not merely a technical adjustment to an algorithm.

A model trained on historical promotion decisions may reproduce earlier bias. Removing protected attributes does not necessarily remove discrimination because correlated variables, such as occupation, location, career interruptions, or work schedule, may act as proxies. Human review is not automatically a safeguard; reviewers can reproduce model outputs or introduce their own bias. Effective governance combines impact assessment, subgroup testing, documentation, monitoring, meaningful human oversight, and an appeal process.

Artificial Intelligence in People Analytics

AI and machine learning can assist with document classification, workforce forecasting, skills inference, conversational support, and anomaly detection. However, automated systems can amplify poor labels, historical inequities, and unstable patterns. Generative AI may also fabricate explanations, expose confidential information, or produce inconsistent recommendations.

Before deploying an AI-enabled HR tool, an organization should establish a clear use case and prohibited uses. It should validate performance on relevant populations, test subgroup error rates, assess privacy and security, document training data and model limitations, and define human accountability. High-impact decisions such as hiring, promotion, discipline, compensation, and termination require especially strong controls. Automation should not eliminate the employee’s opportunity to understand, question, or appeal a consequential decision.

Illustrative Case: Northstar Services

Background

Northstar Services, a fictional regional customer-support company, employs 1,200 representatives across four centers. Annual voluntary turnover is high, especially among employees with less than one year of tenure. Executives believe that a new e-learning platform will solve the problem, but the HR analytics team is asked to investigate before a major investment.

Problem and question

The team formulates the question: “Among newly hired representatives, does structured supervisor coaching during the first 90 days reduce six-month voluntary turnover compared with the existing onboarding process?” The outcome is voluntary separation within six months. The intervention is at least four documented coaching sessions. The proposed comparison is employees hired in centers that implement coaching immediately versus comparable centers scheduled to implement it later.

Analysis

The team first audits data. It discovers that the learning system records completion, but coaching records are inconsistent and one center uses a different definition of voluntary turnover. After harmonizing definitions, the team compares historical trends and checks whether the early-adopting and later-adopting centers had similar pre-intervention turnover. It also examines role, shift, supervisor, hiring cohort, pay, and workload.

A simple pre–post comparison shows turnover falling from 24% to 18% in the early-adopting centers. However, a major hiring campaign and a seasonal reduction in demand occurred simultaneously. The team therefore uses a difference-in-differences design, while clearly stating the parallel-trends assumption. It finds an estimated six-month reduction of 3 percentage points, with uncertainty wide enough that a larger pilot is warranted.

The team interviews employees and supervisors. Employees report that coaching is helpful when it concerns schedules, customer escalation, and career pathways, but not when it is merely a checklist exercise. Supervisors report that coaching time is difficult during peak demand.

Approach and solution

Northstar launches a preregistered pilot across matched teams. Teams are randomly assigned, where operationally feasible, to immediate coaching or delayed implementation. The pilot specifies the primary outcome, secondary outcomes such as service quality and employee experience, implementation measures, subgroup analyses, and stopping rules for serious harm. Coaching content is standardized, but supervisors receive discretion to address role-specific issues.

The company estimates implementation cost at $72,000. After six months, the pilot estimates that eight avoidable departures were prevented, each valued at $15,000 in replacement and lost-productivity costs. Estimated benefit is $120,000; net benefit is $48,000; ROI is ($48,000 ÷ $72,000) × 100 = 66.7%. The organization also observes improved early-tenure experience, but no reliable change in customer-service productivity. Results are less favorable on night shifts because supervisors receive insufficient protected coaching time.

Outcome and learning points

Northstar expands coaching only where protected time and supervisor capability exist. It does not claim that the e-learning platform caused retention improvement, because the evaluated intervention was supervisor coaching. It continues monitoring turnover, quality, employee experience, subgroup effects, and implementation fidelity.

The case demonstrates that credible people analytics begins with a decision question, not a dashboard. It also shows the importance of data definitions, comparison groups, qualitative evidence, implementation measurement, subgroup analysis, and economic sensitivity. A positive average ROI does not justify universal implementation when effects differ by shift or when employee experience and fairness risks are unresolved.

Best Practices for Graduate HR Managers

Effective practice can be summarized as follows:

• Begin with a consequential decision and a theory of change.
• Define constructs and metrics before examining results.
• Use the simplest valid method, not the most sophisticated available method.
• Separate descriptive, predictive, and causal claims.
• Report denominators, missing data, uncertainty, and alternative explanations.
• Combine quantitative evidence with employee and managerial perspectives.
• Test interventions rather than assuming that correlation identifies a solution.
• Evaluate distributional effects and unintended consequences.
• Protect confidentiality and limit individual-level access.
• Document data lineage, model assumptions, decisions, and revisions.
• Monitor models and metrics after implementation because organizational conditions change.
• Give affected people understandable explanations and meaningful avenues for challenge.

Challenges and Limitations

People analytics is constrained by imperfect measurement, incomplete records, changing organizational systems, and ethical limits on experimentation. Employee behavior is context-dependent; a relationship found in one organization may not generalize to another. Models can become obsolete when work processes, labor markets, technology, or policies change. Small subgroups may be impossible to analyze without creating reidentification risk, yet excluding them can conceal inequity.

There is also a danger of managerial reductionism: treating complex human experiences as if they were fully captured by a score. Quantification can create an illusion of objectivity and encourage surveillance. The responsible analyst must know when not to collect data, when not to automate, and when evidence is too weak to support a confident recommendation.

Chapter Summary

People analytics supports HR decision making by organizing workforce data, research evidence, professional judgment, and stakeholder values around important organizational questions. Strong practice moves from decision definition to measurement, data-quality assessment, analysis, intervention, evaluation, and learning.

Descriptive metrics and dashboards are valuable when definitions, denominators, time periods, and limitations are visible. Predictive models estimate risk but do not automatically identify effective interventions. Causal claims require designs that approximate the counterfactual, with randomized experiments providing strong evidence when feasible and ethical and quasi-experiments offering useful alternatives. ROI calculations can clarify economic value but depend on defensible cost and benefit assumptions and should be complemented by nonfinancial and distributional considerations.

Responsible people analytics requires privacy protection, fairness assessment, transparency, governance, security, human accountability, and opportunities for appeal. The most effective HR analytics function is not the one that produces the greatest number of models; it is the one that improves decisions while preserving validity, trust, dignity, and organizational learning.

Key Terms

People analytics: Systematic use of workforce data and analytical methods to improve decisions affecting employees and organizations.

Evidence-based HR: HR decision making that integrates research evidence, internal data, professional expertise, and stakeholder values.

Construct: An abstract concept, such as engagement or commitment, that must be operationalized through measures.

Metric: A quantified representation of an organizational phenomenon.

Key performance indicator: A strategically important metric used to monitor progress and support action.

Dashboard: A structured visual display of selected metrics, trends, targets, and alerts.

Data validity: The degree to which a measure represents the construct it is intended to measure.

Data completeness: The extent to which required records and fields are present.

Correlation: A measure of association between variables; it does not by itself establish causation.

Regression: A statistical method for estimating relationships between an outcome and one or more predictors.

Counterfactual: The outcome that would have occurred for the same unit under an alternative condition.

Randomized controlled trial: An evaluation in which eligible units are assigned by chance to intervention and comparison conditions.

Quasi-experiment: A design that estimates intervention effects without full random assignment.

Difference-in-differences: A method comparing changes over time between intervention and comparison groups.

Predictive analytics: Analysis used to estimate likely future outcomes or risks.

Return on investment: Net financial benefit divided by program cost, expressed as a percentage.

Algorithmic bias: Systematic and unjustified disadvantage produced or amplified by an analytical system.

Data minimization: Limiting collection and use to information necessary for a defined purpose.

Model governance: Processes for documenting, validating, monitoring, controlling, and reviewing analytical models.

Implementation fidelity: The extent to which an intervention is delivered as designed.

Review Questions

Short Answer Questions

1. Distinguish descriptive, diagnostic, predictive, and prescriptive HR analytics.
2. Why is average headcount used in a turnover-rate denominator?
3. Explain why correlation does not establish causation in people analytics.
4. Identify four dimensions of workforce data quality.
5. State two privacy safeguards that should govern people analytics.

Long Answer Questions

1. Explain how a manager can transform a broad HR concern into a decision-relevant analytical question.
2. Discuss the design principles of an effective HR dashboard and the risks of poorly designed dashboards.
3. Compare randomized experiments with quasi-experimental designs for evaluating HR interventions.
4. Explain how ROI analysis should be combined with nonfinancial, ethical, and distributional evaluation.
5. Evaluate the governance requirements for using AI in hiring, promotion, or retention decisions.

Analytical / Application Questions

1. An organization has 500 employees at the start of a year, 560 at the end, and 48 separations during the year. Calculate the approximate turnover rate and explain one limitation of the calculation.
2. A retention program costs $200,000 and is estimated to prevent 12 departures valued at $22,000 each. Calculate ROI and identify two assumptions that should be tested through sensitivity analysis.
3. A department’s engagement score rises after a leadership-training program, but the organization also introduced higher wages and reduced workload at the same time. Design an evaluation strategy that would improve the credibility of a causal claim.

Multiple Choice Questions with Answers

Question 1

Which question is most clearly causal rather than merely descriptive?

A. How many employees resigned last quarter?

B. Which department has the highest absence rate?

C. Does structured supervisor coaching reduce six-month voluntary turnover?

D. What percentage of employees completed training?

Correct Answer: C. Does structured supervisor coaching reduce six-month voluntary turnover?

Question 2

A unit begins with 240 employees, ends with 260, and has 18 separations. Using average headcount, what is the approximate turnover rate?

A. 6.9%

B. 7.2%

C. 7.5%

D. 8.0%

Correct Answer: B. 7.2%

Question 3

Which statement best describes a predictive HR model?

A. It proves that an intervention caused an outcome.

B. It estimates the likelihood of a future outcome.

C. It replaces the need for managerial judgment.

D. It guarantees fair decisions across all groups.

Correct Answer: B. It estimates the likelihood of a future outcome.

Question 4

What is the principal purpose of a comparison group in an intervention evaluation?

A. To increase the number of dashboard colors

B. To provide an estimate of what might have happened without the intervention

C. To eliminate all measurement error

D. To ensure that every employee receives the intervention

Correct Answer: B. To provide an estimate of what might have happened without the intervention

Question 5

Which assumption is especially important in a difference-in-differences design?

A. All employees must have identical salaries

B. The groups must contain exactly the same number of employees

C. The groups would have followed parallel trends without the intervention

D. The intervention must last exactly one year

Correct Answer: C. The groups would have followed parallel trends without the intervention

Question 6

A program costs $120,000 and produces estimated financial benefits of $180,000. What is its ROI?

A. 33.3%

B. 50.0%

C. 60.0%

D. 150.0%

Correct Answer: B. 50.0%

Question 7

Which practice best supports responsible people analytics?

A. Collecting all available data indefinitely

B. Removing protected attributes and assuming bias is eliminated

C. Providing purpose limitation, access controls, documentation, and appeal processes

D. Using predictions as automatic employment decisions

Correct Answer: C. Providing purpose limitation, access controls, documentation, and appeal processes

Question 8

Why might the median be preferred to the mean when analyzing employee pay?

A. The median is always larger than the mean

B. The median is less affected by extreme values

C. The median measures causality

D. The median eliminates missing data

Correct Answer: B. The median is less affected by extreme values

Question 9

What does implementation fidelity measure?

A. Whether an intervention was delivered as designed

B. Whether a model has the highest possible accuracy

C. Whether all employees received identical outcomes

D. Whether a dashboard contains enough metrics

Correct Answer: A. Whether an intervention was delivered as designed

Question 10

Which statement about removing protected attributes from an algorithm is most accurate?

A. It always eliminates discrimination.

B. It guarantees equal error rates.

C. Other variables may still act as proxies for protected characteristics.

D. It makes human oversight unnecessary.

Correct Answer: C. Other variables may still act as proxies for protected characteristics.