Statistics for HR Analytics

Introduction and context

HR analytics uses evidence to understand workforce patterns and improve decisions about recruitment, retention, performance, rewards, absence, learning and workforce planning. Statistics provides the language for describing data, quantifying uncertainty and evaluating whether observed differences are likely to reflect genuine patterns rather than random variation.

A sound HR analyst must distinguish three questions:

1. What happened? Descriptive statistics summarize observed data.
2. How uncertain is the estimate? Probability models and confidence intervals quantify uncertainty.
3. What can reasonably be concluded? Hypothesis tests, effect sizes, correlation and regression support cautious inference.

Statistical significance alone is not sufficient for an HR decision. A very small effect may be statistically significant in a large workforce, while an important practical effect may fail to reach significance in a small sample.

Learning objectives

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

• calculate and interpret common descriptive statistics;
• explain probability, sampling and important probability distributions;
• construct and interpret confidence intervals;
• formulate and evaluate hypotheses;
• distinguish statistical significance from practical significance;
• calculate and interpret effect sizes, correlation and regression;
• identify assumptions, uncertainty, bias and common analytical errors;
• communicate results responsibly in HR contexts.

1. Describing HR data

A population is the complete group of interest, such as all employees in an organization. A sample is a subset used to learn about that population. A parameter describes a population; a statistic describes a sample.

Variables may be categorical (department or contract type), ordinal (performance rating), discrete (number of absences) or continuous (salary or tenure). Measurement scales matter: means are generally inappropriate for nominal categories, whereas frequencies and proportions are appropriate.

1.1 Centre and spread

For observations x₁, x₂, …, xₙ, the sample mean is:

The mean uses every observation but can be distorted by extreme values. The median is the middle ordered value and is more robust for skewed variables such as salary. The mode is the most frequent value.

The sample variance and standard deviation are:

The denominator n − 1 is used when estimating population variability from a sample; this is Bessel’s correction. The interquartile range (IQR) is Q₃ − Q₁ and summarizes the middle 50% of observations. The coefficient of variation is s ÷ x̄, provided the variable has a meaningful non-zero ratio scale.

Worked example: absence

Suppose five employees record 2, 3, 3, 4 and 8 absence days. The mean is 20 ÷ 5 = 4 days. The median is 3 days. Deviations from the mean are −2, −1, −1, 0 and 4; squared deviations total 22. Thus, sample variance = 22 ÷ 4 = 5.5 days² and sample standard deviation ≈ 2.35 days. The difference between mean and median indicates that the employee with 8 days raises the mean.

For categorical outcomes, a proportion is:

where x is the number with the characteristic and n is the total number observed. Percentages should always state their denominator and time period.

2. Probability and distributions

Probability ranges from 0 to 1. For an event A, P(A) expresses its long-run relative frequency or degree of uncertainty. The complement rule is P(not A) = 1 − P(A). For mutually exclusive events A and B, P(A or B) = P(A) + P(B). In general, the addition rule is:

Conditional probability is:

Independence means P(A | B) = P(A), not merely that events appear unrelated in a small dataset.

Common distributions in HR analytics

• Bernoulli distribution: one trial with two outcomes, such as voluntary exit or no exit.
• Binomial distribution: number of successes in n independent trials with constant success probability p. Its mean is np and variance is np(1 − p).
• Normal distribution: continuous, symmetric distribution described by mean μ and standard deviation σ. Standardization is:

• t distribution: used commonly for means when population standard deviation is unknown, especially with smaller samples. It has heavier tails than the normal distribution and depends on degrees of freedom.
• Chi-square distribution: used in tests of categorical association and variance-related procedures.

Distributional assumptions are models, not facts. A normal approximation may be reasonable for a mean under suitable conditions, but individual HR variables such as absence counts may be skewed or zero-inflated.

3. Sampling and the central limit theorem

Sampling error occurs because a sample is not the entire population. The standard error of the mean is:

This formula assumes independent observations and uses s as an estimate of population standard deviation. Increasing n reduces standard error at a rate of √n, so quadrupling the sample size approximately halves it.

The central limit theorem states, under suitable independence and finite-variance conditions, that the sampling distribution of the mean approaches normality as sample size increases. It does not state that the original data become normally distributed. Random sampling supports generalization; a large convenience sample does not automatically remove selection bias.

4. Confidence intervals

A confidence interval combines a point estimate with a margin of error. For a mean with unknown population standard deviation, a two-sided interval is:

Here t* is the critical value from the t distribution with n − 1 degrees of freedom. The assumptions include an appropriate sampling process, independent observations and either approximate normality of the population or a sufficiently large sample.

For a proportion, a simple large-sample interval is:

This approximation can perform poorly with small samples or proportions near 0 or 1; Wilson or exact methods may be preferable.

Interpretation

A 95% confidence interval does not mean there is a 95% probability that a fixed parameter lies in this particular interval. In repeated sampling, the method would capture the true parameter in approximately 95% of intervals, assuming its conditions hold. In practical reporting, one may say: “The estimated mean is 4.0 days, with a 95% confidence interval from 2.1 to 5.9 days.”

Worked example: training completion

In a random sample of 100 employees, 72 complete training. The estimate is p̂ = 0.72. Using z* = 1.96, the standard error is √[0.72 × 0.28 ÷ 100] ≈ 0.0449. The approximate margin of error is 1.96 × 0.0449 ≈ 0.088. The interval is approximately 0.632 to 0.808, or 63.2% to 80.8%. The result should not be generalized to all employees if the sample was self-selected.

5. Hypothesis testing

A hypothesis test begins with a null hypothesis H₀ and an alternative hypothesis H₁. For example, H₀ may state that average turnover intention is equal in two groups, while H₁ states that it differs.

A p-value is the probability, assuming H₀ and the model are true, of obtaining a result at least as extreme as the observed result. It is not the probability that H₀ is true, nor the probability that findings occurred by chance.

A significance level α is selected before analysis, commonly 0.05. If p ≤ α, analysts reject H₀; otherwise they fail to reject H₀. “Fail to reject” is not proof that groups are identical.

Errors and power

• Type I error: rejecting a true H₀.
• Type II error: failing to reject a false H₀.
• Power: 1 − probability of Type II error.

Power depends on effect size, sample size, variability and α. Low power makes non-significant results difficult to interpret. A post hoc search through many outcomes or subgroups increases false-positive risk unless multiplicity is addressed.

Common tests include the independent-samples t test for two means, paired t test for matched before-and-after observations, chi-square test for categorical association, and analysis of variance (ANOVA) for three or more means. Tests should be selected according to design and measurement, not simply because software offers them.

6. Effect sizes and practical significance

Effect size describes the magnitude of a difference or relationship, independently or partly independently of sample size.

For two independent means, Cohen’s d is commonly:

where the pooled standard deviation is:

This pooled form assumes comparable variances and independent groups. If variances differ substantially, a standardized effect using another justified standard deviation or a robust method may be preferable.

For a binary outcome, the risk ratio is:

where a and b are outcome and non-outcome counts in group 1, and c and d are corresponding counts in group 2. The odds ratio is:

Risk ratios are usually easier to communicate. An RR of 1 means equal risks; above 1 indicates higher risk in group 1; below 1 indicates lower risk.

A statistically significant difference may be too small to justify an intervention. Conversely, a meaningful reduction in turnover may be uncertain because the sample is small. HR leaders should consider cost, feasibility, fairness, baseline risk and confidence intervals alongside p-values.

7. Correlation and regression

Pearson correlation measures linear association between two quantitative variables:

Its value lies between −1 and +1. The sign indicates direction and the magnitude indicates linear strength. Correlation does not establish causation. It may be distorted by outliers, restriction of range or non-linear relationships.

Simple linear regression predicts an outcome Y from predictor X:

where b₀ is the intercept and b₁ is the expected change in Y for a one-unit increase in X. In ordinary least squares, coefficients minimize the sum of squared residuals, where residual eᵢ = yᵢ − ŷᵢ.

The coefficient of determination is:

R² is the proportion of sample variation explained by the fitted linear model; it is not the percentage of causation and does not prove that the model is appropriate.

Multiple regression can adjust for several predictors, but adjustment is not automatically causal. Analysts must consider confounding, mediation, multicollinearity, omitted variables, reverse causality and measurement error. For binary outcomes such as exit/no exit, logistic regression is often more suitable than linear regression; its coefficients are commonly interpreted through odds ratios after exponentiation.

Regression assumptions and diagnostics

Important assumptions include appropriate functional form, independent observations, reasonably stable variance of residuals (homoscedasticity), and no influential observations that dominate results. Normal residuals are mainly relevant to small-sample inference, not to unbiasedness of coefficients under every condition. Repeated observations from the same employee violate ordinary independence and may require clustered or mixed-effects methods.

Diagnostics should include residual plots, leverage and influence checks, missing-data assessment, correlation among predictors and out-of-sample validation. Prediction accuracy should be evaluated on data not used to fit the model.

8. A detailed illustrative case: evaluating a retention intervention

Background. A fictional organization introduces a manager-coaching programme in six departments. Six comparable departments continue usual practice. After one year, 108 of 120 coached employees remain, compared with 144 of 180 comparison employees.

Problem. Leaders ask whether the programme reduced turnover and whether it should be expanded.

Analysis. Retention is 108 ÷ 120 = 0.90 in the coached group and 144 ÷ 180 = 0.80 in the comparison group. The retention risk ratio is 0.90 ÷ 0.80 = 1.125. Thus, coached employees had 1.125 times the retention probability, or a relative retention increase of 12.5%. Turnover is 10% versus 20%, an absolute reduction of 10 percentage points. The number needed to treat, under a causal interpretation, would be 1 ÷ 0.10 = 10 employees; this interpretation requires credible comparability and design assumptions.

A chi-square test or a two-proportion method could assess statistical evidence, but the department-level assignment means employees are clustered. Treating all 300 employees as independent may understate uncertainty. A better analysis would account for department clustering, compare baseline turnover, report a confidence interval for the difference or ratio, and examine whether effects vary by tenure or job family.

Approach and solution. The analyst should predefine the outcome, use intention-to-treat assignment where applicable, document missing records, conduct a power assessment, and model department as a cluster or random effect. Qualitative feedback can identify whether coaching improved manager behaviour or merely coincided with other changes.

Outcome. The observed difference is operationally promising, but it is not automatically causal. Expansion should depend on uncertainty intervals, implementation cost, equity effects and replication.

Learning points. Always state denominators; distinguish absolute from relative effects; respect the unit of assignment; and combine statistical evidence with organizational relevance.

9. Common analytical errors and governance

Frequent errors include:

• treating correlation as causation;
• using the mean for highly skewed data without reporting robust summaries;
• ignoring missing-data mechanisms;
• testing many outcomes and reporting only significant findings;
• changing hypotheses after seeing results without disclosure;
• confusing percentage points with percentage change;
• using a model outside its measurement or sampling assumptions;
• interpreting a p-value as effect size;
• extrapolating beyond the observed population;
• allowing a dashboard to conceal uncertainty.

HR data are sensitive personal data. Analysts should apply purpose limitation, data minimization, access controls, retention rules and applicable privacy law. Protected characteristics require special care: models can reproduce historical discrimination, and seemingly neutral variables may act as proxies. Fairness assessment should examine selection rates, error rates and impacts across relevant groups, while recognizing that fairness definitions can conflict. Automated recommendations should be explainable, contestable and subject to human oversight.

Modern tools include statistical software, SQL, dashboards and machine-learning systems. AI can assist with pattern detection and coding, but it does not replace research design, assumption checking, privacy review or professional judgment. Reproducible scripts, versioned data definitions, documented transformations and peer review improve reliability.

Best-practice workflow

1. Define the HR decision and estimand precisely.
2. Identify the population, unit of analysis and sampling or assignment process.
3. Inspect distributions, missingness, outliers and data quality.
4. Report descriptive statistics before inferential tests.
5. Select methods whose assumptions fit the design and measurement.
6. Report estimates, confidence intervals and effect sizes—not p-values alone.
7. Check robustness and relevant subgroup impacts.
8. Distinguish association, prediction and causal claims.
9. Communicate limitations in plain language.
10. Document governance, privacy, fairness and reproducibility controls.

Chapter Summary

Statistics enables HR professionals to describe workforce data, quantify uncertainty and evaluate evidence. Means, medians, proportions and measures of spread summarize observations, while probability distributions provide models for sampling and outcomes. Confidence intervals express estimation uncertainty; hypothesis tests evaluate compatibility with a null model but do not measure importance. Effect sizes, absolute differences and relative risks support practical interpretation. Correlation and regression describe or predict relationships, subject to design and model assumptions, but neither automatically proves causation. Sound HR analytics requires attention to sampling, clustering, missing data, multiplicity, privacy, fairness and reproducibility. Decisions should integrate statistical evidence with organizational cost, context and ethical responsibility.

Key Terms

Population: Complete group about which a conclusion is intended.

Sample: Observed subset of a population.

Mean: Arithmetic average of numerical observations.

Median: Middle ordered observation.

Variance: Average squared deviation, using n − 1 for a sample estimate.

Standard deviation: Square root of variance.

Standard error: Estimated standard deviation of a sampling distribution.

Probability: Quantified uncertainty about an event.

Normal distribution: Symmetric continuous probability distribution described by μ and σ.

Central limit theorem: Result describing approximate normality of sampling means under suitable conditions.

Confidence interval: Interval produced by a procedure designed to capture a parameter at a stated long-run rate.

Null hypothesis: Default statistical claim evaluated by a test.

P-value: Probability, under H₀, of a result at least as extreme as observed.

Statistical power: Probability of detecting an effect of a specified size under assumptions.

Effect size: Quantitative magnitude of a difference or relationship.

Risk ratio: Ratio of two probabilities of an outcome.

Correlation: Standardized measure of linear association.

Regression: Model relating an outcome to one or more predictors.

Confounding: Distortion caused by a variable related to both predictor and outcome.

Practical significance: Importance of an effect for decisions, resources or people.

Review Questions

Short Answer Questions

1. Why might the median be preferable to the mean for salary data?
2. What does a 95% confidence interval mean under repeated sampling?
3. Distinguish Type I error from Type II error.
4. Why does correlation not establish causation?
5. What is the difference between a percentage-point change and a percentage change?

Long Answer Questions

1. Explain how descriptive statistics, confidence intervals and effect sizes should be combined in an HR report.
2. Discuss the assumptions and limitations of independent-samples t tests.
3. Compare risk ratios and odds ratios for communicating turnover outcomes.
4. Explain how clustering can affect inference in department-level interventions.
5. Evaluate ethical and governance risks in predictive HR analytics.

Analytical / Application Questions

1. A survey reports an engagement mean of 3.8 with a standard deviation of 0.6 from 64 independent employees. Calculate the standard error and explain what additional information is needed for a confidence interval.
2. A programme has 18 exits among 150 participants and 30 exits among 150 non-participants. Calculate each exit risk, the absolute risk difference and the risk ratio, then explain why these figures alone do not prove causation.
3. A regression predicts performance from training hours, but residual variance increases with fitted values. Identify the assumption problem and recommend two diagnostic or remedial steps.

Multiple Choice Questions with Answers

Question 1

Which statistic is generally most resistant to an extreme salary value?

A. Mean

B. Median

C. Variance

D. Standard deviation

Correct Answer: B. Median

Question 2

A sample contains 72 training completers out of 100 employees. What is the sample completion proportion?

A. 0.28

B. 0.50

C. 0.72

D. 1.72

Correct Answer: C. 0.72

Question 3

What does a p-value represent?

A. The probability that the null hypothesis is true

B. The probability that the result is practically important

C. The probability, assuming the null model, of a result at least as extreme as observed

D. The probability that the sample was randomly selected

Correct Answer: C. The probability, assuming the null model, of a result at least as extreme as observed

Question 4

If a sample mean is 50, the sample standard deviation is 10, and n = 25, what is the standard error?

A. 0.4

B. 2

C. 5

D. 10

Correct Answer: B. 2

Question 5

A retention rate is 0.90 in one group and 0.80 in another. What is the retention risk ratio for the first group relative to the second?

A. 0.10

B. 0.80

C. 1.125

D. 1.70

Correct Answer: C. 1.125

Question 6

Which statement about statistical significance is correct?

A. It proves that an effect is important to managers

B. It is independent of sample size

C. It indicates evidence against a null model under specified assumptions

D. It proves that the finding is causal

Correct Answer: C. It indicates evidence against a null model under specified assumptions

Question 7

A correlation of −0.75 indicates:

A. A strong negative linear association

B. A weak positive association

C. That one variable causes the other

D. That the variables have identical means

Correct Answer: A. A strong negative linear association

Question 8

In the regression equation ŷ = 4 + 2x, what does 2 represent?

A. The predicted outcome when x is zero

B. The residual standard deviation

C. The expected change in the predicted outcome for a one-unit increase in x

D. The proportion of variance explained

Correct Answer: C. The expected change in the predicted outcome for a one-unit increase in x

Question 9

Which issue most directly violates the ordinary independence assumption?

A. Using a median

B. Having repeated observations from the same employee without accounting for clustering

C. Reporting a confidence interval

D. Using a categorical predictor

Correct Answer: B. Having repeated observations from the same employee without accounting for clustering

Question 10

If a statistically significant turnover reduction is only 0.2 percentage points and implementation is very costly, the appropriate conclusion is:

A. The intervention must be adopted immediately

B. The result has no statistical meaning

C. Statistical significance should be considered alongside practical significance and cost

D. The p-value should be replaced by the mean

Correct Answer: C. Statistical significance should be considered alongside practical significance and cost