Quantitative Data Collection and HR Analytics

Introduction and Context

Human resource management increasingly depends on evidence rather than intuition alone. Organizations collect data about recruitment, attendance, performance, compensation, learning, engagement, turnover and workforce diversity. When these data are systematically prepared and analysed, they can support better decisions about staffing, retention, workforce planning and employee development. This is the purpose of HR analytics: the disciplined use of data, statistical reasoning and contextual knowledge to understand workforce patterns and improve organizational decisions.

This chapter focuses on quantitative data collection and analysis for a bachelor's-level project or dissertation. It connects research design with practical HR applications. The emphasis is not merely on producing numerical outputs, but on asking defensible questions, measuring concepts appropriately, checking assumptions, interpreting uncertainty and communicating findings responsibly.

Learning Objectives

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

1. Design a quantitative HR data-collection process aligned with a research question.
2. Distinguish variables, measurement scales, populations, samples and sampling methods.
3. Prepare a dataset through coding, cleaning, missing-data treatment and documentation.
4. Calculate and interpret descriptive statistics and workforce metrics.
5. Select and interpret correlation, regression and group-comparison procedures.
6. Evaluate assumptions, statistical significance, confidence intervals and effect sizes.
7. Use spreadsheets or statistical software to create valid tables and visualisations.
8. Recognise ethical, legal and practical risks in people analytics.

1. From an HR Question to Quantitative Evidence

A quantitative study begins with a clearly specified problem. “Why are employees leaving?” is too broad for direct analysis. A more researchable question might be: “Is perceived supervisor support associated with employees’ intention to leave among customer-service employees?” The question identifies a population, measurable variables and a relationship that can be tested.

A useful analytical sequence is:

Concept → operational definition → indicator → data collection → analysis → interpretation → decision.

A concept is an abstract idea, such as engagement or perceived fairness. An operational definition explains how the concept will be measured. An indicator is the observed item or metric used to represent it. For example, engagement may be operationalised as the mean score of six Likert-scale items, while turnover may be measured as the number of employees leaving during a period divided by average headcount.

Research questions should identify whether the study is primarily:

• Descriptive: What is the current level or distribution?
• Comparative: Do two or more groups differ?
• Associational: Are variables related?
• Predictive: Can one or more variables predict an outcome?
• Evaluative: Did an intervention produce a measurable change?

A quantitative relationship does not automatically establish causation. A correlation between training participation and performance may reflect self-selection, managerial nomination or prior differences between participants and non-participants.

2. Variables, Measurement and Sampling

2.1 Variables and data types

A variable is a characteristic that can take different values. In an HR study, variables may include age, tenure, department, salary, engagement score, absence days and voluntary turnover status.

A single Likert item is technically ordinal, although a multi-item scale is often analysed approximately as interval data when its construction and reliability justify that treatment. Researchers should state this choice rather than treating measurement level as irrelevant.

A dependent variable is the outcome to be explained, such as turnover intention. An independent variable is a proposed predictor, such as workload. A control variable is included to account for plausible alternative explanations, such as tenure or job level. A confounder is related to both the predictor and outcome and can distort their apparent relationship.

2.2 Population and sample

The population is the complete group to which conclusions are intended to apply. The sample is the group actually observed. A sampling frame is the operational list from which the sample is selected.

Probability sampling gives known, non-zero selection probabilities. Simple random, systematic, stratified and cluster sampling are examples. Non-probability methods include convenience, purposive, quota and snowball sampling. Probability sampling generally supports stronger generalisation, but access constraints often make non-probability sampling common in student projects. In that case, limitations concerning representativeness must be stated explicitly.

A sample-size decision should consider the research design, expected effect size, desired power, significance level, number of predictors, attrition and feasibility. A large sample cannot correct systematic selection bias. Conversely, a small sample may produce imprecise estimates even when the sample is unbiased.

2.3 Questionnaire design

A good questionnaire uses clear, specific and non-leading questions. Avoid double-barrelled items such as “My manager is supportive and fair,” because respondents may assess the two ideas differently. Use balanced response options, consistent coding and a defined recall period. Sensitive questions should be justified and placed carefully.

For a multi-item scale, researchers should document the source, response range, reverse-coded items, scoring rule and reliability evidence. If item scores are x₁, x₂, …, xₖ, a simple mean score for respondent i is:

Here, Sᵢ is the respondent’s scale score, xᵢⱼ is the response to item j, and k is the number of items answered. This mean is appropriate when items measure the same construct, use compatible scales and a justified missing-response rule. Reverse-coded items must be transformed before calculation; on a 1-to-5 scale, the reversed score is 6 − x.

3. Preparing Quantitative HR Data

Data preparation is part of research quality, not a clerical afterthought. A documented data dictionary should identify each variable, label, measurement level, coding scheme, permitted range and missing-value code.

A defensible workflow is:

1. Preserve the original, read-only dataset.
2. Assign an anonymous study ID and remove direct identifiers from the analysis file.
3. Check duplicate records and impossible values.
4. Standardise dates, categories, units and spelling.
5. Examine missingness by variable and by relevant subgroup.
6. Identify potential outliers using substantive knowledge and statistical diagnostics.
7. Recode variables transparently and retain an audit trail.
8. Produce a final analysis dataset and codebook.

Missing data should not automatically be coded as zero. A missing salary is not a salary of zero, and “not applicable” differs from “prefer not to answer.” The researcher should consider whether missingness is plausibly completely random, related to observed variables, or related to the unobserved value itself. Complete-case analysis is simple but can reduce power and introduce bias. Imputation methods may be appropriate in larger, well-designed studies, but imputed values should not be presented as observed facts.

Reliability and scale quality

Internal consistency concerns whether items in a scale produce coherent responses. Cronbach’s alpha is commonly reported, although a high alpha does not prove unidimensionality or validity. Alpha for k items can be expressed as:

Here, sⱼ² is the variance of item j and s_T² is the variance of the total score. Alpha assumes, among other conditions, a reasonably coherent scale and comparable item interpretation. It should be interpreted alongside content validity, construct validity and the number of items; no universal cut-off is sufficient for every context.

4. Descriptive Statistics and Workforce Metrics

Descriptive statistics summarise what was observed. They do not, by themselves, justify claims about a wider population or causes.

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

The sample variance and standard deviation are:

The divisor n − 1 is used when estimating population variability from a sample. The mean is sensitive to extreme values, whereas the median is the middle ordered observation and is often more robust for skewed pay or tenure data. The interquartile range is Q₃ − Q₁ and describes the middle 50% of observations.

A proportion is:

where x is the number with a characteristic and n is the relevant denominator. A percentage is p̂ × 100. The denominator must be defined: “turnover was 12%” is incomplete unless the period and headcount basis are specified.

Common HR metrics include:

Worked example: turnover and descriptive interpretation

A unit has 120 employees at the beginning of a quarter and 130 at the end. It records 15 separations. Average headcount is (120 + 130) ÷ 2 = 125. The quarterly turnover rate is 15 ÷ 125 × 100 = 12%. If the organization annualises this figure by multiplying by four, the result is 48%, but such annualisation assumes comparable rates across quarters and should be labelled as an estimate, not an observed annual rate.

Suppose engagement scores for five employees are 2, 3, 3, 4 and 5 on a five-point scale. The mean is 17 ÷ 5 = 3.4, the median is 3, and the sample standard deviation is approximately 1.14. The difference between mean and median suggests modest asymmetry, but a larger sample and visualisation would be needed for a meaningful distributional assessment.

5. Visualisation and Exploratory Analysis

Visualisations should answer a question rather than decorate a report. Use bar charts for category counts, histograms for distributions, box plots for group distributions, line charts for time trends and scatterplots for relationships between quantitative variables. Label axes, units, denominators and time periods. Avoid truncated axes when they exaggerate differences and avoid three-dimensional effects.

Exploratory analysis may reveal unusual observations, non-linearity, subgroup differences or data-entry errors. It should precede formal modelling. A box plot, for example, can show whether a small number of employees have unusually high absence; the researcher must then investigate whether these are genuine observations or recording errors.

6. Correlation and Association

Correlation measures the direction and strength of association between two variables. Pearson’s sample correlation is:

The value ranges from −1 to +1. A positive value indicates that higher values of one variable tend to accompany higher values of the other; a negative value indicates the opposite. Pearson correlation is most informative for approximately linear relationships between quantitative variables and can be distorted by outliers. A scatterplot is essential.

Spearman’s rank correlation is useful for ordinal data or monotonic relationships and is based on ranked values. Neither correlation establishes causation. Statistical significance depends on sample size and uncertainty, while practical importance depends on the magnitude and consequences of the relationship.

7. Group Comparisons

A two-group comparison may use an independent-samples t-test when the outcome is approximately continuous, observations are independent and group distributions are reasonably suitable. Welch’s t-test is often preferred when group variances differ. A paired-samples t-test is used when the same units are measured twice, such as pre-training and post-training scores.

For more than two independent groups, one-way analysis of variance (ANOVA) tests whether group means differ overall. Its null hypothesis is that all population means are equal. If the overall test is significant, planned contrasts or suitable post-hoc comparisons identify which groups differ. ANOVA assumptions include independence, reasonably appropriate residual distributions and, for the classical equal-variance form, similar group variances. Kruskal–Wallis is a rank-based alternative when assumptions for mean comparisons are seriously unsuitable, although it answers a distributional question rather than automatically testing means.

A test statistic should be reported with an effect size and confidence interval, not only a p-value. For a two-group standardised mean difference, Cohen’s d is:

where sₚ is the pooled standard deviation under the equal-variance interpretation. If equal variances are not credible, a different standardisation or an unstandardised mean difference with its confidence interval may be more appropriate. Conventional labels such as small, medium and large are rough guides, not universal judgments of importance.

8. Regression and Prediction

Linear regression estimates the expected value of a quantitative outcome from one or more predictors. A simple model is:

Here, yᵢ is the outcome for person i, xᵢ is the predictor, β₀ is the intercept, β₁ is the expected change in y for a one-unit increase in x, and εᵢ is the unexplained error. In multiple regression:

Each coefficient is interpreted while holding the other included predictors constant. The model requires careful consideration of linearity, independent observations, influential cases, residual behaviour and multicollinearity. A predictor may be statistically associated with an outcome without being a cause.

The coefficient of determination is:

SS_res is the residual sum of squares and SS_tot is the total sum of squares. R² describes the proportion of sample outcome variation explained by the fitted model relative to a mean-only model. It is not the percentage of individual behaviour explained, nor proof of causal adequacy. Adjusted R² can be useful when comparing models with different numbers of predictors, but model choice should also be guided by theory and validation.

For binary outcomes such as whether an employee leaves, logistic regression is commonly used. It models the log odds:

where p is the predicted probability of the outcome. The quantity exp(βⱼ) is an odds ratio for a one-unit increase in predictor j, conditional on the other model variables. Odds ratios should not be casually described as probability changes, especially when the outcome is common.

9. Assumptions, Uncertainty and Interpretation

Assumption checking is a reasoning process, not a ritual in which a single test determines whether analysis is “allowed.” Inspect plots, study design and substantive knowledge. Important issues include:

• Independence: observations must not be treated as independent when employees are nested within teams or repeated over time.
• Linearity: linear regression and Pearson correlation require an appropriately linear relationship.
• Homoscedasticity: residual variability should be reasonably stable across fitted values.
• Normality: for small-sample inference, the relevant residuals or sampling distributions should be considered; raw variables need not always be normal.
• Multicollinearity: highly correlated predictors make individual regression coefficients unstable.
• Influential observations: a case can have a large effect on results even if its value is not an error.

A confidence interval gives a range of parameter values compatible with the data and modelling assumptions. A 95% confidence interval is not correctly interpreted as a 95% probability that a fixed parameter lies in this particular interval. Rather, across repeated samples using the same method, approximately 95% of such intervals would contain the fixed parameter under the model assumptions.

A p-value is the probability, under the null model and assumptions, of results at least as inconsistent with the null as those observed. It is not the probability that the null hypothesis is true. Researchers should report estimates, uncertainty, effect sizes, sample size and practical context.

Pre-registering hypotheses where feasible, distinguishing confirmatory from exploratory analyses, avoiding repeated unplanned testing and reporting null or inconclusive findings reduce risks of selective reporting and p-hacking.

10. Software and Reproducible Analysis

Excel or Google Sheets can support data entry, basic summaries, pivot tables and charts. Jamovi, JASP, SPSS, R and Python provide progressively richer facilities for statistical analysis and reproducibility. The software does not select a valid design automatically. Researchers remain responsible for variable coding, assumptions, interpretation and reporting.

A reproducible project should contain a raw-data folder, cleaned-data file, syntax or analysis script, codebook, output folder and decision log. Version control or dated file naming helps prevent accidental overwriting. Tables should report n, means or proportions, variability, confidence intervals where appropriate and clearly defined units. Visualisations should be accessible, with readable labels and colour choices that do not rely solely on colour distinctions.

AI tools may help draft code, explain error messages or suggest visualisation options, but generated code must be tested against known results. Confidential employee data should not be uploaded to an unapproved service. AI-generated interpretations can invent patterns or overstate causality; human review, institutional policy and data-protection requirements remain decisive.

11. Ethics, Governance and Legal Responsibility

People analytics concerns identifiable and potentially sensitive information. Ethical practice requires a legitimate purpose, proportional data collection, informed communication where appropriate, secure storage, restricted access, retention limits and responsible disposal. De-identification reduces risk but does not guarantee anonymity, particularly when datasets can be linked.

Researchers should avoid using protected characteristics to justify discriminatory decisions. Fairness assessment may require examining selection or error rates across groups, while recognising that fairness definitions can conflict and that small subgroup samples create uncertainty. Automated screening systems require validation, monitoring for disparate impact and a route for human review.

A dissertation must distinguish research participation from managerial coercion. Employees may feel compelled to respond if their manager distributes a survey. Confidentiality statements should be accurate; anonymity should not be promised if small teams or identifiable links make re-identification possible.

12. Illustrative Case: Predicting Voluntary Turnover Intention

Background

A fictional customer-support organization has 240 employees and reports increasing voluntary turnover. Management suspects that workload and supervisor support are related to turnover intention, but wants evidence before redesigning schedules or management training.

Problem and design

The researcher formulates the question: “After accounting for tenure and job level, are workload and supervisor support associated with turnover intention?” A cross-sectional survey measures workload and support on five-point multi-item scales. Turnover intention is measured by a three-item scale. Administrative records provide tenure and job level. Because the design is cross-sectional, the study is framed as associational rather than causal.

Analysis

After removing duplicate responses and checking ranges, 186 usable records remain. Scale scores are calculated only when at least two of three items are answered, with the rule documented in advance. Descriptive results show mean turnover intention of 3.1, workload of 3.7 and supervisor support of 2.8. A scatterplot suggests a positive workload relationship and a negative support relationship, with some curvature.

The researcher reports Pearson correlations as exploratory, then fits a multiple linear regression. Residual plots show mild heteroscedasticity and two influential observations. A sensitivity analysis reports results with and without those cases, rather than silently deleting them. The workload coefficient is positive, while support is negative; confidence intervals are included. The report avoids saying that workload “caused” intention to leave.

Approach and solution

The organization pilots a workload review and supervisor coaching in selected teams. A later evaluation uses pre- and post-measures and, where feasible, a comparison group. This design is stronger than the original cross-sectional study because temporal change and comparison are considered, although non-random team assignment still limits causal certainty.

Outcome and learning points

The analysis identifies areas for action but not a guaranteed causal mechanism. Managers learn that an average score can conceal team-level variation, that statistical association must be interpreted with effect size and uncertainty, and that workforce interventions should be evaluated rather than assumed to work. The case also demonstrates why combining survey data with administrative data requires clear governance and careful access controls.

Best-Practice Checklist

Before submitting a quantitative HR project, confirm that:

• the research question, variables and population are aligned;
• measurement instruments and coding decisions are documented;
• the sampling process and limitations are transparent;
• missing data, outliers and exclusions are explained;
• descriptive statistics match measurement levels;
• graphs are labelled and not misleading;
• the selected test matches the design and assumptions;
• effect sizes and confidence intervals accompany p-values;
• causal language is justified by the design;
• data are protected and ethical approval requirements are met;
• the analysis can be reproduced from the documented files.

Chapter Summary

Quantitative HR analytics transforms systematically collected workforce information into evidence for organizational decisions. High-quality analysis begins with a precise research question, operational definitions, suitable sampling and valid measurement. Data preparation requires coding, error checking, missing-data decisions and an audit trail. Descriptive statistics and workforce metrics summarise observed patterns, while visualisations support exploration and communication.

Correlation describes association but not causation. Group comparisons and regression can test differences and estimate relationships when their assumptions and design conditions are considered. Statistical significance must be interpreted alongside effect size, confidence intervals, sample size and practical consequences. Ethical governance is essential because HR data can affect privacy, employment opportunities and fairness. Software and AI can increase efficiency, but professional judgment, transparency and human accountability remain central.

Key Terms

Operational definition: A precise statement of how an abstract concept will be measured.

Population: The complete group to which a study intends to generalise.

Sample: The observations actually included in a study.

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

Variable: A characteristic that can take different values.

Dependent variable: The outcome being explained or predicted.

Independent variable: A proposed explanatory or predictive variable.

Confounder: A variable related to both a predictor and an outcome that can distort their relationship.

Reliability: The consistency of a measurement procedure.

Validity: The extent to which a measure supports the intended interpretation.

Missing data: Values not recorded or unavailable for a variable.

Outlier: An observation unusually distant from other observations.

Mean: The arithmetic average of observed values.

Standard deviation: A measure of dispersion around the mean.

Correlation: A numerical description of the direction and strength of association.

Regression: A model estimating an outcome from one or more predictors.

Effect size: A measure of the magnitude of a difference or relationship.

Confidence interval: An interval estimate expressing sampling uncertainty under specified assumptions.

Multicollinearity: Strong association among predictors in a regression model.

People analytics: The systematic use of workforce data, analysis and evidence to support HR decisions.

Review Questions

Short Answer Questions

1. What is the difference between a population and a sample?
2. Why should missing responses not automatically be coded as zero?
3. What does a correlation coefficient describe, and what does it not establish?
4. State two assumptions relevant to linear regression.
5. Why should HR metrics always specify their denominator and time period?

Long Answer Questions

1. Explain how an HR concept is transformed into a measurable variable, using employee engagement as an example.
2. Compare probability and non-probability sampling for a dissertation involving employees.
3. Explain the difference between statistical significance, effect size and practical significance.
4. Describe an appropriate workflow for cleaning and documenting a quantitative HR dataset.
5. Discuss the ethical and governance risks of using AI-based employee screening.

Analytical / Application Questions

1. An organization reports 18 separations, beginning headcount of 200 and ending headcount of 220 during a six-month period. Calculate the turnover rate using average headcount and explain two limitations of the metric.
2. A researcher finds that supervisor support and turnover intention have r = −0.42, p < .01. Interpret the result carefully, including what cannot be concluded.
3. Design a quantitative evaluation of a new manager-training programme. Identify the outcome, sampling or comparison strategy, appropriate analysis and at least three threats to validity.

Multiple Choice Questions with Answers

Question 1

Which measurement scale describes departments such as Finance, Marketing and Operations?

A. Nominal

B. Ordinal

C. Interval

D. Ratio

Correct Answer: A. Nominal

Question 2

A unit has 120 employees at the beginning of a quarter, 130 at the end and 15 separations. Using average headcount, what is the quarterly turnover rate?

A. 10%

B. 12%

C. 15%

D. 18%

Correct Answer: B. 12%

Question 3

Which statement about correlation is correct?

A. It proves that one variable causes another.

B. It can describe the direction and strength of association.

C. It is appropriate only for nominal variables.

D. It eliminates the effect of confounding variables.

Correct Answer: B. It can describe the direction and strength of association.

Question 4

Why is n − 1 used in the usual sample variance formula?

A. To convert a proportion into a percentage

B. To estimate population variability from sample data

C. To remove all outliers

D. To guarantee normally distributed data

Correct Answer: B. To estimate population variability from sample data

Question 5

Which graph is generally most suitable for examining the relationship between two quantitative variables?

A. Pie chart

B. Bar chart

C. Scatterplot

D. Stacked categorical chart

Correct Answer: C. Scatterplot

Question 6

What does a regression coefficient β₁ represent in a simple linear model, subject to its assumptions?

A. The expected change in the outcome for a one-unit predictor increase

B. The percentage of missing observations

C. The sample size required for the study

D. The probability that the null hypothesis is true

Correct Answer: A. The expected change in the outcome for a one-unit predictor increase

Question 7

Which procedure is commonly used for comparing means across more than two independent groups?

A. One-way ANOVA

B. Pearson correlation

C. Cronbach’s alpha

D. Simple random sampling

Correct Answer: A. One-way ANOVA

Question 8

Which statement about a p-value is accurate?

A. It is the probability that the null hypothesis is true.

B. It measures practical importance directly.

C. It is calculated under a null model and describes the extremeness of the observed result.

D. It proves that the result will replicate.

Correct Answer: C. It is calculated under a null model and describes the extremeness of the observed result.

Question 9

What is a major purpose of a data dictionary?

A. To replace the research question

B. To document variable meanings, coding and permitted values

C. To guarantee a representative sample

D. To increase the number of respondents automatically

Correct Answer: B. To document variable meanings, coding and permitted values

Question 10

Which action best reflects responsible HR analytics?

A. Upload identifiable employee data to any free AI tool

B. Report only statistically significant findings

C. Combine effect sizes, uncertainty, design limitations and ethical safeguards

D. Treat a high correlation as proof of a causal relationship

Correct Answer: C. Combine effect sizes, uncertainty, design limitations and ethical safeguards