8. Quantitative Research and Statistical Inference

Introduction and context

Quantitative research converts observations into numerical representations and uses statistical reasoning to describe patterns, estimate unknown quantities, evaluate propositions, and support decisions under uncertainty. In business research, quantitative methods are used to examine questions such as whether employee autonomy predicts innovation, whether a pricing intervention changes demand, whether customer satisfaction mediates the relationship between service quality and loyalty, or whether the effect of a leadership programme differs across organisational units.

The distinctive contribution of quantitative research is not simply the production of numbers. Its value lies in disciplined measurement, explicit assumptions, transparent analytical procedures, and a defensible connection between evidence and inference. Statistical inference allows researchers to move from a sample to claims about a wider population, but such movement is never automatic. It depends on the research design, sampling process, quality of measurement, model specification, and uncertainty surrounding the estimates.

At doctoral level, quantitative analysis should therefore be treated as an integrated research logic rather than a collection of techniques. A statistically significant coefficient may have little managerial importance; a non-significant result may reflect limited statistical power rather than absence of an effect; and a sophisticated model cannot compensate for biased sampling or poor measurement. This chapter develops a complete analytical perspective, from data screening and visualisation to regression, multivariate analysis, mediation, moderation, multilevel models, and structural equation modelling.

Learning objectives

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

1. Explain the logic, assumptions, and limitations of quantitative research and statistical inference.

2. Screen, clean, visualise, and document quantitative data appropriately.

3. Interpret estimation, confidence intervals, hypothesis tests, statistical power, and uncertainty.

4. Distinguish statistical significance from effect size and substantive managerial relevance.

5. Select and interpret regression, multivariate, mediation, moderation, multilevel, and structural equation models.

6. Diagnose common threats, including missing data, outliers, multicollinearity, endogeneity, measurement error, and model misspecification.

7. Apply ethical, governance, and reproducibility principles to quantitative business research.

From constructs to evidence

Quantitative research normally begins with theoretical constructs rather than directly observable facts. Constructs such as organisational commitment, perceived fairness, absorptive capacity, digital maturity, and strategic agility must be operationalised through variables. A variable may be continuous, ordinal, nominal, binary, count-based, or time indexed. The level of measurement affects which summaries and models are defensible.

A construct can be measured with a single indicator or with multiple items. Multi-item measures may improve reliability and represent different dimensions of a construct, but adding items does not automatically create validity. Reliability concerns consistency; validity concerns whether the measure supports the intended interpretation. Content validity asks whether relevant aspects are represented. Criterion-related validity examines association with an external criterion. Construct validity includes convergent and discriminant evidence. In doctoral research, validity is an argument supported by theory and evidence, not a property established by one coefficient.

The unit of analysis must also be explicit. A study may analyse individuals, teams, firms, transactions, projects, countries, or repeated observations over time. An individual-level variable should not be casually interpreted as a firm-level phenomenon. The ecological fallacy occurs when group-level patterns are inferred about individuals; the atomistic fallacy occurs when individual-level associations are used to explain groups. Clear specification of the unit of analysis protects the research from such errors.

A quantitative design may be cross-sectional, longitudinal, experimental, quasi-experimental, or panel-based. Randomised experiments provide strong leverage for causal claims when implementation is valid and attrition is controlled. Observational designs can estimate associations and, with suitable identification strategies, may support stronger causal arguments, but statistical adjustment alone does not guarantee causality. Temporal ordering, confounding, selection, and measurement processes must be considered together.

Preparing the dataset: screening before modelling

Data screening is an analytical stage, not clerical housekeeping. The researcher should preserve the raw dataset, create a documented analysis file, maintain a data dictionary, and record every transformation. Variable names, coding rules, units, missing-value codes, and sources should be transparent to collaborators and future readers.

Initial screening examines impossible values, duplicates, inconsistent coding, range violations, and implausible combinations. For example, an employee age of 250, a negative revenue amount where impossible, or a completion date preceding a start date indicates a data-quality problem. Such values should not be silently deleted. The researcher should determine whether they represent entry errors, legitimate exceptional observations, or missingness disguised by a code such as 999.

Missing data require substantive reasoning. Missing completely at random means that missingness is unrelated to observed and unobserved values. Missing at random means that missingness can be explained by observed variables after those variables are included in the analysis. Missing not at random means that the missingness depends on unobserved or missing values themselves. These conditions cannot usually be proven from the dataset alone, so researchers should combine statistical diagnostics with knowledge of the data-generating process.

Listwise deletion is simple but can reduce power and introduce bias. Mean substitution generally understates variability and distorts relationships. More defensible alternatives may include maximum likelihood methods or multiple imputation, provided the imputation model reflects the analysis model and includes relevant predictors of missingness. A sensitivity analysis using alternative assumptions is often more informative than presenting one apparently precise result.

Outliers are observations that are unusual relative to the distribution or model. They may be errors, influential but valid cases, or evidence that the assumed population is heterogeneous. Univariate outliers can be inspected using boxplots, robust statistics, or standardised scores; multivariate outliers may be assessed with measures such as Mahalanobis distance. Influence diagnostics, including leverage and Cook’s distance, help identify observations that materially change fitted regression estimates. Deletion should be justified substantively and analytically, not used merely to improve significance.

Reliability screening may include internal-consistency coefficients, item-total relationships, and examination of reverse-coded items. A high alpha is not proof of unidimensionality and can be inflated by many redundant items. Factor analysis, theory, and measurement validity should accompany reliability assessment.

Visualisation as reasoning

Visualisation allows the researcher to see distributional shape, heterogeneity, non-linearity, and data errors before imposing a model. Histograms and density plots show concentration and skewness. Boxplots compare distributions across groups. Scatterplots reveal association, curvature, clusters, and influential cases. Line plots are useful for time trends, while interaction plots can clarify whether a relationship differs across levels of a moderator.

Effective visualisation is analytical rather than decorative. Axes should be labelled, units identified, scales chosen honestly, and uncertainty shown where relevant. A truncated axis may exaggerate a small difference; a three-dimensional chart may obscure rather than explain relationships. Graphs should support the research question and should not imply causal direction when the design only supports association.

Descriptive statistics provide the first numerical account of the data. The mean is sensitive to extreme values, whereas the median is more robust in skewed distributions. The standard deviation describes dispersion around the mean; the interquartile range describes the middle half of observations. Frequencies and proportions are appropriate for categorical variables. Descriptive summaries should be reported for the sample as a whole and, where theoretically meaningful, for important subgroups.

Estimation and the logic of uncertainty

A population parameter, such as the mean employee engagement score or the slope relating training hours to productivity, is usually unknown. A sample statistic estimates it. Estimation should be accompanied by an interval expressing uncertainty. A confidence interval gives a range of values generated by a procedure that would contain the true parameter in a specified proportion of repeated samples under the model assumptions. It is not technically the probability that the particular parameter lies in the computed interval.

The width of an interval depends on variability, sample size, and the chosen confidence level. Larger samples generally produce narrower intervals, although a large sample does not correct systematic bias. Precision and validity are different: an estimate can be very precise and consistently wrong if the sample or measurement process is biased.

A hypothesis test begins with a null hypothesis and an alternative hypothesis. The p-value is the probability, under the null model and assumptions, of obtaining a result at least as extreme as that observed. It is not the probability that the null hypothesis is true, nor is it a measure of the size or importance of an effect. A conventional significance threshold may be used as a decision rule, but it should not replace substantive reasoning.

Statistical power is the probability of detecting an effect of a specified size under specified assumptions. It increases with sample size, effect magnitude, measurement reliability, and the chosen significance level, and decreases with noise. Power analysis should be conducted before data collection where possible. Post hoc power calculations based only on a non-significant result add little information; confidence intervals and plausible effect sizes are usually more useful.

Multiple testing creates an increased risk of false-positive findings when many hypotheses, outcomes, subgroups, or model specifications are examined. Researchers should pre-specify primary analyses where possible, distinguish confirmatory from exploratory work, and use appropriate adjustments or multilevel evidence when many tests are unavoidable. Selective reporting, p-hacking, and hypothesising after results are known undermine credible inference.

Effect sizes and managerial relevance

Effect size expresses the magnitude of an association or difference. Examples include a mean difference, standardised mean difference, correlation, regression coefficient, odds ratio, risk ratio, R-squared, or change in predicted probability. The appropriate effect size depends on the outcome, design, and decision context.

A statistically significant effect may be trivial if the sample is very large. Conversely, an economically important effect may fail to reach conventional significance in a small or noisy sample. Managers need to know not only whether an intervention changes an outcome, but by how much, for whom, at what cost, and with what uncertainty. A salary intervention that increases retention by a statistically detectable but financially negligible amount may not be worthwhile. A modest average effect may nevertheless be valuable if concentrated in a strategically important segment.

Interpretation should therefore integrate the point estimate, confidence interval, measurement scale, baseline level, opportunity cost, and organisational consequences. Standardised effects facilitate comparison across measures, but unstandardised effects often communicate more clearly to decision-makers. Researchers should translate results into meaningful units, such as revenue per customer, days of absence, probability of renewal, or percentage-point change.

Regression as a framework for explanation and prediction

Linear regression models an outcome as a function of one or more predictors. In a simple model, the slope represents the expected change in the outcome associated with a one-unit increase in the predictor. In multiple regression, a coefficient represents the association with a predictor while holding other included predictors constant. This conditional interpretation is not automatically causal.

Core assumptions include an appropriate functional form, independent errors where relevant, constant error variance, and an adequately specified error distribution for the intended inference. Residual plots can reveal non-linearity and heteroskedasticity. Robust standard errors may improve inference under heteroskedasticity, but they do not repair omitted-variable bias or incorrect functional form. Transformations, polynomial terms, splines, or generalised linear models may be appropriate when outcomes are skewed, binary, counts, or otherwise non-normal.

Multicollinearity occurs when predictors are strongly related. It can inflate standard errors and make individual coefficients unstable even when the model predicts well. Correlation matrices, variance inflation diagnostics, theoretical justification, and sensitivity analyses help assess the problem. Removing a theoretically essential variable merely to obtain significance is not a sound solution.

Model fit should not be reduced to R-squared. A model with high explanatory power may overfit the sample and perform poorly on new observations. Cross-validation, holdout samples, information criteria, prediction error, and out-of-sample validation are valuable when prediction is an objective. Explanation and prediction are related but distinct goals: a variable can improve prediction without supporting a useful causal explanation.

For binary outcomes, logistic regression models the log odds of an event. Odds ratios above one indicate higher odds per unit increase, but odds are not probabilities and can be difficult to interpret when the outcome is common. Predicted probabilities and marginal effects are often more accessible. Count outcomes may require Poisson or negative binomial models, while time-to-event outcomes may require survival methods.

Multivariate and latent-variable analysis

Multivariate analysis examines several outcomes or predictors jointly. MANOVA can test group differences across multiple correlated dependent variables, although interpretation requires attention to the covariance structure and follow-up analyses. Principal components analysis reduces observed variables into components oriented toward data reduction. Exploratory factor analysis seeks latent dimensions that may account for covariance among observed items. These methods should not be treated as interchangeable.

Confirmatory factor analysis tests a theoretically specified measurement model. It evaluates whether observed indicators plausibly represent latent constructs and whether constructs are distinguishable. Fit indices can inform judgment, but no single threshold establishes validity. Researchers should consider theoretical coherence, residuals, cross-loadings, measurement error, and the consequences of modifications. Modification indices should not be used as a data-driven route to a preferred model without substantive justification.

Structural equation modelling combines measurement and structural models. It can estimate relationships among latent variables while explicitly representing measurement error. A typical workflow specifies the theoretical model, evaluates the measurement model, assesses identification, tests the structural relations, examines indirect effects, and conducts sensitivity or alternative-model analyses. Identification means that the available information is sufficient to estimate the model; underidentified models cannot yield unique estimates.

Fit is not the same as truth. A model can fit well because it is flexible, because the sample is unusual, or because several theoretically different models make similar predictions. SEM is most informative when theory guides model specification and rival explanations are considered. Measurement invariance testing is important when comparing groups or time points. Configural invariance concerns the general pattern of factors; metric invariance concerns comparable loadings; scalar invariance concerns comparable intercepts. Without appropriate invariance, group comparisons may reflect measurement differences rather than substantive differences.

Mediation, moderation, and conditional processes

Mediation addresses how or through what process an antecedent relates to an outcome. For example, a developmental leadership practice may influence retention through perceived career support. The indirect effect is the product of the path from the antecedent to the mediator and the path from the mediator to the outcome. Bootstrap confidence intervals are commonly used because indirect effects often have asymmetric sampling distributions.

Mediation does not by itself establish a causal mechanism. Stronger claims require temporal ordering, credible measurement, and control of confounding. Cross-sectional mediation models are especially limited because all variables are measured at one time and alternative directions remain plausible.

Moderation addresses when or for whom a relationship changes. A product term between a predictor and moderator represents interaction. The coefficient of the predictor is conditional on the moderator’s value, so main effects must be interpreted carefully. Researchers should probe significant interactions using predicted values, simple slopes, or regions of significance, and graph the result. Mean-centering can improve interpretability and reduce nonessential collinearity, but it does not solve substantive collinearity or endogeneity.

Conditional process models combine mediation and moderation, such as a training programme affecting performance through self-efficacy more strongly when supervisory support is high. Such models should be driven by a clear theory rather than assembled because the software permits it. Complex models increase demands on sample size, measurement quality, transparency, and replication.

Multilevel and longitudinal reasoning

Business data are often nested: employees belong to teams, teams belong to firms, customers generate repeated transactions, and observations occur across months or years. Ordinary regression assumes independent observations; ignoring clustering can understate uncertainty and produce misleading inferences. Multilevel models separate within-group and between-group variation and can estimate random intercepts and, where justified, random slopes.

A group-mean predictor and a grand-mean predictor may have different meanings. For example, an employee’s perception of team climate differs from the team’s average climate. Conflating these levels can produce ecological or compositional errors. Intraclass correlation quantifies the proportion of variance attributable to clustering and helps determine whether multilevel modelling is warranted, although substantive design considerations remain important.

Longitudinal models can examine change, temporal ordering, and individual trajectories. Fixed-effects approaches control for time-invariant characteristics in some panel settings, while growth models estimate systematic change and variation in change. Time trends, autocorrelation, attrition, and changing measurement properties require explicit treatment. Repeated observations do not automatically convert an observational design into a causal experiment.

An illustrative case: evaluating a digital sales-coaching programme

Background

This illustrative case concerns Meridian Components, a fictional business-to-business distributor with twelve regional sales teams. Senior leaders introduced a digital coaching platform intended to improve opportunity management and reduce customer churn. The research team was asked to determine whether the programme worked and whether it justified expansion.

Situation and research design

The company assigned six teams to begin using the platform immediately and six to begin after six months. The outcome was quarterly renewal probability at the customer level, with sales representatives and teams forming nested structures. The researchers also measured coaching usage, perceived managerial support, opportunity quality, and representative experience.

A naive comparison of renewal rates after six months suggested a statistically significant advantage for users. However, early-adopting teams had more experienced managers and served larger accounts. The research question therefore required more than a difference in means.

Analysis

The team first checked duplicate customer records, missing renewal outcomes, inconsistent account identifiers, and extreme contract values. Visualisations showed that renewal rates varied substantially by region and account segment. A multilevel logistic model estimated renewal probability while accounting for customers nested within representatives and teams. Baseline renewal history, account size, sector, and representative experience were included as covariates. The researchers reported predicted probabilities rather than only odds ratios.

Usage was treated cautiously. Because representatives who used the platform more may already have been more motivated, usage could not be interpreted as a randomly assigned treatment. The primary comparison therefore focused on assignment to early implementation, while usage analyses were labelled exploratory. An interaction between programme assignment and managerial support tested whether implementation conditions altered the effect.

The estimated average improvement was positive, but its confidence interval included effects ranging from negligible to commercially meaningful. The interaction suggested that benefits were larger in teams where managers discussed coaching outputs regularly. A mediation analysis proposed that the programme might improve renewal through opportunity-quality assessments, but the cross-sectional measurement of the mediator limited causal interpretation.

Decision and outcome

Meridian did not expand the platform solely because the p-value crossed a conventional threshold. Instead, leaders compared the predicted improvement with licence costs, implementation time, and the financial value of retained accounts. They funded a second-stage rollout with stronger randomisation, pre-specified outcomes, manager training, and a twelve-month follow-up. The original evidence supported cautious continuation, not an unconditional claim of success.

Learning points

The case demonstrates the importance of assignment rather than self-selected usage, hierarchical modelling for nested data, uncertainty intervals, interaction analysis, and translation of effects into financial consequences. It also shows that a statistically persuasive result can remain decisionally ambiguous when implementation costs and effect heterogeneity are considered.

Threats, limitations, and common errors

Quantitative research is vulnerable to sampling bias, non-response, common-method bias, measurement error, omitted variables, reverse causality, post-treatment adjustment, selection effects, and inappropriate generalisation. Controlling for a variable affected by the treatment can introduce bias rather than remove it. Statistical controls should be selected from a causal and theoretical model, not from an automatic search for variables that change the coefficient.

Researchers frequently mistake correlation for causation, use stepwise selection without theoretical rationale, interpret p-values as effect sizes, remove outliers selectively, ignore clustering, overfit complex models, and report only significant outcomes. Another error is treating a statistically significant result as evidence that all organisations or subgroups will experience the same effect. External validity depends on the population, context, implementation, and mechanism.

Robustness checks should be purposeful. They may involve alternative operationalisations, plausible missing-data assumptions, nonlinear specifications, clustered or robust standard errors, subgroup analyses specified in advance, and out-of-sample validation. A long list of arbitrary specifications is not a substitute for a clear analytical strategy.

Ethics, governance, and responsible use of analytics

Quantitative research involving employees, customers, or firms must protect privacy, confidentiality, autonomy, and fairness. Data minimisation, secure storage, access controls, de-identification, retention policies, and lawful processing are governance requirements as well as ethical responsibilities. Researchers should explain how data will be used and should avoid repurposing sensitive information without appropriate authorisation.

Automated scoring and predictive models may reproduce historical discrimination. Fairness should be examined across relevant groups, but fairness metrics can conflict and cannot be selected without considering the decision context. A model that predicts employee attrition should not automatically be used to penalise employees or deny opportunities. Human review, contestability, documentation, and monitoring are essential.

Reproducibility is strengthened through preregistration where appropriate, version-controlled code, a transparent analytic record, data dictionaries, documented exclusions, and sharing of materials consistent with confidentiality obligations. AI-assisted analytics can support coding, visualisation, anomaly detection, and model comparison, but generated analyses require human verification. Automation does not remove the need for theoretical judgment, data protection, or accountability.

Current developments and best practice

Contemporary quantitative research increasingly combines causal inference, predictive analytics, longitudinal data, experiments, and machine learning. Machine-learning methods can improve prediction through flexible algorithms, regularisation, and cross-validation, but predictive accuracy does not establish causal explanation. Explainability methods can describe model behaviour, yet they should not be confused with causal mechanisms.

Open science practices are also expanding. Researchers increasingly distinguish exploratory from confirmatory analyses, report all primary outcomes, use registered reports, publish code, and conduct replication or multiverse analyses. Bayesian approaches provide an alternative framework in which prior information is combined with observed data to produce posterior distributions. They can be especially useful for sequential learning and direct probability statements about parameters, provided priors and modelling decisions are transparent.

Strong practice can be summarised as follows:

• Define the unit of analysis, estimand, outcome, and target population before modelling.
• Link measurement choices and model assumptions to theory.
• Inspect and visualise data before testing hypotheses.
• Report estimates, uncertainty intervals, effect sizes, and practical implications.
• Separate confirmatory tests from exploratory discovery.
• Diagnose missingness, influence, dependence, nonlinearity, and measurement quality.
• Use models proportionate to the design and research question.
• Validate predictions out of sample and test plausible alternative explanations.
• Document decisions sufficiently for replication and audit.
• Communicate limitations without treating them as a ceremonial appendix.

Chapter Summary

Quantitative research is a structured approach to measurement, estimation, explanation, prediction, and decision-making under uncertainty. Credible analysis begins with clear constructs, appropriate operationalisation, a defined unit of analysis, and a design capable of supporting the intended claim. Data screening, missing-data analysis, outlier assessment, reliability evaluation, and visualisation are essential before formal modelling.

Statistical inference uses estimates, confidence intervals, and hypothesis tests, but p-values do not communicate effect magnitude or managerial value. Effect sizes, uncertainty, costs, baseline conditions, and heterogeneity must be considered together. Regression and generalised linear models support association and prediction, while multivariate methods, mediation, moderation, multilevel models, and SEM address increasingly complex theoretical and organisational structures. These methods require attention to assumptions, measurement error, clustering, identification, invariance, and causal limitations.

Responsible quantitative research also requires ethical data governance, reproducible documentation, protection against discriminatory automation, and honest communication of uncertainty. The central doctoral-level principle is that analytical sophistication is valuable only when it is matched by valid measurement, appropriate design, transparent reasoning, and substantive interpretation.

Key Terms

Construct: An abstract theoretical concept represented through observable indicators.

Operationalisation: The process of defining how a construct will be measured.

Unit of analysis: The entity about which a study makes its principal claims.

Parameter: A numerical characteristic of a population.

Estimator: A rule or statistic used to estimate a population parameter.

Confidence interval: An interval produced by a procedure designed to contain a parameter at a specified long-run rate.

P-value: The probability, under a null model, of results at least as extreme as the observed result.

Statistical power: The probability of detecting a specified effect under stated assumptions.

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

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

Missing at random: Missingness explainable by observed information included in the analysis.

Multicollinearity: Strong association among predictors that can destabilise individual coefficient estimates.

Heteroskedasticity: Non-constant variance of model errors.

Mediation: A process in which an intermediate variable helps explain an association.

Moderation: A condition in which the strength or direction of a relationship varies by another variable.

Multilevel model: A model accounting for observations nested within higher-level units.

Measurement invariance: Evidence that a measure functions comparably across groups or time.

Structural equation modelling: A framework combining latent-variable measurement and structural relationships.

Endogeneity: Association between a predictor and the unexplained component of an outcome, often producing biased estimates.

Reproducibility: The capacity for analytical results to be regenerated from documented data and procedures.

Review Questions

Short Answer Questions

1. Why is statistical significance insufficient for judging managerial importance?

2. Distinguish missing at random from missing not at random.

3. What is the difference between mediation and moderation?

4. Why can ignoring team-level clustering distort regression inference?

5. What does measurement invariance allow a researcher to assess?

Long Answer Questions

1. Explain how a researcher should move from theoretical constructs to valid quantitative measures.

2. Discuss the role of confidence intervals, effect sizes, and power in interpreting statistical results.

3. Compare explanatory regression with predictive modelling and explain why high R-squared does not prove causality.

4. Describe the stages and assumptions involved in developing a structural equation model.

5. Evaluate the ethical and governance risks of using predictive analytics in employee management.

Analytical / Application Questions

1. A large survey finds that a leadership score predicts retention with p < .001, but the predicted improvement is 0.3 percentage points and the confidence interval ranges from 0.1 to 0.5 points. How should a researcher communicate the result to senior management?

2. A company compares users and non-users of a training platform, but users self-select and belong disproportionately to high-performing teams. Design an analysis that improves the credibility of the comparison and identify remaining limitations.

3. An employee-engagement model includes employees nested within departments and proposes that managerial support mediates the effect of autonomy on innovation, with the effect moderated by workload. What analytical issues must be addressed before estimating this model?

22. Multiple Choice Questions with Answers

Question 1

Which statement best describes a p-value?

A. The probability that the null hypothesis is true

B. The probability that the observed effect is practically important

C. The probability, under the null model, of observing a result at least as extreme as the result obtained

D. The probability that the sample represents the population

Correct Answer: C. The probability, under the null model, of observing a result at least as extreme as the result obtained

Question 2

A researcher finds a statistically significant effect in a very large sample, but the estimated change is too small to affect operating decisions. This illustrates that:

A. Statistical significance and substantive significance are distinct

B. Large samples eliminate measurement error

C. A significant result must be causal

D. Confidence intervals are unnecessary

Correct Answer: A. Statistical significance and substantive significance are distinct

Question 3

Which practice is generally most defensible when missing data are plausibly missing at random?

A. Replace every missing value with zero

B. Use multiple imputation or an appropriate likelihood-based method

C. Delete all variables containing missing values

D. Substitute the sample mean without further assessment

Correct Answer: B. Use multiple imputation or an appropriate likelihood-based method

Question 4

In a regression model, multicollinearity primarily threatens:

A. The existence of the dependent variable

B. The stability and precision of individual coefficient estimates

C. The need to visualise the data

D. The definition of the research population

Correct Answer: B. The stability and precision of individual coefficient estimates

Question 5

A moderation effect exists when:

A. A predictor has no relationship with an outcome

B. An intermediate variable explains every observed association

C. The strength or direction of a relationship varies according to another variable

D. Two variables have identical means

Correct Answer: C. The strength or direction of a relationship varies according to another variable

Question 6

Why is a multilevel model appropriate for employees nested within teams?

A. It guarantees causal conclusions

B. It accounts for dependence and variation at more than one organisational level

C. It removes all missing data

D. It converts ordinal measures into continuous measures

Correct Answer: B. It accounts for dependence and variation at more than one organisational level

Question 7

Which statement about structural equation modelling is most accurate?

A. Good model fit proves that the proposed theory is true

B. SEM can combine latent-variable measurement with structural relationships

C. SEM eliminates the need for theory

D. SEM is appropriate only for experimental data

Correct Answer: B. SEM can combine latent-variable measurement with structural relationships

Question 8

In the illustrative sales-coaching case, why was platform usage not treated as the primary causal comparison?

A. Usage was impossible to measure

B. Usage was randomly assigned to all representatives

C. Usage could reflect pre-existing motivation and therefore involved self-selection

D. Usage had no relationship with any outcome

Correct Answer: C. Usage could reflect pre-existing motivation and therefore involved self-selection

Question 9

Which result would most directly communicate managerial relevance for a binary customer-renewal outcome?

A. Only the model’s p-value

B. A predicted probability change with its uncertainty interval and financial interpretation

C. The number of predictors entered into the model

D. The unlabelled correlation matrix

Correct Answer: B. A predicted probability change with its uncertainty interval and financial interpretation

Question 10

What is the principal danger of interpreting a cross-sectional mediation model as proof of a causal mechanism?

A. Cross-sectional data always have too many observations

B. The temporal order and alternative explanations remain uncertain

C. Mediation cannot involve latent variables

D. Indirect effects are always exactly zero

Correct Answer: B. The temporal order and alternative explanations remain uncertain