Research Technology, Statistical Software and AI

Introduction

Business research is increasingly conducted within digital ecosystems: survey platforms export machine-readable files, statistical software performs complex analyses, cloud repositories preserve code and outputs, and artificial intelligence (AI) assists with searching, coding, prediction and communication. These technologies improve speed and scale, but they do not remove the responsibilities of the researcher. An MBA researcher must still formulate a defensible question, select an appropriate design, inspect data quality, justify analytical choices, protect participants and communicate uncertainty.

This chapter treats software and AI as components of a research workflow rather than as substitutes for methodological judgement. It focuses on SPSS and R, reproducible research, AI-assisted literature review, qualitative coding and analytics, and the governance issues created by algorithmic decision-making.

Learning objectives

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

1. Select and use SPSS or R for appropriate business-data analyses.
2. Build a reproducible workflow linking data, code, outputs and decisions.
3. Evaluate AI assistance in literature review, coding and quantitative analytics.
4. Diagnose algorithmic bias and assess explainability, privacy and validation.
5. Design human oversight and governance for AI-supported research.
6. Interpret software output critically rather than treating it as evidence automatically.

1. The research-technology workflow

A modern workflow commonly contains these stages:

1. Question and design: define the construct, population, unit of analysis and identification strategy.
2. Data acquisition: collect survey, archival, transaction, text or experimental data lawfully.
3. Data management: document variables, coding decisions, missing values, transformations and exclusions.
4. Analysis: use statistical models, visualisations, qualitative coding or machine-learning methods appropriate to the question.
5. Validation: test assumptions, sensitivity, robustness, generalisability and possible leakage.
6. Communication: produce tables, figures, an audit trail and a transparent account of limitations.
7. Preservation: retain an appropriate version of the data, code, metadata and software environment.

A useful distinction is between reproducibility and replicability. Reproducibility means that another analyst can obtain the same results using the same data, code and documented environment. Replicability means that a new study, often with new data, reaches sufficiently consistent conclusions. A perfectly reproducible analysis may still fail to replicate if its sample is unusual, its measurement is weak or its claim is context-dependent.

Software creates three important risks. First, default settings can hide analytical choices. Second, proprietary interfaces may preserve results without preserving the process that generated them. Third, convenient automation can encourage researchers to search for statistically significant results rather than test a pre-specified argument.

2. SPSS and R as research environments

SPSS

SPSS provides a menu-driven environment suited to survey analysis, descriptive statistics, reliability analysis, cross-tabulation, regression, analysis of variance and many common multivariate procedures. Its strengths include accessible output, data-view facilities and widespread use in business schools. Syntax can also be saved, allowing menu operations to be converted into an auditable script.

Its limitations include dependence on proprietary software, less flexible data engineering than a programming language, and the risk that analysts click through multiple models without recording the decision process. An SPSS project should therefore preserve the original data, a cleaned working file, syntax files, output files and a codebook.

R

R is a programming language and statistical environment. It supports data import, transformation, visualisation, classical statistics, machine learning, text analysis and reporting. Packages extend its capabilities, while notebooks and dynamic documents can combine explanation, code and output. R is especially valuable where analysis must be automated, repeated or adapted to large and complex datasets.

R also creates responsibilities. Packages change, functions may have different defaults, and poorly documented scripts can be difficult to audit. A professional project records package versions, random seeds where stochastic procedures are used, operating-system details when relevant, and the sequence of transformations.

Choosing between them

The choice is not a contest between “easy” and “advanced” software. It depends on the research problem, team competence, data structure, reproducibility requirement, budget and governance environment. SPSS may be efficient for a standard organisational survey; R may be preferable for a pipeline combining web data, text, visualisation and repeated modelling. Both can produce excellent or poor research. Validity comes from design, measurement, assumptions and scrutiny—not from the brand of software.

A basic statistical illustration

Suppose a researcher models employee performance score Y_i using training hours X_i and controls Z_1i, ..., Z_ki. An ordinary least-squares model can be written as:

Here β₀ is the intercept, β₁ is the expected change in performance associated with one additional training hour holding the controls constant, β₂ through β_k are control coefficients, and ε_i is the unexplained component for observation i. The model is appropriate only when its substantive assumptions are defensible: variables are suitably measured, observations and error structure are considered, functional form is reasonable, and the interpretation does not exceed the research design. Regression association does not by itself establish causation.

The least-squares estimator selects coefficients that minimise the sum of squared residuals:

Y_i is the observed outcome and Ŷ_i the fitted value. Squaring makes positive and negative errors comparable and gives greater weight to large errors. The expression is a fitting criterion, not proof that the model is true. In managerial research, a statistically significant coefficient can be practically trivial, unstable across samples or confounded by omitted variables.

3. Reproducible workflows

A reproducible project separates raw data, processed data, analysis code, outputs, documentation and communication. Raw files should be read-only. Transformations should be scripted rather than performed only by hand. A data dictionary should identify variable names, labels, units, permissible values, missing-value codes, sources and confidentiality classifications.

A practical folder structure is:

project/
  README.md
  data_raw/
  data_processed/
  code/
  outputs/
  reports/
  references/
  environment/

A README explains the research question, access restrictions, software versions, execution order and known limitations. Version control records changes to code and documentation. It does not automatically make confidential data safe: access controls, encryption, de-identification and approved storage remain necessary.

A robust pipeline follows the principle import → inspect → clean → transform → analyse → validate → report. Each step should have explicit inputs and outputs. For example, a cleaning script should state how duplicate respondents were identified, how implausible values were handled, and whether missing data were deleted, imputed or modelled. A researcher should never describe “cleaning” as a neutral technical act; it changes the effective sample and can alter conclusions.

Validation and uncertainty

Validation includes checking ranges, duplicates, joins, missingness patterns, outliers, variable types and temporal ordering. For predictive models, data should be divided into training and evaluation information without allowing future or test information to influence model development. Data leakage occurs when information unavailable at the time of prediction enters the predictors or preprocessing procedure, producing deceptively strong performance.

Common performance measures have different meanings. For a binary classifier, a confusion matrix distinguishes true positives, false positives, true negatives and false negatives. Accuracy is:

where TP, TN, FP and FN denote the four outcome counts. Accuracy can be misleading with imbalanced classes. Precision, TP ÷ (TP + FP), describes the proportion of predicted positives that are correct; recall, TP ÷ (TP + FN), describes the proportion of actual positives detected. The appropriate metric depends on the costs of errors and the decision context.

4. AI-assisted literature review

AI tools can support discovery, screening, summarisation, translation, citation management and mapping of concepts. They can identify synonyms, cluster abstracts, suggest related papers and help a researcher construct search strings. These functions are useful when a field is large or interdisciplinary.

However, AI-generated literature assistance is not equivalent to scholarship. A system may hallucinate articles, misstate findings, privilege highly visible journals, omit non-English or negative results, reproduce database ranking biases or compress a contested debate into a false consensus. A generated citation must be checked against the original publication. The researcher remains responsible for inclusion criteria, quality appraisal, interpretation and accurate quotation.

A defensible AI-assisted review records:

• databases and dates searched;
• search strings and eligibility criteria;
• the AI tool, model or version and purpose of use;
• prompts and relevant settings;
• human decisions to include, exclude or reclassify studies;
• verification of every important claim and citation.

AI may propose categories, but a systematic review needs an auditable protocol. Automation should increase coverage and consistency without converting a review into an opaque ranking exercise.

5. AI-assisted qualitative coding

AI can assist with transcription, translation, keyword extraction, first-cycle coding, sentiment classification, memo drafting and retrieval of passages. In a customer-experience study, for example, it might label comments as “delivery delay,” “product quality” or “service recovery.” Such labels are hypotheses about meaning, not final findings.

Qualitative validity depends on context, reflexivity and interpretation. Models may misunderstand sarcasm, culturally specific language, power relationships or silence. They may impose categories from dominant language patterns and erase minority voices. A good workflow begins with a human-developed coding purpose and codebook, pilots AI suggestions on a sample, compares AI and human coding, resolves disagreements, and preserves examples supporting each theme.

Agreement is not identical to validity. Two coders can agree on an inadequate codebook, while disagreement can reveal ambiguity that requires theoretical refinement. Researchers should report whether AI generated, suggested or applied codes; how many records were reviewed by humans; and how disagreements were handled.

6. AI-assisted analytics and predictive modelling

AI analytics includes supervised prediction, unsupervised clustering, recommendation, anomaly detection and natural-language processing. The distinction between explanation and prediction is crucial. A predictive model can classify likely churn accurately while offering a weak causal explanation of why customers leave. Conversely, a theoretically interpretable regression may be useful for explanation but less accurate for prediction.

The analyst should define the decision, prediction horizon, population, outcome, acceptable error and intervention before selecting an algorithm. Model comparison should use an evaluation design appropriate to deployment, not merely the highest score on a convenient dataset. Baselines, out-of-sample testing, subgroup analysis and sensitivity analysis are essential.

Explainability has at least three meanings:

1. Transparency: whether the model or procedure is inspectable.
2. Interpretability: whether a human can understand how inputs relate to outputs.
3. Post-hoc explanation: an approximation that describes a particular prediction or overall pattern after a complex model has been fitted.

Post-hoc explanations can be useful but are not guarantees that the model is fair or causally meaningful. A feature-importance ranking does not prove that changing a feature will change the outcome. Explanations should therefore be tested for stability, fidelity to the model and usefulness to the affected decision-maker.

7. Algorithmic bias, privacy and governance

Algorithmic bias is systematic and unjustified difference in model performance or treatment across people or groups. It may enter through historical decisions, non-representative samples, labels reflecting unequal opportunity, measurement error, proxy variables, selective missingness, model objectives or deployment practices. Removing a protected attribute does not necessarily remove bias because other variables may encode it.

Bias must be assessed in context. Relevant questions include: Who is represented in the data? Who is missing? Are outcomes measured equally? Do error costs differ across groups? Is the target itself a legitimate construct? Are decisions being made in a domain where individuals can challenge or correct the result?

Several group metrics are commonly examined. A false-positive rate is FP ÷ (FP + TN) and a false-negative rate is FN ÷ (FN + TP). Equalised odds requires comparable true-positive and false-positive rates across protected groups. Demographic parity requires comparable positive prediction rates. These criteria can conflict, particularly when groups have different base rates; no single metric establishes fairness in every context. Fairness is therefore a socio-technical judgement requiring stakeholder input, not merely a software setting.

Privacy protection begins with data minimisation: collect only what is necessary for the stated purpose. Other safeguards include informed consent, purpose limitation, role-based access, encryption, retention limits, de-identification and secure deletion. De-identification is not an absolute guarantee; combinations of quasi-identifiers can permit re-identification. Sensitive data should not be entered into public AI tools without an approved legal and institutional basis.

Governance should allocate responsibility. A model card or equivalent record can document intended use, training data, performance, subgroup limitations, prohibited uses and monitoring requirements. A data-protection impact assessment may be appropriate where processing is high-risk. Organisations should also address intellectual property, confidential information, cross-border transfer, sector regulation and rights to contest automated decisions.

8. Human oversight and model lifecycle

Human oversight is meaningful only when humans have competence, authority, time and access to relevant evidence. A nominal approval click is not oversight. A strong lifecycle includes:

• Before development: define purpose, lawful basis, stakeholders, risks and success criteria.
• During development: document data provenance, test alternatives, assess subgroup performance and review leakage.
• Before deployment: conduct independent validation, security testing, usability review and an approval gate.
• During use: monitor drift, complaints, error patterns, subgroup outcomes and unexpected incentives.
• After incidents: suspend or roll back the system, investigate, notify affected parties where required and update controls.

Human review should be risk-based. Low-risk descriptive automation may require routine checking, while a system influencing employment, credit, healthcare or disciplinary action requires stronger review, appeal mechanisms and often a prohibition on fully automated final decisions.

9. Illustrative case: AI-supported employee-retention research

Background

A fictional retail company wants to identify employees at risk of leaving within six months. It has payroll records, scheduling data, an engagement survey and exit interviews. Management proposes an AI model to prioritise retention conversations.

Problem

The historical data reflect past managerial practices. Part-time workers and employees with incomplete surveys are underrepresented. Exit interviews are available mainly for employees who chose to participate. A model might therefore learn managerial bias rather than genuine predictors of avoidable turnover. In addition, using absence, schedule and location data may expose sensitive personal information.

Analysis and approach

The research team first defines the outcome and prediction date, excludes information occurring after that date, and creates a data dictionary. It compares a transparent logistic-regression baseline with a tree-based model. Data are split by time so that earlier records train the model and later records evaluate likely deployment performance. The team reports recall, precision and calibration, then examines false-positive and false-negative rates by job type, gender and employment status where lawful and ethically justified.

The model is not used to punish employees. Instead, a flagged employee is offered a voluntary conversation and support, with an explanation that the flag is a prioritisation aid rather than a judgement. Employees can correct inaccurate records. Managers receive training and cannot take adverse action solely on the model output. An independent review checks whether interventions are offered equitably and whether the model remains valid as labour conditions change.

Outcome

The complex model performs only slightly better than the transparent baseline and has poorer calibration for part-time workers. The organisation therefore adopts the baseline for research and limited decision support, improves data collection, and rejects automated disciplinary use. The result demonstrates that the best managerial choice may be a less complex model with clearer accountability.

Learning points

The case illustrates that prediction is not causation, performance averages can conceal subgroup harm, historical labels can encode institutional practices, and human oversight must include the power to reject or correct an algorithmic recommendation.

10. Best practices for MBA researchers

1. Begin with the research question and decision context, not the software.
2. Pre-specify primary outcomes, predictors and major analyses where feasible.
3. Preserve raw data and write transformations as reproducible scripts or syntax.
4. Record software, package, model and prompt versions.
5. Inspect data before modelling; document missingness and exclusions.
6. Distinguish descriptive, explanatory and predictive claims.
7. Validate on data that represent intended use and prevent leakage.
8. Report uncertainty, sensitivity and subgroup performance.
9. Verify every AI-generated citation, quotation, code fragment and factual claim.
10. Minimise personal data and never upload confidential material to unapproved systems.
11. Provide explanations that are accurate, comprehensible and actionable.
12. Establish human appeal, monitoring, rollback and incident procedures.

Chapter Summary

Research technology can improve efficiency, scale and transparency, but only when embedded in a disciplined methodological workflow. SPSS offers accessible menu-driven analysis and syntax-based auditing; R offers flexible, script-centred and extensible analysis. Reproducibility requires preserved data where permissible, code, metadata, environment information and an explicit audit trail. AI can assist literature discovery, qualitative coding and analytics, but generated content requires verification and human interpretation. Predictive performance must be separated from causal explanation, and explainability must not be confused with fairness. Algorithmic bias can arise from data, labels, measurement, modelling and deployment, while privacy requires minimisation, security and purpose limitation. Effective governance combines validation, subgroup assessment, documentation, monitoring, human authority and routes for correction or appeal.

Key Terms

Reproducibility: Obtaining the same results from the same data, code and documented environment.

Replicability: Obtaining consistent findings in a new study or dataset.

Data dictionary: Documentation describing variables, values, units, sources and missingness.

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

Validation: Evaluation of data, assumptions, performance and generalisability.

SPSS: A proprietary statistical environment with menu-driven and syntax-based analysis.

R: An open-source programming language and statistical environment.

Artificial intelligence: Computational systems that perform tasks associated with learning, inference or generation.

Prompt: An instruction or input supplied to a generative AI system.

Algorithmic bias: Systematic and unjustified differences in automated outcomes or errors.

Explainability: The extent to which model behaviour can be understood or communicated.

Interpretability: The degree to which a model’s operation is intelligible to people.

Fairness metric: A quantitative criterion used to compare outcomes or errors across groups.

Calibration: Agreement between predicted probabilities and observed frequencies.

Precision: The proportion of predicted positives that are actually positive.

Recall: The proportion of actual positives that a classifier detects.

Human oversight: Competent human review with authority to intervene, correct or stop a system.

Data minimisation: Collecting and retaining only data necessary for a defined purpose.

Model drift: Decline or change in model performance as data or context changes.

Data provenance: A record of where data came from and how they were transformed.

Review Questions

Short Answer Questions

1. Distinguish reproducibility from replicability.
2. Give two strengths and two limitations of using SPSS.
3. What is data leakage, and why does it distort predictive evaluation?
4. Why must AI-generated literature-review citations be verified?
5. Define meaningful human oversight.

Long Answer Questions

1. Compare SPSS and R for a large organisational survey and justify a software choice.
2. Explain how a reproducible workflow should manage raw data, transformations, code and outputs.
3. Discuss why predictive accuracy does not establish causal explanation.
4. Explain how algorithmic bias can enter a business analytics project at several stages.
5. Design a governance framework for an AI system used in employee selection.

Analytical / Application Questions

1. A churn model has high accuracy but a very low recall for a small customer segment. Analyse possible causes and recommend validation and governance actions.
2. An AI tool proposes themes for 2,000 customer comments. Design a human-AI coding protocol that protects validity, privacy and auditability.
3. A literature-review assistant omits studies from low-citation journals and produces two unverifiable references. Evaluate the review risk and propose a corrective workflow.

Multiple Choice Questions with Answers

Question 1

Which practice most directly improves reproducibility?

A. Saving only the final charts

B. Recording code, data transformations and software versions

C. Using the most complex algorithm available

D. Deleting intermediate files

Correct Answer: B. Recording code, data transformations and software versions

Question 2

In a regression model, β₁ is interpreted as the expected change in the outcome for a one-unit increase in X when:

A. All observations are identical

B. The sample size is zero

C. Other included variables are held constant

D. Causation has automatically been established

Correct Answer: C. Other included variables are held constant

Question 3

Which measure is especially concerned with the proportion of predicted positives that are correct?

A. Precision

B. Recall

C. Accuracy

D. False-negative rate

Correct Answer: A. Precision

Question 4

What is data leakage?

A. Encrypting a dataset

B. Losing a file during transfer

C. Removing duplicate records

D. Allowing unavailable future or test information into model development

Correct Answer: D. Allowing unavailable future or test information into model development

Question 5

Why should an AI-generated citation be checked against the original publication?

A. AI systems always refuse citations

B. Generated citations may be inaccurate or fabricated

C. Original articles cannot contain methods

D. Citation checking eliminates all research bias

Correct Answer: B. Generated citations may be inaccurate or fabricated

Question 6

Which statement best distinguishes prediction from explanation?

A. Prediction always proves causation

B. Explanation never requires theory

C. A predictive model can forecast well without identifying causal mechanisms

D. Explanation is identical to accuracy

Correct Answer: C. A predictive model can forecast well without identifying causal mechanisms

Question 7

Which is an example of data minimisation?

A. Collecting every available employee attribute

B. Retaining personal data indefinitely

C. Sharing raw data publicly by default

D. Collecting only variables necessary for the approved purpose

Correct Answer: D. Collecting only variables necessary for the approved purpose

Question 8

Why can removing a protected attribute fail to eliminate algorithmic bias?

A. Other variables may act as proxies

B. Models cannot use numerical variables

C. Protected attributes are always irrelevant

D. Bias exists only in handwritten decisions

Correct Answer: A. Other variables may act as proxies

Question 9

Meaningful human oversight requires humans to have:

A. Only a formal approval role

B. Authority, competence and evidence to intervene

C. No access to model limitations

D. A duty to accept every prediction

Correct Answer: B. Authority, competence and evidence to intervene

Question 10

A model has TP = 80, TN = 900, FP = 20 and FN = 100. What is its accuracy?

A. 80%

B. 88%

C. 89.1%

D. 90%

Correct Answer: C. 89.1%