Analytics and Evidence-Based Decision-Making

Introduction

Managers make decisions under conditions of incomplete information, competing objectives, time pressure and uncertainty. Analytics improves such decisions by converting data into structured evidence, while evidence-based management (EBM) disciplines the use of that evidence through critical evaluation of its quality, relevance and limitations. Neither analytics nor evidence replaces managerial judgment. Rather, each helps managers clarify a problem, compare alternatives, estimate consequences and recognize when confidence is unwarranted.

Analytics is commonly classified into four forms:

1. Descriptive analytics asks what happened.
2. Diagnostic analytics asks why it happened.
3. Predictive analytics asks what is likely to happen.
4. Prescriptive analytics asks what should be done, given objectives and constraints.

These categories are related but not interchangeable. A dashboard may describe declining sales without explaining their cause. A predictive model may forecast customer churn without identifying an intervention that is profitable. A prescriptive recommendation may optimize resource allocation while remaining unreliable if its demand assumptions are poor.

The central managerial task is therefore not merely to “use data,” but to create a defensible chain from decision problem to evidence, model, action and learning.

Learning Objectives

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

• distinguish descriptive, diagnostic, predictive and prescriptive analytics;
• formulate managerial problems as decisions rather than as vague data requests;
• evaluate data quality, causal claims, uncertainty and model performance;
• apply basic statistical and decision-analysis measures;
• compare alternatives using expected value, sensitivity analysis and constraints;
• interpret dashboards, forecasts and optimization outputs critically;
• design evidence-based decision processes that include human judgment;
• identify ethical, governance and organizational risks in analytics applications.

1. From Data to Decision

Data are recorded observations: transactions, customer interactions, employee records, sensor readings or survey responses. Information is data organized to answer a question. Evidence is information judged sufficiently relevant and credible to support a claim or decision. Insight is an interpretation that changes understanding. A decision is a committed choice among alternatives, normally involving resources, risk and accountability.

A useful analytical chain is:

Decision context → question → data-generating process → analysis → interpretation → action → outcome measurement

Breakdowns can occur at every stage. An organization may measure the wrong outcome, use a biased sample, confuse correlation with causation, optimize an incomplete objective or fail to monitor whether an intervention worked.

Decision framing

A well-framed decision specifies:

• the decision owner;
• the time horizon;
• the alternatives under consideration;
• the objective or objectives;
• relevant constraints;
• affected stakeholders;
• uncertainty and irreversibility;
• the outcome that will determine success.

For example, “How can we increase sales?” is too broad. “Should the firm increase digital advertising, reduce price, or improve delivery reliability over the next quarter, subject to a budget of $200,000 and a minimum gross margin of 30%?” is an analyzable decision.

A decision statement should also distinguish descriptive, causal and predictive questions. “Which customers churned?” is descriptive. “Did the retention campaign reduce churn?” is causal. “Which current customers are most likely to churn next month?” is predictive. These questions require different designs and should not be answered by the same metric automatically.

2. The Four Forms of Analytics

2.1 Descriptive analytics

Descriptive analytics summarizes historical or current data. Typical outputs include frequencies, averages, medians, rates, indexes, dashboards and time-series displays.

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

Here x̄ is the sample mean, xᵢ is observation i, and n is the number of observations. The mean is appropriate when observations are measured on a meaningful numerical scale and extreme values do not make it misleading. The median, the middle ordered observation, is often preferable for skewed measures such as income or delivery time.

A sample proportion is:

where k is the number of observations with the characteristic of interest and n is the total number observed. A proportion must have a clearly defined numerator, denominator and time period. “Conversion rate” is incomplete unless the organization states whether the denominator is impressions, visitors, leads or eligible customers.

Descriptive analytics is valuable for monitoring performance and identifying anomalies, but it does not establish causes. A rise in complaints may reflect poorer service, increased customer volume, a reporting-system change or improved willingness to complain.

2.2 Diagnostic analytics

Diagnostic analytics investigates explanations for observed outcomes. It may involve segmentation, drill-down analysis, cohort comparison, process mining, root-cause analysis, correlation analysis or quasi-experimental methods.

Correlation measures the direction and strength of linear association between two variables. The sample correlation coefficient is:

where xᵢ and yᵢ are paired observations, x̄ and ȳ are their sample means, and n is the number of pairs. The coefficient ranges from −1 to +1 and measures linear association, not causation. It is sensitive to outliers and can be distorted by restricted ranges or nonlinear relationships.

A diagnostic analysis should ask whether the apparent relationship survives plausible alternative explanations. Important threats include confounding, reverse causality, selection bias, measurement error and changes in the population being studied.

2.3 Predictive analytics

Predictive analytics estimates unknown, future or unobserved outcomes. It includes regression, classification, forecasting, simulation and machine learning. Prediction is evaluated by performance on data not used to fit the model.

A simple linear regression model is:

where yᵢ is the outcome, xᵢ is a predictor, β₀ is the intercept, β₁ is the expected change in y for a one-unit change in x under the model, and εᵢ is an error term. The model is not automatically causal. A predictive association may be useful even when it does not explain why the outcome occurs.

For a binary outcome, a logistic model represents the log-odds as a linear function of predictors:

where p is the modeled probability of the outcome and x₁ through xₖ are predictors. The probability must lie between 0 and 1; the log-odds transformation allows the linear predictor to range over all real numbers. Interpretation requires care: e^βⱼ is an odds ratio for a one-unit increase in xⱼ, holding other included variables constant, under the model assumptions.

Forecast accuracy should match the decision context. Common measures include mean absolute error:

and root mean squared error:

Here yᵢ is the observed value and ŷᵢ is the predicted value, in the same units as the outcome. MAE gives equal linear weight to errors; RMSE penalizes large errors more heavily. Neither is universally superior. Inventory decisions may penalize under-forecasting differently from over-forecasting, so business-specific loss functions may be more appropriate.

A model should be assessed using a holdout set or cross-validation, not only training performance. Data leakage—allowing information unavailable at the decision time into model development—can produce deceptively strong results. Calibration, fairness across groups, stability over time and operational usefulness matter in addition to predictive accuracy.

2.4 Prescriptive analytics

Prescriptive analytics evaluates or recommends actions. It combines forecasts, objectives, constraints, preferences and uncertainty. Linear programming is a foundational model.

Here xⱼ is the amount of decision activity j, cⱼ is its contribution to the objective, aᵢⱼ is the use of resource i per unit of activity j, bᵢ is the available amount of resource i, and Z is the objective value. This formulation assumes linear contributions and resource use, known coefficients, divisibility of decision variables unless integrality is added, and correctly specified constraints.

Prescriptive analytics may use optimization, simulation, decision trees, queuing models, robust optimization or reinforcement learning. A mathematically optimal solution is only useful if the objective represents organizational value and the input assumptions are credible. “Optimal” means optimal relative to a model, not necessarily optimal in the real world.

3. Evidence-Based Management and Critical Evaluation

Evidence-based management integrates four sources:

1. Organizational data, such as operational and financial records;
2. Professional expertise, including contextual and tacit knowledge;
3. Stakeholder values and concerns;
4. The best available external evidence, including research and industry evidence.

The approach is not “data versus intuition.” Expertise helps formulate questions and interpret feasibility; data tests claims; stakeholder input identifies consequences that metrics may omit; external evidence challenges local assumptions.

A practical evidence hierarchy is question-dependent. A randomized experiment is powerful for estimating the effect of an intervention under its conditions, but it may be infeasible or unethical for some strategic questions. Administrative data may be comprehensive but measure outcomes imperfectly. Qualitative interviews may reveal mechanisms and meanings that numerical records cannot capture. Quality depends on fit between evidence and question, not on a universal ranking.

Managers should evaluate evidence using the following tests:

• Relevance: Does it address the decision and population?
• Validity: Does the design support the claim being made?
• Reliability: Would measurement be reasonably consistent?
• Timeliness: Does it reflect current conditions?
• Completeness: Are important cases or outcomes missing?
• Transparency: Can assumptions and methods be examined?
• Actionability: Can the organization act on the finding?

Causality and experimentation

The strongest practical method for estimating the effect of an intervention is often a randomized controlled experiment. If eligible units are randomly assigned to treatment and control groups, randomization tends, in expectation, to balance observed and unobserved characteristics. The difference in average outcomes estimates the average treatment effect when implementation, measurement and compliance are appropriately handled.

In an experiment, the estimated difference is:

where ȳₜ is the average outcome for the treatment group and ȳ𝚌 is the average outcome for the control group. The interpretation depends on valid randomization, adequate sample size, limited interference between units, consistent outcome measurement and an appropriate analysis population.

When randomization is impossible, managers may use interrupted time series, difference-in-differences, matching or instrumental variables. Each method has assumptions. For example, difference-in-differences generally requires parallel trends: absent the intervention, treatment and comparison groups would have followed similar outcome trends. This assumption must be examined rather than asserted.

Statistical significance is not the same as managerial importance. A small effect may be statistically precise but economically irrelevant; a large effect may be uncertain because the sample is small. Managers should consider effect size, uncertainty intervals, costs, benefits and consequences of error.

4. Decision-Making Under Uncertainty

Risk involves uncertain outcomes for which probabilities can be estimated or approximated. Deep uncertainty exists when probabilities, models or relevant outcomes are themselves disputed. The distinction matters because a precise-looking expected value can create false confidence.

If alternatives a₁, …, aₘ lead to outcomes under states s₁, …, sₖ, the expected monetary value of alternative aⱼ is:

where P(sᵢ) is the probability of state i, and V(aⱼ, sᵢ) is the value produced by alternative j in that state. The probabilities should sum to 1, and values must use a consistent unit and time basis. EMV is appropriate when probabilities are meaningful and the decision-maker can reasonably evaluate repeated or comparable decisions. It does not imply that a single uncertain outcome will equal its expected value.

Worked example: capacity choice

A retailer must choose a small or large distribution facility. Demand may be high with probability 0.6 or low with probability 0.4. Estimated annual contribution, in thousands of dollars, is:

For the small facility:

For the large facility:

The large facility has the higher EMV by $24,000 per year. However, it also produces a lower contribution in the low-demand state. A risk-averse decision-maker may value downside protection, flexibility or real options more than this expected difference.

The expected value of perfect information is:

In the example, perfect information would yield 0.6 × 540 + 0.4 × 300 = 444. The best current EMV is 396, so EVPI is $48,000. This is the maximum rational amount to pay for perfect information, before considering the cost and feasibility of obtaining it.

Sensitivity and scenario analysis

Sensitivity analysis changes one or more assumptions to identify decision thresholds. In the example, let p be the probability of high demand. The large facility is preferred when:

Simplifying gives 360p + 180 ≥ 120p + 300, so 240p ≥ 120 and p ≥ 0.5. Thus, the large facility is preferred when the probability of high demand is at least 50%, assuming the stated payoffs. This threshold is more informative than simply reporting the base-case recommendation.

Scenario analysis changes a coherent set of assumptions, such as demand, price and costs together. It is useful when relationships among variables matter or when probabilities are not credible. Robust decisions often perform acceptably across several plausible scenarios rather than maximizing one fragile forecast.

5. A Managerial Analytics Workflow

A disciplined workflow contains the following stages.

Stage 1: Define the decision and theory of change

Specify the action, target outcome, time frame and mechanism by which the action is expected to produce value. Identify leading indicators as well as final outcomes. A theory of change prevents teams from optimizing an isolated metric.

Stage 2: Establish measurement and data provenance

Define each variable, unit, source, refresh frequency and ownership. Document missing values, revisions, inclusion rules and transformations. Data lineage is essential when decisions affect customers, employees or regulatory reporting.

Stage 3: Explore and diagnose

Use distributions, trends, segmentation and process analysis. Compare results across relevant groups and time periods. Investigate anomalies before deleting them; an outlier may be an error, a fraud signal or a genuine operational event.

Stage 4: Build and validate models

Select a model that is proportionate to the decision. Separate training and evaluation data, prevent leakage and report uncertainty. Compare with a simple baseline. A complex model should earn its complexity through improved, stable and actionable performance.

Stage 5: Translate outputs into alternatives

A forecast is not a decision. Convert it into options, expected consequences, constraints and implementation requirements. Include nonfinancial outcomes where relevant, such as safety, equity, employee welfare or customer trust.

Stage 6: Decide, document and implement

Record assumptions, evidence, dissenting views, approval authority and triggers for review. Assign implementation ownership. A decision log supports learning and accountability.

Stage 7: Monitor outcomes and update

Compare actual outcomes with forecasts and intended effects. Investigate model drift, changes in behavior and unintended consequences. Post-implementation evaluation closes the loop between analysis and organizational learning.

6. Detailed Applied Case: MetroCare Appointment Access

The following is an illustrative case, created for educational purposes.

Background

MetroCare, a regional outpatient healthcare network, experiences long telephone queues and missed appointments. Management proposes an online scheduling and reminder system. The investment would cost $1.2 million in the first year. Leaders expect fewer missed appointments and better access, but they are concerned that patients with limited digital access could be disadvantaged.

Problem

The initial proposal states: “Implement the system to reduce no-shows.” The analytics team reframes the decision: Should MetroCare pilot online scheduling and automated reminders, for which patient groups, under what support conditions, and with what safeguards?

The relevant outcomes are no-show rate, time to appointment, call-wait time, staff workload, patient satisfaction and access equity. The decision also includes constraints: clinical safety, privacy, accessibility, budget and staff capacity.

Analysis

Historical data show an overall no-show rate of 12%, but the rate is 19% for appointments booked more than 30 days in advance and 8% for appointments booked within 7 days. This descriptive result suggests that lead time may matter, but it does not show that shortening lead time causes attendance to improve. Patients who obtain early appointments may differ systematically from other patients.

A diagnostic review identifies three mechanisms: patients forget appointments, rescheduling is difficult, and reminder messages are unavailable in some preferred languages. The team proposes a randomized pilot among eligible clinics. Clinics continue ordinary booking, while randomly selected appointment slots receive the new reminder and rescheduling process. Patients without smartphones can receive voice calls or staff-assisted booking.

Suppose the pilot estimates a reduction in no-shows from 12% to 10%. If 10,000 eligible appointments are observed, the estimated absolute reduction is 2 percentage points, or approximately 200 appointments. The financial value should not be inferred from 200 alone. The team estimates avoided unused capacity, implementation cost, staff time, accessibility costs and any adverse effects.

The prescriptive question is then which combination of reminder channel, staffing level and booking horizon maximizes net benefit subject to budget and equity constraints. A dashboard alone cannot answer this question.

Approach and solution

MetroCare adopts a six-month pilot with:

• randomized allocation of eligible appointment slots where operationally feasible;
• pre-specified primary and secondary outcomes;
• subgroup monitoring by language, age and digital access;
• an accessibility alternative for patients unable to use online services;
• privacy impact assessment and limited data retention;
• weekly operational monitoring and independent review at month three;
• a decision threshold based on net benefit, not no-show reduction alone.

The pilot uses an intention-to-treat analysis for the effect of being offered the intervention, while also recording actual use. This avoids overstating effects by analyzing only patients who complied. Qualitative interviews investigate why the intervention succeeds or fails.

Outcome

Assume the pilot finds a reduction in no-shows and call volume, but no improvement for patients requiring translated voice reminders until that channel is added. Management therefore scales the system conditionally: digital reminders are expanded, multilingual voice support is funded, and quarterly equity audits are required. The recommendation is not “technology solved the problem”; it is “a tested service design improved selected outcomes when paired with inclusive implementation.”

Learning points

1. The decision was improved by defining alternatives and outcomes before selecting a tool.
2. Historical patterns generated hypotheses but did not establish causality.
3. Predictive targeting would require fairness and calibration checks.
4. Prescriptive analysis had to include budget, accessibility and privacy constraints.
5. Human judgment remained necessary to interpret patient experience and ethical obligations.
6. Evaluation continued after implementation rather than ending with deployment.

7. Governance, Ethics and Organizational Implications

Analytics can reproduce or amplify inequity. Bias may enter through historical decisions, under-represented samples, proxy variables, labels reflecting unequal treatment or optimization objectives that ignore vulnerable groups. Fairness is not a single technical property. Possible criteria include equal error rates, equal opportunity, demographic parity and individual fairness, but these criteria can conflict and may not fit the decision’s moral purpose.

Organizations should establish:

• clear data ownership and access controls;
• lawful, transparent and proportionate data collection;
• privacy-by-design, minimization and retention limits;
• model documentation, version control and audit trails;
• human review for high-impact decisions;
• appeal and correction mechanisms;
• monitoring for drift, disparate impact and unintended effects;
• separation of model development, validation and approval where risk warrants it.

Privacy is not merely a compliance issue. Excessive collection can increase security risk and weaken trust. Anonymization may not be absolute when datasets can be linked. Sensitive analytics should therefore consider purpose limitation, legitimate authority, consent where appropriate and the possibility of re-identification.

Generative and predictive AI tools can accelerate analysis, but they may produce fabricated explanations, leak confidential information, encode bias or obscure accountability. Managers should require source verification, controlled data environments, reproducible workflows and explicit disclosure of AI assistance. AI-generated recommendations should be treated as model outputs requiring validation, not as authoritative evidence.

Organizational adoption also depends on incentives. If employees are rewarded for a metric, they may optimize the metric rather than the underlying objective. Goodhart’s law—when a measure becomes a target, it may cease to be a good measure—explains why performance systems need multiple indicators, qualitative review and periodic redesign.

8. Limitations and Common Failure Modes

Metric fixation. A single KPI omits dimensions such as quality, safety or long-term value.

Confusing association with cause. A relationship in observational data may reflect confounding or selection.

Overfitting. A model learns noise in historical data and performs poorly on new cases.

Data leakage. Information unavailable when a decision is made is accidentally used in prediction.

Base-rate neglect. A high accuracy percentage can conceal poor performance for a rare but important event.

False precision. A forecast with many decimal places may rest on uncertain assumptions.

Automation bias. Decision-makers accept a system recommendation without adequate scrutiny.

Model drift. Relationships change as markets, policies, customer behavior or technology change.

Implementation blindness. A theoretically effective intervention fails because staff cannot execute it or users do not adopt it.

Optimizing the wrong objective. A model can be technically correct yet strategically harmful if the objective is incomplete.

The remedy is not to avoid analytics. It is to use appropriate evidence, transparent assumptions, independent challenge, scenario testing and outcome monitoring.

9. Best Practices for Evidence-Based Decisions

• Begin with a decision, not a dataset or fashionable tool.
• State the estimand: what quantity or effect is being estimated?
• Define success before viewing results where feasible.
• Use simple baselines and meaningful business metrics.
• Separate exploratory analysis from confirmatory claims.
• Report uncertainty and sensitivity, not only point estimates.
• Validate models out of sample and across relevant groups.
• Distinguish statistical significance, practical significance and financial significance.
• Combine quantitative evidence with expertise and stakeholder knowledge.
• Make assumptions, data lineage and limitations visible.
• Pilot consequential interventions and evaluate them prospectively.
• Design for privacy, fairness, security and accountability.
• Establish review triggers for changing conditions.

Chapter Summary

Analytics supports managerial decision-making through four connected forms. Descriptive analytics summarizes what happened; diagnostic analytics investigates why; predictive analytics estimates what may happen; and prescriptive analytics evaluates what should be done under objectives and constraints. Each form answers a different question and requires appropriate data, assumptions and validation.

Evidence-based management combines organizational data, external research, professional expertise and stakeholder values. Correlation does not establish causation, prediction does not explain mechanism, and optimization cannot correct an ill-defined objective. Expected value, sensitivity analysis, experimentation and scenario planning help managers reason under uncertainty, but none eliminates judgment.

A robust analytics workflow frames the decision, defines outcomes, documents data, validates models, compares alternatives, implements responsibly and learns from results. Ethical governance is essential because data and models can affect privacy, access, employment, credit, healthcare and public trust. The managerial standard is therefore not maximum automation; it is defensible, transparent and adaptive judgment informed by credible evidence.

Key Terms

Analytics: Systematic use of data, models and reasoning to generate insight and support decisions.

Descriptive analytics: Analysis that summarizes historical or current conditions.

Diagnostic analytics: Analysis that investigates explanations for observed outcomes.

Predictive analytics: Estimation of future or unknown outcomes.

Prescriptive analytics: Evaluation or recommendation of actions under objectives and constraints.

Evidence-based management: Integration of organizational data, research evidence, expertise and stakeholder values.

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

Confounding: Distortion of an observed relationship by another variable related to both variables of interest.

Causal effect: Change in an outcome attributable to an intervention or exposure under a defined causal framework.

Randomized controlled experiment: Study assigning units randomly to intervention and comparison conditions.

Expected monetary value: Probability-weighted average value of an alternative across possible states.

Expected value of perfect information: Maximum value obtainable from knowing the uncertain state before deciding.

Sensitivity analysis: Examination of how conclusions change when assumptions or inputs vary.

Scenario analysis: Evaluation of outcomes under coherent alternative sets of assumptions.

Overfitting: Modeling noise or idiosyncrasies in development data rather than generalizable patterns.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Model drift: Decline or change in model performance as conditions evolve.

Data governance: Policies, roles and controls governing data quality, access, use and accountability.

Automation bias: Excessive reliance on an automated recommendation or system output.

Goodhart’s law: The tendency of a measure to become less informative when it is used as a target.

Review Questions

Short Answer Questions

1. Distinguish descriptive, diagnostic, predictive and prescriptive analytics.
2. Why does correlation not establish causation?
3. What is data leakage, and why does it threaten predictive validity?
4. Define expected monetary value and state when it is useful.
5. Identify four dimensions on which managerial evidence should be evaluated.

Long Answer Questions

1. Explain how a manager can move from a vague business concern to a well-framed analytical decision.
2. Compare randomized experiments with observational diagnostic analysis, including their assumptions and limitations.
3. Discuss why predictive accuracy alone is insufficient for deploying a model in a managerial setting.
4. Explain how sensitivity analysis and scenario analysis improve decisions under uncertainty.
5. Evaluate the ethical and governance requirements for analytics used in high-impact decisions.

Analytical / Application Questions

1. A company reports that customers receiving more service calls have lower satisfaction. Design an analysis that distinguishes whether calls cause dissatisfaction, dissatisfaction causes calls, or a third factor causes both.
2. Using the MetroCare case, propose a set of primary and secondary measures for evaluating the pilot. Explain how you would address accessibility, privacy and unintended consequences.
3. A firm’s forecasting model has a lower RMSE but systematically under-forecasts demand for a critical product. Recommend whether to adopt it and justify your answer using loss, risk and operational considerations.

Multiple Choice Questions with Answers

Question 1

Which question is primarily addressed by predictive analytics?

A. What happened last quarter?

B. Why did the failure occur?

C. What is likely to happen next month?

D. Which action best satisfies the constraints?

Correct Answer: C. What is likely to happen next month?

Question 2

A sample contains 40 conversions from 500 visitors. What is the sample conversion proportion?

A. 0.04

B. 0.08

C. 0.40

D. 8.00

Correct Answer: B. 0.08

Question 3

What is the principal limitation of interpreting a correlation as a causal effect?

A. Correlation cannot be calculated from numerical data

B. Correlation always has a negative value

C. Confounding or reverse causality may explain the relationship

D. Correlation requires a randomized experiment

Correct Answer: C. Confounding or reverse causality may explain the relationship

Question 4

Which practice best prevents evaluating a predictive model only on memorized training patterns?

A. Increasing the number of dashboard colors

B. Using a holdout set or cross-validation

C. Removing all unusual observations automatically

D. Adding more predictors without validation

Correct Answer: B. Using a holdout set or cross-validation

Question 5

In the capacity example, what is the expected monetary value of the large facility?

A. $324,000

B. $372,000

C. $396,000

D. $444,000

Correct Answer: C. $396,000

Question 6

What does the expected value of perfect information represent?

A. The guaranteed profit from a decision

B. The maximum value of knowing the uncertain state before deciding

C. The cost of collecting any business data

D. The difference between training and test accuracy

Correct Answer: B. The maximum value of knowing the uncertain state before deciding

Question 7

Which condition is especially important for interpreting a difference-in-differences estimate?

A. The outcome must be binary

B. The treatment group must be larger than the comparison group

C. The groups should have followed parallel trends absent the intervention

D. All observations must have identical outcomes before treatment

Correct Answer: C. The groups should have followed parallel trends absent the intervention

Question 8

Why might MAE be preferred to RMSE in some forecasting applications?

A. MAE is always smaller in monetary value

B. MAE penalizes large errors more heavily

C. MAE gives errors linear rather than squared weighting

D. MAE can be used only for categorical outcomes

Correct Answer: C. MAE gives errors linear rather than squared weighting

Question 9

Which statement best describes prescriptive analytics?

A. It only reports historical totals

B. It recommends or evaluates actions using objectives and constraints

C. It proves that a predictor causes an outcome

D. It eliminates the need for managerial judgment

Correct Answer: B. It recommends or evaluates actions using objectives and constraints

Question 10

Which is an example of automation bias?

A. Comparing a model with a simple baseline

B. Reviewing outcomes across demographic groups

C. Rejecting a recommendation because its data are incomplete

D. Accepting a system recommendation without adequate scrutiny

Correct Answer: D. Accepting a system recommendation without adequate scrutiny