Problem Framing and the Data Science Lifecycle

Introduction

Data science projects rarely fail because an algorithm cannot be trained. They more often fail because the organization has not clearly defined the problem, the decision to be improved, the population affected, or the evidence required to justify action. A technically accurate model can therefore produce little value if it answers the wrong question, uses information unavailable at decision time, optimizes a misleading metric, or ignores organizational constraints.

Problem framing is the disciplined process of translating an organizational challenge into an analytical problem that can be investigated, measured, and acted upon. It connects strategy, domain knowledge, data, statistical reasoning, experimentation, and implementation. The data science lifecycle provides the structure through which this translation becomes an iterative project rather than a one-time modelling exercise.

This chapter examines four complementary perspectives: CRISP-DM, OSEMN, design thinking, and scientific research methods. It also distinguishes descriptive, diagnostic, predictive, and prescriptive analytics; develops a method for writing analytical questions and hypotheses; and applies these ideas to an illustrative business case.

Learning Objectives

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

1. Translate an organizational challenge into a decision-oriented analytical problem.
2. Distinguish business objectives, analytical questions, hypotheses, targets, features, constraints, and outcomes.
3. Explain and compare CRISP-DM, OSEMN, design thinking, and the scientific research cycle.
4. Differentiate descriptive, diagnostic, predictive, and prescriptive analytics.
5. Define measurable success criteria and select appropriate evaluation metrics.
6. Recognize causal, ethical, operational, and governance risks in problem formulation.
7. Develop a defensible project charter and lifecycle plan.

1. From Organizational Challenge to Analytical Problem

An organizational challenge is usually expressed in broad and ambiguous language: “reduce customer churn,” “improve hospital capacity,” “increase loan approval efficiency,” or “make marketing more effective.” Such statements identify a concern but do not yet specify what must be measured, predicted, changed, or decided.

A well-framed problem identifies five elements:

• Decision: What action or choice may change?
• Decision-maker: Who will use the analysis?
• Population and unit of analysis: Which people, transactions, products, or time periods are considered?
• Outcome: What measurable result represents success?
• Constraints: What legal, ethical, financial, technical, and operational limits apply?

For example, “reduce churn” is incomplete. A more precise formulation is: “For customers with an active subscription, can the service team identify customers with a high probability of cancellation within the next 30 days, early enough to offer an approved retention intervention, while maintaining a false-alert rate below 20% and avoiding discriminatory treatment?”

This formulation is better because it specifies a time horizon, population, intervention, operational user, and constraints. It also reveals that the project may involve two separate questions:

1. Prediction: Who is likely to cancel?
2. Decision or causal effectiveness: Which intervention is likely to prevent cancellation?

These questions should not be confused. A customer may be highly likely to churn but unlikely to respond to a discount. Prediction estimates risk; causal or uplift analysis helps select an effective action.

1.1 Levels of Problem Definition

A useful hierarchy is:

1.2 Unit of Analysis and Time Design

The unit of analysis is the entity represented by one row in the analytical dataset. It may be a customer, customer-month, claim, delivery, machine-hour, or patient encounter. Choosing the wrong unit can create biased conclusions or operationally unusable predictions.

Time must also be explicit. A prediction generally has:

• an observation window, during which features are collected;
• a prediction point, when the decision is made; and
• an outcome window, during which the target is observed.

Features must be available before or at the prediction point. Using information that becomes available only after the outcome begins is a form of target leakage. For example, using a cancellation survey completed after a customer has already decided to leave would make a model appear accurate while being useless for early intervention.

2. Analytical Questions, Hypotheses, and Outcomes

An analytical question should be answerable with evidence and should lead to a decision. Common forms include:

• Descriptive: What happened, to whom, where, and when?
• Diagnostic: What factors are associated with the observed pattern?
• Predictive: What is likely to happen next?
• Prescriptive: Which action should be taken under specified constraints?
• Causal: What change would result from an intervention?

A hypothesis is a testable proposition, not merely a topic. “Customer engagement matters” is vague. “Customers whose weekly product sessions fall by at least 50% during the observation window have a higher 30-day cancellation rate than comparable customers whose usage remains stable” is testable, provided the variables and comparison group are defined.

A strong hypothesis specifies:

1. the population;
2. the exposure, treatment, or explanatory variable;
3. the outcome;
4. the expected direction or difference;
5. the time relationship; and
6. relevant comparison conditions.

In formal testing, a null hypothesis often represents no difference or no association, while an alternative hypothesis represents the effect of interest. Statistical significance does not by itself establish business importance, causality, or fairness. A small effect may be statistically detectable in a large dataset but operationally irrelevant.

2.1 Operational Definitions

An operational definition translates a concept into a reproducible measurement. “Loyal customer” might mean a customer with at least three purchases in the previous six months, whereas “service failure” might mean a delivery more than 24 hours beyond the promised time. Different definitions produce different datasets and conclusions.

Every important variable should have a data dictionary containing its name, definition, type, unit, source, valid range, missing-value rule, time availability, and owner. For a binary outcome, define which value represents the event. For a financial outcome, specify whether revenue is gross or net of refunds, discounts, and variable costs.

3. The Analytics Taxonomy

3.1 Descriptive Analytics

Descriptive analytics summarizes observed data. It includes counts, averages, distributions, dashboards, cohort tables, and time-series summaries. It answers “what happened?” but does not necessarily explain why it happened.

For observations x₁, x₂, ..., x_n, the arithmetic mean is:

Here, n is the number of observations and xᵢ is observation i. The mean is appropriate when values are numerically meaningful and extreme values do not make it misleading. The median may better represent a skewed distribution such as income or claim cost.

3.2 Diagnostic Analytics

Diagnostic analytics investigates patterns, differences, and possible explanations. It may use segmentation, drill-down analysis, correlation, regression, process mining, or controlled comparisons. Association is not automatically causation. A relationship may be confounded by a third variable, caused in the reverse direction, or produced by selection bias.

The Pearson correlation between variables X and Y is:

It measures linear association and lies between −1 and +1 under the usual definition. It does not establish a causal effect and can be distorted by outliers, nonlinear relationships, or restricted ranges.

3.3 Predictive Analytics

Predictive analytics estimates an unknown or future outcome using information available at prediction time. Outputs may be a class, probability, count, ranking, or forecast. A probability is not a guarantee; it should be interpreted in relation to calibration, base rates, and decision thresholds.

For binary classification, a confusion matrix contains true positives, false positives, true negatives, and false negatives. Precision and recall are:

These formulas assume the denominators are nonzero. Precision answers, “Among cases predicted positive, how many were positive?” Recall answers, “Among actual positive cases, how many were identified?” The appropriate balance depends on the cost of false alerts and missed cases.

3.4 Prescriptive Analytics

Prescriptive analytics recommends actions, often under constraints. It may use optimization, simulation, decision analysis, reinforcement learning, or causal uplift modelling. A recommendation should state the objective, decision variables, constraints, uncertainty, and implementation conditions.

A simple expected-value formulation is:

Here, A is the set of feasible actions, a is one action, X is available information, Y is an uncertain outcome, and U is utility. This expression is only useful when the organization can estimate outcome consequences and the action set is operationally feasible.

3.5 Causal and Decision Analytics

Causal analysis asks what would happen under an intervention. The fundamental challenge is that, for the same unit, the outcome under treatment and the outcome without treatment cannot normally be observed simultaneously. Randomized experiments address this by assigning treatment independently of potential outcomes, while observational methods require assumptions about confounding, selection, measurement, and interference.

A predictive model can identify high-risk individuals without identifying who benefits from an intervention. For example, a credit customer may be likely to default but not responsive to a reminder. Prescriptive systems therefore require more than predictive accuracy: they require credible estimates of action consequences and safeguards against harmful recommendations.

4. Lifecycle Frameworks

No lifecycle framework is a rigid sequence. Real projects iterate because new evidence changes the question, data understanding, or feasibility assessment.

4.1 CRISP-DM

CRISP-DM—Cross-Industry Standard Process for Data Mining—contains six interconnected phases:

1. Business understanding: Define objectives, resources, risks, stakeholders, and success criteria.
2. Data understanding: Collect initial data, describe it, explore quality, and identify problems.
3. Data preparation: Select, clean, integrate, construct, and format data.
4. Modelling: Select techniques, design tests, train models, and tune parameters.
5. Evaluation: Assess whether results satisfy the business objective and whether the model is valid and acceptable.
6. Deployment: Deliver reports, scoring services, workflows, or other usable outputs; monitor them after release.

CRISP-DM’s strength is its accessible project structure and explicit recognition that deployment is part of the lifecycle. Its limitation is that it may underemphasize user research, causal identification, continuous monitoring, and modern machine-learning operations unless the team adds them deliberately.

4.2 OSEMN

OSEMN emphasizes practical data work:

• Obtain: Acquire data lawfully and document provenance.
• Scrub: Clean errors, duplicates, missingness, inconsistent formats, and implausible values.
• Explore: Examine distributions, relationships, subgroups, and anomalies.
• Model: Build statistical or machine-learning models.
• iNterpret: Explain findings, uncertainty, limitations, and implications.

OSEMN is memorable and useful for hands-on analysis, but it gives less explicit attention to organizational change, stakeholder discovery, and deployment governance than a broader lifecycle model.

4.3 Design Thinking

Design thinking contributes a human-centred perspective:

1. Empathize: Understand users, affected communities, and work practices.
2. Define: Synthesize observations into a specific problem statement.
3. Ideate: Generate alternative interventions and analytical approaches.
4. Prototype: Create a low-cost representation of a dashboard, workflow, or model output.
5. Test: Obtain feedback, observe use, and revise.

Design thinking is especially important when the analytical output changes employee behaviour, customer treatment, or public services. A technically strong system may fail because users do not trust it, cannot interpret it, or have no time to act on it.

4.4 Scientific Research Methods

Scientific reasoning contributes disciplined inquiry:

1. formulate a question;
2. review prior evidence and theory;
3. specify hypotheses and constructs;
4. design a study or experiment;
5. collect and analyze evidence;
6. assess uncertainty and threats to validity;
7. communicate, replicate, and revise.

Important validity types include:

• Construct validity: Does the measurement represent the concept?
• Internal validity: Is the observed relationship credible within the study?
• External validity: Does it generalize to other settings or populations?
• Statistical conclusion validity: Are the analysis and uncertainty assessment appropriate?

Scientific methods discourage post hoc storytelling. If a team searches many variables and reports only the strongest relationship, the apparent finding may be a chance result. Pre-specification, holdout evaluation, replication, and transparent reporting reduce this risk.

4.5 Integrating the Frameworks

The frameworks are complementary rather than competing. A mature project may use design thinking to understand the decision context, CRISP-DM to manage the overall project, OSEMN to organize data work, and scientific methods to evaluate claims.

5. Success Criteria and Evaluation Design

Success must be defined at multiple levels:

• Business success: revenue, cost, service quality, safety, retention, or social outcome.
• Analytical success: accuracy, calibration, recall, forecast error, interpretability, or causal precision.
• Operational success: latency, adoption, workflow completion, intervention capacity, and reliability.
• Responsible-use success: fairness, privacy, security, accessibility, and absence of unacceptable harm.

A model metric is not a substitute for a business outcome. Suppose a retention model has high recall but produces more alerts than the service team can handle. Its operational value may be low. A capacity constraint should be included in the decision design, perhaps by ranking customers and selecting only the top available cases.

For a binary outcome, accuracy is:

Accuracy can be misleading when classes are imbalanced. If only 2% of transactions are fraudulent, a model that labels every transaction legitimate achieves 98% accuracy but detects no fraud. Metrics should be selected according to the decision costs, prevalence, and intervention process.

For a continuous prediction, mean absolute error is:

Here, yᵢ is the observed value, ŷᵢ is the prediction, and n is the number of cases. MAE has the same units as the target and treats each absolute error linearly. It is applicable when the target is numerical and errors can reasonably be averaged.

Evaluation should reflect deployment. A random train-test split may be inappropriate for time-dependent data because it can allow future patterns to influence evaluation. Temporal validation, group-aware validation, or prospective testing may be necessary. The test set should remain separate until model choices are substantially complete.

6. Detailed Illustrative Case: Reducing Missed Outpatient Appointments

This is a fictional illustrative case designed for educational purposes.

Background

A regional outpatient clinic reports that missed appointments reduce access, increase administrative costs, and lengthen waiting times. Leadership asks the data science team to “build a no-show model.”

Problem Framing

The team interviews patients, reception staff, clinicians, and scheduling managers. They discover that the real decision is not simply prediction. Staff need a daily list of appointments for which a reminder, transport assistance, or rescheduling call may be useful. The clinic has limited calling capacity and does not want a model to penalize patients by restricting access.

The project charter states:

• Population: appointments scheduled at the clinic during the next six months.
• Unit: one scheduled appointment.
• Observation point: 48 hours before the appointment.
• Target: whether the appointment is missed without timely cancellation.
• Decision: which appointments should receive an additional approved reminder or support action.
• Constraints: no use of protected attributes for punitive decisions; privacy controls; minimum accessibility standards; staff capacity of 200 calls per day.
• Business outcome: reduce avoidable missed appointments without increasing inequitable treatment.

Analytical Questions and Hypotheses

The team asks:

1. What is the historical no-show rate by clinic, appointment type, lead time, and reminder status?
2. Which information available 48 hours before an appointment predicts a no-show?
3. Does an additional reminder reduce no-shows, and for which patients?
4. Can the clinic act on the predictions within its staffing capacity?

A predictive hypothesis is: “Longer scheduling lead time and previous missed appointments are associated with increased no-show probability.” A causal hypothesis is: “For eligible appointments, an additional reminder reduces the probability of a no-show compared with usual care.” These require different designs. The first may be studied with historical data; the second is best evaluated through a randomized or carefully designed quasi-experimental study.

Data Understanding and Preparation

Potential sources include appointment records, reminder logs, scheduling changes, and attendance records. The team checks whether a “no-show” is consistently recorded and whether cancellations are incorrectly classified. It identifies missing phone numbers, duplicate appointments, changes made after the prediction point, and variables that indirectly encode sensitive information.

The team creates an appointment-level dataset. Features are restricted to information available at 48 hours. A patient’s eventual attendance status is used only as the outcome. The data is split temporally: earlier appointments support training and validation, while later appointments form a final test period.

Modelling and Evaluation

A transparent baseline, such as a regularized logistic regression, is compared with a tree-based model. The baseline is valuable because it is interpretable and provides a reference point. Evaluation includes discrimination, calibration, subgroup performance, and operational capacity.

Suppose the clinic’s test set contains 1,000 appointments, of which 100 are no-shows. A selected threshold produces:

• TP = 60
• FP = 140
• TN = 760
• FN = 40

Then:

The model identifies 60% of no-shows, but only 30% of flagged appointments are no-shows. Whether this is acceptable depends on the cost and benefit of the reminder. If staff can make only 200 calls, this threshold uses the full capacity. If reminders are inexpensive and automated, a different threshold may be appropriate.

Causal Evaluation

The clinic randomly assigns eligible appointments to usual care or an additional reminder, subject to ethical and operational review. The primary outcome is the no-show rate. The estimated absolute treatment effect can be expressed as:

Here, T = 1 denotes the intervention group, T = 0 the comparison group, and Y the outcome. If Y is coded as 1 for a no-show, a negative ATE indicates that the intervention reduces no-shows. Randomization supports a causal interpretation when assignment is properly implemented, interference is limited, and outcome measurement is consistent.

The team may later estimate heterogeneous treatment effects, but it must avoid presenting unstable subgroup differences as established facts. Small groups require uncertainty intervals and cautious interpretation.

Deployment and Monitoring

The final system sends a ranked list to staff rather than automatically cancelling or deprioritizing appointments. Staff record whether the recommended action was completed. Monitoring includes data freshness, missingness, calibration, alert volume, attendance outcomes, subgroup disparities, and changes in scheduling policy.

Outcome and Learning Points

After a pilot, the clinic observes fewer missed appointments among the intervention group, but the effect is concentrated in patients who receive accessible reminders in their preferred communication channel. The project therefore changes the workflow: prediction identifies cases for consideration, while causal evidence guides which intervention to offer.

The case demonstrates that:

• the business problem is broader than model construction;
• prediction and intervention effectiveness are different questions;
• capacity constraints affect threshold selection;
• operational definitions determine the target;
• fairness and privacy must be addressed during framing, not after deployment; and
• monitoring is necessary because populations, policies, and behaviour change.

7. Challenges, Ethics, and Governance

Ambiguous Objectives

Different stakeholders may use the same words to mean different outcomes. “Efficiency” may mean lower cost to finance, shorter waiting time to patients, or reduced workload to staff. A written decision log and agreed success criteria prevent silent changes in scope.

Measurement Error and Proxy Variables

Administrative data is collected for operational purposes, not necessarily research. Outcomes may be missing, inconsistently recorded, or affected by prior decisions. Features such as postcode, language, device type, or purchasing channel may act as proxies for protected characteristics. Removing a protected variable does not guarantee fairness if correlated proxies remain.

Fairness

Fairness is context-dependent. Possible assessments include differences in false-positive rates, false-negative rates, calibration, selection rates, or error costs across groups. These criteria can conflict mathematically and operationally, especially when base rates differ. Teams should identify the harm being prevented, consult affected groups, document trade-offs, and provide appeal or review mechanisms where appropriate.

Privacy and Security

Problem framing should apply data minimization: collect and use only information necessary for the legitimate purpose. Access controls, retention limits, encryption, audit logs, de-identification where suitable, and secure handling of outputs are part of the lifecycle. Privacy law and sector-specific regulation may impose additional requirements concerning lawful basis, consent, automated decision-making, and data subject rights.

Reproducibility and Accountability

A defensible project maintains versioned data definitions, code, model artifacts, evaluation results, assumptions, and decision records. Reproducibility does not mean that every system must be open to the public; it means that authorized reviewers can understand how conclusions were produced.

Drift and Feedback Loops

Once a model influences decisions, its outputs can change future data. A fraud model may cause investigators to examine certain cases more often, creating a biased record of detected fraud. A hiring model may reproduce historical hiring patterns. Monitoring should therefore examine both technical drift and social consequences.

8. Contemporary Practice and Best Practices

Modern data science increasingly treats the lifecycle as a socio-technical system. Cloud platforms, automated pipelines, foundation models, and real-time analytics can accelerate experimentation, but they do not remove the need for problem definition or governance. Generative AI may assist with brainstorming, coding, documentation, and interface design, yet generated content must be checked for fabricated assumptions, privacy leakage, security weaknesses, and unsupported claims.

Recommended practices include:

1. Begin with the decision and affected users, not with an available dataset.
2. Write a one-page project charter before substantial modelling.
3. Define the unit, time windows, target, intervention, and constraints explicitly.
4. Separate prediction, explanation, and causal claims.
5. Establish a simple baseline before using complex models.
6. Match validation to time, groups, deployment conditions, and data-generating processes.
7. Evaluate technical, business, operational, and responsible-use outcomes together.
8. Document assumptions, exclusions, missingness, leakage checks, and uncertainty.
9. Prototype the workflow and obtain user feedback before full deployment.
10. Monitor performance, data quality, fairness, security, and unintended effects after release.
11. Establish a stop, review, or rollback procedure when risks exceed acceptable limits.
12. Treat the lifecycle as iterative: evidence may require reframing the problem.

Chapter Summary

Problem framing transforms a broad organizational concern into a precise decision problem with defined users, populations, time windows, outcomes, hypotheses, constraints, and success criteria. The unit of analysis and operational definitions are foundational because they determine what the data represents and whether a model can be used at decision time.

Descriptive analytics summarizes what happened; diagnostic analytics investigates patterns and possible explanations; predictive analytics estimates future or unknown outcomes; and prescriptive analytics recommends actions under constraints. Causal analysis addresses the different question of what would happen under an intervention. Prediction alone does not establish causation or identify the most effective treatment.

CRISP-DM organizes business understanding, data understanding, preparation, modelling, evaluation, and deployment. OSEMN emphasizes obtaining, scrubbing, exploring, modelling, and interpreting data. Design thinking centres user needs and workflow adoption, while scientific research methods strengthen hypothesis formation, validity, uncertainty assessment, and reproducibility. These approaches are best integrated rather than used mechanically.

A successful data science project must satisfy more than a model metric. It should create measurable organizational value, fit operational capacity, respect privacy and fairness, and remain reliable under changing conditions. Governance, documentation, monitoring, and human accountability are therefore integral parts of the data science lifecycle.

Key Terms

Problem framing: Defining an organizational challenge as a precise, measurable decision and analytical problem.

Decision problem: A choice or action that analysis is intended to improve.

Unit of analysis: The entity represented by one observation in a dataset.

Observation window: The period during which information used as features is collected.

Outcome window: The period during which the target outcome is measured.

Target leakage: Use of information unavailable at prediction time or created after the outcome process begins.

Operational definition: A reproducible specification of how an abstract concept is measured.

Hypothesis: A testable proposition about a relationship, difference, or intervention effect.

Descriptive analytics: Analysis that summarizes observed events or conditions.

Diagnostic analytics: Analysis that investigates patterns and possible explanations.

Predictive analytics: Estimation of an unknown or future outcome.

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

Causal inference: Estimation of outcomes attributable to an intervention or exposure.

CRISP-DM: An iterative lifecycle comprising business understanding, data understanding, preparation, modelling, evaluation, and deployment.

OSEMN: A workflow of obtaining, scrubbing, exploring, modelling, and interpreting data.

Design thinking: A human-centred approach involving empathy, definition, ideation, prototyping, and testing.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Concept drift: Change in the relationship between inputs and outcomes over time.

Data governance: Policies and practices governing data quality, access, security, use, and accountability.

Review Questions

Short Answer Questions

1. Why is “reduce customer churn” an incomplete analytical problem statement?
2. Define the unit of analysis and explain why it matters.
3. Distinguish predictive analytics from prescriptive analytics.
4. What is target leakage, and why does it make evaluation misleading?
5. Name the six phases of CRISP-DM.

Long Answer Questions

1. Compare CRISP-DM, OSEMN, design thinking, and scientific research methods. Explain how they can be integrated in one project.
2. Explain the difference between prediction, association, and causation, using an organizational example.
3. Discuss how business, analytical, operational, and responsible-use success criteria should be combined.
4. Explain why time windows, operational definitions, and data dictionaries are essential to reliable analytics.
5. Discuss the ethical and governance issues that should be addressed before deploying a high-impact predictive system.

Analytical / Application Questions

1. A university wants to “improve student success.” Frame this as a decision problem, define a unit of analysis and outcome window, propose two hypotheses, and identify at least three risks.
2. A fraud model detects 90% of confirmed fraud but flags 25% of all transactions. Design an evaluation and deployment approach that considers investigation capacity, false positives, calibration, and fairness.
3. A hospital predicts which patients will be readmitted but wants to know whom to contact with a support intervention. Explain why a predictive model alone is insufficient and propose a causal evaluation design.

Multiple Choice Questions with Answers

Question 1

Which statement best represents a well-framed analytical problem?

A. Improve customer satisfaction using artificial intelligence.

B. Build the most accurate model possible from all available data.

C. Identify eligible customers likely to cancel within 30 days so the service team can offer an approved intervention within its capacity.

D. Analyze customer data to find interesting patterns.

Correct Answer: C. Identify eligible customers likely to cancel within 30 days so the service team can offer an approved intervention within its capacity.

Question 2

What is the primary purpose of defining the unit of analysis?

A. To select the programming language

B. To specify what each observation represents

C. To guarantee causal validity

D. To eliminate missing values

Correct Answer: B. To specify what each observation represents

Question 3

Which activity belongs most directly to the CRISP-DM evaluation phase?

A. Agreeing on the organizational objective

B. Checking whether results satisfy business and analytical requirements

C. Writing database extraction queries only

D. Deploying a model without monitoring

Correct Answer: B. Checking whether results satisfy business and analytical requirements

Question 4

Which question is predictive?

A. How many claims were submitted last quarter?

B. Why did delivery delays increase last month?

C. Which shipments are likely to be delayed next week?

D. Which routing policy maximizes service under a fuel constraint?

Correct Answer: C. Which shipments are likely to be delayed next week?

Question 5

A model produces TP = 60 and FP = 140. What is its precision?

A. 0.20

B. 0.30

C. 0.60

D. 0.70

Correct Answer: B. 0.30

Question 6

Why is using a cancellation survey completed after account closure a problem in an early-warning model?

A. It reduces the number of features to zero

B. It creates target leakage

C. It guarantees underfitting

D. It converts a classification problem into clustering

Correct Answer: B. It creates target leakage

Question 7

Which framework most explicitly emphasizes empathy, prototyping, and user testing?

A. OSEMN

B. CRISP-DM

C. Design thinking

D. Confusion-matrix analysis

Correct Answer: C. Design thinking

Question 8

What does recall measure in binary classification?

A. The proportion of predicted positives that are correct

B. The proportion of actual positives that are identified

C. The proportion of all predictions that are correct

D. The proportion of actual negatives that are identified

Correct Answer: B. The proportion of actual positives that are identified

Question 9

Which statement about a predictive model is correct?

A. High predictive accuracy automatically proves causation.

B. A model can identify high-risk cases without identifying who benefits from treatment.

C. Removing protected attributes guarantees fairness.

D. A model metric is always more important than operational capacity.

Correct Answer: B. A model can identify high-risk cases without identifying who benefits from treatment.

Question 10

In the appointment case, why would a randomized reminder study be useful?

A. It estimates whether the reminder itself changes no-show outcomes under the study conditions.

B. It guarantees that the prediction model has perfect calibration.

C. It removes the need to define the outcome.

D. It ensures every patient receives the intervention.

Correct Answer: A. It estimates whether the reminder itself changes no-show outcomes under the study conditions.