Visualization, Presentation, and Decision Communication

Introduction

Artificial intelligence and machine-learning projects produce evidence in many forms: experimental results, error analyses, fairness measurements, model predictions, dashboards, technical reports, and policy recommendations. However, evidence has value only when intended audiences can understand it, evaluate its quality, and use it appropriately. Visualization and presentation are therefore not merely matters of appearance. They are research and professional practices that influence interpretation, trust, resource allocation, and public decisions.

A technically accurate graph may still mislead if its axes are truncated, its denominator is hidden, or uncertainty is omitted. A well-performing model may still be unsuitable for deployment if its limitations, data dependencies, and consequences are not communicated. Similarly, a dashboard can increase awareness while weakening decision quality if it displays too many indicators without explaining what action each indicator supports.

This chapter examines how to transform research evidence into tables, charts, model cards, dashboards, posters, and presentations. It also considers uncertainty, limitations, audience adaptation, ethical communication, and decision support. The central principle is that communication should preserve the meaning of evidence while making its implications usable for a particular audience.

Learning Objectives

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

1. Select tables and visualizations that match research questions and data types.

2. Design accurate and accessible charts, dashboards, posters, and presentations.

3. Communicate uncertainty, assumptions, limitations, and model performance responsibly.

4. Adapt evidence for technical, managerial, policy, and public audiences.

5. Use model cards and decision-oriented communication frameworks.

6. Critically evaluate misleading or ineffective visual communication.

7. Support evidence-based decisions without overstating what research demonstrates.

1. Communication as Part of the Research Process

Research communication is often treated as the final stage of a project, but communication decisions begin much earlier. The choice of outcome measure, comparison group, evaluation period, and reporting format affects what readers can infer. A researcher who records only average accuracy, for example, may be unable to communicate subgroup disparities later.

A useful communication chain is:

Data → Analysis → Evidence claim → Audience interpretation → Decision or action

Errors can occur at every stage. Data may be incomplete; analysis may use an unsuitable metric; the claim may exceed the evidence; the audience may misunderstand probability; or a decision-maker may apply a result outside its intended context. Effective communication reduces these risks by making definitions, assumptions, uncertainty, and scope explicit.

A communication artifact should answer four questions:

• What is being shown?
• How was it measured or produced?
• How certain or limited is the evidence?
• What decision, if any, should the audience consider?

The last question must not be confused with advocacy. Evidence-based communication can recommend an action, but it should distinguish observed results from interpretation, prediction, and value judgment.

2. Choosing the Right Representation

2.1 Tables

Tables are appropriate when readers need exact values, comparisons across several dimensions, or detailed methodological information. A table should have a descriptive title, clearly labeled columns, units, an identifiable population or sample, and notes for abbreviations and statistical conventions.

A table of model results should normally identify the data split, evaluation metric, number of observations, and whether values are averages across repeated runs. Reporting “accuracy = 0.91” without stating the dataset or class distribution is incomplete.

Good practice includes:

• ordering rows according to a meaningful rule, such as baseline followed by improved models;
• using consistent decimal precision;
• reporting sample sizes;
• avoiding excessive precision that implies unjustified certainty;
• explaining missing values;
• separating absolute values from percentages;
• using footnotes for definitions rather than forcing long text into cells.

Tables should not imitate charts unnecessarily. If the purpose is to reveal a trend, a chart may be more effective; if the purpose is to support exact lookup, a table is usually preferable.

2.2 Charts

The chart type should follow the analytical relationship.

• Bar charts compare discrete categories. Bars should normally begin at zero when length represents magnitude.
• Line charts show change across an ordered sequence, especially time. Connecting unrelated categories with lines can imply a trend that does not exist.
• Scatter plots show association between two quantitative variables and can reveal clusters, outliers, and non-linear patterns.
• Histograms show the distribution of a quantitative variable using contiguous intervals. They are different from bar charts because the bins represent ranges.
• Box plots summarize distributions using quartiles, median, and possible outliers, but require explanation for audiences unfamiliar with them.
• Heat maps encode values through color and are useful for matrices such as confusion matrices or correlation displays. Color scales must be perceptually ordered and labeled.
• Maps display geographic patterns but may exaggerate areas or obscure population denominators. Counts and rates should not be confused.
• Network diagrams display relationships, but they become difficult to interpret when there are too many nodes or edges.

A visualization encodes data through visual variables such as position, length, angle, area, color, and shape. Position along a common scale is generally easier to compare accurately than area, volume, or decorative perspective. Consequently, a simple dot plot may communicate differences more precisely than a three-dimensional graphic.

2.3 Choosing a Chart from the Question

The research question should precede chart selection. “How did performance change across model versions?” suggests a line chart or grouped bar chart. “Which groups experience the highest false-negative rate?” suggests a sorted bar chart with uncertainty or sample sizes. “Are errors concentrated in particular regions of feature space?” suggests a scatter plot colored by error status.

A chart should have a visual purpose. If removing a graphical element does not reduce understanding, that element may be unnecessary. Decorative images, gradients, unnecessary gridlines, and three-dimensional effects can compete with the evidence.

3. Accuracy, Integrity, and Accessibility

3.1 Scale and Baseline

Axis choices strongly influence perception. Truncating a bar-chart axis can make a small difference appear large because bar length is interpreted relative to the baseline. Truncation is not always forbidden—for example, line charts may use a focused scale to show small changes—but it must be clearly indicated and should not distort the comparison.

The scale must also be appropriate to the measurement. A percentage axis should not be labeled as a proportion, and a logarithmic scale should be identified because equal distances represent multiplicative rather than additive changes.

3.2 Denominators and Normalization

Counts can be misleading when group sizes differ. Suppose a system produces 100 errors for Group A and 50 errors for Group B. If Group A contains 10,000 cases and Group B contains 1,000, the error rates are 1% and 5%, respectively. The raw count is larger for Group A, but the rate is higher for Group B. A responsible visualization either shows both count and rate or explains why one is being used.

Rates should include their denominators, particularly in health, safety, education, and demographic applications. A dashboard displaying “complaints increased by 40%” should indicate the baseline count, observation period, and whether the increase reflects more users, changed reporting behavior, or a genuine change in incidence.

3.3 Color and Accessibility

Color should not be the only channel used to distinguish categories. Designs should remain interpretable in grayscale and by people with color-vision deficiencies. Labels, patterns, line styles, or direct annotation can supplement color.

Sequential color scales are suitable for ordered magnitudes; diverging scales are suitable when values depart from a meaningful midpoint; qualitative palettes are suitable for categories without an inherent order. Red-green combinations are often problematic. Text and important graphical elements should have sufficient contrast, and digital artifacts should include alternative text or an accessible textual description.

3.4 Uncertainty and Variation

A point estimate is not the whole result. Uncertainty may arise from sampling, measurement error, model instability, missing data, or future variation. It can be communicated using confidence intervals, prediction intervals, credible intervals, error bars, distributions, or scenario ranges, provided the chosen representation is defined.

A confidence interval is an interval procedure with a repeated-sampling interpretation. A 95% confidence interval does not mean that there is a 95% probability that a fixed parameter lies inside this particular interval under the classical interpretation. Rather, if the procedure were repeated under its assumptions, approximately 95% of intervals would contain the parameter. In applied communication, the interval should be accompanied by a plain-language explanation and the method used to calculate it.

A prediction interval concerns a future observation and is often wider than an interval for a mean because it includes individual-level variation. These concepts should not be used interchangeably.

For machine-learning evaluation, uncertainty can be estimated through repeated resampling, cross-validation, or bootstrap procedures. The method, number of repetitions, and dependence among observations should be reported. An error bar without a definition can create false confidence.

4. Communicating Machine-Learning Evidence

4.1 Metrics and Context

No single metric adequately describes every model. Classification reports may include accuracy, precision, recall, F1 score, specificity, area under a receiver operating characteristic curve, or area under a precision-recall curve. The appropriate metric depends on class prevalence, error costs, and the decision being supported.

For a binary classifier:

• Precision = TP / (TP + FP), when the denominator is non-zero.
• Recall or sensitivity = TP / (TP + FN), when the denominator is non-zero.
• Specificity = TN / (TN + FP), when the denominator is non-zero.
• Accuracy = (TP + TN) / (TP + TN + FP + FN), when the total is non-zero.

Here TP, FP, TN, and FN denote true positives, false positives, true negatives, and false negatives. These metrics assume a defined binary outcome and a specified classification threshold. They can change when prevalence, threshold, or data population changes.

A model report should therefore identify the threshold, evaluation population, outcome definition, time period, missing-data treatment, and baseline. A high accuracy score may be uninformative for a highly imbalanced dataset. A confusion matrix often communicates operational consequences more clearly than a single score.

4.2 Calibration and Uncertainty

If a model outputs probabilities, calibration matters. Among cases assigned a predicted probability near 0.7, approximately 70% should experience the outcome for the model to be well calibrated in that population and context. Discrimination and calibration are different: a model can rank cases effectively while producing probabilities that are systematically too high or too low.

Communicators should distinguish:

• aleatory uncertainty, arising from inherent variability or randomness;
• epistemic uncertainty, arising from limited knowledge, data, or model specification;
• distribution shift, occurring when deployment data differ from development data.

These distinctions affect action. More data may reduce some epistemic uncertainty, but it cannot eliminate inherent variability. Monitoring may be necessary when population or operational conditions change.

4.3 Model Cards

A model card is a structured document describing a model’s intended use, performance, limitations, and evaluation conditions. A useful model card includes:

• model identity, version, and development organization;
• intended uses and foreseeable misuse;
• training data characteristics and important exclusions;
• evaluation datasets and metrics;
• subgroup performance where justified and ethically permissible;
• known limitations, failure modes, and out-of-distribution risks;
• human oversight and recommended operating procedures;
• privacy, security, safety, and fairness considerations;
• maintenance, update, and contact information.

A model card is not a guarantee of safety or fairness. It is an accountability and communication instrument. Its usefulness depends on truthful disclosure, meaningful evaluation, and governance beyond documentation.

5. Dashboards and Decision Communication

A dashboard is an interactive display designed to support monitoring, exploration, or action. Dashboards should not merely collect every available metric. They should be built around decisions and user tasks.

A practical dashboard design process is:

1. Identify the decisions and users.

2. Specify the indicators that inform those decisions.

3. Define each indicator, denominator, update frequency, and data source.

4. Set meaningful reference values, targets, or alert thresholds.

5. Display current status and relevant historical context.

6. Provide drill-downs without hiding the headline message.

7. Show data quality and uncertainty.

8. Test the dashboard with representative users.

A monitoring dashboard may emphasize current values, alerts, and changes since the previous period. An analytical dashboard may support filtering and comparison. An executive dashboard may summarize strategic indicators, risks, and recommended decisions. These purposes should not be mixed without clear navigation.

Alert thresholds should be based on operational or statistical reasoning, not arbitrary colors. A red indicator should mean that a defined response is warranted. Otherwise, repeated alerts can produce alert fatigue. Dashboards should also show when data were last updated and whether a missing feed makes the apparent status unreliable.

6. Posters and Presentations

6.1 Research Posters

A poster is a constrained communication environment. Its purpose is to attract attention, support a short explanation, and make the central contribution memorable. A strong poster has a clear title, research question, concise methods, selected results, limitations, conclusion, and contact or reference information.

The visual hierarchy should guide reading from the main claim to supporting evidence. Paragraphs should be short, figures should be large enough to read at viewing distance, and every figure should have a direct caption. A poster should not reproduce an entire paper in small type. The author should be able to state the contribution in one or two sentences.

6.2 Oral Presentations

A technical presentation should have a narrative structure:

1. Why the problem matters.

2. What question was investigated.

3. How the study or system addressed it.

4. What evidence was obtained.

5. What the evidence does and does not support.

6. What action, research, or monitoring is appropriate.

Slides are visual aids, not a manuscript displayed on a screen. Each slide should have one dominant message. Charts should be simplified for projection, labels should be readable, and animations should serve explanation rather than decoration. Presenters should explain unfamiliar terms and anticipate questions about data quality, baselines, generalizability, and limitations.

7. Audience Adaptation

The same evidence may require different emphasis for different audiences, but adaptation must not change the underlying facts.

Technical Audience

Technical readers need methods, data provenance, evaluation design, baselines, uncertainty, reproducibility details, and failure analysis. They may require equations, model architecture, hyperparameters, and code or data availability statements.

Managerial Audience

Managers usually need implications for cost, performance, risk, workflow, resources, and implementation time. Technical detail should be translated into operational consequences. A manager may need to know that a model reduces manual review by an estimated amount, but also what human oversight and retraining costs remain.

Policy Audience

Policy audiences require attention to legal authority, equity, accountability, affected populations, public interest, enforceability, and unintended consequences. Evidence should be accompanied by assumptions and alternative scenarios rather than presented as an automatic policy command.

Public Audience

Public communication should use plain language, explain uncertainty without condescension, identify who may be affected, and avoid sensational claims. Percentages should be paired with understandable frequencies where helpful. Technical vocabulary should be defined or replaced.

Audience adaptation can be represented as a translation matrix:

| Evidence element | Technical emphasis | Managerial emphasis | Policy emphasis | Public emphasis |

|---|---|---|---|---|

| Performance | Metric definitions and validation | Operational benefit | Reliability and accountability | What the system can and cannot do |

| Risk | Error analysis and uncertainty | Cost and mitigation | Rights, equity, and governance | Potential harm and safeguards |

| Deployment | Architecture and monitoring | Workflow and resources | Oversight and compliance | Human control and recourse |

8. From Evidence to Decision

A decision-oriented communication should distinguish four layers:

1. Observation: What was measured?

2. Inference: What conclusion is justified under the study assumptions?

3. Recommendation: What action is proposed?

4. Value judgment: Why is that action desirable or acceptable?

For example, an evaluation may observe that a classifier has lower false-negative rates than a baseline. It may infer that the classifier improves detection in the evaluated sample. A recommendation may be to conduct a limited pilot with human review. Whether the pilot is acceptable may depend on ethical, legal, financial, and social values beyond the metric itself.

A useful decision brief identifies the decision owner, available options, evidence for each option, uncertainty, consequences, reversibility, and monitoring plan. It should also state what new information would change the recommendation.

Evidence should not be presented as more causal than the design permits. An observational association may inform a decision, but it does not automatically demonstrate that changing one variable will produce the observed outcome. Randomized experiments, quasi-experiments, simulations, expert judgment, and descriptive studies answer different questions.

9. Worked Example: Evaluating an Automated Student-Risk Model

Consider an illustrative case in which a university develops a model to identify students who may benefit from academic support. The model is not intended to determine discipline or deny services.

Background and Problem

Advisers have limited time and want to prioritize outreach. The university has historical records on course performance, attendance, and prior support use. The project team proposes a binary model that flags students for optional advising.

Analysis

The team evaluates the model on a future-semester holdout set rather than only on training data. Because students who need support represent a minority of the population, accuracy alone is inadequate. The team reports precision, recall, the confusion matrix, subgroup results, and calibration. It also compares the model with a simple baseline based on prior academic performance.

Suppose the holdout set contains 1,000 students, of whom 200 later request or are assessed as needing support. At the selected threshold, the model produces:

• TP = 140
• FN = 60
• FP = 160
• TN = 640

The total is 1,000. Precision is 140 / (140 + 160) = 140 / 300 = 0.467, or 46.7%. Recall is 140 / (140 + 60) = 140 / 200 = 0.70, or 70%. Accuracy is (140 + 640) / 1,000 = 0.78, or 78%. The results show that the model identifies many students needing support, but more than half of flagged students do not fall into the measured outcome group. That does not necessarily make the model useless: if outreach is low-cost and beneficial, the decision may tolerate false positives. If adviser capacity is very limited, precision and resource constraints become more important.

The team also discovers that recall is lower for students who study part-time and that the label “needed support” is partly based on whether a student previously contacted an adviser. This introduces a measurement problem: the label may reflect access and help-seeking behavior rather than need alone.

Approach and Solution

The university revises the project in five ways:

1. It defines the intended use as prioritizing optional outreach, not making high-stakes eligibility decisions.

2. It adds subgroup evaluation and investigates why performance differs.

3. It provides advisers with explanations, but warns that explanations are not proof of causation.

4. It establishes human review, student notification, and a process to correct inaccurate records.

5. It monitors outcomes after deployment, including whether outreach improves academic experience and whether some groups are disproportionately contacted or missed.

The model card states the data period, excluded populations, threshold, metrics, known label limitations, and prohibited uses. A dashboard displays the number flagged, contact rate, response rate, outcome measures, subgroup comparisons, and data freshness. It does not display a single “risk score” as if it were an objective measure of a student’s worth or ability.

Outcome

The model is approved only for a monitored pilot. After one term, the university finds that outreach increased adviser contacts but has not yet established a causal improvement in academic outcomes. The pilot is continued with a stronger evaluation design and a review of whether outreach itself is equally accessible.

Learning Points

The case demonstrates that communication is inseparable from responsible deployment. A confusion matrix is more informative than accuracy alone; labels require critical examination; subgroup results may reveal structural issues; and a model’s appropriate use depends on workflow, capacity, human oversight, and governance. The model card and dashboard support accountability, but neither replaces institutional responsibility.

10. Ethical, Legal, and Governance Considerations

Visual and written communication can cause harm even when numerical values are technically correct. Selective reporting may hide poor subgroup performance. A map may stigmatize a community. A dashboard may expose personal information. A risk score may be treated as a judgment rather than a probabilistic aid.

Important safeguards include:

• protecting privacy and applying data minimization;
• avoiding unnecessary personal or sensitive attributes in public displays;
• documenting consent, lawful basis, access controls, and retention where applicable;
• assessing fairness and disparate impact in context;
• providing meaningful human review and avenues for appeal;
• distinguishing anonymization from stronger privacy guarantees;
• recording version history and changes to definitions or metrics;
• retaining audit trails for consequential decisions;
• identifying who is accountable for errors and remediation.

Legal requirements vary by jurisdiction and application. Researchers should consult relevant institutional policies and laws rather than assuming that publication or visualization makes data use permissible. Ethical communication also requires respect for affected communities, especially when findings could increase stigma or surveillance.

11. Contemporary Developments and Best Practices

Interactive visualization, automated reporting, generative AI, and real-time data systems are changing communication practice. These technologies can improve accessibility and speed, but they introduce new risks. Generative systems may fabricate citations, alter chart values, omit caveats, or produce persuasive but unsupported summaries. Automated dashboards may silently change when data pipelines or definitions change.

Best practices include:

• preserve a reproducible link between source data, analysis, and artifact;
• validate automatically generated charts and text against authoritative results;
• version datasets, code, models, and dashboard definitions;
• maintain human review for consequential communication;
• record the provenance of every important number;
• use plain-language explanations alongside technical detail;
• test accessibility with users and assistive technologies;
• display uncertainty and data freshness;
• state limitations near the relevant claim, not only in an appendix;
• separate exploratory visualizations from decision-ready outputs;
• evaluate whether the communication changes decisions in the intended direction.

Challenges and Limitations

No visualization can eliminate ambiguity in imperfect evidence. Simplification is necessary, but simplification can hide heterogeneity. Interactive tools may allow exploration while making it difficult to reproduce the exact view used in a decision. Uncertainty intervals may be misunderstood, and subgroup analysis may be statistically unstable when sample sizes are small. A compelling presentation can cause audiences to overvalue confidence, fluency, or visual polish.

Communication also involves trade-offs. Complete disclosure may overwhelm a public audience; excessive simplification may become misleading. A dashboard designed for rapid action may not support causal interpretation. A model card can improve transparency but may become a compliance document that few people read. These limitations reinforce the need for audience testing, layered communication, and accountable decision processes.

Chapter Summary

Visualization and presentation are integral parts of research methodology because they shape how evidence is interpreted and used. Tables support precise lookup, while charts reveal comparisons, distributions, relationships, and trends. Chart selection should follow the research question, and design should preserve accurate scales, denominators, units, accessibility, and context.

Machine-learning communication requires more than reporting a single performance score. Confusion matrices, metric definitions, calibration, subgroup evaluation, uncertainty, baselines, and failure modes provide a more realistic account of model behavior. Model cards organize intended uses, data, performance, limitations, and governance information.

Dashboards should be designed around decisions and user tasks, with clear definitions, thresholds, data freshness, uncertainty, and monitoring plans. Posters and presentations should communicate one central contribution through a clear narrative and readable evidence. Technical, managerial, policy, and public audiences require different levels of detail, but adaptation must not alter the underlying evidence.

Responsible decision communication separates observation, inference, recommendation, and value judgment. It acknowledges uncertainty, avoids causal overclaiming, protects privacy, considers fairness, and identifies accountability. Effective communication is therefore not decoration added after analysis; it is a disciplined method for making evidence understandable, honest, and useful.

Key Terms

Visualization: A graphical representation designed to communicate data or evidence.

Dashboard: An interactive or static display that supports monitoring, exploration, or decisions.

Model card: A document describing a model’s intended uses, performance, limitations, and evaluation conditions.

Confusion matrix: A table showing true and predicted classes, including TP, FP, TN, and FN.

Precision: The proportion of predicted positives that are true positives.

Recall: The proportion of actual positives that are correctly identified.

Specificity: The proportion of actual negatives that are correctly identified.

Calibration: Agreement between predicted probabilities and observed outcome frequencies.

Uncertainty: Imperfect knowledge about a quantity, estimate, prediction, or conclusion.

Confidence interval: An interval produced by a repeated-sampling procedure with a stated coverage interpretation.

Prediction interval: An interval intended to describe uncertainty for a future individual observation.

Denominator: The reference quantity used to calculate a rate or proportion.

Baseline: A simple or established method used for comparison.

Data provenance: Information about the origin, processing, and lineage of data.

Distribution shift: A change between development data and deployment data.

Accessibility: The extent to which an artifact can be used by people with diverse abilities and technologies.

Evidence claim: A statement whose strength and scope should be supported by research results.

Decision brief: A concise document connecting evidence, options, uncertainty, and recommended action.

Data freshness: Information about how recently displayed data were collected or updated.

Visual hierarchy: The ordering of visual elements to guide attention and interpretation.

Review Questions

Short Answer Questions

1. Why should chart selection begin with the research question rather than the software being used?

2. Distinguish precision from recall in binary classification.

3. What information should a model card normally contain?

4. Why are denominators important when comparing groups?

5. State two ways to improve visualization accessibility.

Long Answer Questions

1. Explain how a misleading axis, hidden denominator, and omitted uncertainty can each distort interpretation of a machine-learning result.

2. Describe a complete process for designing a decision-oriented dashboard.

3. Compare the communication needs of technical, managerial, policy, and public audiences.

4. Explain why accuracy alone may be inadequate for evaluating a classifier on an imbalanced dataset.

5. Discuss the ethical and governance responsibilities involved in communicating model performance.

Analytical / Application Questions

1. A classifier is evaluated on 2,000 cases: TP = 180, FN = 20, FP = 300, and TN = 1,500. Calculate precision, recall, and accuracy, and explain which metric a service manager might prioritize under limited review capacity.

2. A public dashboard reports that complaints rose from 80 to 120 while the number of registered users rose from 1,000 to 2,000. Explain why the count alone may be misleading and propose an appropriate visualization.

3. Design a communication plan for deploying a model that flags patients for optional preventive outreach. Identify the evidence to show, the limitations to disclose, the audiences to address, and the safeguards required before deployment.

Multiple Choice Questions with Answers

Question 1

Which chart is generally most appropriate for showing change in a metric across ordered time periods?

A. Line chart

B. Pie chart

C. Network diagram

D. Unordered qualitative color map

Correct Answer: A. Line chart

Question 2

A classifier has TP = 80 and FP = 20. What is its precision?

A. 20%

B. 50%

C. 80%

D. 100%

Correct Answer: C. 80%

Question 3

What is the main purpose of a model card?

A. To replace all model testing

B. To document intended use, performance, limitations, and evaluation conditions

C. To guarantee that a model is fair

D. To hide implementation details from users

Correct Answer: B. To document intended use, performance, limitations, and evaluation conditions

Question 4

Why can raw error counts be misleading when comparing two groups?

A. Counts never represent real observations

B. Groups may have different sizes, so rates may differ from counts

C. Rates cannot be calculated for groups

D. Charts cannot display counts

Correct Answer: B. Groups may have different sizes, so rates may differ from counts

Question 5

Which visual variable usually supports the most accurate comparison of quantitative values?

A. Three-dimensional volume

B. Decorative texture

C. Position along a common scale

D. Perspective shading

Correct Answer: C. Position along a common scale

Question 6

What does recall measure in binary classification?

A. The proportion of predicted positives that are correct

B. The proportion of actual positives that are correctly identified

C. The proportion of actual negatives that are correctly identified

D. The proportion of all cases that are correctly classified

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

Question 7

Which statement correctly distinguishes a prediction interval from a confidence interval for a mean?

A. A prediction interval concerns a future observation and is often wider

B. A prediction interval always has zero width

C. A confidence interval describes only individual future observations

D. The two terms always have identical interpretations

Correct Answer: A. A prediction interval concerns a future observation and is often wider

Question 8

What should a decision brief distinguish?

A. Observation, inference, recommendation, and value judgment

B. Color, font, animation, and layout only

C. Training data from all forms of evidence

D. Public audiences from all technical audiences

Correct Answer: A. Observation, inference, recommendation, and value judgment

Question 9

Which practice improves accessibility of a chart?

A. Using color as the only category distinction

B. Removing all labels

C. Adding direct labels or line styles alongside color

D. Using low-contrast text

Correct Answer: C. Adding direct labels or line styles alongside color

Question 10

Why is a baseline important in model evaluation?

A. It provides a comparison against a simple or established method

B. It guarantees deployment success

C. It eliminates the need for a test set

D. It makes uncertainty irrelevant

Correct Answer: A. It provides a comparison against a simple or established method