Exploratory Data Analysis and Uncertainty

Introduction

Exploratory Data Analysis (EDA) is the systematic use of summaries, visual displays, and interactive investigation to understand a dataset before formal modelling or decision-making. It is not merely the production of attractive charts. EDA is a reasoning process through which an analyst discovers structure, identifies data-quality problems, detects unusual observations, evaluates assumptions, generates hypotheses, and communicates what is known and what remains uncertain.

In postgraduate data visualization, EDA must be understood together with uncertainty. A visual pattern may result from a genuine phenomenon, sampling variation, measurement error, missing data, data-processing decisions, or random coincidence. Responsible visualization therefore represents not only central values but also variation, confidence, probability, missingness, and model uncertainty.

The central principle of this chapter is:

> A visualization should make the strength, limitations, and uncertainty of evidence as visible as the pattern itself.

EDA is exploratory rather than confirmatory. It can suggest hypotheses, but a pattern discovered by repeatedly examining many variables, groups, and chart types should not automatically be treated as statistically significant or causally meaningful.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Explain the purpose, workflow, and epistemic limits of exploratory data analysis.

2. Use visualizations for data profiling, distributional analysis, anomaly detection, and pattern discovery.

3. Distinguish variability, uncertainty, confidence intervals, prediction intervals, and probability.

4. Select appropriate visual encodings for uncertainty and missingness.

5. Identify misleading presentations of error bars, scales, smoothing, and incomplete data.

6. Use robust and model-based methods to investigate anomalies.

7. Generate and document hypotheses without overstating exploratory evidence.

8. Evaluate ethical, governance, and professional implications of uncertain visual evidence.

1. EDA as an Analytical Workflow

A useful EDA workflow is iterative rather than strictly linear.

1.1 Establish the analytical context

Before plotting, define the unit of analysis, population, time period, measurement process, and intended decision. A row may represent a customer, transaction, sensor reading, patient visit, or daily aggregate. Confusing these units can produce pseudo-replication: many rows may represent the same underlying subject.

The analyst should ask:

• What question motivated data collection?
• Which population does the dataset represent?
• Are observations independent, clustered, repeated, or time-dependent?
• What variables are measured directly and which are constructed?
• What decisions could be affected by the analysis?

1.2 Profile the dataset

Data profiling combines structural, statistical, and semantic inspection. It commonly includes:

• row and column counts;
• data types and valid ranges;
• unique-value counts and duplicate records;
• missing-value counts and patterns;
• minimum, maximum, mean, median, standard deviation, and quantiles;
• category frequencies and rare levels;
• timestamps, time zones, and temporal coverage;
• units of measurement and coding conventions.

A profile table is often the first visualization or analytical artifact. For each variable, it can show type, completeness, cardinality, representative values, and quality warnings. Profiling should precede interpretation because an apparently interesting relationship may be caused by duplicated records, unit conversion errors, or a change in data collection.

1.3 Inspect univariate structure

Univariate EDA examines one variable at a time. Histograms and density plots display quantitative distributions; box plots summarize quartiles and potential extreme values; empirical cumulative distribution functions show the proportion of observations below each value. Bar charts are appropriate for categorical frequencies, while ordered dot plots can reveal differences among categories with less visual clutter.

No single display is sufficient. Histogram appearance depends on bin width and origin. A very wide bin can conceal multimodality, whereas a very narrow bin can exaggerate random fluctuations. Kernel density estimates depend on bandwidth and can create or remove apparent modes. These are analytical choices, not neutral windows onto the data.

1.4 Examine relationships

For two quantitative variables, scatterplots reveal direction, form, spread, clusters, and influential observations. For quantitative and categorical variables, jittered dot plots, box plots, violin plots, or interval plots can compare groups. Time-series plots expose trend, seasonality, level changes, and serial dependence. Heat maps and association matrices can summarize many relationships, but they should be accompanied by careful attention to scale, missingness, and multiple comparisons.

For high-dimensional data, pair plots, projection methods, linked views, and coordinated multiple views are useful. However, dimensionality-reduction displays such as principal component analysis or t-distributed stochastic neighbour embedding are representations of transformed data. Distances and clusters in the projection may not have a simple interpretation in the original feature space.

1.5 Investigate, revise, and document

EDA is recursive. An unexpected pattern leads to a data check, a revised plot, a subgroup analysis, or a new question. Every important transformation should be recorded: filtering, imputation, aggregation, scaling, treatment of outliers, and choice of smoothing parameters. Reproducible notebooks, versioned data extracts, and chart metadata help distinguish an observed feature from an artefact of analysis.

2. Understanding Variation

Variation is the observed difference among values. It is not synonymous with error. A dataset may vary because individuals genuinely differ, a process fluctuates over time, instruments are imprecise, or samples differ from the population.

2.1 Measures of spread

For observations \(x_1, x_2,\ldots,x_n\), the arithmetic mean is

\[

\bar{x}=\frac{1}{n}\sum_{i=1}^{n}x_i.

\]

The sample variance is

\[

s^2=\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2,

\]

and the sample standard deviation is \(s=\sqrt{s^2}\). The denominator \(n-1\) is used when estimating the variance of a larger population from a sample under the usual independent-sampling framework. Standard deviation is expressed in the same units as the observations, whereas variance is expressed in squared units.

The interquartile range is

\[

IQR=Q_3-Q_1,

\]

where \(Q_1\) and \(Q_3\) are the first and third quartiles. The IQR is robust to extreme values and is particularly useful for skewed distributions. The coefficient of variation, \(s/\bar{x}\), is sometimes used to compare relative dispersion, but it is inappropriate when the mean is near zero, negative, or lacks a meaningful ratio interpretation.

2.2 Distributional shape

Important features include symmetry, skewness, heavy tails, multimodality, boundedness, and zero inflation. A mean may be a poor summary for a strongly skewed distribution. For example, income, response time, and insurance claims frequently have long right tails; medians and quantiles may communicate the typical experience more honestly.

A box plot conventionally marks the median, quartiles, and possible observations beyond a rule based on 1.5 IQR from the quartiles. Such points are not automatically errors. The rule is a screening convention, not a proof that an observation is invalid.

2.3 Dependence and heterogeneity

Variation may differ across groups or levels of a predictor. This is heteroscedasticity. In a scatterplot, a funnel-shaped cloud indicates that the spread of the response changes with the explanatory variable. Aggregating data can hide this structure. A smooth average line may conceal subgroups with different trends, a phenomenon related to Simpson’s paradox. Analysts should therefore inspect stratified and individual-level displays before relying on aggregate patterns.

3. Anomaly Detection Through Visualization

An anomaly is an observation, event, or pattern that is unusual relative to a defined reference. “Unusual” must be specified: unusual globally, within a subgroup, at a particular time, or under a predictive model.

3.1 Types of anomalies

• Point anomaly: one observation is unusual relative to the rest.
• Contextual anomaly: an observation is unusual in a specific context, such as a temperature that is normal globally but abnormal for a particular season.
• Collective anomaly: a sequence or group is unusual even though individual values may appear ordinary.
• Data-quality anomaly: a value violates a known rule, such as an impossible date or negative age.

3.2 Visual tools

Time-series plots with reference bands can reveal sudden changes, missing intervals, and sensor drift. Scatterplots can expose isolated points, leverage, or clusters. Small multiples support comparison across sites or products. Control charts distinguish ordinary process variation from signals requiring investigation, but their limits depend on assumptions about the process and sampling scheme.

A practical workflow is:

1. Define the reference population and context.

2. Visualize the raw observations before deleting or winsorizing values.

3. Check units, timestamps, duplication, and data lineage.

4. Compare the observation with domain constraints and related variables.

5. Determine whether it represents error, a rare valid case, or a change in process.

6. Record the decision and perform sensitivity analysis.

3.3 Robust and model-based detection

The standardized score \(z_i=(x_i-\bar{x})/s\) is useful for approximately symmetric data without severe outliers, but the mean and standard deviation themselves can be distorted by extreme values. A robust alternative uses the median and median absolute deviation (MAD), defined as

\[

MAD=\operatorname{median}(|x_i-\operatorname{median}(x)|).

\]

A robust standardized score is commonly formed by scaling the difference from the median by MAD; the exact scaling convention should be stated because different software uses different constants.

Model-based anomaly scores can be derived from regression residuals, density estimates, clustering, or isolation algorithms. Such scores are conditional on the model, features, tuning choices, and training data. A model that labels a minority group as anomalous because it is underrepresented may encode sampling bias rather than operational risk.

4. Hypothesis Generation and the Limits of Exploration

EDA supports inductive reasoning: repeated observations may suggest a general explanation, but they do not guarantee it. A visible association may be confounded, non-linear, generated by selection, or produced by chance.

A sound exploratory hypothesis should identify:

• the population and variables involved;
• the expected direction or form of association;
• the relevant time and subgroup context;
• a plausible mechanism;
• a proposed confirmatory test or follow-up study.

Exploring many variables creates a multiple-comparisons problem. Even if every null hypothesis is true, examining enough relationships will produce apparently strong patterns by chance. Therefore, exploratory findings should be labelled as hypotheses, replicated on new data, or evaluated using a pre-specified confirmatory analysis. Splitting data into exploration and confirmation sets can reduce, but does not eliminate, analytical flexibility.

Correlation does not establish causation. Causal interpretation requires design or assumptions concerning confounding, temporal order, intervention, and measurement. A trend line should not be described as an intervention effect unless the design supports that claim.

5. Visualizing Uncertainty

Uncertainty concerns incomplete knowledge about a quantity, prediction, or conclusion. It has several sources:

• Sampling uncertainty: the sample is one of many possible samples.
• Measurement uncertainty: instruments or procedures are imperfect.
• Process variability: the phenomenon itself fluctuates.
• Model uncertainty: different plausible models yield different results.
• Data uncertainty: values are missing, censored, duplicated, or ambiguously coded.
• Scenario uncertainty: future conditions or assumptions are unknown.

These sources should not be collapsed into one unexplained error bar.

5.1 Standard errors and confidence intervals

For independent observations with finite variance, the standard error of the sample mean is commonly estimated as

\[

SE(\bar{x})=\frac{s}{\sqrt{n}}.

\]

This describes uncertainty in estimating a population mean, not the spread of individual observations. Under an approximately normal sampling distribution, a two-sided 95% confidence interval for a mean is often written

\[

\bar{x}\pm t_{0.975,n-1}\frac{s}{\sqrt{n}},

\]

where \(t_{0.975,n-1}\) is the relevant Student’s t critical value with \(n-1\) degrees of freedom. The formula assumes independent observations and is most reliable when the sampling distribution is adequately approximated by the t framework. Complex surveys, clustered observations, serial dependence, and severe skewness require appropriate alternatives, such as design-based, cluster-robust, bootstrap, or transformation-based methods.

A frequentist 95% confidence interval is not correctly interpreted as a 95% probability that this fixed interval contains the parameter. Rather, under repeated sampling and valid assumptions, the procedure would capture the parameter in approximately 95% of comparable samples.

5.2 Prediction intervals and credible intervals

A confidence interval for a mean or regression parameter expresses uncertainty about an estimated population quantity. A prediction interval concerns a future individual or observation and is usually wider because it includes both uncertainty in the estimated mean and individual-level variation.

A Bayesian credible interval is an interval containing a specified posterior probability, conditional on the model, prior distribution, and observed data. It may be expressed as a 95% posterior credible interval. Confidence and credible intervals answer different questions and should not be described interchangeably.

5.3 Visual encodings

Uncertainty can be shown with:

• vertical or horizontal intervals around points;
• shaded confidence or prediction bands;
• quantile ribbons for distributions over time;
• fan charts for multiple future probability levels;
• violin plots, box plots, and interval plots;
• ensemble or spaghetti plots for model and scenario variation;
• gradient or opacity, provided the scale remains interpretable.

Intervals should have explicit legends and labels. If overlap is used to suggest comparison, analysts should avoid the simplistic rule that overlapping 95% intervals prove no difference, or that non-overlap alone proves a particular significance level. The relevant uncertainty of the contrast should be calculated directly.

5.4 Uncertainty in data and models

A model can produce narrow intervals while being systematically wrong. This is false precision: low estimated variance conditional on an inadequate model. Analysts should display residual diagnostics, alternative specifications, sensitivity analyses, and, where possible, uncertainty from model selection and data preprocessing.

For forecasts, prediction intervals should widen when uncertainty increases with the forecast horizon. For simulations, multiple runs can be shown as ensembles rather than only a mean trajectory. Scenario bands should not be presented as probabilities unless scenario probabilities have been justified.

6. Missingness and Incomplete Evidence

Missingness is an analytical feature, not merely a formatting problem. Values may be missing completely at random, missing at random conditional on observed variables, or missing not at random when the probability of missingness depends on the unobserved value itself. These assumptions cannot usually be verified from observed data alone.

Visualization should show both values and missingness. Useful displays include missingness matrices, bar charts of completeness by variable and group, time-series gaps, and plots comparing missingness indicators with observed characteristics. A single overall percentage can hide systematic absence in a vulnerable subgroup.

Common treatments include complete-case analysis, single imputation, multiple imputation, model-based methods, and explicit “unknown” categories for categorical fields. Each has conditions and consequences. Mean imputation reduces visible variance and can create artificial concentrations; treating missing as zero changes the meaning of the variable; deleting records can bias group comparisons. When multiple imputation is used, uncertainty across imputed datasets should be propagated rather than treating one completed dataset as observed truth.

A visualization should distinguish zero, not applicable, not recorded, censored, and unknown. Blank space without a legend is ambiguous.

7. A Detailed Applied Case: Monitoring Hospital Emergency Demand

The following is an illustrative case based on a hypothetical regional hospital network.

Background

A hospital network wants to understand emergency-department demand and determine whether staffing should be adjusted. The dataset contains hourly arrivals for 18 months across four hospitals, together with triage category, waiting time, day of week, weather indicators, and staffing levels.

Problem

Initial management reports show a rising monthly average in arrivals. Leaders request a forecast and ask whether staffing shortages are causing longer waits. The analysis must distinguish genuine demand growth from seasonal variation, data-quality problems, and changes in recording practice.

Analysis

The team first profiles the data. Two hospitals have duplicate arrival records during a software migration, and one hospital has a three-week gap in staffing data. Missing waiting-time values are concentrated in the highest-acuity category because those patients were transferred rapidly to treatment. Treating missing waiting times as zero would therefore produce a misleadingly favourable result.

A time-series plot of hourly arrivals, separated by hospital, reveals strong day-of-week and seasonal patterns. A monthly average alone conceals the fact that winter increases occur mainly during evenings. Small multiples show that one hospital has a level shift immediately after a triage-system change. The shift is investigated rather than interpreted automatically as a demand increase.

Scatterplots of staffing and waiting time show a negative association, but staffing is lower during unusually busy periods. This is a potential reverse-causality and confounding problem: high demand may cause both longer waits and emergency staffing redeployment. A smoothed line is retained for description, but the report does not claim that increasing staffing alone will produce the observed association.

For anomaly detection, the team uses robust baselines by hospital, hour, and day type. Several extreme arrival counts are found to be valid mass-casualty events; others are duplicate uploads. Valid extreme events are retained and annotated. Duplicate records are corrected through data lineage procedures. Forecast displays show median predictions together with 50% and 95% prediction bands, not only a single line.

Approach and solution

The final dashboard includes:

• a completeness panel by hospital and triage category;
• hourly and weekly distributions rather than monthly means alone;
• observed data distinguished from imputed or corrected values;
• annotated operational events;
• prediction intervals for demand forecasts;
• separate displays for typical demand and high-acuity demand;
• a methodological note describing exclusions, missingness, and uncertainty.

A follow-up study is designed to evaluate staffing effects using a time-aware causal design and pre-specified covariates, rather than relying on the exploratory association.

Outcome

Managers discover that apparent annual growth is partly due to the software-related level shift and seasonal concentration. Staffing plans are changed for evening winter periods, while a separate evaluation is commissioned for causal staffing effects. The dashboard is judged more useful than the original report because it shows when evidence is strong, incomplete, or dependent on assumptions.

Learning points

1. Averages can conceal operationally important temporal and subgroup structure.

2. Anomalies require domain investigation, not automatic deletion.

3. Missingness can be systematically related to severity or workflow.

4. Prediction intervals are more appropriate than confidence bands for staffing decisions about future arrivals.

5. Exploratory association should generate a study design, not substitute for one.

8. Professional, Ethical, and Governance Considerations

Uncertainty communication has ethical consequences. A narrow interval may encourage overconfident action; an unexplained missing-data pattern may disadvantage a subgroup; an anomaly detector may disproportionately flag rare but legitimate cases. In high-impact contexts such as healthcare, finance, employment, and public services, visual analytics should support contestability and auditability.

Important practices include:

• documenting provenance, transformations, and analytical versions;
• reporting denominators and sample sizes alongside percentages;
• protecting confidential and personally identifiable information;
• evaluating subgroup performance and missingness;
• separating descriptive evidence from causal or predictive claims;
• preserving raw records and correction histories;
• providing accessible colour, labels, and non-visual alternatives;
• retaining uncertainty in executive summaries rather than presenting only point estimates.

AI-assisted EDA can rapidly identify clusters, correlations, and unusual records, but automated suggestions are not evidence by themselves. Models may reproduce historical bias, infer sensitive attributes, or create spurious patterns through extensive search. Analysts should inspect provenance, use holdout or replication data where appropriate, and disclose material use of automated methods.

9. Contemporary Developments and Best Practices

Modern EDA increasingly combines interactive visualization, statistical computing, and reproducible pipelines. Progressive disclosure allows users to begin with a summary and drill into distributions, subgroups, and records. Uncertainty-aware dashboards can support simulation, scenario comparison, and probabilistic forecasting. However, interactivity increases analytical flexibility and therefore increases the need for logging, pre-registration of confirmatory analyses, or disciplined separation between exploration and reporting.

Recommended practices are:

1. Begin with the data-generating process, not with a preferred chart.

2. Inspect raw and transformed data separately.

3. Use multiple complementary views for important claims.

4. Label units, denominators, time windows, and uncertainty types.

5. Display distributions and sample sizes, not only averages.

6. Investigate missingness before imputation or deletion.

7. Treat outliers as questions requiring context.

8. Distinguish confidence, prediction, and credible intervals.

9. Test sensitivity to scales, smoothing, exclusions, and models.

10. Record exploratory decisions and validate important findings on new or held-out data.

11. Use colour, opacity, and animation conservatively so uncertainty is not hidden.

12. Write conclusions at the strength supported by the evidence.

Challenges and Limitations

EDA is vulnerable to analyst degrees of freedom, confirmation bias, selective reporting, overplotting, and visual pattern-seeking. Large datasets can make negligible differences appear visually prominent or statistically precise. Small datasets can make meaningful effects appear unstable. Aggregation can hide heterogeneity, while excessive stratification can produce noisy comparisons. Interactive tools may encourage repeated searching without a clear record of how a conclusion was reached.

No visualization can recover information that was never measured. Uncertainty intervals quantify only the uncertainty represented by their method; they do not automatically include measurement bias, selection bias, omitted variables, or a wrong model. Responsible EDA therefore combines visual reasoning with domain knowledge, statistical diagnostics, reproducibility, and independent validation.

Chapter Summary

Exploratory Data Analysis is an iterative process for profiling data, understanding distributions and relationships, detecting anomalies, and generating hypotheses. Effective EDA begins with the data-generating context and examines units, types, completeness, dependence, and measurement quality before interpretation. Variation should be described with suitable distributional summaries, while anomalies must be investigated in context rather than mechanically removed.

Uncertainty arises from sampling, measurement, process variation, missingness, modelling, and future scenarios. Standard errors, confidence intervals, prediction intervals, and credible intervals express different ideas and must be labelled accurately. Visualizations should communicate distributions, interval meanings, sample sizes, missingness, and sensitivity to assumptions. Exploratory patterns are valuable for developing hypotheses but do not by themselves establish significance or causation. Ethical practice requires transparent provenance, attention to subgroup effects, privacy protection, accessibility, and resistance to false precision. The most reliable EDA combines complementary visual views, robust statistical reasoning, domain investigation, and validation on new evidence.

Key Terms

Exploratory Data Analysis: Iterative investigation of data to discover structure, quality issues, and hypotheses.

Data profiling: Structural and statistical description of variables, records, types, ranges, and completeness.

Variation: Differences among observations arising from real heterogeneity, measurement, process, or sampling.

Outlier: An observation unusually distant from a relevant reference distribution.

Anomaly: An unusual observation, event, or pattern defined within a specified context.

Robust statistic: A statistic relatively resistant to extreme values or assumption violations.

Heteroscedasticity: Variation in spread that changes across groups or predictor values.

Sampling uncertainty: Uncertainty caused by observing a sample rather than the full population.

Standard error: Estimated standard deviation of a sampling distribution or estimator.

Confidence interval: An interval produced by a procedure with a stated long-run coverage property.

Prediction interval: An interval expressing uncertainty about a future individual observation or outcome.

Credible interval: A Bayesian posterior interval containing a specified probability conditional on a model and prior.

Model uncertainty: Uncertainty resulting from alternative plausible model structures or specifications.

Missingness: Absence of a recorded value, with a mechanism that may itself be informative.

Multiple imputation: Creation and analysis of multiple plausible completed datasets to reflect imputation uncertainty.

Data dredging: Extensive searching that increases the chance of finding coincidental patterns.

Simpson’s paradox: Reversal or alteration of an aggregate association after stratification.

False precision: Appearance of exactness or certainty unsupported by data quality or model validity.

Prediction band: A graphical representation of uncertainty around predicted outcomes.

Sensitivity analysis: Examination of how conclusions change under alternative assumptions or analytic choices.

Review Questions

Short Answer Questions

1. What is the difference between variation and uncertainty?

2. Why is an observation beyond 1.5 IQR not automatically an error?

3. Distinguish a confidence interval from a prediction interval.

4. Why should missingness be visualized by subgroup and time?

5. What is the purpose of separating exploratory and confirmatory analysis?

Long Answer Questions

1. Describe a complete EDA workflow from data profiling to hypothesis generation, including the role of reproducibility.

2. Explain how histograms, density plots, box plots, and empirical cumulative distribution functions provide complementary information.

3. Discuss the main sources of uncertainty in a data visualization and recommend appropriate visual encodings for each.

4. Explain why correlation and a smooth trend line do not establish causation. Use confounding or reverse causality in your discussion.

5. Evaluate the ethical and governance responsibilities involved in visualizing missing data, anomalies, and model predictions in a high-impact domain.

Analytical / Application Questions

1. A dashboard reports the mean waiting time for three hospitals but omits the number of observations and displays only standard-error bars. Design a replacement visualization and explain what additional information should be reported.

2. A fraud-detection model flags a rare transaction type as anomalous. Propose an EDA and validation workflow that distinguishes fraud, data error, and legitimate rarity.

3. A monthly sales chart shows a strong increase after a new software system is introduced. Develop an analysis plan to determine whether the increase reflects genuine demand, a recording change, seasonality, or an aggregation artefact.

Multiple Choice Questions with Answers

Question 1

Which statement best describes the purpose of exploratory data analysis?

A. To prove a causal theory without further testing

B. To discover structure, quality issues, and hypotheses in data

C. To replace all statistical modelling

D. To guarantee that every unusual value is erroneous

Correct Answer: B. To discover structure, quality issues, and hypotheses in data

Question 2

For a sample mean under the usual independent-observation framework, the estimated standard error is:

A. s multiplied by the square root of n

B. s divided by n

C. s divided by the square root of n

D. The variance multiplied by n

Correct Answer: C. s divided by the square root of n

Question 3

Which measure is generally most resistant to extreme values?

A. Sample mean

B. Sample variance

C. Interquartile range

D. Maximum value

Correct Answer: C. Interquartile range

Question 4

A prediction interval differs from a confidence interval for a population mean because it:

A. Usually includes individual-level outcome variation

B. Can only be used for categorical variables

C. Never depends on a model

D. Always has a narrower width

Correct Answer: A. Usually includes individual-level outcome variation

Question 5

Which is an example of a contextual anomaly?

A. A duplicated database row

B. A temperature that is unusual for the season but ordinary globally

C. A valid measurement repeated twice

D. A missing column name

Correct Answer: B. A temperature that is unusual for the season but ordinary globally

Question 6

What is the correct interpretation of a frequentist 95% confidence interval?

A. There is a 95% probability that this fixed interval contains the parameter

B. The procedure captures the parameter in about 95% of comparable repeated samples

C. Exactly 95% of observations lie inside the interval

D. The model has a 95% probability of being correct

Correct Answer: B. The procedure captures the parameter in about 95% of comparable repeated samples

Question 7

Why can mean imputation be problematic?

A. It always increases observed variance

B. It may reduce variance and create artificial concentrations

C. It converts all values into categorical data

D. It proves that missingness is random

Correct Answer: B. It may reduce variance and create artificial concentrations

Question 8

A scatterplot shows that staffing is lower during periods of unusually high demand and waiting times are longer. Which issue is especially relevant before claiming that staffing causes waiting times?

A. Reverse causality or confounding

B. Colour accessibility only

C. The definition of the median

D. The use of an IQR

Correct Answer: A. Reverse causality or confounding

Question 9

Which practice best supports responsible uncertainty communication?

A. Showing only point estimates to avoid confusion

B. Using unexplained error bars on every chart

C. Labelling whether intervals are confidence, prediction, or credible intervals

D. Removing all extreme observations before plotting

Correct Answer: C. Labelling whether intervals are confidence, prediction, or credible intervals

Question 10

Why should exploratory findings often be tested on new or held-out data?

A. To increase the number of visual effects in the original sample

B. To reduce the risk that a discovered pattern is a chance result of flexible searching

C. To eliminate all measurement error

D. To make missing values equal to zero

Correct Answer: B. To reduce the risk that a discovered pattern is a chance result of flexible searching