7. Data Management and Statistical Foundations

Introduction: From Responses to Evidence

Advertising research rarely fails because respondents were unavailable or because a statistical formula was unknown. It more often fails because data were poorly defined, inconsistently recorded, inadequately documented, or interpreted without regard to the decision being made. A campaign-tracking study may contain thousands of observations, yet a single coding error can reverse a conclusion about brand preference. Similarly, a statistically significant difference may be too small to justify a change in media allocation.

Data management and statistical reasoning therefore form the bridge between research design and managerial action. Data management concerns the disciplined preparation, protection, transformation, and documentation of information. Statistical foundations provide the language for describing patterns, quantifying uncertainty, comparing groups, and judging the strength and practical importance of findings.

For advertising and media professionals, these capabilities are especially important because research data are heterogeneous. They may include survey responses, media-exposure records, digital-behaviour logs, sales outcomes, brand-lift measures, social-media indicators, and qualitative codes. The analyst must decide what each variable means, how it should be represented, which observations are trustworthy, and what conclusions the evidence can support.

This chapter develops a complete analytical workflow: preparing and documenting data; cleaning and transforming variables; summarising distributions; using probability and sampling logic; estimating confidence intervals; conducting hypothesis tests; interpreting cross-tabulations; and distinguishing statistical significance from managerial significance.

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the role of data management in advertising and media research.

2. Distinguish cases, variables, values, measurement levels, and data types.

3. Design a data dictionary and a reproducible data-preparation workflow.

4. Identify and handle missing, invalid, duplicate, inconsistent, and extreme observations.

5. Apply suitable descriptive statistics and interpret distributions.

6. Explain probability, sampling error, confidence intervals, null hypotheses, p-values, and statistical power.

7. Select and interpret common hypothesis tests, cross-tabulations, and effect-size measures.

8. Evaluate whether a finding is managerially meaningful, not merely statistically significant.

9. Recognise ethical, technical, and governance risks in the use of research data.

10. Communicate analytical results accurately to academic and professional audiences.

1. The Analytical Data Lifecycle

A useful way to understand data work is as a lifecycle rather than a single act of analysis. The lifecycle usually includes:

Specification. Before collecting or importing data, the researcher defines the unit of analysis, variables, coding rules, population, time frame, and intended decisions. For example, the unit may be an individual respondent, a household, an advertisement impression, a campaign week, or a purchase transaction. Confusing units is a serious error: an analysis of impressions cannot automatically be interpreted as an analysis of people.

Acquisition and integration. Data may arrive from survey software, spreadsheets, customer relationship systems, advertising platforms, panels, or third-party providers. Integration requires matching identifiers, aligning time periods, and checking whether measures have comparable definitions.

Preparation. Raw files are inspected, labelled, cleaned, transformed, and documented. The original raw data should normally be preserved as read-only, while cleaned and analysis-ready versions are saved separately.

Analysis. Descriptive and inferential methods are selected according to the research question, measurement level, design, and assumptions.

Interpretation and communication. Results are connected to the research objective and business decision. Tables and visualisations should make uncertainty and limitations visible rather than presenting figures with false precision.

Archiving and governance. Syntax, versions, codebooks, approvals, access permissions, and final outputs are retained so that analyses can be audited or reproduced.

This lifecycle is iterative. An unusual result may reveal an error in a questionnaire code, an unexpected but genuine subgroup, or a measurement problem. Statistical analysis can therefore lead back to data inspection and revision.

2. What a Research Dataset Contains

A case is one unit of observation. In a survey, it is often one respondent. A variable is a measured characteristic, such as age, brand awareness, platform used, or purchase intention. A value is the recorded response for a particular case and variable. A dataset is commonly arranged with cases as rows and variables as columns, although long-format data may contain multiple rows per case—for example, one row per respondent-platform combination.

Measurement levels

Measurement level determines which summaries and tests are defensible.

• Nominal variables classify observations without order, such as gender category, region, campaign cell, or preferred platform.
• Ordinal variables have an order, but the distances between categories cannot be assumed equal. Agreement categories and purchase-intention ratings are common examples.
• Interval variables have ordered values with approximately equal intervals but no meaningful absolute zero. Some index scores are treated as interval measures for practical analysis.
• Ratio variables have equal intervals and a meaningful zero, such as impressions, expenditure, time, clicks, or number of purchases.

The distinction is not merely theoretical. A mean is usually natural for expenditure, but less informative for a nominal variable such as platform. A median may be preferable for an ordinal scale or a highly skewed income variable. The measurement level should guide both analysis and interpretation.

Data dictionaries and metadata

A data dictionary records how variables are defined and coded. A useful entry includes variable name, label, unit, measurement level, permissible values, missing-value codes, source, collection date, transformation history, and confidentiality classification. For example:

`brand_awareness`: aided awareness; 1 = aware, 0 = not aware; binary; source = post-campaign survey; missing = 99; denominator excludes missing responses.

Metadata are part of the evidence. Without them, analysts may mistake the value 99 for a genuine response, confuse percentages based on all respondents with percentages based on valid respondents, or fail to recognise that a variable changed wording between waves.

3. Preparing and Cleaning Data

Cleaning is not the removal of inconvenient results. It is the systematic correction of errors and the transparent treatment of observations that do not meet predefined standards.

Structural checks

The first inspection examines file structure. Analysts should verify the number of rows and columns, variable names, data types, duplicate identifiers, date formats, and whether imported values have been truncated or converted. A numeric variable accidentally imported as text can prevent valid calculations; a date stored in different regional formats can misorder a time series.

Validity and range checks

Every variable should have an expected range or set of categories. A seven-point attitude item containing a value of 17 indicates an entry or coding error. A campaign date before the campaign launch may indicate a time-zone issue, a wrong file, or an incorrectly linked record. Validation rules should be established before inspecting outcomes where possible, reducing the risk of selectively editing data to obtain a desired result.

Duplicate records

Duplicates may arise from repeated exports, respondent re-entry, or system failures. They should not be removed automatically. Two records with the same respondent ID may represent duplicate submissions, but they may also represent legitimate repeated observations. The decision requires knowledge of the design and should be recorded.

Missing data

Missingness can be classified as:

• Unit nonresponse: an entire case is absent because a selected person did not participate.
• Item nonresponse: a participant answered the study but omitted a particular question.
• Structural missingness: a question does not apply, such as a media-spend question for a non-advertiser.
• System missingness: a value was lost through technical or processing failure.

The pattern and likely mechanism of missingness matter. If respondents with low brand involvement are less likely to answer a purchase-intention question, simply replacing missing values with the mean may bias results. Common treatments include complete-case analysis, explicit “not applicable” categories, imputation, weighting, and sensitivity analysis. No method removes uncertainty; each should be justified and documented.

Outliers and influential observations

An outlier is unusually distant from other observations. It may be an error, a legitimate extreme, or evidence of a distinct segment. A high media-spend value could be a misplaced decimal or a genuine national campaign. Analysts should inspect the source, compare with related variables, and evaluate the effect of retaining or excluding the observation. Deleting a value solely because it weakens a result is unacceptable. Robust summaries such as the median and interquartile range can reduce the influence of extreme values.

Transformations

Transformation changes the form of a variable for analytical or interpretive purposes. Common examples include:

• recoding age into meaningful segments;
• creating a total attitude score from several items;
• calculating click-through rate as clicks divided by impressions;
• converting expenditure into logarithmic form when values are highly skewed;
• standardising values into z-scores;
• creating campaign-exposure groups from frequency measures.

Transformations can clarify patterns but may also discard information or introduce arbitrary thresholds. The original variable should be retained, and the transformation formula should be documented.

Reproducibility and version control

A professional workflow separates raw, cleaned, and analysis-ready files. Cleaning steps should be performed through documented syntax, formulas, or scripts rather than undocumented manual editing. File names should include dates or version numbers, and the final dataset should be accompanied by a data dictionary, cleaning log, and statement of exclusions. Reproducibility is not limited to academic research; it protects agencies and advertisers when clients question how a result was produced.

4. Describing Data Before Testing Claims

Descriptive statistics answer the question: what does this dataset look like? They do not, by themselves, establish whether a pattern generalises to a wider population.

Frequencies and proportions

For categorical variables, frequency tables show counts and percentages. Percentages require a declared denominator. A table may report 48% of all respondents, 48% of valid respondents, or 48% of exposed respondents; these are not equivalent. Small denominators should be displayed cautiously because percentages can appear dramatic while being based on very few cases.

Centre and spread

The mean is the arithmetic average and is sensitive to extreme values. The median is the middle observation after ordering and is more resistant to skew. The mode is the most frequent value and is particularly useful for categorical data.

Measures of dispersion include the range, variance, standard deviation, and interquartile range. The standard deviation describes the typical distance from the mean in the units of the variable. The interquartile range covers the middle 50% and is often useful for skewed distributions.

A mean campaign frequency of 5.2 exposures does not imply that every person saw the campaign approximately five times. The standard deviation, distribution, or percentiles are needed to understand concentration and reach-frequency balance.

Shape and visual inspection

Histograms, bar charts, box plots, scatterplots, and time-series plots can reveal skewness, clusters, gaps, nonlinearity, and influential observations. Visual inspection should precede formal testing. Two groups can have the same mean but very different distributions; a correlation can be driven by a small cluster; and a high overall conversion rate can conceal poor performance in a key audience.

Standard scores and association

A z-score expresses a value’s distance from the mean in standard-deviation units:

`z = (x − mean) / standard deviation`.

Correlation describes the direction and strength of linear association between two quantitative variables, commonly ranging from −1 to +1. Association is not proof of causation. In advertising, exposure may correlate with purchase because effective campaigns increase exposure, because interested consumers seek the advertisements, or because both are related to a third factor.

5. Probability, Samples, and Uncertainty

Probability provides a framework for reasoning about uncertain outcomes. In research, it helps analysts distinguish a pattern that could plausibly arise from sampling variation from one that would be unusual if a specified assumption were true.

A population is the full set of units of interest; a sample is the subset observed. A parameter describes a population, whereas a statistic describes a sample. Sampling error arises because different samples would produce somewhat different statistics. It is not the same as poor measurement, nonresponse bias, coverage error, or data-entry error.

A probability distribution describes the possible values of a statistic and their probabilities. With sufficiently large and appropriate samples, the sampling distribution of a mean often approaches a normal shape, which supports confidence intervals and many tests. This approximation does not repair a biased sample. A very large convenience sample can produce highly precise estimates of the wrong population.

6. Confidence Intervals and Estimation

A point estimate, such as a sample proportion of 0.42, is incomplete without an indication of uncertainty. A confidence interval provides a range of plausible values for the population parameter under the assumptions of the method.

For a sample proportion, a simplified interval is often expressed as:

`estimate ± critical value × standard error`.

The standard error reflects sampling variability and generally decreases as sample size increases. A 95% confidence interval is constructed by a procedure that would capture the true parameter in approximately 95% of repeated samples under the stated assumptions. It is not strictly correct to say that there is a 95% probability that a fixed parameter lies in this particular interval, although that wording is common in informal communication.

Confidence intervals communicate both direction and precision. If campaign awareness is estimated at 42% with a reasonably narrow interval, management can plan with greater confidence than if the estimate is 42% with a very wide interval. Intervals for differences are particularly useful: if an interval for the difference in conversion rates excludes zero, it corresponds, under standard conditions, to a two-sided test at the 5% level.

7. Hypothesis Testing as Structured Decision-Making

Hypothesis testing begins with a research claim translated into a null hypothesis and an alternative hypothesis. The null hypothesis commonly states that there is no difference, association, or effect in the population. The alternative hypothesis states what would count as evidence against that position.

A test statistic measures how far the observed result lies from what would be expected under the null hypothesis. The p-value is the probability, assuming the null hypothesis and model are correct, of obtaining a result at least as extreme as the one observed. It is not the probability that the null hypothesis is true, nor is it the probability that the finding will replicate.

A conventional threshold such as 0.05 is a decision rule, not a law of nature. A p-value below the threshold is often described as statistically significant, but it does not establish practical importance. Multiple testing, optional stopping, poor measurement, and violated assumptions can make p-values misleading.

Two errors are central. A Type I error rejects a true null hypothesis. A Type II error fails to reject a false null hypothesis. Statistical power is the probability of detecting an effect of a specified size under a specified design. Power depends on sample size, variability, significance threshold, and effect size. Planning should focus on the smallest effect that would matter to the decision, not merely on achieving a large sample.

Selecting a test

The method should follow the question and design:

• A one-sample test compares a sample estimate with a benchmark.
• An independent-groups comparison examines different groups, such as exposed and unexposed respondents.
• A paired comparison examines the same units before and after an intervention.
• A chi-square test examines association between categorical variables.
• Correlation or regression examines relationships involving quantitative variables.
• Nonparametric methods may be appropriate when scale properties, outliers, or distributional assumptions make conventional methods unsuitable.

The test choice must also consider independence. Repeated observations from the same person are not equivalent to observations from unrelated people. Ignoring clustering can make standard errors too small and findings appear more certain than they are.

8. Cross-Tabulation and Effect Size

Cross-tabulation is one of the most useful tools in advertising research because it connects audience characteristics with behaviours or outcomes. A table might compare aided awareness by campaign exposure, platform preference by age group, or purchase intention by region.

Percentages must be oriented correctly. In a table of exposure by awareness, row percentages answer a different question from column percentages. The analyst should state whether percentages are calculated across rows or columns and identify the base for each calculation.

The chi-square statistic evaluates whether the observed pattern differs from what would be expected if the two categorical variables were independent. Its result is affected by sample size. With a very large sample, a trivial difference may be significant; with a small sample, a meaningful pattern may fail to reach significance.

Effect sizes address magnitude. For a difference in means, Cohen’s d expresses the difference relative to pooled standard deviation. For categorical association, phi or Cramér’s V can describe strength of association. For binary outcomes, risk difference, risk ratio, and odds ratio provide different interpretations. An odds ratio should not casually be described as a probability ratio, especially when the outcome is common.

Consider a campaign that raises purchase incidence from 10% to 12%. The relative increase is 20%, but the absolute increase is only two percentage points. If the campaign reaches one million relevant consumers, two percentage points may have substantial commercial value. If it reaches a small niche at high cost, the same effect may not justify investment. Effect size must therefore be interpreted with reach, cost, baseline performance, and strategic importance.

9. Statistical Significance and Managerial Significance

Statistical significance asks whether the data are sufficiently inconsistent with a specified null model. Managerial significance asks whether the magnitude, reliability, cost, and implications of the result justify action.

A statistically significant lift of 0.2 percentage points may be irrelevant if implementation costs exceed incremental profit. Conversely, a non-significant lift may deserve further investigation when the study is underpowered, the potential value is high, and the estimate is directionally promising. “Not statistically significant” does not mean “no effect”; it means that the evidence did not meet the chosen evidential standard under the model and data available.

A managerial interpretation should consider:

• absolute and relative effect size;
• confidence-interval width;
• audience size and commercial value;
• media and production costs;
• duration and scalability;
• equity, brand, and reputational consequences;
• study quality and possible bias;
• whether the result is consistent with theory and prior evidence.

A useful reporting sentence might be: “The exposed group showed a 2.1-percentage-point higher awareness rate, with a confidence interval that includes effects from near zero to a modest positive lift. The result is not conclusive, but the potential value in the priority segment warrants a larger controlled test.” This is more responsible than either declaring victory or dismissing the campaign solely because a threshold was not crossed.

10. Applied Case: Evaluating a Streaming-Video Campaign

Background

This illustrative case concerns a fictional consumer-finance brand, Northstar, which ran a six-week streaming-video campaign aimed at adults aged 25–44. The research team conducted a post-campaign survey and linked respondents’ consented exposure records to survey responses. The main outcome was aided brand awareness; secondary outcomes were consideration and intended application.

Situation and problem

The marketing team observed that awareness was 46% among exposed respondents and 43% among unexposed respondents. The agency initially described the three-point difference as proof of campaign success. The client asked whether the result was statistically reliable and large enough to justify expanding the media budget.

Data preparation

The research unit was the individual respondent. The team first preserved the raw files and created a data dictionary. It found 2,410 survey records, including 37 duplicate IDs, 19 respondents outside the target age range, and a code of 99 used for both “not asked” and “prefer not to say.” The duplicates were investigated using timestamps and response completeness; 29 were confirmed repeat submissions and removed according to the pre-specified rule. The age exclusions were retained in a separate file and excluded from the target-population analysis. The missing codes were separated into structural and substantive missingness.

Exposure was defined as at least one verified impression during the campaign period. Respondents with incomplete linkage information were not automatically treated as unexposed. Instead, the team reported the linked sample separately and conducted a sensitivity analysis. Awareness was coded 1 for aided recognition of Northstar and 0 otherwise.

Analysis

Descriptive results showed awareness of 46% among 612 exposed respondents and 43% among 1,084 unexposed respondents. The absolute difference was three percentage points. A cross-tabulation indicated that exposed respondents were also more likely to live in metropolitan areas, where Northstar had stronger distribution. This suggested possible confounding: exposure was not randomly assigned.

A two-group comparison produced a confidence interval for the awareness difference that ranged from a small negative value to a positive value of approximately eight percentage points. The corresponding test did not meet the conventional 0.05 threshold. The team therefore did not claim definitive incremental lift. It also noted that the observed association could reflect targeting, self-selection, or pre-existing brand interest.

The analysts then examined consideration and found a larger difference among respondents with three or more verified exposures, but the subgroup was small and less precise. Rather than presenting it as confirmation, they labelled it exploratory. They recommended a controlled follow-up design using matched markets or randomised audience assignment where feasible.

Managerial decision

The client did not cancel the campaign. Instead, it maintained limited activity while commissioning a better-designed test. The decision reflected the distinction between evidence and action: the initial study did not establish causal lift, but the possible upside justified a more rigorous experiment. Media expansion was tied to pre-specified success criteria, including an absolute awareness lift meaningful enough to cover incremental cost.

Learning points

1. A difference in percentages is not automatically an attributable campaign effect.

2. Cleaning and variable definitions can materially change the analysis base.

3. Confidence intervals communicate uncertainty more effectively than a p-value alone.

4. Cross-tabulation can reveal imbalance between exposed and comparison groups.

5. Statistical evidence should be integrated with cost, reach, scale, and study design before budget decisions are made.

11. Risks, Ethics, and Professional Governance

Data quality and data ethics are connected. A technically accurate analysis can still be professionally unacceptable if respondents were misled, data were used beyond consent, or vulnerable audiences were targeted irresponsibly.

Researchers should collect only data necessary for the stated purpose, protect personally identifiable information, restrict access, and apply appropriate retention and deletion rules. Identifiers should be separated from analytical files where possible. Linkage between survey responses and digital exposure records requires clear consent, lawful processing, and careful communication of risks. Applicable privacy and data-protection requirements vary by jurisdiction, so organisations should obtain specialist legal and governance advice rather than assuming that anonymisation is automatic.

Analysts must disclose exclusions, missing-data treatment, weighting, subgroup analyses, and deviations from the original plan. Selective reporting, p-hacking, HARKing—presenting post hoc hypotheses as if planned—and suppression of unfavourable outcomes undermine research integrity. Automated systems also require review for biased classification, opaque scoring, data leakage, and inappropriate use of sensitive attributes.

12. Digital Tools, Automation, and Emerging Practice

Spreadsheets remain useful for small, transparent tasks, but they become risky when datasets are large, transformations are complex, or multiple analysts collaborate. Statistical packages, SQL databases, notebooks, and version-control systems support reproducible workflows. Automated validation can flag invalid ranges, unexpected category changes, duplicate IDs, and sudden shifts in response distributions.

Machine-learning tools can assist with anomaly detection, imputation, segmentation, and prediction, but prediction is not the same as explanation or causal inference. An algorithm that identifies people likely to respond to an advertisement does not prove that advertising caused the response. Models may reproduce historical inequality or optimise a narrow metric at the expense of brand safety, privacy, or long-term value. Human review, documentation, holdout validation, and monitoring are essential.

Emerging practice increasingly emphasises data lineage: the ability to trace a reported metric back to its source, transformations, assumptions, and responsible owner. Analysts are also moving toward interval estimates, preregistered analysis plans, causal experiments, and decision-focused dashboards rather than isolated headline metrics.

Best-Practice Workflow

A defensible workflow can be summarised as follows:

1. Define the decision, population, unit of analysis, and outcome.

2. Preserve raw data and establish access controls.

3. Create a data dictionary and record coding conventions.

4. Inspect structure, ranges, duplicates, dates, and missingness.

5. Apply pre-specified cleaning rules and retain an audit log.

6. Preserve original variables when transforming or recoding.

7. Describe distributions before conducting inferential tests.

8. Match tests to the research design, measurement level, and assumptions.

9. Report estimates, confidence intervals, effect sizes, sample bases, and limitations.

10. Interpret statistical evidence alongside cost, scale, risk, and strategic relevance.

11. Use reproducible files, scripts, and version control.

12. Protect participants and communicate uncertainty honestly.

Chapter Summary

Data management is the foundation of credible advertising analytics. Researchers must define units and variables, preserve raw data, document coding, inspect quality, treat missingness carefully, investigate duplicates and outliers, and maintain reproducible workflows. Descriptive statistics reveal the centre, spread, shape, and composition of data, while probability and sampling theory explain uncertainty. Confidence intervals provide estimates with precision; hypothesis tests evaluate evidence against a null model; and cross-tabulations reveal relationships among categorical variables. Effect sizes indicate magnitude and should accompany p-values. Most importantly, statistical significance does not determine managerial significance. Decisions require consideration of effect size, uncertainty, cost, reach, causality, risk, and strategic value. Ethical governance, privacy protection, transparent reporting, and responsible use of digital and automated tools are integral to professional research practice.

Key Terms

Case: A unit of observation represented in a dataset.

Variable: A measurable characteristic that can take different values across cases.

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

Nominal measurement: Classification into categories without intrinsic order.

Ordinal measurement: Ordered categories without assured equal intervals.

Ratio measurement: Numerical measurement with equal intervals and a meaningful zero.

Missing data: Values unavailable because of nonresponse, inapplicability, system failure, or other causes.

Outlier: An observation unusually distant from the rest of a distribution.

Transformation: A documented change to a variable’s form, scale, or coding.

Population: The complete set of units to which a study seeks to refer.

Sample: The observed subset drawn from or representing a population.

Parameter: A numerical characteristic of a population.

Statistic: A numerical characteristic calculated from a sample.

Standard error: The estimated sampling variability of a statistic.

Confidence interval: An interval produced by a procedure designed to capture a population parameter at a stated long-run confidence level.

Null hypothesis: A formal statement of no specified effect, difference, or association.

P-value: The probability, under the null model, of results at least as extreme as those observed.

Effect size: A measure of the magnitude of a difference or relationship.

Cross-tabulation: A table showing the joint distribution of two or more categorical variables.

Managerial significance: The practical and strategic importance of a finding for decision-making.

Review Questions

Short Answer Questions

1. Why should raw, cleaned, and analysis-ready data files be stored separately?

2. Distinguish unit nonresponse, item nonresponse, and structural missingness.

3. What information should a data dictionary contain?

4. What does a confidence interval add to a point estimate?

5. Why is statistical significance not equivalent to managerial significance?

Long Answer Questions

1. Explain the data-management lifecycle for an advertising research project, from specification to archiving.

2. Discuss how measurement levels influence the choice of descriptive statistics and analytical tests.

3. Explain Type I error, Type II error, statistical power, and p-values, including their relevance to campaign evaluation.

4. Describe how cross-tabulation, chi-square testing, and effect sizes can be used together in audience analysis.

5. Discuss ethical and governance requirements when linking survey responses with digital advertising exposure data.

Analytical / Application Questions

1. A campaign produces a 1.5-percentage-point increase in conversion with a p-value of 0.01. What additional information should a manager request before increasing the budget?

2. In a survey, the exposed group is younger and more metropolitan than the unexposed group. Explain why a simple awareness comparison may not estimate causal campaign lift and propose a stronger design.

3. A dataset contains a value of 99 in an attitude variable coded from 1 to 7. Describe the checks and documentation required before calculating the mean and testing group differences.

22. Multiple Choice Questions with Answers

Question 1

Which action best supports reproducibility in a research-data workflow?

A. Replacing all unusual values with the sample mean

B. Editing the original spreadsheet directly without recording changes

C. Preserving raw data and documenting transformations in a cleaning log or script

D. Reporting only the final percentages to the client

Correct Answer: C. Preserving raw data and documenting transformations in a cleaning log or script

Question 2

A variable recording preferred social-media platform is usually measured at which level?

A. Nominal

B. Ordinal

C. Interval

D. Ratio

Correct Answer: A. Nominal

Question 3

What is the most appropriate interpretation of a p-value?

A. The probability that the null hypothesis is true

B. The probability of observing results at least as extreme as those found, assuming the null model is true

C. The probability that the campaign caused the observed result

D. The percentage of the population affected by the campaign

Correct Answer: B. The probability of observing results at least as extreme as those found, assuming the null model is true

Question 4

Why can a statistically significant result still be managerially unimportant?

A. Significance tests never use sample size

B. A small effect can be detected in a large sample but may not justify its cost

C. Confidence intervals always exaggerate effects

D. Managers should ignore effect sizes

Correct Answer: B. A small effect can be detected in a large sample but may not justify its cost

Question 5

Which measure is generally most resistant to extreme values in a skewed expenditure distribution?

A. Mean

B. Variance

C. Median

D. Range

Correct Answer: C. Median

Question 6

A cross-tabulation reports percentages by column. What does each percentage describe?

A. The distribution of column categories within each row category

B. The distribution of row categories within each column category

C. The percentage of missing observations only

D. The causal effect of the row variable on the column variable

Correct Answer: B. The distribution of row categories within each column category

Question 7

An observed campaign lift has a confidence interval that includes zero. Which conclusion is most defensible?

A. The campaign definitely had no effect

B. The campaign definitely succeeded

C. The data do not establish a non-zero effect at the corresponding confidence level

D. The sample must be invalid

Correct Answer: C. The data do not establish a non-zero effect at the corresponding confidence level

Question 8

Which situation is most likely to create confounding in a post-campaign exposure comparison?

A. Exposure is randomly assigned to respondents

B. Exposed and unexposed groups are identical on relevant pre-campaign factors

C. People already interested in the brand are more likely to receive or seek exposure

D. The outcome is measured using a binary variable

Correct Answer: C. People already interested in the brand are more likely to receive or seek exposure

Question 9

What does an effect size primarily communicate?

A. The magnitude of a difference or relationship

B. The number of variables in a dataset

C. The probability that respondents answered honestly

D. The percentage of missing values

Correct Answer: A. The magnitude of a difference or relationship

Question 10

A value of 99 appears in a seven-point attitude item and the codebook identifies 99 as ‘not asked’. What should the analyst do?

A. Treat 99 as the highest attitude score

B. Convert 99 to 7 because it is near the upper end

C. Treat it according to the documented missing-data rule and exclude or otherwise handle it appropriately

D. Delete the entire dataset

Correct Answer: C. Treat it according to the documented missing-data rule and exclude or otherwise handle it appropriately