Data Preparation and Management

Introduction

Data preparation and management convert raw observations into a trustworthy, documented, secure, and research-ready dataset. In business research, raw data may arrive as survey forms, interview schedules, transaction records, spreadsheets, administrative databases, sensor feeds, or exported online-platform files. Such data frequently contain omissions, inconsistent coding, duplicate records, invalid values, ambiguous labels, and formatting errors. Statistical software can process a file rapidly, but it cannot determine whether the file represents the intended sample or whether a value was entered correctly. Those are methodological and managerial responsibilities.

Data preparation is therefore not a clerical stage performed after research design. It is a controlled part of the research process. Decisions about coding, missing data, transformations, exclusions, and corrections can alter descriptive results, relationships among variables, and ultimately managerial recommendations. A defensible dataset must allow another qualified researcher to understand what was collected, what was changed, why it was changed, and which version was used for analysis.

Learning Objectives

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

1. Explain the purpose and sequence of editing, coding, data entry, cleaning, missing-data treatment, and transformation.
2. Design a codebook and metadata structure for a research-ready dataset.
3. Detect and resolve invalid, inconsistent, duplicate, and implausible records.
4. Select defensible approaches to missing data and distinguish missingness from legitimate zero values.
5. Apply transformations without obscuring the original observations.
6. Maintain secure documentation, version control, and audit trails.
7. Evaluate the quality and limitations of a prepared dataset before analysis.

1. The Data-Preparation Workflow

A useful workflow is sequential but iterative:

instrument and file inspection → editing → coding → data entry or import → validation and cleaning → missing-data assessment → transformation and derived variables → documentation and quality assurance → analysis dataset.

The sequence is not completely linear. For example, a cleaning check may reveal that a response category was poorly designed, requiring a return to the codebook. Similarly, a missing-data pattern may reveal that a survey question was misunderstood. The researcher should preserve each stage rather than overwriting earlier files.

A dataset normally contains:

• Rows or cases: individual respondents, firms, transactions, branches, or other units of analysis.
• Columns or variables: measured characteristics such as age, revenue, satisfaction, or industry.
• Values: the observations recorded for each case-variable combination.
• Identifiers: fields used to distinguish cases, preferably using study IDs rather than names.
• Metadata: information describing variables, coding, measurement, provenance, and processing decisions.

A research-ready dataset is not merely a rectangular spreadsheet. It is a dataset accompanied by a codebook, source documentation, processing scripts or syntax where possible, quality-control records, and a clear statement of what the cases and variables represent.

2. Editing: Reviewing Data for Completeness and Credibility

Editing is the systematic examination of collected data for completeness, legibility, consistency, and apparent credibility before formal analysis. It may occur in the field, immediately after collection, or centrally after all records have been received.

2.1 Field and office editing

Field editing occurs close to the point of collection. An interviewer may identify an unanswered question, clarify an illegible response, or confirm that a skip instruction was followed. Field editing must not become fabrication. A researcher should never infer a respondent’s answer merely to complete a form.

Office editing is a more systematic review conducted by the research team. It checks whether:

• every record has a valid study identifier;
• required questions were answered;
• responses follow the questionnaire’s skip patterns;
• values lie within permitted ranges;
• open-ended answers are intelligible and relevant;
• dates, currencies, units, and decimal conventions are consistent;
• records appear duplicated;
• contradictory responses require verification.

Editing decisions should be classified. A useful classification is: accepted as recorded, corrected from an authoritative source, recoded according to a documented rule, treated as missing, or excluded with a stated reason.

2.2 Logical and range checks

A range check identifies values outside the permitted set. If a five-point satisfaction scale is coded from 1 to 5, a value of 8 is invalid unless the instrument explicitly permits another code. A logical check assesses relationships between variables. For example, a respondent recorded as not owning a vehicle should not have a positive annual vehicle-maintenance expenditure unless the question referred to another vehicle.

Logical checks should distinguish definite errors from unusual but possible observations. A very high purchase value may be genuine, especially in business-to-business data. Automatically deleting outliers because they seem unusual can introduce serious bias.

3. Coding and the Codebook

Coding assigns meaningful symbols or numbers to responses so that they can be stored, retrieved, compared, and analyzed. Coding does not make a variable quantitative. For example, industry codes 1 = manufacturing, 2 = services, and 3 = retail are labels; their numerical order has no arithmetic meaning.

3.1 Closed-ended and open-ended responses

Closed-ended questions are usually pre-coded. A gender or employment-status question may have a defined set of categories, including a carefully considered option such as prefer not to answer. Likert responses should retain their conceptual order, for example 1 = strongly disagree through 5 = strongly agree.

Open-ended responses require coding after collection or a combination of inductive and deductive coding. The researcher should first read a sample of responses, develop categories, define inclusion rules, and test whether different coders apply the scheme consistently. Categories should be mutually exclusive where one answer is expected and collectively exhaustive enough to represent plausible responses. An other category may be necessary, but it should be reviewed rather than used as a substitute for thoughtful category design.

For qualitative text converted into categories, retain the original text. The coded variable is an analytic representation, not a replacement for the evidence.

3.2 Code values and special codes

A codebook should distinguish substantive values from administrative codes. For example:

Special codes should not be confused with valid measurements. Treating 999 years of tenure as a real value would distort means and correlations. In many statistical systems, missing values must be explicitly declared; otherwise special codes may be included in calculations.

3.3 Coding reliability

When multiple people code the same material, training and a coding manual are essential. Coders should practice on the same subset, discuss disagreements, refine definitions, and then independently code a reliability sample. Percentage agreement is easy to understand but does not account for agreement expected by chance. For nominal categories, Cohen’s kappa is commonly used for two coders:

Here, Pₒ is the observed proportion of agreement and Pₑ is the proportion of agreement expected by chance under the chosen marginal-distribution assumptions. Kappa is informative but should not be interpreted mechanically: its value can be affected by category prevalence, and substantive review of disagreements remains necessary. For ordered categories, weighted approaches may be more appropriate.

4. Data Entry and Import

Data entry is the transfer of responses into a digital structure; data import is the transfer from another digital system. Both can introduce errors. Manual entry may produce transposed digits, skipped rows, or incorrect decimal points. Import may create problems through incompatible date formats, character encoding, truncated labels, changed variable types, or altered leading zeros.

4.1 Designing the entry structure

Each row should represent one defined unit of analysis and each column one variable. A separate data dictionary should record the variable name, label, type, allowable values, missing-value rules, source question, and any transformation. Variable names should be short, unique, stable, and free from ambiguous spaces or symbols.

Data-validation controls can restrict entries to permitted ranges, require dates in a standard format, and enforce skip patterns. Such controls reduce errors but do not eliminate the need for checking.

4.2 Verification procedures

Useful procedures include:

• Double entry: two independent entries are compared and discrepancies resolved against the source.
• Verification sampling: a selected proportion of records is checked against original forms.
• Automated validation: scripts identify impossible values, duplicated IDs, unexpected categories, and inconsistent formats.
• Frequency inspection: frequencies reveal rare or unexpected codes.
• Independent totals: entered totals can be compared with known source totals where applicable.

The appropriate level of verification depends on risk. Sensitive, high-value, or regulatory data justify stronger controls than a low-risk exploratory exercise.

5. Data Cleaning and Validation

Data cleaning identifies and resolves errors, inconsistencies, and inappropriate representations while preserving an evidence trail. It should be rule-based and reproducible rather than an informal search for results that appear attractive.

5.1 Common problems

Important classes of problems include:

• invalid values outside the code frame;
• inconsistent spellings, such as London, LONDON, and Londn;
• mixed units, such as dollars and thousands of dollars;
• incompatible dates, such as day-month and month-day formats;
• duplicate cases;
• impossible combinations of responses;
• outliers and influential observations;
• accidental text in numeric fields;
• truncation or rounding introduced during export.

5.2 Duplicates and identity

A duplicate is not simply two identical rows. A customer may legitimately place two orders with the same amount. Duplicate detection should use the unit of analysis and available identifiers. Exact duplicates can often be removed after verification. Near duplicates require comparison of dates, names, contact details, transaction numbers, or other keys. When two records represent the same case but contain different information, the researcher should consult the source and document which record was retained or how it was consolidated.

5.3 Outliers and implausible values

An outlier is an observation that differs markedly from others; it is not automatically an error. The researcher should ask whether it is:

1. a data-entry or measurement error;
2. a valid member of the target population;
3. a case from a different population;
4. a value that disproportionately influences a model.

Possible actions include correcting a verified error, retaining the value, transforming a skewed variable, using a robust method, conducting sensitivity analysis, or excluding the case under a pre-specified and defensible rule. Deletion should be the last option, not the default response to an inconvenient result.

6. Missing Data

Missing data occur when a value that could have been observed is unavailable. Missingness may arise because a respondent refuses, does not know, is not asked due to a skip pattern, drops out, or because of technical or administrative failure. A legitimate zero is not missing: zero employees, zero sales in a specified period, and zero purchases may all be substantive values.

6.1 Types of missingness

A practical classification is:

• Unit nonresponse: no usable data are obtained from a sampled unit.
• Item nonresponse: a unit participates but omits particular variables.
• Structural missingness: a question does not apply because of a valid skip pattern.
• Censored or unavailable information: the value exists conceptually but cannot be observed, such as income withheld by a respondent.

Missingness mechanisms are often discussed as:

• MCAR, missing completely at random: missingness is unrelated to observed and unobserved values.
• MAR, missing at random conditional on observed information: missingness may depend on variables observed in the dataset, but not on the missing value after those variables are controlled.
• MNAR, missing not at random: missingness depends on the unobserved value itself or on unobserved factors, even after observed information is considered.

These are assumptions about the data-generating process, not labels that can be proven solely by inspecting a spreadsheet. For example, higher-income respondents may be less likely to report income, suggesting a plausible MNAR concern. A respondent’s age and job level may help explain observed income nonresponse, supporting a MAR-based model, but cannot establish MAR with certainty.

6.2 Diagnosing missingness

The researcher should report the amount and pattern of missing data by variable and, where relevant, by important groups. Useful diagnostics include frequency tables, cross-tabulations, response-order analysis, and comparisons of respondents with and without missing values on observed characteristics. The reason for missingness is often more informative than the percentage alone.

A useful descriptive measure is the item-missingness rate:

where rⱼ is the percentage missing for variable j, mⱼ is the number of missing values for that variable, and nⱼ is the number of eligible cases. The denominator should be defined carefully when structural skips exist.

6.3 Treatment options

Complete-case analysis uses only cases with all variables required for a particular analysis. It is simple and transparent, but it reduces sample size and can produce bias when complete cases differ systematically from incomplete cases.

Available-case analysis uses all observed values for each calculation. It may use more data, but different statistics can be based on different subsets, making comparisons difficult and sometimes producing unstable covariance matrices.

Single imputation replaces missing values with one value, such as a mean, median, regression prediction, or hot-deck value. Mean substitution is particularly problematic because it reduces variation and can distort relationships. If single imputation is used for operational reasons, the method and limitations must be disclosed.

Multiple imputation replaces each missing value with several plausible values drawn from a model that uses observed information. Each completed dataset is analyzed separately, and results are combined so that uncertainty from missingness is represented. Multiple imputation is generally more defensible than single imputation when its model is appropriate, but it does not solve severe measurement error or an implausible missingness model.

Weighting or adjustment may be used when nonresponse patterns are known and suitable auxiliary information is available. Weights must be constructed and documented carefully because extreme weights can increase variance.

No universal percentage determines the correct treatment. The decision depends on the research design, missingness mechanism, variable type, sample size, analysis model, and consequences of error. Sensitivity analysis should compare plausible treatments where missingness could materially affect conclusions.

7. Transformation and Derived Variables

Transformation changes the representation or scale of a variable while preserving a documented link to the source value. It may be required to harmonize units, reduce skewness, create analytical indicators, or operationalize a construct.

7.1 Common transformations

• Unit conversion: converting annual revenue from dollars to thousands of dollars.
• Aggregation: calculating a mean or sum across items measuring a construct, after checking that the items are conceptually compatible.
• Reverse coding: changing the direction of a negatively worded item so that higher values consistently indicate more of the construct.
• Standardization: expressing a value relative to a sample mean and standard deviation.
• Logarithmic transformation: reducing strong right skew in variables such as revenue, provided zero and negative values are handled explicitly.
• Categorization: converting a continuous variable into bands, such as firm-size classes.
• Ratios and rates: creating measures such as revenue per employee, with attention to zero denominators and interpretation.

For standardization, the usual sample z-score is:

where xᵢ is an observation, x̄ is the sample mean, and s is the sample standard deviation. The resulting value has no measurement unit and indicates how many sample standard deviations an observation lies above or below the sample mean. The transformation is appropriate when a relative scale is useful; it does not make a variable normally distributed.

For a natural-log transformation of a strictly positive variable:

where xᵢ > 0. If zero values are meaningful, adding a constant such as 1 may be used in some applications, but that changes interpretation and should not be applied automatically. Negative values require a different, substantively justified approach.

7.2 Rules for defensible transformation

Retain the raw variable and create a new transformed variable. Record the formula, units, date, software or syntax used, and rationale. Do not transform merely because a result becomes statistically significant. Examine whether the transformation improves interpretability, satisfies model requirements, or addresses a known measurement issue. Report the scale used in tables and explain it in the methodology.

8. Metadata, Documentation, and Audit Trails

Metadata are data about data. They describe provenance, meaning, structure, ownership, measurement, coding, and processing. At minimum, a research project should maintain:

• study title, purpose, and unit of analysis;
• source instrument and collection dates;
• variable names, labels, types, units, and permissible values;
• question wording and skip logic;
• missing-value definitions;
• transformations and derived-variable formulas;
• data-quality rules and exception decisions;
• file versions, dates, and responsible persons;
• links between raw, intermediate, and analysis files.

An audit trail is a chronological record of changes and decisions. A useful entry states the date, file or variable affected, issue identified, action taken, reason, person responsible, and approval where required. Version names such as raw_01, cleaned_02, and final_latest are weak unless supported by a change log. Prefer immutable raw data, version control, meaningful commit messages, and reproducible scripts.

A data dictionary may include the following fields:

9. Security, Confidentiality, and Governance

Data preparation must protect participants, employees, customers, and organizations. Security is not limited to a password. Appropriate controls may include:

• collecting only data necessary for the research purpose;
• separating direct identifiers from analytical data;
• using pseudonymous study IDs;
• encrypting data in storage and transfer;
• applying role-based access and multifactor authentication;
• maintaining secure backups and recovery procedures;
• recording access and changes;
• securely deleting temporary exports;
• retaining data only for an approved period;
• preventing confidential values from appearing in filenames, logs, or screenshots.

A de-identified file may still carry disclosure risk if combinations of variables identify a rare executive, branch, or firm. Small-cell suppression, aggregation, masking, or controlled access may be necessary. Legal requirements vary by jurisdiction and sector, so researchers must follow applicable privacy, employment, contractual, and institutional requirements. Ethical approval does not remove the obligation to minimize harm throughout the data life cycle.

AI-assisted cleaning can suggest duplicate records, classify text, or identify anomalies. It should not silently alter a research file. Sensitive data should not be uploaded to an unapproved service. Any AI-assisted decision requires human review, documented prompts or rules where relevant, validation against source data, and an audit trail. Reproducibility and confidentiality take priority over convenience.

10. Quality Assurance and Readiness for Analysis

Before analysis, the researcher should conduct a documented release check. The checklist should confirm that:

1. the unit of analysis and population are defined;
2. the number of cases agrees with the sampling and collection records;
3. identifiers are unique or duplicates are explained;
4. all variables have labels, types, units, and coding rules;
5. invalid values and impossible combinations have been reviewed;
6. missing values are distinguished from legitimate zeros and structural skips;
7. transformations are reproducible and raw variables remain available;
8. exclusions and corrections are recorded;
9. confidentiality controls are active;
10. the released analysis file can be recreated from the raw file and documented procedures.

A quality decision should consider accuracy, completeness, consistency, validity, timeliness, uniqueness, and traceability. No dataset is perfect. The objective is to make limitations visible and ensure that analytical claims do not exceed data quality.

Applied Case: Preparing a Multi-Branch Employee Survey

Background

A fictional retail company, Northstar Stores, surveys 1,200 employees across 24 branches to examine whether training is associated with service confidence. The survey contains a five-point confidence scale, training hours, tenure, branch, and employment status.

Problem

The initial export contains 1,247 rows. Twenty-seven employee IDs are duplicated, training hours include values of −4 and 999, branch names use three spelling conventions, and 14% of respondents did not report training hours. Some part-time employees were not eligible for the training question, but they were incorrectly coded in the export as blank. The raw file also contains names and email addresses.

Analysis and approach

The research team first preserves the original export in a restricted, read-only location and creates pseudonymous study IDs. It compares the row count with the survey platform’s completion report and investigates duplicated IDs against submission timestamps. Twelve rows are confirmed as duplicate submissions; the remaining apparent duplicates represent two legitimate survey waves and are retained with a wave indicator.

The team consults the questionnaire. Training hours must be zero or greater, and 999 is an accidental missing-value code. The value −4 is checked against the source form and corrected to 4. Branch names are standardized through a controlled lookup table, while the original text is retained. Structural noneligibility is coded separately from item nonresponse. The codebook is updated to distinguish 97 = not applicable, 98 = do not know, and 99 = no answer.

The team examines whether missing training hours are associated with employment status, branch, and tenure. Missingness is substantially higher among part-time employees and in two branches, suggesting that treating missing values as random may be unsafe. Because eligibility and observed characteristics explain part of the pattern, the team uses a documented multiple-imputation sensitivity analysis for eligible respondents and reports a complete-case comparison. It does not impute structurally inapplicable cases.

A composite confidence score is created only after checking that the confidence items have the same direction and acceptable internal coherence for the intended analysis. The raw items remain in the file. A derived training indicator distinguishes no training from positive training, while training hours are retained as a continuous variable for alternative analyses.

Outcome

The final analytical file contains 1,208 eligible, unique employee-wave records. A data dictionary, cleaning syntax, exception log, missingness report, and version history accompany it. Results are reported with sensitivity analyses because nonresponse is concentrated in particular groups.

Learning points

The case demonstrates that cleaning is not indiscriminate deletion. It combines source verification, codebook design, structural-missingness recognition, secure handling, reproducible transformation, and explicit uncertainty. The most important managerial result is not simply a clean spreadsheet; it is a defensible chain from employee response to reported conclusion.

Challenges and Limitations

Data preparation cannot repair a poorly designed construct, biased sampling frame, misleading question, or systematically unobserved population. Coding open-ended responses may reduce nuance. Imputation relies on assumptions that may be uncertain. Transformations can improve statistical behavior while reducing intuitive interpretation. Automated validation may reject legitimate exceptional cases, whereas overly permissive rules may allow errors through. Documentation also requires time and expertise, particularly when several systems and external suppliers are involved.

Researchers should therefore distinguish data error from data variation, missingness from nonapplicability, and statistical convenience from methodological justification. Decisions should be made before inspecting final results where possible, and all material alternatives should be tested or disclosed.

Chapter Summary

Data preparation converts raw observations into a reliable and documented research resource. Editing checks completeness, credibility, ranges, and logical consistency. Coding assigns systematic labels and values, while the codebook explains their meaning and distinguishes valid values from special missing codes. Data entry and import require validation, verification, and attention to formats and units.

Cleaning addresses invalid values, duplicates, inconsistent representations, impossible combinations, and unusual observations. Outliers must be investigated rather than automatically deleted. Missing data should be classified, diagnosed, and treated in a way consistent with the research design and plausible missingness mechanism. Complete-case analysis, available-case analysis, single imputation, multiple imputation, and weighting each have conditions and limitations.

Transformations should be substantively justified, reproducible, and created as new variables while preserving raw data. Metadata, data dictionaries, version control, and audit trails make decisions traceable. Security requires minimization, pseudonymization, access control, encryption, retention discipline, and awareness of re-identification risk. A research-ready dataset is released only after quality checks establish its validity, completeness, consistency, confidentiality, and reproducibility.

Key Terms

Audit trail: A chronological record of data changes, decisions, reasons, and responsible persons.

Available-case analysis: An analysis using all observed values available for each calculation.

Codebook: A structured description of variables, labels, codes, units, and missing-value rules.

Coding: Assigning symbols or categories to responses for storage and analysis.

Complete-case analysis: An analysis restricted to cases with required variables observed.

Data cleaning: Identifying and resolving errors, inconsistencies, and inappropriate representations.

Data dictionary: Metadata describing the structure, meaning, provenance, and permissible values of a dataset.

Data entry: Transferring observations from a source into a digital data structure.

De-identification: Removing or modifying identifying information to reduce disclosure risk.

Field editing: Reviewing data close to the point of collection.

Item nonresponse: Missingness in a particular variable when other information was obtained.

MCAR: Missing completely at random; missingness is unrelated to observed and unobserved values.

MAR: Missing at random conditional on observed information.

MNAR: Missing not at random; missingness depends on unobserved values or factors after observed information is considered.

Metadata: Information about the origin, structure, meaning, and processing of data.

Multiple imputation: Creating several plausible completed datasets to represent uncertainty from missing values.

Outlier: An observation unusually distant from other observations.

Pseudonymization: Replacing direct identifiers with codes while retaining controlled re-identification capability.

Structural missingness: A value that is absent because the question does not apply under the study design.

Transformation: A documented change in variable representation, scale, unit, or construction.

Review Questions

Short Answer Questions

1. What is the difference between field editing and office editing?
2. Why should special missing-value codes be distinguished from valid numerical values?
3. Define unit nonresponse, item nonresponse, and structural missingness.
4. Why should raw variables be retained when transformed variables are created?
5. Identify four components of a useful data dictionary.

Long Answer Questions

1. Explain the complete workflow from raw data receipt to release of a research-ready analytical dataset.
2. Compare complete-case analysis, single imputation, and multiple imputation, including their assumptions and limitations.
3. Discuss how coding decisions for open-ended responses can affect validity and reproducibility.
4. Explain the role of metadata, version control, and audit trails in defensible business research.
5. Develop a security and confidentiality protocol for a dataset containing employee survey responses.

Analytical / Application Questions

1. A customer dataset contains negative ages, duplicated customer IDs, blank income values, and income coded as 999999 for some respondents. Design a sequence of checks and documented decisions for preparing the file.
2. In a survey, high-income managers are less likely to report salary, while salary nonresponse is also related to observed job grade and age. Evaluate plausible missingness mechanisms and recommend a treatment and sensitivity analysis.
3. A researcher deletes all observations above the 99th percentile because they weaken a regression result. Critically assess this decision and propose a more defensible outlier protocol.

Multiple Choice Questions with Answers

Question 1

Which document primarily defines variable labels, allowable values, units, and missing-value rules?

A. A codebook

B. A consent form

C. A sampling frame

D. A presentation slide

Correct Answer: A. A codebook

Question 2

A respondent is not asked a question because a valid skip pattern makes it irrelevant. This is best classified as:

A. Unit nonresponse

B. Structural missingness

C. A duplicate record

D. An outlier

Correct Answer: B. Structural missingness

Question 3

Which action best protects reproducibility when creating a transformed variable?

A. Overwrite the raw variable

B. Create a new variable and document the formula

C. Delete all unusual observations

D. Change the variable label without recording the change

Correct Answer: B. Create a new variable and document the formula

Question 4

In the expression zᵢ = (xᵢ − x̄) ÷ s, what does s represent?

A. The sample size

B. The sample standard deviation

C. The missingness rate

D. The study identifier

Correct Answer: B. The sample standard deviation

Question 5

Which statement about an outlier is most defensible?

A. Every outlier is a data-entry error

B. Every outlier must be deleted

C. An outlier should be investigated before a treatment decision is made

D. Outliers can never influence analysis

Correct Answer: C. An outlier should be investigated before a treatment decision is made

Question 6

What is the main purpose of a pseudonymous study ID?

A. To increase the number of respondents

B. To replace direct identifiers in analytical files

C. To convert ordinal data into ratio data

D. To guarantee that missing data are MCAR

Correct Answer: B. To replace direct identifiers in analytical files

Question 7

Which missing-data approach creates several plausible completed datasets?

A. Mean substitution

B. Complete-case analysis

C. Multiple imputation

D. Range checking

Correct Answer: C. Multiple imputation

Question 8

A value of 999 is used to mean no answer on a variable measuring years of tenure. What is the principal risk if it is not declared missing?

A. It may be treated as a genuine tenure value and distort analysis

B. It will automatically become a duplicate ID

C. It will convert all variables to text

D. It will eliminate structural missingness

Correct Answer: A. It may be treated as a genuine tenure value and distort analysis

Question 9

Which practice is most appropriate for sensitive research data?

A. Place names and email addresses in the analysis file

B. Upload the raw file to any convenient online tool

C. Use access controls and separate direct identifiers from analytical data

D. Use descriptive filenames containing participant names

Correct Answer: C. Use access controls and separate direct identifiers from analytical data

Question 10

Under the usual definition of MAR, missingness may depend on:

A. Observed information, after conditioning on that information

B. Nothing whatsoever

C. Only the missing value and never observed variables

D. The researcher’s preferred hypothesis

Correct Answer: A. Observed information, after conditioning on that information