Data Cleaning Before Statistical Analysis: An Analysis-Readiness Checklist for Researchers

Data cleaning before analysis is not a clerical exercise. Its purpose is to establish whether the dataset faithfully represents the observations, variables, measurements, and research design that the statistical analysis is supposed to address. A dataset can contain no obvious blank cells or duplicate rows and still be analytically unready because variables are miscoded, missingness is consequential, extreme observations have been mishandled, measurement direction is wrong, relationships are nonlinear, or the planned model does not match the data structure. (Newton & Rudestam, 1999; Sekaran & Bougie, 2016; Tabachnick & Fidell, 2013).

A defensible data cleaning checklist for research therefore moves beyond “remove duplicates and fill missing cells.” It works from the research question through the structure and measurement of the data, then through univariate and bivariate screening, and only then toward modeling. Newton and Rudestam (1999) describe data screening as a quality-control process that begins with individual variables and builds toward more complex analysis; Sekaran and Bougie (2016) similarly place coding, editing, transformation, descriptive examination, relationships between variables, reliability, and validity before hypothesis testing.

What Does “Analysis Ready” Mean?

An analysis-ready dataset is not necessarily complete, normally distributed, free of outliers, or perfectly tidy. Instead, the researcher should understand what each case and variable represents, how values were coded, where data are missing, whether unusual observations are errors or genuine observations, how variables are distributed and related, whether measurements support their intended interpretation, and whether the resulting data structure is compatible with the planned analysis. (Newton & Rudestam, 1999; Sekaran & Bougie, 2016; Tabachnick & Fidell, 2013).

This distinction matters because statistical software can process a poorly prepared dataset without recognizing that the substantive meaning of the analysis has been compromised. Tabachnick and Fidell (2013) emphasize that errors and anomalies that would be obvious in small-scale manual work can become difficult to see in computer-based multivariate analysis, making preliminary descriptive and graphical examination essential.

Data cleaning should therefore be treated as analysis preparation, not as an attempt to manufacture a convenient dataset.

First Make the Critical Distinction: Data Error, Unusual Observation, or Model Problem?

Many poor cleaning decisions begin when these three problems are treated as if they were interchangeable.

Before Cleaning

1. Preserve the Original Data

Keep an unchanged version of the original dataset or source records before recoding, excluding observations, constructing variables, or otherwise altering the working data.

This is a practical safeguard implied by a fundamental recommendation in both Newton and Rudestam (1999) and Tabachnick and Fidell (2013): suspicious values should be checked against the original data whenever possible. If cleaning overwrites the only available version, that verification becomes impossible.

Use the preserved source as the reference dataset and conduct cleaning on a separate working copy or through reproducible code.

Checklist

• Preserve an unchanged source version of the data.
• Create a separate working or analysis-preparation version.
• Retain identifiers needed to trace questionable observations back to source records where ethically and procedurally appropriate.
• Avoid manually overwriting suspicious values without retaining what was originally recorded.
• Establish a cleaning log for corrections, exclusions, recoding, transformations, and derived variables.

2. Re-establish the Research Question Before Touching the Data

Cleaning decisions should be made in relation to what the study is intended to measure and analyze. The unit of analysis, variable roles, measurement levels, design, and intended statistical question determine what constitutes a meaningful record and what preparation is required. Newton and Rudestam (1999) begin data preparation with cases, variables, values, coding, and data organization and later connect statistical-test selection to the relationships and variable structures implied by the research question. Adams and Lawrence (2018) similarly integrate research design, measurement, descriptive analysis, relationships, and method selection.

Before cleaning, write down:

• What is the primary research question?
• What is the unit of analysis?
• Which variable is the primary outcome, if applicable?
• Which variables are predictors, exposures, grouping variables, covariates, or measurements?
• Which observations are supposed to belong in the analysis population?
• Are observations independent, paired, repeated, grouped, or otherwise structured?
• What statistical analysis is currently anticipated?
• What measurements or composite scores must be constructed before analysis?

The goal is to prevent the cleaning process from quietly redefining the study.

3. Build or Verify the Variable Dictionary

A numerical column has no statistical meaning until its definition and coding are understood.

Newton and Rudestam (1999) emphasize defining cases, variables, values, coding rules, missing-value codes, and a codebook. Sekaran and Bougie (2016) likewise describe cases as rows and variables as columns and emphasize consistent coding of responses and nonresponses.

For every analysis variable, record:

• Variable name.
• Plain-language definition.
• Unit of measurement.
• Measurement level or response structure.
• Permitted values or categories.
• Meaning of each categorical code.
• Missing-value representation.
• Direction of scoring.
• Whether the variable is observed or derived.
• Formula or rule for derived variables.
• Planned role in the analysis.

During Cleaning

4. Check Coding Before Checking Statistical Assumptions

Before asking whether a variable is normally distributed, establish whether its values mean what you think they mean.

Coding converts observations or response categories into numerical representations for analysis. The numerical code itself does not change the substantive nature of the variable; it represents the underlying category or measurement. (Newton & Rudestam, 1999).

Check:

• Categorical codes against the codebook.
• Missing-value codes.
• Binary/dummy coding.
• Reverse-scored questionnaire items.
• Derived-variable formulas.
• Units and decimal placement.
• Consistency of coding across files, waves, sites, or groups.
• Whether identical codes have accidentally been assigned different meanings.

For multi-item measures, direction matters. Sekaran and Bougie (2016) show that negatively or oppositely worded items may require reverse scoring before items can be meaningfully combined.

5. Find Impossible, Illegal, and Out-of-Range Values

Frequency distributions, minima, maxima, and descriptive statistics are basic quality-control tools.

Newton and Rudestam (1999) recommend routinely examining the frequency distribution of every variable because implausible codes can reveal data-entry problems. Tabachnick and Fidell (2013) recommend checking whether continuous values fall within their permissible ranges, whether discrete variables contain out-of-range numbers, and whether missing-value codes have been correctly specified. Sekaran and Bougie (2016) likewise recommend frequency distributions for identifying illegal codes.

Ask:

• Does every categorical value correspond to a permitted category?
• Are continuous values within possible or instrument-defined limits?
• Are minimum and maximum values plausible?
• Are means and standard deviations plausible given the scale?
• Are missing codes appearing as genuine observations?
• Are there impossible combinations of variables?

A value should not be corrected merely because it looks surprising. If the value is possible, investigate before changing it.

6. Check Duplicated Records—But Define “Duplicate” First

A repeated identifier can indicate that a case was accidentally entered twice. Newton and Rudestam (1999) specifically recommend examining identification numbers because repeated IDs may expose duplicate entry or another data-entry error.

But identical rows are not automatically duplicates. Repeated observations can be legitimate in longitudinal, repeated-measures, transactional, clustered, or event-level datasets.

Check:

• Are supposedly unique identifiers actually unique?
• If an identifier repeats, should it repeat under the study design?
• Do repeated records represent separate measurement occasions?
• Are records identical because of duplicate import or entry?
• Would deleting one record alter the intended unit of analysis?

The cleaning decision must follow the study structure, not a generic “remove duplicates” command.

7. Investigate Missingness Rather Than Merely Counting Blank Cells

Missing data are not one problem with one solution.

Newton and Rudestam (1999) emphasize first determining the extent of missing data and investigating why information is missing. Some missingness is structurally appropriate—for example, when a follow-up question does not apply because of an earlier response—whereas other missingness may reflect refusal, data-collection failure, data-entry problems, or systematic differences among participants.

Tabachnick and Fidell (2013) stress that the pattern and reason for missingness can matter more than the simple percentage missing. Deleting observations with missing values can distort the analyzed sample when missingness is related to other variables.

For every important variable:

• Count missing observations.
• Calculate the usable sample for relevant analyses.
• Distinguish structural/not-applicable missingness from unobserved data.
• Check whether missingness clusters by variable, group, wave, site, or participant.
• Investigate plausible reasons for missingness.
• Check whether cases with missing data differ systematically on observed variables.
• Determine what the intended analysis does with missing observations.
• Document the chosen missing-data strategy.

8. Find Outliers—but Do Not Automatically Delete Them

Outlier screening is an investigation, not a deletion rule.

An extreme value can arise from data-entry error, measurement problems, an unusual but genuine participant, or the natural tail of a population distribution. (Newton & Rudestam, 1999; Sekaran & Bougie, 2016; Adams & Lawrence, 2018).

Use appropriate descriptive and graphical tools to locate potentially unusual observations. Frequency distributions, histograms, boxplots, and related displays can reveal univariate extremes; scatterplots can reveal unusual bivariate combinations that would not appear abnormal when either variable is examined alone. (Newton & Rudestam, 1999; Tabachnick & Fidell, 2013).

For every potential outlier:

• Verify the recorded value against source information where possible.
• Check units, coding, and decimal placement.
• Determine whether the value is impossible, improbable, or merely extreme.
• Examine whether the observation is unusual on one variable or on a combination of variables.
• Examine its influence on relevant summaries or models.
• Consider whether a robust summary, transformation, or different model addresses the problem better than deletion.
• Record any exclusion and its justification.

Do not do this

9. Examine Every Important Distribution

Researchers should become familiar with their variables before advanced modeling.

Newton and Rudestam (1999) recommend beginning data screening with frequency distributions for every variable and then using graphical displays to understand distribution shape. Sekaran and Bougie (2016) describe obtaining frequencies, central tendency, dispersion, and visual summaries as a necessary first step in data analysis. Tabachnick and Fidell (2013) similarly use descriptive statistics and graphics to screen data before multivariate analysis.

Depending on the variable, examine:

• Frequencies and percentages.
• Minimum and maximum.
• Mean, median, or mode where meaningful.
• Standard deviation, range, or interquartile spread where meaningful.
• Histogram or comparable distributional display.
• Boxplot where useful.
• Skewness or other distributional features relevant to the planned method.
• Floor or ceiling concentration.
• Sparse categories.
• Unexpectedly little variability.

A strange distribution is not automatically dirty data. It may accurately describe the sample.

10. Examine Bivariate Relationships Before Multivariable Modeling

Univariate checks cannot detect every problem.

A value can be ordinary on two separate variables but highly unusual in combination. Newton and Rudestam (1999) use the example of a young age and several divorces: neither value alone must be an outlier, while their combination can be highly unusual. Scatterplots can reveal such patterns.

Bivariate examination also helps detect nonlinear relationships, unusual clusters, restricted ranges, unexpected associations, and potential redundancy among variables. Sekaran and Bougie (2016) recommend examining relationships among study variables after obtaining their descriptive statistics, while Tabachnick and Fidell (2013) emphasize that multivariate analyses often depend on correlation or covariance patterns.

Before modeling:

• Examine scatterplots for important continuous-variable pairs.
• Examine cross-tabulations for important categorical pairs.
• Review appropriate correlations or associations.
• Look for nonlinear patterns.
• Look for unusual bivariate observations.
• Look for clusters or subgroup structure.
• Identify near-duplicate or highly redundant variables.
• Check whether observed relationships make substantive and measurement sense.

11. Check Measurement Before Creating Scores

Data preparation cannot repair a measure that does not represent the intended construct.

Reliability concerns the consistency of measurement, whereas validity concerns whether the measurement supports its intended interpretation. Both are relevant before researchers treat measured variables as if they were error-free representations of theoretical constructs. (Meier et al., 2014; Adams & Lawrence, 2018).

For multi-item measures:

• Confirm that all items are coded in the intended direction.
• Reverse-score applicable items before constructing a score.
• Verify that the intended items—and only those items—enter the composite.
• Establish the relevant reliability or measurement evidence before combining items when that is required by the measurement strategy.
• Define how missing items affect score construction.
• Do not create a composite solely because several items share the same response scale.

Sekaran and Bougie (2016) specifically place reverse scoring before composite-score calculation and state that combining items should follow satisfactory interitem consistency.

12. Transform Data for an Analytical Reason, Not a Cosmetic One

Transformation changes the numerical representation of a variable and can alter interpretation.

Sekaran and Bougie (2016) describe transformations as changes undertaken to address issues relevant to subsequent analysis, including transformations of uneven distributions and reverse scoring. Tabachnick and Fidell (2013) discuss transformations as possible tools for improving distributional properties and reducing the influence of univariate outliers.

Before transforming:

• Identify the specific problem the transformation is intended to address.
• Confirm that transformation is compatible with the intended interpretation.
• Preserve the original variable.
• Create the transformed variable separately.
• Re-examine its distribution and relationships afterward.
• Check how the transformation handles zeros and missing values.
• Document the transformation exactly.

Do not transform variables simply until an assumption test becomes nonsignificant or a preferred model produces the desired result.

Before Modeling

13. Produce Descriptive Statistics Before Advanced Analysis

Descriptive analysis is not an optional reporting exercise performed after the “real” statistics. It is part of understanding whether advanced analysis is sensible.

Sekaran and Bougie (2016) describe “getting a feel for the data” as the necessary first step in data analysis and recommend frequencies, measures of central tendency, dispersion, and examination of relationships. Newton and Rudestam (1999) likewise argue that researchers should become comfortable with their data before full-scale hypothesis testing.

Before fitting the primary model:

• Report the number of observations available.
• Describe key categorical variables.
• Summarize important continuous variables.
• Inspect distributions.
• Review key bivariate relationships.
• Confirm that the effective sample size is what you expect after missing-data handling.
• Confirm that derived variables were constructed correctly.
• Check that group sizes and category frequencies support the intended analysis.

If the descriptive results are surprising, investigate them before moving on.

14. Separate Cleaning Checks From Model Diagnostics

Not every failed diagnostic calls for another round of deleting data.

Once data errors have been addressed and valid observations identified, remaining issues may concern the assumptions or specification of the statistical model. Tabachnick and Fidell (2013) emphasize examination of normality, linearity, homoscedasticity, multicollinearity, singularity, and multivariate outliers because these characteristics affect multivariate analyses.

Use this decision sequence:

1. Is the value wrong?
   → Correct a verified data error.
2. Is the value unusual but credible?
   → Treat it as an unusual but valid observation and investigate influence.
3. Does the planned model poorly represent otherwise valid data?
   → Treat it as a model problem and reconsider specification, transformation, diagnostics, or statistical method.

This prevents model convenience from becoming an undocumented exclusion criterion.

15. Recheck the Dataset Against the Research Question

A technically cleaner dataset can be scientifically worse if cleaning has changed the population, measurement, outcome, or design being analyzed.

After exclusions, recoding, missing-data handling, transformations, and score construction, ask again:

• Does each row still represent the intended unit of analysis?
• Does the analyzed sample still correspond to the population or sample the research question concerns?
• Do variables still represent the intended constructs or quantities?
• Has exclusion disproportionately altered an important subgroup?
• Has missing-data handling changed the effective analysis sample?
• Are repeated or clustered observations still represented correctly?
• Does the planned statistical method match the cleaned data structure?
• Does the model answer the original research question?

Research design, measurement, and statistical analysis are linked; method selection should follow the question and the structure of the data rather than the availability of a convenient software procedure. (Newton & Rudestam, 1999; Adams & Lawrence, 2018).

Before Final Analysis

16. Freeze and Document the Analysis Dataset

The final analysis dataset should be reproducible from the preserved source data through explicit cleaning and transformation decisions.

At minimum, document:

• Corrections of verified errors.
• Duplicate-record decisions.
• Recoding rules.
• Reverse scoring.
• Derived variables and composite-score rules.
• Missing-data handling.
• Exclusion criteria and excluded cases.
• Outlier investigations and decisions.
• Transformations.
• Changes to the originally planned analysis prompted by diagnostics.
• Final sample size for each major analysis.

This documentation matters because coding, editing, transformations, missing-data treatment, and exclusions can change the observations and quantities that enter statistical analysis. (Newton & Rudestam, 1999; Sekaran & Bougie, 2016; Tabachnick & Fidell, 2013).

A useful decision log has four fields:

Issue → Evidence → Decision → Consequence for analysis

The StatsAlly Analysis-Readiness Checklist

Before Cleaning

• Preserve the original/source data unchanged.
• Create a separate working dataset.
• State the primary research question.
• Identify the unit of analysis.
• Identify the intended analysis population.
• Define outcome, predictors, groups, covariates, and measurement variables.
• Build or verify the variable dictionary/codebook.
• Define valid values and categories.
• Define missing-value codes.
• Identify reverse-scored and derived variables.
• Establish a cleaning decision log.

During Cleaning

• Verify coding against the codebook.
• Check minimum and maximum values.
• Produce frequency distributions for categorical and coded variables.
• Identify impossible, illegal, or out-of-range values.
• Trace suspicious values to source records where possible.
• Check supposedly unique identifiers.
• Investigate apparent duplicate records in light of the study design.
• Quantify missingness by variable and case.
• Distinguish structural missingness from unavailable observations.
• Investigate patterns and plausible reasons for missingness.
• Identify potential univariate outliers.
• Verify extreme observations before considering exclusion.
• Inspect distributions graphically and numerically.
• Examine important bivariate relationships.
• Look for unusual combinations of otherwise valid values.
• Check scoring direction for multi-item measures.
• Evaluate relevant measurement quality before constructing composites.
• Apply transformations only for an explicit analytical reason.
• Preserve original variables alongside transformed versions.
• Record every material correction, recode, transformation, and exclusion.

Before Modeling

• Produce descriptive statistics for the analysis variables.
• Confirm the actual sample size entering each analysis.
• Examine distributions relevant to the intended model.
• Examine key bivariate relationships.
• Check for nonlinearity where relevant.
• Check for problematic redundancy among variables where relevant.
• Investigate influential or multivariate observations without assuming they are errors.
• Separate data-quality problems from model-assumption problems.
• Reconsider the model rather than automatically deleting valid observations.
• Confirm that the cleaned data still match the intended unit of analysis and design.
• Confirm that constructed variables still represent their intended measurements.
• Confirm that the proposed method answers the original research question.

Before Final Analysis

• Resolve or explicitly document every material flagged issue.
• Freeze/version the analysis dataset.
• Retain the cleaning code, syntax, or reproducible sequence of operations.
• Record all exclusions and their reasons.
• Record all recoding and transformations.
• Record the missing-data strategy.
• Record final analysis sample sizes.
• Re-run descriptive checks on the final dataset.
• Verify that no cleaning step was chosen because it produced a preferred inferential result.
• Confirm that the final analysis dataset still corresponds to the intended research question.

Do Not Do This

Do not delete observations because they prevent statistical significance

Whether an observation supports the hypothesis is not evidence that it is erroneous. Adams and Lawrence (2018) explicitly caution against omitting data simply because they do not align with the researcher's hypothesis.

Do not label every extreme value an error

Outliers can be genuine observations. Investigate them before deciding what they mean. (Newton & Rudestam, 1999; Sekaran & Bougie, 2016; Adams & Lawrence, 2018).

Do not “clean” away nonnormality automatically

Nonnormality may be a genuine characteristic of the variable. Whether it presents a problem depends on the intended statistical procedure and the role of the variable in that procedure. (Newton & Rudestam, 1999; Tabachnick & Fidell, 2013).

Do not let missing-data defaults silently choose your sample

Default case deletion can reduce the analyzed sample without the researcher fully appreciating which observations disappeared. (Tabachnick & Fidell, 2013).

Do not recode inconsistent responses merely because another answer looks more plausible

Sekaran and Bougie (2016) warn that apparently inconsistent responses may nevertheless be deliberate; changing them without adequate justification can introduce bias.

Do not construct a composite before checking scoring direction and measurement quality

Reverse-scored items must be handled correctly, and combining multiple items should follow the relevant evidence that they can defensibly function together. (Sekaran & Bougie, 2016).

Do not transform until the result looks statistically convenient

A transformation needs an analytical justification and changes the interpretation of the variable. (Sekaran & Bougie, 2016; Tabachnick & Fidell, 2013).

Do not confuse a model problem with dirty data

Nonlinearity, heteroscedasticity, redundancy among predictors, or influential observations can indicate that the proposed model requires reconsideration. They do not by themselves prove that valid observations should be removed. (Tabachnick & Fidell, 2013).

A Better Definition of Statistical Data Cleaning

That definition changes the workflow.

Instead of:

import → delete duplicates → fill missing values → remove outliers → model

use:

research question → unit of analysis → variable definitions → coding verification → error checks → missingness → distributions → outliers → bivariate relationships → measurement → transformations where justified → descriptive statistics → model diagnostics → research-question recheck → final analysis

This sequence reflects a recurring message across the methodological sources: researchers should understand the data before asking sophisticated statistical procedures to summarize them. (Newton & Rudestam, 1999; Sekaran & Bougie, 2016; Tabachnick & Fidell, 2013).

The Bottom Line

The objective of data preparation for statistical analysis is not to produce a dataset with no missing values, no extreme observations, and perfectly shaped distributions. It is to produce an analysis dataset whose construction is understood, whose errors have been investigated, whose measurement and coding are defensible, and whose remaining characteristics are represented appropriately in the statistical analysis. (Newton & Rudestam, 1999; Sekaran & Bougie, 2016; Tabachnick & Fidell, 2013).

The most important distinction throughout the process is:

A dataset becomes analysis ready not when every irregularity has disappeared, but when the researcher knows what the irregularities mean, has made defensible decisions about them, and can explain exactly how the source data became the dataset used for analysis.

References

Adams, K. A., & Lawrence, E. K. (2018). Research methods, statistics, and applications (2nd ed.). SAGE Publications.

Meier, K. J., Brudney, J. L., & Bohte, J. (2014). Applied statistics for public and nonprofit administration (9th ed.). Cengage Learning.

Newton, R. R., & Rudestam, K. E. (1999). Your statistical consultant: Answers to your data analysis questions. SAGE Publications.

Sekaran, U., & Bougie, R. (2016). Research methods for business: A skill-building approach (7th ed.). John Wiley & Sons.

Tabachnick, B. G., & Fidell, L. S. (2013). Using multivariate statistics (6th ed.). Pearson.