Preparing data and selecting statistical tests: from variable types to thesis reporting
Software can return a number even when the calculation is inappropriate. Selecting a suitable test begins with identifying data types and checking the data matrix before running the analysis.
Data types guide appropriate calculations. Coding, missing-data handling and error checking affect trustworthiness. Test assumptions guide the choice of parametric or nonparametric methods. Once a p-value is available, effect size still matters because large samples can make small differences statistically significant.
1. Data types and measurement precision
Saunders et al. (2023) distinguish categorical data, including nominal (with binary data as a two-category case) and ordinal data, from numerical interval and ratio data. Numerical variables can also be classified as continuous or discrete. More precise measurement supports a wider range of analyses.
| Type | Characteristics | Example |
|---|---|---|
| Binary | Two categories | Pass/fail, male/female in a binary-coded variable |
| Nominal | Unordered categories | Commuting mode or branch |
| Ordinal | Order is known, distances are not. | Likert agreement or ranks |
| Interval | Meaningful differences, without a true zero for ratios | Degrees Celsius |
| Ratio | Meaningful differences and ratios | Revenue, profit and years of service, as listed in the supplied article |
The textbook gives two reasons to classify correctly. Software can easily calculate an inappropriate statistic, such as a mean of 3.74 for six numerically coded hotels. A mode is meaningful for those nominal codes. Also, more precise data can be grouped into less precise categories, but the reverse operation cannot recover lost detail. If unsure about future requirements, collect at the highest practical precision (Saunders et al., 2023).
2. Coding and the codebook
Numerical codes generally facilitate recoding and derived variables (Saunders et al., 2023). Established occupational, industry or social-category classifications can save time, offer tested coding and support comparison with other research.
For open questions, the textbook suggests examining the first 50 to 100 cases before constructing a coding scheme. The process has five steps.
Identify broad categories
Read responses and establish the main groups.
Develop subcategories
Refine them to the detail required by the planned analysis.
Assign codes
Code each category at the necessary level.
Record actual examples
Document responses assigned to each group to build the codebook.
Use adjacent codes for related groups
Make later merging and recoding easier.
Two coders can independently code the same subset to assess consistency. Percentage agreement equals agreements divided by coded segments, multiplied by 100. McHugh, cited in Saunders et al. (2023), discusses 80% as a commonly accepted level. Cohen's kappa provides a more stringent two-coder measure.
3. Missing data
When a case contains some data, each variable needs a recorded value or missing code. Codes may distinguish reasons for missingness. Systematic missingness concentrated in particular questions or respondent groups can affect representativeness and needs investigation. Saunders et al. (2023), drawing on Little and Rubin, contrast this with random missingness.
Questionnaire missingness may arise from screening skips, refusal, misunderstanding, not knowing or having no opinion, and accidental omissions. “Do not know” can be treated as a response or as missing, depending on the question. Decide consistently in advance.
4. Data matrix, error checks and weights
Most software expects variables in columns and cases in rows. One questionnaire question need not equal one column. Multiple-answer questions can use multiple-response coding, with as many variables as the maximum answers per person, or multiple-dichotomy coding, one yes/no variable per option. The latter makes counting easier but can create many variables (Saunders et al., 2023, pp. 586–588).
Use variable and value labels matching the questionnaire, save regularly and retain secure backups. Check the following.
- Invalid codes, such as 7 on a five-point scale or letter O instead of zero
- Implausible combinations, such as a senior-management occupation with a manual-work job description
- Screening logic, so cases screened out have appropriately missing later responses
Assign a unique identifier to link each row to its original questionnaire. The textbook stresses that checking takes time but skipping it can lead to incorrect conclusions.
5. Test assumptions and normality
The textbook contrasts nonparametric methods, used extensively with categorical or ranked data, with parametric analyses of numerical outcomes. Its overview discusses independence, normality and equal variances, alongside the required measurement level (Saunders et al., 2023, p. 614). Requirements depend on the particular test.
Inspect histograms, box plots or frequency polygons alongside Kolmogorov–Smirnov or Shapiro–Wilk tests. A p-value ≤ .05 indicates evidence against normality under the test. Large samples can detect small departures, so graphical assessment matters (Saunders et al., 2023, p. 615).
6. Hypothesis tests and practical importance
Testing starts with a null hypothesis, H0, such as no relationship. Software reports a statistic, degrees of freedom and p-value. Wasserstein and Lazar (2016) describe the p-value as the probability, under the test assumptions, of a statistic as extreme as or more extreme than the observed result. A value below a prespecified threshold, commonly .05, provides evidence against H0. The ASA statement emphasises that conclusions also require attention to design, data quality and external evidence.
A directional alternative, such as “greater than”, uses a one-sided probability. A nondirectional alternative uses a two-sided probability (Saunders et al., 2023, p. 618).
Statistical significance and practical importance
Small samples can have limited sensitivity, whereas very large samples can detect very small differences. Small expected effects require more observations. Report effect sizes to assess the practical importance of significant results (Saunders et al., 2023).
Type I and type II errors
Type I error rejects a true H0. The significance level controls its probability under the testing assumptions. Type II error fails to reject a false H0. Lowering alpha from .05 to .01 reduces type I error but can increase type II error when other conditions are held constant. Saunders et al. (2023, pp. 618–619) discuss the conventional emphasis on avoiding type I error.
7. Selecting tests by data type
| Purpose | Categorical or ordinal data | Numerical data |
|---|---|---|
| Normality assessment | Not applicable | Kolmogorov–Smirnov, Shapiro–Wilk |
| Independence of two variables | Chi-square, possibly with justified category merging | Chi-square after categorising |
| Association between categories | Cramer's V, phi for two binary variables | Not applicable |
| Differences between two groups | Kolmogorov–Smirnov or Mann–Whitney U for ordered data | Independent or paired t-tests. The supplied overview suggests Mann–Whitney U for skewed or small independent samples. |
| Differences between three or more groups | Not covered in this table | One-way ANOVA |
| Strength of relationship | Spearman's rho or Kendall's tau for ordered data | Pearson correlation |
| Explained variation in a specified model | Not covered in this table | r² for one predictor, R² for several |
| Predicting an outcome | Not covered in this table | Regression equation |
8. Conditions for individual tests
| Test | Main conditions in the supplied guidance | Suggested response to problems |
|---|---|---|
| Chi-square | Mutually exclusive categories. The cited guidance allows at most 25% of expected counts below five and prefers no expected count below ten in a 2×2 table. | Combine rows or columns only when substantively justified. |
| t-test | Normality, with greater robustness in larger samples. The overview notes reduced concern about unequal variances with similar group sizes. | Consider Mann–Whitney U for skewed or small independent samples. |
| One-way ANOVA | Independent observations, within-group normality (less critical at around 30 per group), and comparable variances (less consequential with a largest-to-smallest size ratio below 1.5) | Use an appropriate repeated-measures method when the same cases are measured repeatedly. |
| Pearson correlation | Numerical variables, random sampling and relevant distributional assumptions | Spearman or Kendall for ordinal variables |
| Regression | Linearity, homoscedasticity, no problematic multicollinearity and normally distributed residuals for the relevant inference | Investigate outliers and transformations and consult multivariate-method references. |
Interpreting correlations
Figure 12.15 of the textbook, developed from Hair and colleagues, describes approximately .20 as weak, .35 moderate, .60 strong, .80 very strong and 1 perfect, with analogous negative magnitudes (Saunders et al., 2023, p. 627). Correlation describes covariation and does not identify causal direction.
Multicollinearity in regression
Hair and colleagues, cited in Saunders et al. (2023, p. 633), flag predictor correlations of .90 or above, tolerance ≤ .10 or VIF ≥ 10 as substantial collinearity. Levene's test assesses equality of group variances, while residual plots help assess regression assumptions.
9. Reporting example, common errors and checklist
Example data-preparation account
Data were entered into [software] as a matrix with a unique questionnaire identifier. Invalid codes, implausible combinations and screening logic were checked. [Number] errors were corrected against the original forms. Missingness accounted for [percentage]%, concentrated in [questions/groups], and was handled through [method]. [Number]-point Likert items were treated as interval when constructing composites. Shapiro–Wilk results for [variable] were [W = …, df = …, p …], informing selection of [parametric/nonparametric test]. Results include the test statistic, degrees of freedom, p-value and effect size.
| Weak practice | Problem | Improvement |
|---|---|---|
| Means of nominal codes, such as subject area or workplace | Arbitrary codes do not support that interpretation. | Report frequencies, percentages and modes. |
| “p = .000” | Rounded output | Report p < .001. |
| Only p-values reported | Large samples can detect small differences. | Add effect sizes. |
| Chi-square with many low expected counts | Conditions may fail. | Consider substantively justified merging and reassess. |
| No data checks | Entry errors can distort conclusions. | Check codes, plausibility and screening. |
| Deleting every incomplete case without investigation | Potential sample distortion | Assess missingness before choosing treatment. |
Before analysis
- Classify each variable and match labels to the questionnaire.
- Prepare an open-question codebook and assess coding agreement.
- Define missing codes and inspect missingness patterns.
- Check invalid codes, implausible combinations and screening logic.
- Use graphical and statistical normality assessments.
- Match the test to data types and its conditions.
- Report statistics, degrees of freedom, p-values and effect sizes.
Conclusion
Many quantitative errors begin with classifying variables incorrectly, overlooking entry errors or missingness, and choosing tests without assessing their conditions.
A convincing results chapter explains preparation, the reason for each analytical choice, and the statistical and practical importance of the findings.
References
- Saunders, M. N. K., Lewis, P., & Thornhill, A. (2023). Research methods for business students (9th ed.). Pearson.
- Wasserstein, R. L., & Lazar, N. A. (2016). The ASA statement on p-values: Context, process, and purpose. The American Statistician, 70(2), 129–133. https://doi.org/10.1080/00031305.2016.1154108
This article synthesises the listed academic sources for methodological reference. Numerical guidelines should be interpreted in context and checked against institutional and target-journal requirements.
