Statistical Analysis in a Thesis — A Practical Guide

Statistical analysis in a thesis represents the greatest challenge of the entire writing process for many students. This applies to master's theses as well as bachelor's and engineering theses where the research section is based on quantitative data. Properly conducted statistical analysis not only strengthens the credibility of results but also demonstrates the author's research competence and can determine the final grade. The problem is that most social science, humanities, or medical students haven't had in-depth prior contact with statistics, and methodological courses at universities often discuss methods too generally. In this comprehensive guide, we explain step by step how to approach data analysis in a thesis — from planning, through tool and method selection, to presentation and interpretation of results.

Statistical Research Planning

Before you begin collecting data, you must plan the research from a statistical perspective. This is a crucial stage that many students skip, yet it helps avoid costly mistakes in later stages. Planning includes: operationalization of variables (transitioning from theoretical concepts to measurable indicators), determining variable types (quantitative continuous/discrete, qualitative nominal/ordinal), formulating research hypotheses, and selecting an appropriate sample size.

Variable operationalization is the process of transforming abstract concepts (e.g., 'job satisfaction,' 'digital competencies,' 'quality of life') into specific, measurable indicators. Every variable in your work must have a clearly defined measurement method — e.g., 'job satisfaction measured by the MSQ questionnaire on a 1-5 scale.' This is not a formality — precise operationalization allows other researchers to evaluate and replicate your study.

Sample size directly impacts the statistical power of the study. A sample too small may fail to detect an existing effect (Type II error — false negative), while one too large generates unnecessary costs and time. Statistical power analysis before data collection will help determine the minimum number of observations. Tools like G*Power (free) allow calculating the required sample size for different tests and assumed effect sizes. Cohen's convention: small effect (d = 0.2) requires n ≈ 400 per group, medium (d = 0.5) requires n ≈ 65, large (d = 0.8) requires n ≈ 26 — for a two-tailed t-test at power 0.8.

Data Preparation for Analysis

After collecting data, before proceeding to actual statistical analysis, you must properly prepare the data. This stage — though often omitted from methodological descriptions — is critical for result quality. Preparation includes: data coding (assigning numerical values to categories, e.g., female = 1, male = 2), import to statistical software, quality control (searching for erroneous, missing, and outlying values), and verification of test assumptions.

Missing values are a common problem, especially in survey research. The strategy for handling missing data depends on their nature: if missing values are random (MCAR — Missing Completely at Random), they can be removed without risk of distorting results (listwise deletion); if they depend on observed variables (MAR), consider imputation (replacing values with mean, median, or using the MICE algorithm). If the percentage of missing values exceeds 5-10%, make sure to report this in your work and describe the strategy used.

Outliers can distort parametric test results. Identify them using box plots or the Z > 3 rule (value more than 3 standard deviations from the mean). Before removing an outlier, ensure it's not the result of a data entry error — if it's real, consider using a non-parametric test instead of removing the observation.

Checking Statistical Test Assumptions

Checking assumptions is the step that distinguishes solid statistical analysis from amateur work. The main assumptions of parametric tests are: normality of the dependent variable distribution, homogeneity of variance (homoscedasticity), and independence of observations. Normality is checked using the Shapiro-Wilk test (sample n < 50) or Kolmogorov-Smirnov (n ≥ 50), and also visually: a histogram with a normal curve overlay and Q-Q (quantile-quantile) plot.

Homogeneity of variance is checked using Levene's test (for ANOVA and t-test). If Levene's test is significant (p < 0.05), variances are not homogeneous — use the Welch variant of the t-test or a non-parametric counterpart. Independence of observations is a design assumption, not a statistical one — if your data contain repeated measurements on the same individuals or nesting (e.g., students in classes), you must use appropriate tests (paired t-test, repeated measures ANOVA, mixed models).

Remember the principle: with large samples (n > 30-50 per group), parametric tests are robust to moderate departures from normality thanks to the central limit theorem. However, with small samples or clearly skewed distributions, non-parametric tests are the safer choice.

Selecting Statistical Tests — Decision Algorithm

Statistical test selection depends on several factors: dependent variable type, independent variable type, number of compared groups, and dependencies between observations. Here is a practical decision algorithm you can apply in your thesis. For one quantitative variable and two independent groups: Student's t-test (normality met) or Mann-Whitney U test (normality not met). For one quantitative variable and two dependent groups: paired t-test or Wilcoxon test.

For one quantitative variable and three or more independent groups: one-way ANOVA (parametric) or Kruskal-Wallis test (non-parametric), with post-hoc tests (Tukey, Bonferroni) to identify differing pairs of groups. For one quantitative variable and three or more dependent measurements: repeated measures ANOVA or Friedman test. For two quantitative variables: Pearson correlation (normality) or Spearman (no normality), linear regression. For two qualitative variables: chi-square test (expected frequencies ≥ 5) or Fisher's test (small frequencies).

Statistical Analysis Tools — Overview

Statistical analysis tools are more accessible today than ever before. SPSS (IBM SPSS Statistics) remains the most popular program among social science and medical students due to its intuitive graphical interface (point-and-click). University or student licenses are usually available. SPSS is ideal for basic analyses: descriptive statistics, t-tests, ANOVA, correlation, regression, chi-square.

Jamovi is a free alternative to SPSS with a similar graphical interface, built on R. It offers most tests needed in a thesis and generates results in a tabular format close to APA standards. JASP is another free option, distinguished by its support for Bayesian statistics. Both programs are excellent choices for students who don't have access to SPSS.

R and Python offer the greatest flexibility and are preferred in exact sciences, bioinformatics, and advanced analytics. R with packages ggplot2 (visualization), dplyr (data manipulation), car (diagnostics), psych (psychometrics), and lavaan (structural equation modeling) covers nearly every analytical need. Python with libraries pandas, scipy, statsmodels, pingouin, and seaborn offers similar capabilities. Tool choice should be guided by research specifics, supervisor requirements, and your programming skills.

Excel, though widely available, has limited statistical capabilities. It's suitable for descriptive statistics and simple tests (t-test, chi-square) but can't handle advanced analyses. If your supervisor accepts Excel, use it for simple analyses but be aware of its limitations.

Presenting Results in a Thesis

Presenting statistical analysis results in a thesis requires following several principles that stem from reporting standards (APA, CONSORT, STROBE). Each result should include: the test statistic value (t, F, χ², U, r...), degrees of freedom (in parentheses — e.g., t(48) = 2.31), p-value (exact, e.g., p = .021, not 'p < 0.05'), effect size measure (Cohen's d, η², r), and — ideally — confidence interval.

Tables and charts are key elements of results presentation. Tables should be readable, properly labeled (title above the table, notes below), and formatted according to APA standards. Charts (histograms, box plots, bar charts, scatter plots) should have readable axes, a legend, and not be overloaded with data. Every table and chart must be discussed in the text — don't place them without reference.

Describe results objectively in the results section, and leave interpretation for the discussion. Correct description: 'Analysis revealed a statistically significant difference in satisfaction levels between the experimental group (M = 4.2, SD = 0.8) and control group (M = 3.5, SD = 0.9), t(98) = 4.12, p < .001, d = 0.82.' Incorrect: 'The test result showed that the method works' (no values, no effect size).

Results Interpretation — Statistical Discussion

Interpreting statistical results in the discussion requires the ability to connect numbers with theoretical context. Don't repeat numerical values — refer to them at a substantive level. Questions the discussion should answer: Do results confirm or reject hypotheses? How do they compare to previous research findings? What is the practical significance of the observed effects? What are possible explanations for unexpected results?

A statistically significant result (p < 0.05) doesn't always mean a practically significant result. With very large samples, even trivial differences can be statistically significant. This is why effect size is so important — a small Cohen's d (< 0.2) with a significant p suggests that the effect, while statistically detectable, has no practical significance. Conversely, a statistically non-significant result doesn't mean there is no effect — it may result from insufficient sample power.

Common Statistical Errors in Theses

Based on years of experience reviewing and consulting on theses, the most common errors are: failure to check test assumptions (normality, homoscedasticity) — this is reviewers' most common criticism; applying parametric tests to ordinal data (Likert scale without justification for treating it as continuous); multiple t-tests instead of ANOVA; ignoring multiple comparisons (no Bonferroni correction); confusing correlation with causation ('students who exercise more pass exams better' doesn't mean exercise improves results — there may be a third variable).

Additional errors: reporting only p-values without effect size measures; using phrases like 'the result showed a tendency' when p > 0.05 (a result is either significant or not — 'tendency' is an unjustified interpretation); confusing simple correlation with regression; failure to describe missing data and outliers; running too many tests on a small sample (fishing for significance); and — most seriously — fabricating or manipulating data to achieve expected results.

If statistical analysis is not your strong suit, consider consulting with a specialist. A statistician will help select appropriate methods, properly configure analyses, and interpret results — it's an investment that pays off in the quality of the entire thesis. The best consultation is at the research planning stage (before data collection), not after the fact.

We offer professional statistical analysis for theses — from test selection to methodological description. Also read our guide to choosing statistical tests to understand the logic behind each test. Contact us for a personalized quote.