Statistical analysis for a psychology master’s thesis - what to do with collected data

You have collected responses for your psychology master’s thesis, but the spreadsheet still does not look like a results chapter? The next step is to check what the columns mean, calculate scores using the instrument’s instructions and connect each analysis to a specific question. Only then can the numbers become an answer you can explain clearly in your thesis.

Perhaps you have a form export, a statistical software file and a message from your supervisor asking you to prepare the results. Do not start by running one test after another. First reconstruct the route from a person’s responses to the score you intend to analyse. This guide covers the stage after data collection: checking the material, working through two different analyses and writing the chapter, rather than designing another questionnaire.

My name is Wioletta Boćkowska and I have run StudentSupport since 2018. In the subject area Psychology, I help connect questions, instruments, calculations and writing. If you already have data, you can request Statistical analysis for bachelor’s and master’s theses within an agreed scope, from preparing the dataset to calculations and reporting. I agree that scope after reviewing the material, before work begins.

First establish what is actually in the file

Keep an unchanged source export and work on a copy. Separate the responses, variable dictionary and decision record. In the dictionary, explain each column, its allowed values, unit and missing-value code. A code of 9 might be a valid response to one question and a missing-value marker in another. The software cannot recognise that distinction for you.

Check the unit of observation. If one person completed the questionnaire twice, two rows do not mean two independent participants. You need a way to match the measurements and an analysis that accounts for that structure. Responses from people in the same class or team may also require attention to their shared context. Column names alone cannot establish independence.

Your Polish university’s requirements and agreements with your supervisor still matter. The UKW Faculty of Psychology’s 2025 guide distinguishes the method, analysis of results and discussion. This is an example of local guidance, not a single chapter structure compulsory across Poland. Check your own department’s current requirements instead of importing page counts or a structure from the internet.

From responses to scale scores: check the scoring key

In a psychology thesis, several items often contribute to one score, but the spreadsheet itself does not tell you how to calculate it. You need the instructions for the specific instrument, language version and scoring procedure. Establish which items belong to each subscale, which require reversing and whether the final indicator is a sum, mean or transformed score. Do not combine different subscales just because their response options look alike.

A simple example shows how an error can enter the analysis. For a teaching item with responses from 1 to 5, reversing means calculating 6 minus the response: 1 becomes 5 and 4 becomes 2. That rule applies to this range, not every questionnaire. Reversing an already transformed column a second time restores the original values. Give raw and recoded responses different, clear names.

Reliability in your sample and the properties reported in the manual are different pieces of information. APA’s JARS-Quant recommends reporting coefficients for the analysed scores where possible. Copying a value from the adaptation paper is not enough. A high alpha also does not establish that every item measures one construct or that the score has a valid interpretation (Sijtsma, 2009). Choose the assessment to suit the instrument’s structure, rather than assuming alpha is always required.

Do not delete items merely to improve a coefficient and then describe the modified set as the complete original scale. If the analysis reveals a problem, first check coding and compliance with the instructions. Any further change needs a reason and a record. Do not derive individual diagnoses from a student research score or assign categories that the instrument does not establish.

Missing and unusual values: decide before choosing a convenient result

An empty cell is not zero. First count missing values for each variable and check whether they cluster in particular items or groups. Distinguish one unanswered item from an entirely missing measurement. Your sample description may therefore include more people than a particular analysis uses. For each result, the actual denominator must be clear.

Do not adopt a universal rule such as replacing up to 20% of missing responses with the mean. Check the scale instructions, the pattern of missingness and the rationale for your procedure. Deleting every incomplete row is also a decision that may change the composition of the analysed sample. JARS-Quant calls for transparency about missing data and outliers. In practice, record the rule, the reason and the number of observations affected.

An unusually high value might be an entry error, but it might also be a valid response. A value of 55 on a 1-to-5 item requires investigation; a valid score near a scale boundary is not automatically grounds for exclusion. Examine the influence of unusual observations and report a justified sensitivity analysis. Do not select the version with or without observations according to which produces a smaller p value.

Connect the hypothesis to columns and an answer

For each hypothesis, write a short chain: question, variables, scoring procedure, study design, analysis and the form of the answer. This often exposes a column labelled stress that contains one item, although the method describes a multi-item scale. Or a supposed before-and-after comparison for which the file contains only the final measurement. A more advanced test cannot repair that gap.

The question and data structure determine the analysis, not the outcome of a normality test alone. Examine distributions, plots, relationships and independence in the context of the particular method. A broader comparison is available in Statistical analysis in a master’s thesis: choosing tests and reporting results. Here, we will work through two prepared datasets rather than build another catalogue of tests.

If a new question emerges after you inspect the data, you can investigate it, but label the analysis exploratory. Retain the results of previously planned analyses, including unexpected findings. When running many tests, establish how you will address the risk of false positives. Adding subgroups until significance appears changes the meaning of the inference and must not disappear from your account of the analysis.

Example 1: the relationship between two scale scores

The following numbers come from author-created synthetic teaching data. They are not client results or a study of students. A is an illustrative overload score and B an illustrative study-organisation score. They do not represent specific questionnaires and have no norms or interpretation thresholds. The example begins after valid composite scores have been calculated; it does not pretend to assess reliability from totals alone.

We ask whether a higher A score is associated with a lower B score. The prepared dataset contains 60 independent pairs with no missing values. A has a mean of 50.30 and a standard deviation of 9.78; the corresponding values for B are 49.34 and 9.25. A scatterplot is needed to check whether a linear summary makes sense and whether one point dominates the relationship. Pearson’s coefficient describes a linear association here; ordered responses alone would not justify that model.

Here, r is the effect-size measure: it expresses the direction and strength of the linear association. Replacing it with a label such as moderate is not a complete interpretation. Its meaning depends on theory, measurement and context. This example does not establish that better study organisation reduces overload. Measurement at one time point cannot establish the direction of the relationship or the role of other characteristics.

Example 2: comparing two independent groups

A second, separate teaching dataset asks whether the mean A score differs between people reporting paid work and those reporting no paid work. Each group contains 20 observations and each person belongs to only one group. We are not splitting a continuous score into low and high categories at its median. Group membership has its own meaning, defined independently of the outcome being analysed.

After checking distributions and unusual observations, this example uses a two-sided Welch t test. It compares means without assuming equal group variances. That does not make it immune to every data problem: independence, distribution shape and influential extremes still matter. If the same people had been measured before and after an event, this analysis would not match the study structure.

The difference in points is the effect-size measure here, not just a supplementary detail beside p. The interval includes zero as well as differences in both directions. Do not write that the groups are identical or that paid work makes no difference. The data do not resolve the issue with that precision. They also do not establish an effect of employment: people were not randomly assigned to groups and may have differed in other ways.

Both examples demonstrate a way of reasoning, not the results you should expect to obtain. The first effect is a correlation; the second is a mean difference in score units. In an actual thesis, use suitable numerical precision and state how intervals were calculated. Exporting a software table alone does not explain why an analysis answers the hypothesis.

What should you say about p and a confidence interval?

The American Statistical Association’s 2016 statement explains that p is not the probability that the hypothesis is true and does not measure effect size. Therefore, p = .003 does not mean 99.7% certainty that a theory is true, and p = .216 does not mean a 21.6% chance that a relationship exists. Interpretation requires the model, an effect estimate and the study context.

A confidence interval helps you assess estimation precision. In the second example, the span from -2.20 to 9.42 points shows why the label no differences would be too strong. In the frequentist interpretation, 95% refers to the long-run coverage of the interval procedure under its assumptions. It is not the probability that a fixed parameter lies inside this one interval after it has been calculated.

How do you turn the analyses into a results chapter?

Arrange the chapter around research questions, not the order in which you clicked through the software. Start by explaining which material ultimately entered the analyses. Then provide the necessary variable descriptions and move to the answers. Tables and figures may be appropriate in the thesis if your requirements allow them. The prose should explain the main finding instead of repeating every cell of an output table.

A useful paragraph begins clearly: to answer the question about the relationship between A and B. It then identifies the method, reports the result, effect and uncertainty, and ends with a plain-language answer. Keep signs consistent: a positive coefficient cannot be described as a negative association, and the direction of first group minus second group must not change halfway through. Check that the prose and attached output use the same dataset version.

JARS-Quant structures analysis reporting and distinguishes primary hypotheses, secondary hypotheses and exploration. It is not a ready-made template for every master’s thesis. To review your own writing, ask a simpler question: can the reader understand what was calculated, using which observations and with what limitation? A list saying only supported or unsupported does not provide enough information to assess the finding.

Where do results end and discussion begin?

Results describe what was found in your material. Discussion compares it with theory and other research, considering explanations and limitations. For the first example, you might consider whether organisation and overload refer to the same period and whether both scores are self-reports. You cannot write that an alternative explanation was tested if the relevant variable was never collected.

For the second example, one limitation is the imprecise difference estimate. Explain its consequence: the material cannot confidently distinguish a small difference from a larger one in the direction indicated by the means. That is more informative than saying the sample was small. For more on this chapter, see How to write the discussion in a bachelor’s or master’s thesis.

What should you prepare to request analysis from StudentSupport?

The most useful package lets someone connect each question to the data. Include the topic and outline, questions and original hypotheses, a dataset without identifying information, a column dictionary, a blank questionnaire and scoring instructions. If you already have output, include the analysis file or settings, rather than selected screenshots alone. Identify previous exclusions, transformations and supervisor comments.

Before sending material, remove names, addresses, emails and other details that could identify participants, including free-text responses. Replacing a name with a number may be insufficient if the remaining details identify the person. A dictionary, blank questionnaire and study description are often enough to discuss the initial scope. You do not need to send the complete material simply to establish what support to request.

If you study in Poland, include your Polish university’s current requirements and agreements with your supervisor. The language of this guide does not determine the language of the commissioned text: agree that alongside scope, format, price and deadline. If the problem extends beyond data to connecting theory with several chapters, a larger project may fall within Bachelor’s and master’s thesis writing. Requesting analysis does not automatically extend the scope in that way.

When reviewing the work, return to the original questions. Check that the answers concern the right variables, sample counts agree, decisions are described and files arrive in the agreed format. An unexpected result may still come from a correctly performed analysis. Quality means being able to follow the reasoning and calculations, not obtaining support for every hypothesis.

Questions about data and reporting results

Sources and how they are used

American Psychological Association (2024). JARS-Quant, Table 1: Information Recommended for Inclusion in Manuscripts That Report New Data Collections Regardless of Research Design. Reporting standard, particularly Psychometrics, Data Diagnostics, and Statistics and Data Analysis. Address: https://apastyle.apa.org/jars/quant-table-1.pdf. Accessed: 6 November 2026.

Wasserstein R.L., Lazar N.A. (2016). The ASA’s Statement on p-Values: Context, Process, and Purpose. The American Statistician, 70(2), 129-133. DOI: 10.1080/00031305.2016.1154108. The six principles appear in the official ASA release: https://www.amstat.org/asa/files/pdfs/p-valuestatement.pdf. Used for p-value interpretation and transparent reporting, not as a universal analysis algorithm.

Faculty of Psychology, Kazimierz Wielki University in Bydgoszcz (2025). Jak przygotować pracę magisterską? Wskazówki dla Studenta. An example of Polish university guidance, especially the sections on structure and chapter content. Address: https://psychologia.ukw.edu.pl/download/70232/Poradnik-praca-magisterska.pdf. Local recommendations are not presented as requirements of all universities.

Sijtsma K. (2009). On the Use, the Misuse, and the Very Limited Usefulness of Cronbach’s Alpha. Psychometrika, 74(1), 107-120. DOI: 10.1007/s11336-008-9101-0. Used to distinguish alpha, scale structure and validity of interpretation.

NIST/SEMATECH (n.d.). e-Handbook of Statistical Methods, 1.3.5.3: Two-Sample t-Test for Equal Means. Address: https://www.itl.nist.gov/div898/handbook/eda/section3/eda353.htm. NIST Dataplot: Correlation Confidence Limits. Address: https://itl.nist.gov/div898/software/dataplot/refman1/auxillar/corrconf.htm. Documentation of the example methods; accessed: 6 November 2026.

The numerical examples are author-created demonstrations using two separate synthetic datasets. Tests are two-sided; the correlation interval uses Fisher’s transformation and the mean-difference interval uses the t distribution with Welch degrees of freedom. They are not two tests of one hypothesis in a real study. The data provide no evidence about students, employment or the properties of psychological scales.