This guide explains how to choose, check, run and report a chi-square test in a psychology dissertation. It focuses on defensible reasoning rather than mechanical software clicks. You will learn how to distinguish independence from goodness-of-fit tests, inspect expected counts, choose exact or model-based alternatives, interpret residuals, report effect sizes and connect the analysis to your research question.
psychology dissertation chi-square test decisions begin with a simple question: are you analysing counts in categories, and does the design make those counts independent? A chi-square analysis can test whether two categorical variables are associated or whether one categorical variable follows a specified distribution. It cannot show that one variable caused another, and a significant result does not identify which cells created the pattern.
What does a psychology dissertation chi-square test answer?
The Pearson chi-square statistic compares observed counts with counts expected under a null model. Each cell contributes a squared observed-minus-expected difference divided by its expected count. Large discrepancies increase the statistic. The p-value then describes how incompatible the observed table is with the null model, assuming the sampling design and test conditions are appropriate.
Table of Contents
Psychology dissertations usually use one of two forms. A test of independence asks whether two categorical variables are associated. For example, is help-seeking preference associated with study mode? A goodness-of-fit test asks whether the distribution of one categorical variable matches prespecified proportions. For example, do four coping-strategy choices occur in equal proportions, or in proportions predicted by a theory?
| Research question | Data structure | Likely test |
|---|---|---|
| Is therapy preference associated with age group? | Two categorical variables from independent participants | Pearson chi-square test of independence |
| Do observed attachment classifications match theoretical proportions? | One categorical variable with prespecified expected proportions | Chi-square goodness-of-fit test |
| Did the same participants change yes/no responses after an intervention? | Paired binary responses | McNemar test, not an ordinary independence test |
| Does a binary outcome differ after adjusting for age and baseline severity? | Categorical outcome with covariates | Logistic regression |
The distinction matters because hypotheses, expected counts and degrees of freedom are formed differently. Write the research question first, then build the analysis around it. The site’s psychology dissertation hypotheses guide can help align the null and alternative hypotheses with the design.
When is a psychology dissertation chi-square test appropriate?
A defensible test requires categorical outcomes represented as frequencies. Nominal categories have no inherent order, such as preferred support channel. Ordinal categories have an order, such as low, moderate and high distress, but Pearson chi-square treats them as labels and does not use that order. If the ordering is central to the hypothesis, an ordinal model or a trend test may be more informative.
Use raw counts, not percentages, means or scale scores. A table displaying percentages is useful for interpretation, but the test itself must be based on the number of observations in each cell. If a multi-item scale produces a continuous score, categorising it solely to use chi-square usually discards information and can create arbitrary boundaries. Analyse the original score with a method suited to its distribution unless theory, validation or a clinical convention justifies categories.
Independence of observations
Ordinary Pearson chi-square assumes that each observation contributes to one cell and observations are independent. A student sampled twice, a participant who selects multiple response options, pupils clustered within classrooms, or matched cases violate that simple structure. Repeated binary responses may require McNemar’s test. Clustered or longitudinal categorical outcomes may require a multilevel, marginal or mixed-effects model. Review the multilevel modelling guide when people are nested in settings or measured repeatedly.
Mutually exclusive and meaningful categories
Every participant should have a clear coding rule. Categories must not overlap unless a specialist multiple-response analysis is planned. Combine categories only when theory and interpretation support the new grouping, not simply because a p-value is inconvenient. Document any recoding before looking at the test result where possible. This protects the construct meaning described in your operational definitions.
A prespecified null model
For a test of independence, the null model says row and column variables are independent. Expected counts are calculated from the marginal totals. For goodness of fit, expected proportions must come from theory, prior evidence or a clearly stated equal-distribution hypothesis. They should not be estimated opportunistically from the same observed frequencies. The NIST chi-square goodness-of-fit guide explains that the result depends on how observations are grouped and that adequate sample size is needed for the approximation.
Build and inspect the contingency table first
Begin with a frequency table before requesting any inferential output. Check category labels, totals, missing codes and impossible combinations. Compare table totals with the analysed sample. Values such as 99 or a blank string can accidentally become genuine categories if missing data were not defined correctly.
Display both counts and an interpretable percentage. Row percentages answer questions such as “within each study mode, what proportion preferred each form of support?” Column percentages reverse the conditioning. State which denominator you use. Overall percentages can conceal the comparison that the hypothesis concerns.
Observed and expected counts
An observed count is the number actually recorded. Under independence, an expected count equals the row total multiplied by the column total and divided by the grand total. It is not a desirable target or a predicted individual response. It is the cell frequency implied by the null model and the table margins.
Suppose 160 independently sampled students are classified by study mode and preferred wellbeing support. The observed table may show that online students choose digital support more often than campus students. The chi-square statistic summarises how far the whole table departs from independence. It does not test each cell separately, so cell-level follow-up requires residuals and multiplicity-aware interpretation.
| Check | Why it matters | Action if problematic |
|---|---|---|
| Raw cell counts | Chi-square is a count-based test | Recreate the table from individual-level data |
| Expected counts | Sparse cells weaken the asymptotic approximation | Consider exact, Monte Carlo or model-based methods |
| Independent units | Repeated or clustered records overstate information | Use paired, marginal or multilevel analysis |
| Missing values | Silent exclusions can change margins and interpretation | Report analysed N and missingness by variable |
| Category definitions | Post hoc collapsing can change the hypothesis | Justify and disclose recoding |
Expected counts and sparse data
The Pearson test uses a chi-square reference distribution as an approximation. Sparse expected counts can make that approximation unreliable. A common diagnostic says that no expected cell should be below 1 and at least 80% should be 5 or more. This is a practical rule, not a universal law. Table dimensions, imbalance and the inferential method all matter.
Do not confuse observed counts with expected counts. A cell can contain three observed participants yet have an expected count above five, or the reverse. Ask software to display expected frequencies and report the nature of any sparsity. The official R documentation for Fisher’s exact test describes the 1-and-5 diagnostic used by its hybrid procedure and explains exact and Monte Carlo options for larger tables.
What to do when cells are sparse
For a 2 by 2 table, Fisher’s exact test is a standard option when the chi-square approximation is questionable. It tests conditional independence with fixed margins and can provide an odds-ratio estimate and confidence interval. For larger tables, exact computation may be intensive, while Monte Carlo inference can estimate the conditional p-value. State the number of simulations and, ideally, set and record a random seed for reproducibility.
Combining sparse categories can be defensible when categories are substantively similar and the combined construct still answers the same question. It is not defensible to merge groups only until significance appears. If important groups are rare, preserve them and consider a suitable model, obtain more data if ethical and feasible, or present careful descriptive evidence with honest uncertainty.
When covariates, interactions or complex sampling are central, regression often answers the scientific question better. Binary logistic regression can estimate adjusted associations; multinomial or ordinal models support outcomes with more categories. A significant unadjusted chi-square association may reflect confounding rather than a direct relation. Use the confounding variables guide to decide which adjustments are justified.
Choose the right categorical-data test
| Design feature | Preferred starting method | Important note |
|---|---|---|
| Two independent categorical variables | Pearson chi-square independence test | Inspect expected counts and residuals |
| One categorical variable versus fixed proportions | Chi-square goodness-of-fit test | Expected proportions must be justified |
| Small or sparse 2 by 2 table | Fisher’s exact test | Report odds ratio and interval where appropriate |
| Same people measured twice on a binary outcome | McNemar test | Analysis depends on discordant pairs |
| Ordered categories with a directional hypothesis | Ordinal model or trend test | Uses ordering ignored by ordinary chi-square |
| Categorical outcome with covariates | Logistic, multinomial or ordinal regression | Supports adjusted estimates |
| Clustered categorical outcome | Multilevel or marginal categorical model | Accounts for dependence within clusters |
Calling chi-square “nonparametric” does not make it assumption-free. The nonparametric tests guide explains why design, measurement level and sampling still determine whether a procedure is credible.
Plan sample size before collecting data
A minimum total sample is not enough because power and approximation quality depend on how cases are distributed among cells. Plan using expected category proportions, the smallest effect worth detecting, alpha, desired power and degrees of freedom. Rare categories may produce sparse cells even when the overall sample seems large.
For an independence test, Cohen’s w is often used for prospective power calculations. Define the smallest effect of substantive interest rather than choosing a generic “medium” effect without context. Sensitivity analysis is useful when prior proportions are uncertain: vary plausible margins and effects, then explain how the final target protects the main comparison. The psychology dissertation power analysis guide covers defensible planning and attrition allowances.
Recruitment targets do not repair non-probability sampling or measurement bias. A very large convenience sample can produce a small p-value for a trivial association while remaining unrepresentative. Describe the sampling frame, eligibility, recruitment channels and exclusions, and keep population claims proportionate.
Run the analysis in SPSS or R
SPSS workflow
For a test of independence, select Analyze, Descriptive Statistics and Crosstabs. Put the explanatory grouping variable in rows and the outcome in columns if that orientation makes percentages easier to read. Under Statistics, request Chi-square and an association measure. Under Cells, request observed counts, expected counts, relevant row or column percentages and adjusted standardized residuals when available.
Review the case-processing summary before the test table. Confirm the analysed N and missing cases. Read the footnote describing small expected counts. SPSS may display continuity correction and Fisher’s exact test for a 2 by 2 table; report the method selected in advance for the data conditions rather than choosing whichever p-value is smallest. The SPSS psychology dissertation workflow provides further guidance on preserving syntax and output decisions.
R workflow
In R, create a table from factors and inspect it before testing. The official R documentation for chisq.test() covers independence and goodness-of-fit tests, expected counts, Pearson residuals and simulated p-values. A transparent workflow might use tab <- table(data$study_mode, data$support_choice), print tab, run fit <- chisq.test(tab), then inspect fit$expected and fit$stdres.
For a prespecified goodness-of-fit test, provide the expected probability vector with p= and ensure it sums to one. For a sparse 2 by 2 table, use fisher.test(tab). Save the script, package versions and data-cleaning decisions. Code does not replace reasoning, but it makes the route from raw records to reported table auditable.
Interpret the omnibus result without overstating it
A significant independence test indicates evidence against the null of independence for the table as a whole. It does not prove a causal effect, show that every category differs, or reveal the direction of association. A non-significant result does not prove equality or absence of a meaningful relationship. It may reflect limited precision, sparse cells or an effect smaller than the study could detect.
The American Statistical Association advises that conclusions should not depend only on whether a p-value crosses a threshold. Its summary of the ASA p-value statement emphasises proper interpretation and transparent reporting. Connect the statistical result to estimates, uncertainty, design quality and the psychological question.
Use residuals to locate the pattern
Cell residuals compare observed and expected counts. Adjusted standardized residuals place discrepancies on an approximately standardised scale under suitable conditions. Positive residuals indicate more observations than expected under independence; negative residuals indicate fewer. Inspect their pattern after the omnibus test, not as an unplanned collection of independent tests.
If you use residual thresholds to flag cells, recognise that examining many cells creates a multiplicity problem. Prespecify key contrasts when theory permits, or apply an appropriate adjustment and label exploratory follow-up clearly. The existing multiple comparisons guide explains how to define a family of tests and avoid selective reporting.
Report effect size and interpretable estimates
A p-value is affected by sample size and does not measure association strength. Report an effect size. Phi is suitable for a 2 by 2 table, while Cramér’s V extends the idea to larger tables. Cramér’s V ranges from zero to one, but interpretation depends on table dimensions, base rates, measurement quality and practical context. UCLA’s annotated contingency-table guide explains the calculation and why Cramér’s V is useful for nominal association.
For a 2 by 2 table, an odds ratio with a confidence interval can be more directly interpretable than V. State which category is the reference and avoid describing an odds ratio as a risk ratio. When outcomes are common, those quantities can differ substantially. If the design supports risk estimates, present absolute proportions alongside relative measures.
Do not apply universal labels mechanically. A V of .10 might matter for a common low-cost educational decision but be negligible for another question. Compare the estimate with theory, prior evidence and a smallest effect of interest. Discuss uncertainty and avoid treating a point estimate as exact.
Worked psychology dissertation example
Imagine a cross-sectional dissertation asking whether preferred wellbeing support is associated with study mode among 240 students. Study mode has two levels, campus and online. Support preference has three levels, individual counselling, peer group and digital self-help. Each participant selects one option, so the table is 2 by 3 and each person contributes once.
The student first verifies coding, missingness and totals. Counts and row percentages show that digital self-help is more common among online students, while peer-group support is more common among campus students. Expected counts all exceed five, so the Pearson approximation is reasonable. The omnibus test is significant. Cramér’s V indicates a modest association.
Adjusted residuals suggest that the clearest discrepancies occur in digital self-help and peer-group cells. Because these follow-ups were planned as descriptive interpretation of one omnibus table, the student reports the full residual pattern and avoids presenting every cell as a separate discovery. The discussion notes that study mode and support preference were measured together, so direction and causality cannot be established. Access, age or work commitments may confound the association.
A concise results sentence could read: “Support preference was associated with study mode, χ²(2, N = 240) = 10.84, p = .004, Cramér’s V = .21. Online students selected digital self-help more often than expected under independence, whereas campus students selected peer-group support more often.” Replace these illustrative values with the saved output, and add the observed table so readers can verify the interpretation.
Common mistakes and how to repair them
Testing percentages instead of counts
Percentages describe a table but do not replace frequencies in the test. Return to participant-level data, generate counts and state the denominator for every percentage.
Ignoring repeated or clustered observations
An ordinary test treats records as independent. If the same participant appears more than once or people are nested in groups, choose a method that models dependence. Do not solve the problem by deleting meaningful records without justification.
Collapsing categories after seeing the p-value
Post hoc merging can change the construct and inflate researcher flexibility. Base grouping on theory, measurement documentation and an analysis plan. If exploratory recoding is necessary, disclose the original categories and run a sensitivity analysis where feasible.
Reporting only chi-square and p
Readers need the contingency table, percentages, analysed N, expected-count diagnostics, degrees of freedom, exact test name, effect size and follow-up rationale. Add confidence intervals for interpretable estimates when the selected method supports them.
Claiming causation from association
Chi-square measures evidence about a distributional association, not a causal mechanism. Tie causal language to design features such as random assignment, temporal ordering and control of plausible alternatives. Most cross-sectional tables support association language only.
Reporting checklist
- Name the test and state whether it assessed independence or goodness of fit.
- Define every category and explain any recoding or exclusions.
- Report the analysed sample and missing-data handling.
- Present observed counts and clearly labelled percentages.
- Describe expected-count diagnostics and any exact or simulated method.
- Report χ², degrees of freedom, N and the exact p-value where practical.
- Report Cramér’s V, phi, odds ratio or another justified effect measure.
- Explain residual or contrast follow-up and any multiplicity adjustment.
- Interpret association strength in psychological and design context.
- Keep causal claims proportionate to the design.
For quantitative reporting more broadly, consult the psychology dissertation results section guide. Preserve the full software output and syntax in your project records, but include only decision-relevant material in the dissertation.
Frequently asked questions
Can I use chi-square with Likert items?
You can analyse response-category counts, but ordinary Pearson chi-square ignores the ordering between Likert categories. If the hypothesis concerns an ordered shift, an ordinal model or trend-sensitive method may use the information more effectively. Avoid turning a validated multi-item scale into arbitrary categories without justification.
Is chi-square suitable for a small sample?
Suitability depends on expected counts and table structure, not only total N. Inspect the expected frequencies. For sparse 2 by 2 tables, Fisher’s exact test is often appropriate. Larger sparse tables may need exact, Monte Carlo or model-based methods.
Should I report Yates’ continuity correction?
Some software displays it automatically for 2 by 2 tables. It can be conservative and should not be selected merely because it changes significance. Follow disciplinary guidance, prespecify the approach and consider exact inference when approximation is doubtful.
What does a significant chi-square test mean?
It means the observed counts are difficult to reconcile with the specified null model under the test assumptions. For independence, it supports an association between the categorical variables. It does not identify causality or automatically show which cells differ.
How do I know which cells matter?
Inspect observed versus expected counts and adjusted standardized residuals. Interpret the whole pattern, control or acknowledge multiplicity, and avoid selecting only cells that support the preferred story.
What effect size should I report?
Use phi for a 2 by 2 table or Cramér’s V for a larger independence table. For 2 by 2 data, an odds ratio with a confidence interval may be especially informative. Choose an estimate that matches the design and explain its practical meaning.
Can chi-square control for confounders?
A basic two-way chi-square test cannot adjust for several covariates. Stratified analysis or logistic, multinomial or ordinal regression may be more suitable. Select covariates using substantive reasoning rather than automatic significance screening.
Conclusion
A strong psychology dissertation chi-square test starts with categorical counts, independent observations and a clearly defined null model. Build and inspect the contingency table, check expected frequencies, choose an exact or model-based alternative when needed, and interpret the omnibus result through residuals and effect estimates. Report the full table and keep claims aligned with the design.
If you need ethical academic support, use feedback to check the logic connecting your question, coding, assumptions, analysis and conclusions. Retain responsibility for your data, decisions and final writing, and follow your institution’s rules on permitted assistance.
