ACADEMIC WRITING GUIDE
How to Conduct and Report Exploratory Factor Analysis in a Dissertation (With Worked Examples)
A questionnaire can contain twenty questions but measure only a few broader ideas. Some questions may reflect confidence, others may reflect access to support, and others may reflect motivation. Exploratory factor analysis (EFA) helps researchers investigate whether patterns of correlations among observed items can be described by a smaller set of underlying common factors.
For dissertation students, the challenge is not simply generating a factor-loading table in software. It is deciding whether EFA fits the research question, preparing suitable data, choosing an extraction and rotation method, deciding how many factors to retain, evaluating the resulting pattern, and reporting decisions transparently.
This guide walks through those decisions with three worked teaching examples. All datasets, statistics, and factor loadings below are hypothetical illustrations, not results from collected research. Your institution's requirements and the characteristics of your actual data must determine the analysis you report.
What is exploratory factor analysis?
EFA is a statistical method that examines relationships among measured variables to explore whether their shared variation can be represented by a smaller number of latent factors. A latent factor is not measured directly; its presence is inferred from how observed items covary.
Imagine a questionnaire with statements such as “I can organise my study time” and “I can complete academic tasks without reminders.” If responses to those items are related, they may reflect a common dimension such as academic self-management. But the factor's meaning must be supported by the content of its items, not invented merely because the software groups them.
EFA is typically appropriate when the underlying structure is uncertain or being explored. If you already have a strongly specified measurement model and want to test its fit, confirmatory factor analysis (CFA) may be more suitable. EFA findings are exploratory evidence, not proof that the proposed constructs have been definitively validated.
EFA, PCA and CFA: what is the difference?
EFA and principal component analysis (PCA) are often confused because both can reduce many variables to fewer dimensions. But PCA constructs components from total observed variance, while common-factor EFA models covariance attributable to shared factors, with item-specific variation treated separately. CFA tests a specified measurement structure under an explicit statistical model.
| Approach | Main purpose | Typical research question | Key caution |
|---|---|---|---|
| Exploratory factor analysis | Explore common latent dimensions | What underlying factors may explain correlations among questionnaire items? | An interpretable pattern is not automatic confirmation of a construct |
| Principal component analysis | Summarise observed variables as components | Can many variables be compressed for descriptive or predictive use? | Components are not necessarily latent common factors |
| Confirmatory factor analysis | Test a previously specified factor model | Does a proposed measurement model fit the observed data? | Model fit alone does not establish all forms of validity |
If you are deciding how a measurement strategy fits the broader study, see our dissertation data analysis plan guide.
Step 1: Establish the question and item set
Start by identifying the construct or domain that the questionnaire is intended to investigate. Explain where the items came from: an established instrument, a justified adaptation, interviews, theory, or a documented item-development process.
Do not conduct EFA on every variable in a dataset just because the software permits it. Demographic variables, outcomes and unrelated administrative variables may not be sensible indicators of one latent measurement structure.
Before analysis, prepare a codebook identifying each item's meaning, response scale, reverse-coded items and planned treatment of missing values. If survey design is still underway, our guide to designing a useful research questionnaire explains why clear items matter before statistical analysis.
Step 2: Assess whether the data are suitable
EFA depends on the correlation structure. Inspect the data before extracting factors: missing responses, unusual response distributions, items with little variability, near-duplicate questions, sample composition and the type of response scale all matter.
For continuous items, a Pearson correlation matrix may be reasonable where its assumptions fit. Ordered categorical responses, especially with few categories or marked skew, may warrant polychoric correlations and an estimation method appropriate to ordinal data. Do not assume that assigning numbers from 1 to 5 automatically turns an ordinal response into a well-behaved continuous measurement.
KMO and Bartlett's test
The Kaiser–Meyer–Olkin (KMO) index describes the relative importance of observed correlations compared with partial correlations. Researchers often use it as a diagnostic of sampling adequacy for factor analysis. Low values invite investigation of the item set and correlation structure. Inspect item-level measures where available rather than relying solely on one overall statistic.
Bartlett's test of sphericity evaluates the null hypothesis that the population correlation matrix is an identity matrix. A significant test suggests correlations are not all zero; it does not prove that the proposed number of factors, instrument or interpretations are correct. Its significance can also be sensitive to sample size.
How many participants do you need?
There is no universally valid “five participants per item” or “300 responses is always enough” rule. Sample adequacy depends on communalities, how strongly items define each factor, the number of indicators per factor, estimation choices and model complexity. Report an evidence-based justification appropriate to your design and acknowledge sampling limitations. See our sampling method and sample-size guide for broader planning considerations.
Step 3: Choose a common-factor extraction method
Popular EFA extraction approaches include principal axis factoring and maximum likelihood factor analysis. They do not make identical assumptions. Principal axis factoring focuses on modelling common variance; maximum likelihood adds stronger distributional modelling assumptions that may support some forms of inference when justified.
State your chosen method by name. Writing “factor analysis was performed in software” is insufficient. Do not call a PCA output a common-factor EFA without recognising the difference.
Extraction should follow the purpose and the properties of the observed data. If you change the method after inspecting results, explain why, and avoid presenting an exploratory decision as if it had been planned in advance.
Step 4: Decide how many factors to retain
Factor retention is one of the most consequential choices in EFA. An overly simple solution may combine distinct constructs, while an overly complex one can create fragile factors defined by very few items.
Review several forms of evidence: theory and item content, parallel analysis, the scree plot, eigenvalues, communalities, factor interpretability and the stability of a candidate solution. Do not depend mechanically on the “eigenvalue greater than one” rule.
Parallel analysis compares eigenvalues from the observed data with eigenvalues generated from suitable reference data. The comparison helps identify factors whose observed structure exceeds what might be expected from random variation. Its implementation should be appropriate to the extraction method and data type.
Worked example 1: Reading a parallel-analysis decision
Suppose a student investigates a hypothetical 12-item academic-support questionnaire. The following values are invented for explaining a decision, not computed from real participants.
| Candidate factor | Observed eigenvalue | Reference eigenvalue | Initial reading |
|---|---|---|---|
| Factor 1 | 3.80 | 1.42 | Observed exceeds reference |
| Factor 2 | 1.91 | 1.27 | Observed exceeds reference |
| Factor 3 | 1.13 | 1.17 | Observed does not exceed reference |
| Factor 4 | 0.86 | 1.08 | Observed does not exceed reference |
In this hypothetical comparison, two factors meet the illustrated parallel-analysis criterion. That makes a two-factor solution a reasonable candidate for evaluation, not an automatic final answer. The student should inspect the full pattern matrix, factor correlations, item content and whether the solution is stable and meaningful. If another retention method suggests a different number, report the disagreement and justify the final choice.
An appropriate dissertation sentence would be: “Parallel analysis suggested retaining two factors. The two-factor solution was then evaluated for interpretability, salient item loadings and consistency with the proposed domains.” That sentence should be used only if those steps were actually performed.
Step 5: Select and justify factor rotation
Rotation changes how the extracted factors are represented to make the loading pattern easier to interpret. It is not a way to rescue a fundamentally unsuitable item set.
Oblique rotation permits factors to correlate. Methods such as oblimin or promax may be appropriate when psychological, educational or social dimensions plausibly overlap. Orthogonal rotation, such as varimax, constrains factors to be uncorrelated. Choose it only where that constraint is defensible.
With oblique rotation, distinguish the pattern matrix from the structure matrix: the pattern matrix provides regression-like coefficients showing each item's unique relationship with factors, while the structure matrix shows item–factor correlations. Reporting only the numbers above an arbitrary loading threshold can conceal cross-loadings or weak items.
Step 6: Interpret loadings, communalities and cross-loadings
A factor loading helps describe an item's relationship to an extracted factor. Its sign, size and meaning should be interpreted in context. A negative loading is not inherently an error: check item wording and any reverse coding.
A communality summarises the proportion of an item's variance accounted for by the common-factor solution under the fitted model. Low communalities may indicate that an item is not well represented by the retained factors, but any removal should also consider substantive coverage of the construct.
A cross-loading occurs when an item relates meaningfully to more than one factor. Such items can indicate ambiguous wording, genuinely overlapping content, or a structure that does not match the intended distinction. There is no universally correct numerical cutoff for every discipline and sample.
Worked example 2: Interpret a hypothetical pattern matrix
Imagine a researcher explores six items about undergraduate academic support. After appropriate diagnostics, a candidate two-factor oblique solution yields the following illustrative pattern coefficients.
| Item | Factor A | Factor B | Interpretation to investigate |
|---|---|---|---|
| I know where to find academic help | 0.76 | 0.11 | Primarily Factor A |
| I can identify the right support service | 0.71 | 0.08 | Primarily Factor A |
| I can access learning resources when needed | 0.63 | 0.19 | Primarily Factor A |
| I feel confident asking for help | 0.15 | 0.73 | Primarily Factor B |
| I am comfortable discussing difficulties | 0.09 | 0.69 | Primarily Factor B |
| I can approach a lecturer about problems | 0.32 | 0.48 | Possible overlap; review item |
Based on the wording, Factor A could provisionally be called Knowledge of support access, while Factor B could be called Comfort seeking support. These are researcher interpretations, not names automatically supplied by software.
The final item is not as cleanly aligned as the others. Its pattern suggests the researcher should investigate whether it captures both knowing how and feeling comfortable asking. Simply deleting it to make a table look tidy could damage content coverage. Check the item's purpose, communality, substantive relevance and results from justified sensitivity analyses.
These are demonstration loadings, not a claim that six items would necessarily support a stable two-factor solution. A real dissertation would also report the sample, factorability checks, extraction and rotation, retention rationale, appropriate communalities and factor correlations.
Step 7: Decide whether to revise or retain questionnaire items
An item should not be removed solely to improve one statistic. Consider the item's theoretical importance, wording quality, weak or complex loadings, missingness, communalities and the consequences for measuring the intended domain.
Document every substantive revision: which item was removed or reworded, why, whether the factor solution was re-estimated, and what changed afterward. Repeatedly deleting items until a desirable pattern appears can capitalise on chance and inflate confidence in a sample-specific solution.
Where possible, test the emergent structure with independent data or a justified later confirmatory analysis. One exploratory dataset cannot establish a universally valid questionnaire.
Step 8: Evaluate reliability without confusing it with factor analysis
EFA explores common structure; internal-consistency statistics describe a different measurement property. A high Cronbach's alpha does not demonstrate that a scale is unidimensional, and a clean-looking factor pattern does not automatically establish reliability, construct validity or predictive usefulness.
After identifying candidate subscales, evaluate reliability using methods suitable for the items and measurement model. Our Cronbach's alpha guide explains internal consistency and its limitations.
If factor scores will be calculated for later analyses, state the scoring method and be cautious about treating estimated latent constructs as perfectly measured observed variables.
Worked example 3: A dissertation methodology and results write-up
The following fictional case shows how methodological decisions and results can be communicated coherently. No actual EFA was performed to generate these figures. Copy the structure, not the invented statistics.
Example methodology paragraph — hypothetical
“An exploratory common-factor analysis was planned for responses to a 12-item questionnaire intended to capture aspects of academic support experience. Before extraction, missingness, item distributions, inter-item correlations and the suitability of the correlation matrix would be reviewed. The planned correlation type and extraction estimator would be selected according to the response scales and observed distributions. Factor retention would consider parallel analysis, the scree plot and interpretability. An oblique rotation was proposed because access-related and help-seeking experiences might be correlated. Candidate items with weak representation or complex loadings would be examined in relation to their wording and conceptual importance; any removal would be documented rather than chosen solely to maximise a statistic.”
This is appropriate for a proposal because it describes intentions without inventing completed analyses. A completed dissertation must replace the future-oriented wording with what was actually done.
Example results paragraph — hypothetical
“In this fictional demonstration, the correlation matrix was considered suitable for exploratory factor analysis after reviewing item distributions, correlations, KMO and Bartlett's test. Parallel analysis suggested two factors, and the two-factor oblique solution was provisionally interpreted as Knowledge of support access and Comfort seeking support. Most items showed their strongest pattern loading on the proposed factor. One question about approaching a lecturer loaded on both dimensions and was retained provisionally pending content review. The two-factor interpretation would require further evaluation and independent replication; these exploratory findings would not alone establish construct validity.”
The wording demonstrates restraint, but it is not a substitute for actual numerical results. In a real results chapter you would report the observed sample size, KMO, Bartlett statistic and degrees of freedom, extraction and rotation methods, factor-retention evidence, loading and communality tables, interfactor correlations and any modifications.
Example reporting inventory
| Reporting element | What to include | Why readers need it |
|---|---|---|
| Data and sample | Analytical N, item types, missing-data handling | Shows the evidence underlying the analysis |
| Factorability | Correlations, overall or item-level KMO where relevant, Bartlett's test | Helps assess suitability without treating one test as proof |
| Extraction and correlation matrix | Named extraction method, Pearson or polychoric correlations when appropriate | Makes assumptions and analytic choices visible |
| Number of factors | Parallel analysis and other relevant criteria, disagreements | Explains the retention decision |
| Rotation | Oblique or orthogonal method and rationale | Shows whether factor correlations were permitted |
| Factor interpretation | Pattern matrix, communalities, factor correlations, item wording | Allows readers to assess proposed factor labels |
| Revisions and limitations | Removed items, exploratory choices, uncertainty, future validation | Prevents overstatement and hidden selective decisions |
Common mistakes when conducting EFA
Using PCA but calling it EFA. Specify whether the analysis modelled common factors or simply constructed components.
Following default settings without justification. Software settings are not a substitute for decisions about extraction, retention, rotation and correlation type.
Using Bartlett's p-value as proof of validity. It tells you something limited about correlations, not whether a questionnaire accurately measures its intended construct.
Applying fixed sample-size rules mechanically. Evaluate item communalities, factor definition, data quality and model conditions rather than relying on a universal participant-to-item ratio.
Calling every loading above a cutoff acceptable. Cross-loadings, weak communalities, unstable factors and item meaning also matter.
Renaming factors without reading the items. Thematic labels should represent their substantive content and remain provisional when evidence is weak.
Claiming the instrument is validated. A single exploratory analysis contributes evidence about structure, but measurement validity requires a broader programme of evidence.
Frequently asked questions
Is exploratory factor analysis suitable for every dissertation?
No. It is most appropriate when the research question concerns latent structure or measurement dimensions and when the available items and data make such an analysis defensible. It should not be added to a dissertation simply to make the methods appear advanced.
Can I run EFA on five-point Likert questions?
Possibly, but consider the ordinal response structure, category frequencies, skew and correlation estimator. A polychoric approach may be preferable in some settings. The right choice depends on the data and estimator rather than a blanket rule.
Is a significant Bartlett's test enough to proceed?
No. Also inspect the correlation matrix, KMO and item characteristics, sampling, and the interpretability of a plausible factor model.
Should I always keep factors with eigenvalues above one?
No. That rule can produce misleading retention decisions. Consider parallel analysis and other evidence, while explaining disagreements and substantive judgment.
Do factor loadings prove that a questionnaire is valid?
No. They help describe an exploratory measurement structure, but evidence about reliability, content, relationships with other variables and replication may also be needed.
Must I report all items, including poor-loading items?
Report enough detail for readers to evaluate the decisions, including original items considered and any removals. If you suppress low loadings in a compact presentation, state the threshold and ensure the complete relevant pattern is available where required.
Checklist before submitting your EFA section
Can you explain why exploring common latent factors answers your research question?
Are the questionnaire items conceptually relevant to the intended domains?
Have you described analytical N, missing data, item types and distributions?
Have you justified your correlation matrix and extraction method?
Did you examine factorability without treating KMO or Bartlett's test as automatic validation?
Is the number of factors supported by an explained retention strategy?
Does rotation match assumptions about relationships between factors?
Are factor labels supported by actual item wording?
Have you reported loadings, communalities, correlations and item changes transparently?
Have you distinguished exploratory structure from reliability and wider validity evidence?
Does the methods description match what was actually done?
Final thoughts
Exploratory factor analysis is useful when it is driven by a real measurement question, not a desire to produce a sophisticated table. A strong dissertation explains the decisions made before, during and after extraction, then presents the resulting dimensions with justified caution.
If you need help understanding methodological feedback, reviewing the clarity of a factor-analysis write-up or strengthening the presentation of your own findings, explore Acadelyra's academic support services. Your research decisions, evidence and academic responsibilities remain your own.
Sources and further reading
Watkins (2018), Exploratory Factor Analysis: A Guide to Best Practice — methodological decision-making and reporting transparency.
Hayton, Allen and Scarpello (2004), Factor Retention Decisions in Exploratory Factor Analysis — parallel analysis and factor retention.
Simon Fraser University, Measurement and Psychometrics — instructional discussion of factorability and interpreting factors.
These resources support methodological reasoning; follow your university's approved requirements and report only analyses actually performed.