The questionnaire contains forty statements evaluating a brand, while the report must fit on a single slide for the management board. Calculating means for each item separately will not answer the question of how many actual attitudinal dimensions the questionnaire measures or which statements essentially convey the same thing. Factor analysis is a technique that organizes such a set of variables: it identifies the underlying dimensions behind respondents’ answers and makes it possible to replace dozens of questions with a few interpretable constructs.
When does factor analysis make sense in a survey?
The starting point is an observation familiar to every analyst working with survey data: responses to questions concerning similar areas are correlated. A respondent who rates a consultant’s politeness highly will usually also rate their patience and willingness to help highly. These three items do not provide three independent pieces of information – they provide one piece of information about the quality of interaction with customer service, measured in three ways. Factor analysis formalizes this intuition: it searches the correlation matrix for a structure in which observed questionnaire items are explained by a smaller number of latent variables that cannot be directly observed, namely factors.
In market research practice, factor analysis is used in several recurring situations. It is not a universal method, and not every questionnaire requires it – it is worth recognizing the situations in which it actually solves a problem:
- Brand image batteries. A brand evaluated on several dozen attributes generates a report that cannot be summarized. Dimensionality reduction makes it possible to show that, for example, four dimensions of brand perception underlie these attributes and to compare brands on those dimensions rather than on each attribute separately.
- Developing and validating measurement scales. If a study is intended to measure a construct such as loyalty, employee engagement, or risk propensity, it is necessary to demonstrate that the questionnaire items actually measure one common dimension rather than several independent things.
- Preparing data for further analysis. Segmentation based on forty correlated variables may produce an unstable or difficult-to-interpret solution. Segmentation based on a few factor dimensions can reduce redundancy and the effect of including the same dimension multiple times.
- Shortening questionnaires in tracking studies. The analysis shows which items contribute unique information and which can be removed without a material loss of measurement value – directly affecting interview length and data quality in subsequent waves.
An important distinction: factor analysis answers a question about the structure of a set of variables, rather than their impact on a business outcome. This is a different class of problem from key driver analysis, which estimates which elements of the customer experience are most strongly associated with satisfaction or NPS. In many projects, the two techniques are used sequentially: factor analysis first reduces a battery of attributes to dimensions, then a driver model assesses which of those dimensions is most strongly associated with recommendation. Confusing them leads to the misunderstanding that the size of a factor loading indicates the strength of its impact on customer behavior.
How do you conduct exploratory factor analysis and assess its validity?
In market research, exploratory factor analysis (EFA) is often used when the dimensional structure is not assumed in advance. Its counterpart within measurement modeling is confirmatory factor analysis (CFA), which tests a previously specified measurement model. It is used, among other things, to validate established instruments and confirm structures obtained through EFA. Below is the logic of the exploratory approach.
1. Checking whether the data are suitable for analysis at all
Before running the procedure, it is necessary to confirm that there is something to reduce in the correlation matrix. Two standard measures are used for this purpose: the Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy and Bartlett’s test of sphericity. KMO assesses the extent to which correlations between variables can be explained by common factors – values close to one indicate data well suited to analysis, while low values mean that the variables are too weakly related. Bartlett’s test examines the hypothesis that the correlation matrix is an identity matrix, meaning that the variables are mutually uncorrelated; a significant result supports further analysis, although in large samples the test may also detect very weak relationships. It is also worth examining the correlation matrix itself: if most coefficients are close to zero, no procedure will fix that.
2. Choosing the extraction method
Principal component analysis (PCA) and principal axis factoring (PAF) are sometimes used interchangeably, but they differ in their assumptions. PCA reduces total variance and treats components as linear combinations of observed variables – it is a purely data-reduction technique. PAF and maximum likelihood factor analysis model only common variance and assume the existence of latent factors generating responses. If the aim is to identify psychological constructs underlying a battery of statements, a factor-analytic approach is more appropriate; if the aim is solely to compress data before segmentation, PCA may be sufficient.
3. Deciding on the number of factors
This is the most subjective point in the entire procedure. The oldest criterion – retaining factors with eigenvalues greater than one – tends to overestimate the number of dimensions and should not be the sole determinant. A scree plot makes it possible to visually identify the point of inflection. Today, Horn’s parallel analysis is considered one of the most reliable criteria, as it compares eigenvalues obtained from actual data with values from random data containing the same number of variables and observations. Alongside statistical criteria, there is also a substantive criterion: a solution with factors that cannot be meaningfully named is useless for a research project, even if it meets formal requirements.
4. Rotation and interpreting loadings
An unrotated solution is rarely interpretable because the first factor usually “collects” most variables. Rotation transforms the arrangement of axes to make the loading pattern easier to interpret. Orthogonal rotations (varimax) force factors to be uncorrelated, while oblique rotations (oblimin, promax) allow correlations between them. In research on attitudes and brand image, dimensions are usually related, so oblique rotation is often closer to reality – an additional advantage is the factor correlation matrix, which is informative in its own right.
When interpreting the loading matrix, analysts look for what is known as a simple structure: each item has one clearly dominant loading. Two situations are problematic – items with low loadings on all factors, which do not fit any dimension, and items that load similarly on two factors (cross-loadings), which may indicate ambiguous question wording or overlapping constructs. Both groups require verification; they may be candidates for removal and re-estimation of the model.
5. Verifying reliability and naming factors
The extracted dimensions require an assessment of internal consistency. The most commonly reported measure is Cronbach’s alpha, although its limitations should be kept in mind: it increases with the number of items in a scale and assumes tau equivalence, which broadly means that all items have equal loadings, a condition that is rarely met. For this reason, McDonald’s omega is increasingly reported alongside it. A very high alpha is not unconditionally good news – it may signal that the items are nearly identical paraphrases of the same statement, reducing the informational value of the scale. A broader discussion of this issue is provided by measurement reliability.
The final step is both the most difficult and the most frequently neglected. A factor must be given a name that reflects the content of the items with the highest loadings and is understandable outside the analytics department. As Hume’s Institute experts point out, a factor has value only when it can be named unambiguously and translated into the language of business decisions – statistical correctness alone is not enough. A solution with exemplary parameters, in which dimensions appear in the report as “Factor 1” and “Factor 2,” will not be used by anyone other than the author of the analysis. A well-named dimension sounds like “service predictability” or “perceived modernity of the offering,” not like a label from a software output.
What errors most often undermine factor analysis?
Problems with factor analysis rarely arise from the calculation procedure itself – they most often emerge earlier, at the questionnaire design stage, or later, when interpreting the results. Below are common pitfalls in survey projects:
- Sample size too small relative to the number of variables. Factor loadings are estimated from the correlation matrix, and correlations in small samples are unstable. The result is a solution that does not replicate in the next wave of the study. Rather than adhering to a single rule of thumb, it is worth considering the ratio of respondents to items, the level of communalities, and the magnitude of loadings – when loadings are strong, sample requirements are less stringent.
- Including questions of different types in the model. Standard factor analysis based on Pearson correlations is designed primarily for continuous variables. For ordinal items, such as Likert scales, polychoric correlations and appropriate estimators are often used. Combining questions with different formats without accounting for their properties may lead to artifacts, including a factor that reflects the question format rather than the content of the attitude.
- Ignoring item wording effects. Reverse-coded items (negatively worded) may form a separate factor that is not a substantive dimension but an artifact of response style. Before interpretation, it is necessary to check whether the extracted factor simply groups all negatively worded statements.
- Treating the number of factors as an objective result. Different criteria produce different solutions. Sound practice involves estimating several variants and selecting the one that is both statistically acceptable and substantively meaningful, with the decision described in the methodological report.
- Interpreting loadings as strength of impact. A loading indicates the relationship between an item and a factor, not the importance of that factor for customer behavior. Models of relationships, rather than dimensionality reduction, are used for this purpose.
- Omitting validation on an independent sample. A structure identified exploratorily in one dataset should be confirmed through confirmatory analysis in another, especially if the scale is to be used in subsequent waves or in other markets.
A separate limitation is the cultural and linguistic context. A scale translated from an international study will not necessarily reproduce the same factor structure in a Polish sample – statements that form a coherent dimension in one language may split into two in another. In multinational research, this requires tests of measurement invariance, without which comparing latent variable means across markets is unjustified.
How does factor analysis differ from other data reduction methods?
A manager commissioning an analysis often hears several method names and does not know which one to choose. The comparison below clarifies the key distinctions:
- Factor analysis versus cluster analysis. The former describes the structure of relationships between variables (questions), while the latter groups observations (respondents). Segmentation answers the question “what types of customers are in the sample,” while factor analysis answers “what dimensions does the questionnaire measure.”
- Factor analysis versus regression analysis. Dimensionality reduction has no dependent variable – it describes the structure of a dataset. Regression and driver models have a clearly defined outcome, such as NPS, and estimate the relationships between predictors and that variable.
- EFA versus CFA. Exploratory factor analysis discovers a structure, while confirmatory factor analysis tests it against an assumed model and provides fit indices. In the lifecycle of a research instrument, EFA is often an earlier stage.
- Factor analysis versus multidimensional scaling. MDS works with a matrix of similarities or distances between objects, such as brands, and is used to build perceptual maps; factor analysis works with correlations between variables.
In project practice, the choice follows the objective, not the analyst’s preference. If the question is “how many actual dimensions does our battery measure?”, factor analysis is the appropriate answer. If it is “who do we have among our customers?” – segmentation. If it is “what should we improve first?” – a driver model, although in the latter case factor analysis can serve as a preparatory step by reducing predictor multicollinearity.
Frequently asked questions
What is factor analysis?
Factor analysis is a family of statistical techniques used to identify latent variables that explain the pattern of correlations among observed questionnaire items. Instead of analyzing each question separately, the method makes it possible to describe them using a smaller number of dimensions that reflect shared content. In market research, it is used primarily to organize brand image batteries, validate measurement scales, and prepare data for segmentation.
When should factor analysis be used in a survey?
It should be used when a questionnaire contains a dozen or more items measured in a similar format and there is reason to suspect that some of them measure the same thing. A prerequisite is clear correlations between variables – if the items concern unrelated issues, the analysis will not extract a meaningful structure. The method does not apply to individual factual questions or to multiple-choice questions in their raw form.
What sample size is needed for factor analysis?
There is no single universal value – the required sample size depends on the number of items in the battery, the strength of factor loadings, communalities, and the number of variables per factor. With a clear structure and high loadings, a stable solution can be obtained with smaller samples than in the case of a weak structure and numerous borderline items. In project practice, replication is an important test: if a similar structure is reproduced in randomly split parts of the dataset or in the next wave of the study, this supports its stability.
Ask about statistical analysis of data from your study
Ask about statistical analysis of data from your study. If the questionnaire is already in the field or the data are awaiting analysis, Hume’s Institute’s analytics team will select a dimensionality reduction method and verify the reliability of the scales – simply describe the scope of the project in your message.