Factor Analysis Cross Loading SPSS: What Should You Remove for Low Communalities Too?

If your factor analysis cross loading SPSS output looks messy right now, you are not alone, and you are not doing anything wrong. Every dissertation student hits this exact wall.

A cross-loading is a problem when an item loads above 0.32 on more than one factor with less than a 0.1 gap between the two values. Communality below 0.40 is generally too low, though 0.30 can still be defensible with a strong KMO. Delete only one problem item at a time, then rerun the whole analysis before touching anything else.

In twelve years of doing this with students across the US and UK, I have never seen a clean Rotated Component Matrix on the first try. Let me walk you through what these numbers actually mean, and more importantly, what you should remove and what you should leave alone.

What a Cross-Loading and a Low Communality Actually Mean (Before You Touch Delete)

Before deleting anything, understand that cross-loading and low communality are two separate problems. This is the cross loading factor analysis SPSS confusion every dissertation student eventually runs into. Fixing one does not automatically fix the other, so treat them as two different diagnoses.

What is a cross-loading in your Rotated Component Matrix?

A cross-loading happens when a single item shows a meaningfully high loading value on more than one factor in your Rotated Component Matrix, instead of loading cleanly on just one. This is the exact definition I give every student, because it matches what SPSS actually shows you on screen.

This is a classic rotated component matrix cross loadings pattern: Item 7 loading 0.41 on Factor 1 and 0.36 on Factor 2. Neither number is wrong on its own. The problem is that the item cannot decide which construct it belongs to.

What is communality, and where do you find it in SPSS output?

Communality is the proportion of an item’s total variance that is explained by the factors you extracted, combined. You will find it in the Communalities table, under the “Extraction” column, not in the Rotated Component Matrix. A low communality means the factor solution is barely explaining that item at all, even if its loading looks fine elsewhere.

Why these are two different problems, not the same fix

An item can have a perfectly clean, single factor loading and still have terrible communality. Another item can load on one factor beautifully and cross-load on another at the same time. Both numbers come from the same starting point though, the correlation matrix between all your items. If you are ever unsure why two items are grouping together at all, it helps to understand that underlying logic, which I cover in my piece on Pearson vs Spearman correlation and which one to use.

Both problems also eat into your construct validity, since a scale is only as valid as the items that actually belong to it. This is also separate from deciding how many factors to extract in the first place. That earlier decision comes from eigenvalues and the scree plot, using Kaiser’s eigenvalue-greater-than-1 rule, not from cross-loadings or communalities. If you are still finalising how many factors to retain, that step comes before anything in this article.

Factor Loading Cutoff SPSS Values You Need (With Sample Size Context)

Most blog posts give you one number and move on. That is not how this actually works, and honestly, that oversimplification is where a lot of dissertation students get into trouble with their examiners.

The 0.3, 0.4 and 0.5 rule, and where it actually comes from

The widely used factor loading cutoff SPSS guideline comes from Hair, Black, Babin and Anderson’s “Multivariate Data Analysis,” not from SPSS itself. SPSS will never tell you what counts as acceptable, it just gives you the raw numbers. Hair’s table links the cutoff directly to your sample size, which is the part most articles conveniently skip.

Sample size	Minimum loading needed
50	0.75
100	0.55
200	0.45
350	0.30

A smaller sample needs a stronger loading before you can trust it is real and not sampling noise. If your dissertation has 120 responses and you are using a flat 0.3 cutoff for everything, an examiner who knows this table will question it.

I will be honest about where I disagree with Hair et al. here. Their table was built with continuous-scale data in mind, not the 5-point Likert scales almost every dissertation questionnaire actually uses. In my experience running EFA on Likert data for over a decade, the same 0.30 cutoff at n=350 feels too lenient for ordinal data. I personally push students toward 0.35 as a floor even at larger samples.

Costello and Osborne (2005), published in Practical Assessment, Research & Evaluation, take a stricter view. They recommend a flat 0.32 minimum loading regardless of sample size. I do not fully agree with that either. A loading of 0.30 at n=400 is statistically far more stable than the same 0.30 at n=80, and their one-size rule ignores that difference. Tabachnick and Fidell, in “Using Multivariate Statistics,” suggest a similar 0.32 floor but allow more flexibility when an item is theoretically central to the construct, a nuance I find more useful in real dissertation work than any flat number.

PCA cutoff versus Common Factor Analysis cutoff in SPSS

One thing almost nobody explains clearly: SPSS’s Principal Component Analysis extraction and Common Factor Analysis (Principal Axis Factoring) extraction produce slightly different loading values for the same data, because PCA includes all variance while common factor analysis only uses shared variance. If you switch extraction methods midway because your first attempt looked messy, your cutoffs need to shift slightly too. Check which one you actually selected in Extraction before you panic about a “bad” number.

What Communality Value Is Too Low in SPSS?

This is the question that comes up again and again in any low communality factor analysis SPSS discussion, and the honest answer has a range, not a single number.

The 0.4 benchmark, and when 0.3 is still fine

For a what communality is too low SPSS question, the standard academic benchmark is 0.40. Below that, the item is not contributing meaningfully to your factor structure. That said, if your overall KMO value is strong, above 0.80, and your sample size is reasonably large, a communality as low as 0.30 is sometimes still defensible, and I have seen examiners accept it when the researcher explains this reasoning clearly.

What a communality below 0.3 is telling you about that item

A communality under 0.30 is a clear signal. It usually means the item is poorly worded, measuring something unrelated to your other items, or was translated or adapted badly if you borrowed the scale from another study. I had a student last year whose adapted leadership scale had one item with communality of 0.19, it turned out the item had been mistranslated during adaptation and was genuinely asking something different from the rest of the scale.

Factor Analysis Cross Loading SPSS: Should You Delete the Item or Keep It?

Delete the item if the cross-loading gap is small and there is no theoretical reason to keep it. Keep it if theory clearly supports the item relating to both constructs. That is the rule in one line.

Here is where I disagree with how most statistics tutorials frame this decision. They treat it as purely numeric. It is actually numeric plus theoretical, and ignoring the second half is exactly how students end up deleting items they should have kept. What you are really aiming for across your whole scale is simple structure, where every item loads cleanly on one factor and nowhere else, not just a fix for one problem item in isolation.

When theory justifies keeping a cross-loading item

If an item loads on two factors SPSS shows as close, say 0.38 and 0.34, and there is a genuine theoretical reason the concept could relate to both constructs, you can keep it and discuss this openly in your limitations section. Reviewers respect a reasoned decision far more than a silently deleted item.

When to delete without hesitation

If the gap between the two loadings is small, there is no theoretical justification, and your communality for that same item is also weak, delete it. This combination is close to definitive factor analysis item removal criteria that most methodologists agree on.

The delete-one-at-a-time, rerun rule

Here is the step-by-step process I make every student follow, because skipping steps here is the single most common mistake I see:

1. Identify the single worst offending item, highest cross-loading gap or lowest communality, not multiple items at once
2. Delete only that one item from the analysis
3. Rerun the entire EFA from scratch, including KMO and Bartlett’s test
4. Check the new Rotated Component Matrix and Communalities table fresh
5. Repeat only if another clear problem item appears

Deleting three or four items in one go is how students end up with a factor structure that bears no resemblance to their original questionnaire. One item changes the whole solution. I have watched it happen more times than I can count.

I have turned this exact cutoff table and decision checklist into a one-page reference I hand to every student I mentor. If you would like the same, my guide on finding the right dissertation writing mentor explains how that mentoring actually works.

What Happens After You Delete an Item (How to Handle Cross Loadings SPSS Properly)

Deletion is not the end of the job. There is a how to handle cross loadings SPSS workflow after the deletion itself that most tutorials never mention.

Checking your Cronbach’s alpha did not drop

After any deletion, rerun your reliability analysis. If alpha drops noticeably, the item you removed was actually contributing to internal consistency even though it looked messy in the factor analysis, and you need to reconsider.

How many items is too few for a factor

A factor with only two items left after deletion is weak and generally not recommended for publication or defence. Three items per factor is the usual minimum most supervisors will accept without pushback.

SPSS Factor Analysis Help Dissertation Students Actually Need: Writing It Up

Getting the numbers right is only half the job. You also need to justify the decision in words your examiner will accept.

Language your supervisor will accept

Instead of writing “the item was deleted because the loading was low,” write something like: “Item 7 was removed due to a cross-loading of 0.41 on Factor 1 and 0.36 on Factor 2, combined with a communality of 0.28, falling below the 0.40 threshold recommended by Hair et al. (2019).” This shows you applied a rule, not a hunch.

If organising your entire results chapter around this kind of evidence-based language feels overwhelming right now, my dissertation chapter organising guide breaks the whole structure down section by section.

A Short Case From My Own Practice

A psychology dissertation student I worked with last year ran a 25-item resilience scale through SPSS and got four items cross-loading across two factors. She wanted to delete all four at once because the deadline was close.

We went through the one-at-a-time process above instead. Only one item genuinely needed removal, the other three resolved themselves once that single item was gone and the analysis was rerun. Her final factor structure matched the original theoretical model almost exactly, and her examiner raised zero questions about the methodology chapter.

This is the entire reason the rerun-one-at-a-time rule exists. If you are not confident reading SPSS output like this on your own yet, my SPSS tutor page is built exactly for this kind of hands-on guidance.

Before You Finalise Your Factor Structure

A few related checks are worth doing at the same time, since examiners often look at these together:

• If your scale also feeds into a regression model later, run a VIF check on your final factor scores to rule out multicollinearity.
• If you are unsure whether your overall data even needs an expert second opinion before submission, this guide on how to know if you need a dissertation expert is a fair, honest starting point, including when you genuinely do not need one.

If none of this fully resolves your specific dataset, a direct look at your actual output usually sorts it out within a single session, that is genuinely the fastest path once the general rules stop being enough.

FAQ

Author: Written by Siddharth, 12 years of dissertation and statistical consulting experience. Connect on LinkedIn.