Categorical Predictor Is Significant but Dummy Variables Are Not in SPSS: How Do You Interpret It?

I get this exact question from dissertation students almost every week: categorical predictor significant but dummy variables not, so what exactly do I write in my results chapter? You run your regression, the omnibus test for your categorical variable comes back significant, and then you look at the parameter estimates table and not a single dummy variable crosses the 0.05 line. Nothing is broken in your SPSS file. This is a real, well documented statistical pattern, and once you understand why it happens, writing it up becomes straightforward.

This happens because SPSS runs two completely different tests on the same categorical variable. The omnibus test checks whether the categories differ from each other as a group. The individual dummy variable tests only check whether each category differs from your one chosen reference category. A variable can show real variation across its categories without any single category being far enough from the reference group to clear significance on its own.

How to Interpret Dummy Variables in SPSS: The Two Tables You’re Comparing

Before going further, it helps to be clear on what SPSS is actually showing you, because this is where most confusion starts.

When you put a categorical variable with more than two groups into a regression, SPSS (or you, manually) converts it into k−1 dummy variables. One category becomes the reference category, and every dummy compares one group against that reference only. This is standard, correct, and not an error on your part.

SPSS calls the first number the omnibus test, and this is also what people mean by categorical variable overall significance in SPSS output. It answers one question only: do the categories differ from each other as a group? It appears as “Tests of Model Effects” in GENLIN, or as the categorical row in the Omnibus Tests table elsewhere, with degrees of freedom equal to k−1.

Separately, the Parameter Estimates table gives each dummy its own individual Wald test, compared against the reference group only.

These two tables answer two different questions, not the same question twice. That single sentence is the one thing I wish every SPSS textbook led with instead of burying it on page nine.

Why a Categorical Predictor Is Significant but Dummy Variables Are Not in SPSS

There are four real reasons why the overall categorical variable is significant while individual categories are not, and in my experience reviewing dissertation chapters, it is usually a mix of the first two.

How Reference Category Placement Affects SPSS Logistic Regression Results

Sometimes your reference category sits in the middle of the effect rather than at one extreme. None of the dummy comparisons against it may then be large enough to reach significance individually, even though the groups clearly differ overall. I have seen this exact scenario play out with income brackets, where the middle bracket was set as reference and every comparison looked flat. Once the reference was moved to the lowest bracket, two comparisons became clearly significant.

Multicollinearity Among the Dummy Variables

When your categories are unevenly sized, say one group has 180 cases and another has 14, the dummy variables can become correlated with each other. This inflates the standard errors on individual coefficients, which is the same multicollinearity logic that shows up in continuous predictors. If you have not already checked your model for it, my guide to checking multicollinearity with VIF covers how to spot it, though VIF for a dummy set needs a slightly different read than for a single continuous variable.

Small Sample Size Per Category

An omnibus test with multiple degrees of freedom pools information across all categories and is, statistically speaking, more powerful than a single one-degree-of-freedom comparison. If one or two of your subgroups are small, the individual Wald test for that group specifically is underpowered even when the pooled, overall effect is not.

The Wald Test Itself Can Misbehave

The individual significance tests in the parameter estimates table default to a Wald test for each categorical predictor in SPSS, and this specific test can misbehave in ways most guidance never mentions. Wald tests estimate significance from the curvature of the likelihood function around your coefficient. When that curvature flattens out, which is common with small samples, near-perfect separation, or extreme coefficients, the standard error gets overestimated and the p-value comes out non-significant even for a genuinely strong effect.

Statisticians named this breakdown the Hauck-Donner effect, after a 1977 paper examining exactly this flaw in Wald-based hypothesis tests. It is widely discussed in R and SAS modelling circles and barely mentioned in SPSS-specific guidance, which is one of the clearer information gaps on this exact topic. If your parameter estimates look unstable, or your confidence interval around an odds ratio looks implausibly wide, suspect this effect before you suspect your data.

Likelihood Ratio Test Significant, Parameter Estimates Not: Which One Do You Trust

If the omnibus significance result for your categorical variable disagrees with the Wald-based parameter estimates, trust the likelihood ratio test over the Wald test, every time the two disagree. The LR test compares the model’s fit with and without the whole categorical term by looking at the actual change in log-likelihood, and it is not vulnerable to the flattening problem that trips up Wald estimates.

Here is the catch for SPSS specifically, one that IBM’s own documentation reflects but which most dissertation-level tutorials skip entirely:

  • In GENLIN, the “Tests of Model Effects” table already gives you a proper Type III test, and you can request this as either Wald or likelihood ratio chi-square directly from the Model menu.
  • In standard Binary Logistic Regression, SPSS does not hand you a clean per-term likelihood ratio test the way GENLIN does. You have to build it manually: run the full model, note the −2 Log Likelihood, remove the categorical variable, rerun, and compare the chi-square difference against the change in degrees of freedom by hand.

A lot of SPSS-focused content talks about “the LR test” as if it is one button away in every procedure. It is not. If your categorical predictor SPSS logistic regression interpretation needs to lean on the LR test, check which procedure you actually ran before promising your supervisor a number you cannot easily reproduce.

What to Actually Report in Your Results Chapter

This is the part my clients actually came for, so here is the decision logic I use, in order:

  1. Omnibus test significant, no individual dummy significant: Report the omnibus effect as significant, explicitly state that no individual category differed from the reference category at p < .05, and name the reference category you used. Do not claim “no effect of X” anywhere in this scenario; that contradicts your own omnibus test.
  2. Omnibus test significant, some dummies significant: Report the omnibus test first, then report only the dummies that cleared significance, with their odds ratios or unstandardised coefficients and confidence intervals.
  3. Omnibus test not significant at all: You can report the variable as not contributing meaningfully to the model, and you do not need to walk through the individual dummies in detail.

When you do report a significant dummy, give the odds ratio its proper read. An odds ratio of 1.8 for “part-time vs distance learner” means part-time students have 1.8 times the odds of the outcome versus distance learners. It does not mean versus students in general. Always name the reference category next to the odds ratio, or the number means nothing to your reader.

An APA 7th edition style sentence you can adapt directly for your write-up:
“The omnibus effect of [variable] was statistically significant, χ²(df) = X.XX, p = .0XX; however, no individual category differed significantly from the reference category ([name]) at the 5% level, suggesting the overall difference reflects variation across the full set of categories rather than any single pairwise comparison.”

If you want someone to sanity-check this decision against your actual SPSS output before you commit to it in your results chapter, a quick SPSS output review can usually confirm it in minutes.

A Short Case From My Desk

A composite example based on a pattern I see often: a client last year ran a logistic regression predicting dropout risk, with “enrolment mode” (full-time, part-time, distance) as a three-level categorical predictor. The omnibus test came back at p = .021. Every individual dummy against the full-time reference group sat between p = .08 and p = .14. She was convinced her analysis was wrong three days before her viva.

We reran it with distance learners set as the reference category instead of full-time. One comparison, part-time vs distance, immediately came out significant at p = .019, and the pattern made substantive sense once we saw it from that angle. Nothing about the underlying data changed. Only the comparison point did.

Should You Change the Reference Category to Fix This?

Sometimes, yes. Sometimes, no. The honest answer depends on why you are doing it.

It’s legitimate when: your theoretical question is naturally framed around a different comparison, such as comparing everyone against the most common group or against a clinically meaningful baseline, and you decide this before looking at which reshuffle produces significance.

It edges into p-hacking when: you try every possible reference category until something turns green, then write up only that version without mentioning the others.

A quick checklist before you choose, or change, a reference category:

  • Pick the group your research question is naturally built around, not whichever group happens to be largest.
  • Decide this before you run the model, not after seeing which version turns significant.
  • If you change it afterward for a genuine theoretical reason, say so in your methods section.
  • Report results from one reference category only; do not present multiple versions and let the reader pick.

Should You Drop the Non-Significant Dummy Variable From Regression?

Generally, no, not on its own. Dummy variables not significant regression results from a single categorical variable should be dropped as a complete set only, based on the omnibus test, never one dummy at a time. Removing one dummy while keeping the others from the same categorical variable breaks the comparison structure. It also leaves your reference category implicit and confusing to a reader.

Statology’s own SPSS dummy variable tutorial, for instance, walks through creating dummies and interpreting the regression output, but stops at the individual p-values. It recommends dropping a categorical predictor from the model the moment its dummy variables individually miss significance, without checking the omnibus test first.

That is precisely the shortcut I would push back on. A predictor can carry real explanatory value even when no single dummy clears 0.05 on its own, and skipping the omnibus check is exactly how real effects get discarded from dissertations by mistake.

If theory says the variable matters, keep the whole set in the model regardless of the omnibus p-value and report it transparently. A middle option worth considering: if two or three categories are all non-significant and theoretically similar (say “divorced” and “widowed” inside a marital status variable), you can merge them into one combined category and re-run the model, provided the merge makes substantive sense and is not chosen just to manufacture significance.

Other Ways to Handle This: Effect Coding and Pairwise Comparisons

Dummy coding against one reference category is not the only option, and knowing the alternatives helps you explain your choices with more authority in a viva.

Effect coding compares each category against the overall sample mean instead of against one reference group. This removes the reference-category sensitivity problem described above entirely, though it changes how you interpret each coefficient, so it is worth choosing deliberately rather than as a panic fix mid-analysis.

Estimated marginal means with pairwise comparisons, available through SPSS’s EMMEANS command in GENLIN and the General Linear Model procedure, give you every category compared against every other category, not just each one against the reference. Apply a Bonferroni or Sidak adjustment when you do this, since running multiple pairwise comparisons inflates your false positive risk otherwise. This is the fuller picture when your dissertation question genuinely needs “which specific groups differ,” rather than just “does this variable matter overall.”

Neither approach is compulsory, and both take more work than the dummy coding SPSS gives you by default. But if your examiner asks why you did not run pairwise comparisons, these are the two answers you should already know.

Common Mistakes Students Make With This Output

  • Reporting “no significant difference” when the omnibus test was actually significant.
  • Quietly swapping reference categories without disclosing it in the methods section.
  • Treating the Wald test as infallible when the likelihood ratio test disagrees with it.
  • Running Binary Logistic Regression and assuming it gave them a true LR test for each term, when it did not.
  • Dropping individual dummies instead of the whole categorical set.

This same omnibus-versus-individual confusion shows up outside logistic regression too, most famously in ANOVA, where you can get a significant overall F but no significant pairwise comparison in your post-hoc table. My comparison of Tukey HSD and Games-Howell post-hoc tests walks through the equivalent situation for ANOVA designs, if that is closer to what you are running. If you are still deciding which correlation or comparison test fits your design in the first place, my guide to choosing between Pearson and Spearman correlation and guide to interpreting the Durbin-Watson test cover two other SPSS outputs that get misread just as often as this one.

If you would rather have someone check your actual SPSS output line by line instead of cross-referencing five different guides at midnight, that is exactly the SPSS logistic regression categorical variable help I provide through one-to-one SPSS tutoring.

FAQ

What is the reference category in SPSS logistic regression?

It is the one category from your categorical variable that does not get its own dummy variable and instead serves as the comparison baseline for every other category’s coefficient. SPSS assigns this automatically unless you specify otherwise, usually the last category by value or alphabetical order, depending on the procedure.

What is a Wald test for a categorical predictor in SPSS?

It is a statistical test that checks whether an individual dummy variable’s coefficient is significantly different from zero, calculated by squaring the ratio of the coefficient to its standard error. In SPSS parameter estimates tables, each dummy gets its own Wald chi-square and p-value, compared against the reference category only.

Can I report a categorical variable as significant if no individual dummy variable is?

Yes. Report the omnibus test result as significant, state clearly that no individual category differed from the reference category, and name which category was used as reference so your reader can follow the comparison.

Should I remove a non-significant dummy variable from my regression model?

Not on its own. Decide on the whole categorical variable as a set based on the omnibus test, not by deleting individual dummies one at a time, since that distorts the reference comparison for the rest.

How do I get a likelihood ratio test for a categorical predictor in SPSS?

In GENLIN, select likelihood ratio chi-square directly from the Model Effects options. In standard Binary Logistic Regression, run the full model and the reduced model without the categorical term separately, and compare the change in −2 Log Likelihood by hand.

Does changing the reference category count as manipulating results?

Only if you try multiple reference categories after seeing the output and report just the one that turned significant without disclosing it. Chosen in advance on theoretical grounds and reported transparently, it is a standard analytical decision.

What’s the difference between the omnibus test and parameter estimates in SPSS categorical covariates?

The omnibus test checks whether the categories differ from each other as a group, using degrees of freedom equal to the number of categories minus one. The parameter estimates table checks each category against the reference category individually, one comparison at a time.

About the author: I’m Siddharth, and I have spent over a decade helping dissertation and thesis students make sense of exactly this kind of confusing SPSS output. Connect with me on LinkedIn.

Perfect for students, researchers, and professionals looking to build real statistical skills.