When I was a psychology student, interpreting SPSS output felt like deciphering hieroglyphics. Endless tables, acronyms without context, asterisks that appeared (or didn't) next to p-values. I remember staring at an ANOVA table and not knowing where to start reading. Nobody had explained it properly, and statistics textbooks weren't much help.
After years working with research data, what once seemed opaque has become almost automatic. The key wasn't studying more theory, but learning to ask the same questions to every table: Is it significant? How large is the effect? Can I trust these data? In this guide I teach you exactly that. And if you are in a hurry and the output is already on your screen, paste it into the free output interpreter and it asks those three questions for you, APA 7 sentence at the end.
The logic behind every SPSS table
Regardless of the analysis you run, SPSS always returns the same basic structure. If you understand this logic, you can interpret any table:
1. The test statistic: A number (t, F, U, chi-squared) summarising how much your data deviate from what you'd expect by chance.
2. Degrees of freedom (df): Indicate how much information your data have to estimate the statistic. More participants = more df = more power.
3. The p-value (Sig.): The probability of obtaining a result this extreme or more if the null hypothesis were true. If p < .05, you reject the null.
4. Effect size: What SPSS often does NOT give you directly, and what actually matters. A result can be significant but irrelevant if the effect is tiny.
That's the structure. Now let's go test by test.
Interpreting a t-test
SPSS returns two tables: group statistics (means, standard deviations) and the test itself.
What to look for in the test table:
Levene's test: Check the "Sig." column for Levene's first. If p > .05, use the "Equal variances assumed" row. If p < .05, use "Equal variances not assumed". Many researchers skip this and always use the first row, which is wrong.
t and df: Report both. Example: t(45) = 2.34.
Sig. (2-tailed): This is your p-value.
Mean difference and 95% CI: Much more informative than the p-value alone.
What SPSS doesn't give you: Cohen's d. You need to calculate it separately. You can use our effect size calculator by entering t and df values.
Interpreting one-way ANOVA
In the ANOVA table, look for the "Between Groups" row. There you'll find F, df (two values), and Sig. If p < .05, there is at least one significant difference between groups.
But be careful: a significant ANOVA only tells you there's a difference somewhere, not where. You need post-hoc tests for that.
If variances are equal (Levene not significant): use Tukey HSD.
If variances are unequal: use Games-Howell.
If you have specific a priori hypotheses: use Bonferroni.
Interpreting Pearson correlation
The SPSS correlation table is a symmetric matrix. For each pair of variables it gives you r, Sig., and N.
Practical interpretation of r (Cohen, 1988): r = .10 small, r = .30 moderate, r = .50 large. In psychology, a correlation of .30 is already considered relevant.
Interpreting chi-square
In the "Chi-Square Tests" table, look for the "Pearson Chi-Square" row: Value (chi-squared), df, and Asymp. Sig.
Important rule: If any expected cell count is less than 5, chi-square is unreliable. Use Fisher's exact test instead (SPSS calculates it automatically for 2x2 tables).
Interpreting linear regression
SPSS gives you three main tables: Model Summary (R-squared), ANOVA (overall model significance), and Coefficients (individual predictors with B, Beta, t, Sig., and confidence intervals).
The 5 most common mistakes I see
1. Not checking assumptions.
2. Reporting only the p-value without effect sizes or confidence intervals.
3. Always using the first row of the t-test without checking Levene's first.
4. Confusing r-squared with r.
5. Not reporting N for each analysis.
My personal recommendation: consider moving away from SPSS
After years using SPSS professionally, I now recommend it less and less. SPSS forces you to calculate many things manually (effect sizes, confidence intervals, robust post-hocs) that other software gives you directly.
My preferred alternatives: R (with packages like effectsize, report, and papaja) and jamovi (free, visual interface similar to SPSS, but returns APA-formatted tables with effect sizes included).
Reading the Coefficients table in multiple regression in detail
The Coefficients table is where most of the regression interpretation actually happens, and where I see the most confusion. SPSS gives you, for each predictor, the unstandardized coefficient B, its standard error, the standardized coefficient Beta, t, Sig., and the 95% confidence interval. Each one answers a slightly different question.
The unstandardized B tells you how much the dependent variable changes for each unit increase in the predictor, holding the other predictors constant. It is in the original units of the variable and is therefore directly interpretable when the variable has a meaningful scale (years, kilograms, points on a clinical scale). The standardized Beta, on the other hand, expresses the change in standard deviations of the dependent variable for each standard deviation of the predictor. It is useful for comparing the relative importance of predictors measured on different scales, but it should not be used to compare across different samples because it depends on the variability of each sample.
What the table does not tell you directly is whether the model meets the assumptions. Before interpreting the coefficients, you must ask SPSS for the diagnostic plots (Analyze, Regression, Linear, Plots), specifically the histogram of residuals, the normal P-P plot, and the scatterplot of standardized residuals against standardized predicted values. Without this, you cannot trust the standard errors or the p-values. There is a more detailed discussion in my article on how to check statistical assumptions.
Another critical point is multicollinearity. In Statistics, Collinearity diagnostics, SPSS gives you Tolerance and VIF. A VIF greater than 5 (and certainly greater than 10) indicates that the predictor is largely explained by other predictors in the model, which makes the coefficient unstable. In that case, you have several options: remove one of the redundant predictors, combine them into a composite index, or move to a regularization technique such as ridge regression (not available in standard SPSS but trivial in R).
Effect sizes that SPSS does not give you and how to calculate them
The systematic absence of effect sizes in SPSS output is, in my opinion, one of the program's main weaknesses. Below is a quick summary of how to compute them manually for the most common analyses.
Independent t-test: Cohen's d = (M1 minus M2) divided by pooled SD. SPSS gives you the mean difference and the SDs of each group, but you must calculate the pooled SD with the formula sqrt(((n1 minus 1) times SD1 squared plus (n2 minus 1) times SD2 squared) divided by (n1 plus n2 minus 2)).
One-way ANOVA: partial eta squared = SS_between divided by SS_total. SPSS gives it to you if you check "Estimates of effect size" in Options. As a rule of thumb (Cohen): 0.01 small, 0.06 medium, 0.14 large.
Repeated-measures ANOVA: use partial eta squared from the Within-Subjects Effects table, but be aware that it inflates with multiple factors. Generalized eta squared (Bakeman, 2005) is preferable for cumulative literature, although SPSS does not calculate it directly.
Chi-square: Cramer's V or phi (the latter only for 2x2 tables). SPSS calculates them if you ask for them in Statistics. Cramer's V interpretation depends on degrees of freedom.
Correlation: r itself is the effect size. If you want r squared (proportion of shared variance), simply square the value.
Two common pitfalls when reading SPSS output
Beyond the errors already mentioned, two specific situations cause confusion in practically every consulting session. The first is reading Mauchly's test of sphericity in repeated-measures ANOVA. If Mauchly's p is less than 0.05, sphericity is violated and you must use the corrected estimates (Greenhouse-Geisser if epsilon is below 0.75, Huynh-Feldt if epsilon is above 0.75). Many researchers ignore Mauchly entirely and report the uncorrected F, which inflates the type I error rate.
The second is misreading the Hosmer and Lemeshow test in logistic regression. Many believe that a non-significant Hosmer-Lemeshow proves good model fit, but the test has very low power with small samples and excessive sensitivity with large samples. A non-significant p does not prove good fit, only that you did not have enough power to detect lack of fit. The Nagelkerke pseudo-R squared is more informative as a description of explanatory capacity, although it should not be interpreted with the same standards as the R squared of linear regression.
References
American Psychological Association. (2020). Publication manual of the American Psychological Association (7th ed.).
Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Lawrence Erlbaum.
Field, A. (2024). Discovering statistics using IBM SPSS statistics (6th ed.). Sage.