Written and maintained by CASRAI Editorial Board
Last updated
SPSS runs multiple linear regression through Analyze > Regression > Linear, and the single most consequential choice in that dialog is the one researchers skip past fastest: the Method dropdown. Enter, Stepwise, and Hierarchical (sequential blocks) produce genuinely different models from the same data — not just different output formatting — and the right one is determined by your research question, not by which option happens to be highlighted by default. This guide covers choosing among the three, requesting and reading the collinearity diagnostics (VIF and Tolerance) and residual plots most default runs skip, and turning the output into a reportable full-model write-up.
Choosing an entry method by research question
SPSS’s Method dropdown is not a matter of preference — each option answers a different question, and reviewers can usually tell which one you actually needed from the write-up alone.
| Method | What it does | Use it when… |
|---|---|---|
| Enter (Standard) | Forces every predictor into the model in a single block, regardless of significance. | You have a pre-specified, theory-driven set of predictors and want to test the full model as a whole — the default choice for confirmatory research. |
| Hierarchical (sequential blocks) | Predictors are entered in researcher-defined blocks (Block 1, Block 2…), each block still using Enter internally; SPSS reports the R² change each new block contributes. | You want to test whether a set of predictors explains variance beyond variables already accounted for — controlling for demographics before entering variables of theoretical interest, for example. |
| Stepwise (Forward, Backward, or Stepwise) | SPSS adds or removes predictors automatically based on statistical criteria (each variable’s F-to-enter/F-to-remove significance), with no researcher input on order. | The analysis is genuinely exploratory and you say so explicitly — it is a variable-selection tool, not a hypothesis test. |
The practical guidance most methods reviewers apply: default to Enter unless you have a specific, defensible reason to do otherwise. Hierarchical is the right tool when the research question is explicitly about incremental variance — “does X predict Y after controlling for Z?” — and the block order should follow theoretical or temporal logic (control variables first, variables of interest later), decided before looking at the data, not chosen after the fact to produce a cleaner result. Stepwise methods capitalize on chance associations in the specific sample (the selected variable set rarely replicates in a new sample) and are treated with real skepticism in published research; if you use one, disclose it as exploratory and do not present the resulting model as if it were confirmatory. This is the same caution SPSS’s own logistic regression procedure carries — see binary logistic regression in SPSS for how the identical Enter-vs-stepwise choice plays out for a categorical outcome.
Running Enter (standard) multiple regression
Go to Analyze > Regression > Linear. Move your outcome into the Dependent box and your predictors into Independent(s). Leave Method on its default, Enter. Before running, confirm your outcome is genuinely continuous (interval or ratio) — see levels of measurement if you’re unsure — and that any categorical predictor with more than two categories has been recoded into dummy variables first; SPSS’s Linear Regression dialog, unlike Binary Logistic, has no built-in Categorical button to do this for you.
Click Statistics and check Estimates, Confidence intervals, Model fit, Collinearity diagnostics, and Durbin-Watson — only Estimates and Model fit are on by default, and the others are exactly the checks a complete write-up needs. Click Plots and request the residual plots described below. Click Save if you want standardized or Cook’s-distance values written back to the data view for outlier screening.
Running Hierarchical (sequential-block) regression
The dialog is the same, but instead of moving every predictor into Independent(s) at once, enter your first block (e.g., control variables), click Next, then enter the second block (e.g., your predictors of theoretical interest) into the now-empty Independent(s) box. Repeat Next for additional blocks. Each block keeps Method set to Enter. Under Statistics, check R squared change specifically — this is the table hierarchical regression exists to produce, and it is not checked by default.
The output adds a row per block to the Model Summary table, each with its own R², R² change, and an F change significance test for whether that block’s addition improved the model significantly. This F-change test, not the R² change value alone, is what supports the incremental-variance claim in a write-up: “Adding [Block 2 variables] explained an additional 8% of variance in [outcome], ΔR² = .08, F(2, 145) = 6.42, p = .002.”
Running Stepwise regression — and reporting it honestly
Set Method to Stepwise (or Forward/Backward if you want the process to only add or only remove, respectively). SPSS uses default entry/removal significance thresholds (typically p ≤ .05 to enter, p ≥ .10 to remove) that you can adjust under Options, though changing them without a stated reason invites the same “why this cutoff” question as any other post hoc choice. The output’s Excluded Variables table lists every predictor SPSS considered but did not retain, alongside its partial correlation and significance at the point it was excluded — worth reporting alongside the final model so a reader can see what was tested, not just what survived.
Because stepwise selection is sample-specific and known to inflate Type I error across the many implicit comparisons it runs, most journals and thesis committees expect it labeled explicitly as exploratory, and some reject it outright for confirmatory claims. If your research question is genuinely confirmatory, use Enter with a pre-specified predictor set instead — see regression analysis: assumptions, interpretation, and how to report it for the model-selection logic that applies across software.
Requesting and reading collinearity diagnostics: VIF and Tolerance
With Collinearity diagnostics checked under Statistics, SPSS appends Tolerance and VIF (Variance Inflation Factor) columns to the Coefficients table — one pair of values per predictor, regardless of which entry method you used. Tolerance is the proportion of a predictor’s variance not explained by the other predictors in the model; VIF is its reciprocal (1 ÷ Tolerance), so the two always tell the same story from opposite directions.
The commonly applied rule of thumb: a VIF above 10 (equivalently, Tolerance below 0.10) signals problematic multicollinearity, though some methodologists apply a more conservative VIF > 5 / Tolerance < 0.20 cutoff, particularly with smaller samples. High multicollinearity does not bias the model’s overall predictions, but it inflates the standard errors of the individual affected coefficients — which is why a predictor can show a large, theoretically meaningful B with a nonsignificant p-value once its variance is being shared with a correlated predictor in the same model. The fix is never to simply drop the “problem” variable without justification; combining correlated predictors into a composite, choosing the more theoretically important one and removing the other with a stated rationale, or centering variables before creating interaction terms are the standard responses — see multicollinearity and VIF in regression for the full detection-and-remedy walkthrough, which applies identically regardless of which software produced the VIF value.
Reading the residual plots
Under Plots, move *ZRESID (standardized residuals) to the Y axis and *ZPRED (standardized predicted values) to the X axis, then check Normal probability plot. This produces two diagnostic outputs SPSS does not generate by default:
- The ZPRED-by-ZRESID scatterplot checks two assumptions at once: linearity and homoscedasticity. A random, roughly rectangular scatter of points around the horizontal zero line is consistent with both. A curved band indicates a nonlinear relationship the linear model is not capturing; a cone or funnel shape (residual spread widening or narrowing across the range of predicted values) indicates heteroscedasticity — unequal error variance that undermines the standard errors and significance tests, even though the coefficient estimates themselves stay unbiased.
- The Normal P-P Plot of Regression Standardized Residual checks whether the residuals are normally distributed — not whether the raw variables are, which is a common misreading. Points falling close to the diagonal reference line support normality; systematic curvature away from it, especially in the tails, does not. See testing normality in SPSS for the Shapiro-Wilk/Kolmogorov-Smirnov route if you need a formal test alongside the visual check.
Also check the Durbin-Watson statistic in the Model Summary table (requested under Statistics): it tests for autocorrelation among residuals and matters most for time-ordered data. Values close to 2 indicate no autocorrelation; values below roughly 1.5 or above roughly 2.5 are commonly flagged as worth investigating, though the appropriate bounds depend on sample size and number of predictors.
Reading the Model Summary, ANOVA, and Coefficients tables
SPSS prints the same three core tables regardless of entry method:
- Model Summary — R (the multiple correlation coefficient), R² (proportion of variance in the outcome explained by the predictor set), and Adjusted R², which corrects R² for the number of predictors and sample size and is the value to report when comparing models with different predictor counts, since R² alone mechanically rises every time a predictor is added regardless of whether it genuinely helps.
- ANOVA — the F-test of whether the model as a whole explains significantly more variance than an intercept-only model. A significant F says the predictor set collectively matters; it says nothing about which individual predictors are doing the work, which is what the Coefficients table is for.
- Coefficients — unstandardized B (the predicted change in the outcome per one-unit change in that predictor, holding others constant, in the outcome’s original units), Std. Error, standardized Beta (B converted to standard-deviation units, which makes predictors on different scales comparable to each other), t and Sig. for each predictor’s individual significance, and — once requested — the 95% confidence interval for B and the Tolerance/VIF pair described above.
Reporting the full model
A complete write-up reports the overall model fit, then each predictor, in a form a reader can check against the tables directly:
“A multiple regression was conducted to predict [outcome] from [predictor 1], [predictor 2], and [predictor 3]. The overall model was significant, F(3, 146) = 18.74, p < .001, R² = .28, adjusted R² = .27, indicating the predictor set explained approximately 28% of the variance in [outcome]. [Predictor 1] was a significant positive predictor (B = 0.42, SE = 0.09, β = .35, t = 4.67, p < .001), as was [predictor 2] (B = 0.18, SE = 0.07, β = .19, t = 2.57, p = .011); [predictor 3] was not a significant predictor once the others were accounted for (B = 0.05, SE = 0.06, β = .06, t = 0.83, p = .408). All VIF values were below 2, indicating multicollinearity was not a concern.”
For a hierarchical model, add the block-level R² change statistic before the final block’s coefficients, in the format shown in the hierarchical section above. See how to report p values for the current APA formatting conventions this example follows, and effect size in research for framing R² and Cohen’s f² as effect sizes rather than reporting significance alone. If your predictor count relative to sample size is a concern, a sample-size-and-power check before running the model — not after a nonsignificant result — is the more defensible order of operations.
Frequently asked questions
Do I need to run Enter, Hierarchical, and Stepwise on the same data to compare them?
No — the entry method should follow from the research question decided before analysis, not be chosen by seeing which produces the best-looking result. Running several and reporting whichever fits best is a form of post hoc model selection that inflates false-positive risk exactly like unreported stepwise selection does.
What sample size does multiple regression need?
Rules of thumb (commonly cited: roughly 10–20 cases per predictor) are a rough floor, not a substitute for an actual power analysis based on your expected effect size, number of predictors, and desired power — run one before collecting data rather than checking adequacy after the fact.
Can I use multiple regression with a categorical outcome?
No — standard multiple linear regression assumes a continuous outcome. A binary outcome calls for logistic regression instead; see binary logistic regression in SPSS. Categorical predictors, by contrast, are fine in linear regression once dummy-coded.
My VIF values are all low but I still have a nonsignificant model — what does that mean?
Low VIF rules out multicollinearity as the explanation, which narrows it down: either the predictors genuinely have limited relationship to the outcome in this sample, the sample is underpowered to detect the true effect sizes involved, or an assumption other than collinearity (linearity, an unaddressed outlier, a measurement issue) is distorting the result — check the residual plots above before concluding the predictors simply don’t matter.
For the broader menu of what SPSS can and can’t do beyond regression, see what is SPSS; for the same regression logic run in Stata’s command syntax instead of SPSS’s menus, see regression in Stata.
