Testing Regression Assumptions in SPSS (Linearity, Residuals, Homoscedasticity)

Running a regression in SPSS takes about two minutes. Knowing whether you are allowed to trust the result takes a little longer — and it is the part most students skip. A regression model rests on assumptions about the relationship between your variables and about the errors the model makes. When those assumptions are badly broken, your p-values, confidence intervals and predictions can all mislead you, even if the output looks clean.

This guide shows you how to test the core regression assumptions in SPSS, click by click: linearity, normality of residuals, homoscedasticity, independence of errors, and — as a quick final check — multicollinearity. You will use one running example throughout: predicting students’ exam scores from hours studied and lecture attendance. Your variables will differ, but the steps and the plots to look for are exactly the same.

Why assumptions matter more than the p-value

A p-value tells you whether a result would be surprising if there were no effect. It does not tell you whether the model that produced it fits your data. Think of regression assumptions as the fine print of that p-value. If the relationship is curved rather than straight, a linear model will understate or overstate the effect in different parts of the range. If the errors fan out as predicted values rise, the standard errors are wrong and so are the significance tests built on them. Checking assumptions first does not make your analysis slower in any meaningful sense; it stops you from writing up a result you later have to retract.

The six assumptions at a glance

Before you open any dialog box, it helps to know what you are checking and where each check lives in the output:

1. Linearity: each predictor has an approximately straight-line relationship with the outcome.
2. Normality of residuals: the residuals (the gaps between observed and predicted values) are approximately normally distributed.
3. Homoscedasticity: the residuals have roughly constant spread across all levels of the predicted values.
4. Independence of errors: one case’s residual does not predict another’s — especially important with time-ordered or clustered data.
5. No perfect multicollinearity: predictors are not so strongly related to each other that their separate effects cannot be untangled.
6. Correct measurement level: the outcome is continuous, and predictors are continuous or properly coded categories. SPSS will happily run a regression on almost anything, so this one is on you.

This guide works through checks 1 to 4 in detail, because those are the ones students most often miss, and adds a quick check for number 5 at the end.

Before you start: prepare your variables

Open your data file and confirm three things. First, your outcome variable is numeric and continuous — exam score, salary, reaction time, or something similar. Second, your predictors are numeric or coded as dummy variables if they are categories. Third, you have no obvious data-entry errors: run a quick frequency check on each variable and look for impossible values, such as an exam score of 999 or negative hours studied. Fix those before you test anything else, because a single wild value can distort every plot in this guide.

Step 1: Test linearity

What to look for: for each predictor, the relationship with the outcome should look roughly like a straight line, not a curve, a U-shape, or a flat line that suddenly jumps.

Click-by-click in SPSS:
1. Go to Graphs, then Legacy Dialogs, then Scatter/Dot.
2. Choose Simple Scatter and click Define.
3. Move your outcome variable into the Y Axis box and your first predictor into the X Axis box, then click OK.
4. Repeat for each predictor.
5. Optionally, double-click the chart, then use Elements and Add Fit Line at Total to add a straight line and judge how well it follows the cloud of points.

How to interpret it: if the points form a roughly straight band that rises or falls steadily, linearity is reasonable. If the points clearly curve — rising then falling, or flattening out — a straight-line model is the wrong shape. Common fixes are transforming a variable, for example using a log transformation for a strongly skewed predictor, or adding a squared term for a genuinely curved relationship. Do not just ignore a clear curve and report the linear result anyway; the slope you report will be an average that describes no part of the data well.

You can also check linearity from the regression residual plot in Step 3: a curved pattern in that plot is another warning sign that the straight-line form does not fit.

Step 2: Run the regression and check normality of residuals

Now run the regression itself, asking SPSS to save the plots you need.

Click-by-click in SPSS:
1. Go to Analyze, then Regression, then Linear.
2. Move your outcome variable into the Dependent box and your predictors into the Independent(s) box.
3. Click Plots. Move ZRESID into the Y box and ZPRED into the X box. Tick Histogram and Normal probability plot, then click Continue.
4. Click Statistics. Tick Durbin-Watson and Collinearity diagnostics, then click Continue.
5. Click OK to run the analysis.

What to look for in the normality plots: the histogram of standardized residuals should look roughly bell-shaped and centred on zero, without a long tail on one side or two separate humps. In the Normal P-P plot, the points should hug the diagonal line fairly closely. Small wiggles at the ends are normal, especially with smaller samples. A clear S-shape or points drifting far from the line at one end suggests the residuals are not normal.

What to do if normality fails: first look for outliers — a handful of extreme cases can wreck an otherwise normal distribution. Check whether those cases are data errors or genuine but unusual observations, and report what you did with them rather than quietly deleting them. With large samples, regression is fairly robust to mild non-normality, because the sampling distribution of the coefficients tends toward normality anyway. With small samples and strong skew, consider transforming the outcome or using a method suited to the distribution, and say so plainly in your write-up.

Step 3: Check homoscedasticity — the scatterplot that matters most

Homoscedasticity is an awkward word for a simple idea: the errors should be about equally spread out no matter what value the model predicts. Its opposite, heteroscedasticity, means the spread grows or shrinks across the range — for example, predictions are tight for low scorers but wildly scattered for high scorers.

You already asked SPSS for the key plot in Step 2: standardized residuals (ZRESID) on the vertical axis against standardized predicted values (ZPRED) on the horizontal axis. Find it in the output under Charts.

How to interpret it: a healthy plot looks like a formless cloud — points scattered evenly above and below zero, with roughly the same vertical spread from left to right. That is what you want. A funnel or cone shape, where the cloud widens as you move right, signals heteroscedasticity. A curved band signals a linearity problem rather than a spread problem, so go back to Step 1 if you see that shape.

A common mistake is to test homoscedasticity with Levene’s test here. Levene’s test compares variances across groups and belongs with t-tests and ANOVA, not with a regression residual plot. In regression, the ZPRED-versus-ZRESID scatterplot is the standard visual check, and it is the one examiners and reviewers expect to see described.

What to do if the spread is uneven: check first whether an outlier or a skewed predictor is driving the pattern. A transformation of the outcome or a predictor sometimes evens the spread. If the pattern persists, robust standard errors or weighted least squares are the standard remedies — both are beyond a first regression course, but naming the problem honestly and using a remedy beats reporting standard errors you know are wrong.

Step 4: Check independence of errors with Durbin-Watson

Independence means that knowing one case’s error tells you nothing about the next case’s error. It is most at risk when cases are ordered in time, such as monthly sales figures, or grouped, such as students nested within classes. If errors are correlated, SPSS will tend to understate uncertainty and overstate significance.

Find the Durbin-Watson statistic in the Model Summary table, which you requested under Statistics in Step 2. Values range from 0 to 4, and a value near 2 indicates no serious autocorrelation. As a rough guide, values between about 1.5 and 2.5 are generally treated as acceptable in student work, while values close to 0 suggest positive autocorrelation and values close to 4 suggest negative autocorrelation.

If independence is violated, the fix depends on the design: time-series methods for time-ordered data, or a multilevel model for clustered data. At minimum, do not report the ordinary regression p-values as if the problem did not exist.

Step 5: A quick multicollinearity check

Multicollinearity is what happens when two predictors overlap so much that SPSS cannot tell which one is doing the work. You asked for collinearity diagnostics in Step 2, so look at the Coefficients table: the Tolerance and VIF columns. A variance inflation factor (VIF) above 10, or a tolerance below 0.10, is the conventional warning threshold, and values above 5 deserve a closer look. If you find a problem, consider dropping one of the overlapping predictors or combining them into a single score, and explain the choice. A fuller walkthrough of VIF and tolerance is a separate topic; the point here is that this one-minute check belongs in every regression run.

Paste-ready SPSS syntax

If you prefer syntax, or you need to reproduce the analysis exactly, this block runs the same regression with all the checks above. Replace the variable names with your own:

REGRESSION
/MISSING LISTWISE
/STATISTICS COEFF OUTS R ANOVA COLLIN TOL
/CRITERIA=PIN(.05) POUT(.10)
/NOORIGIN
/DEPENDENT exam_score
/METHOD=ENTER hours_studied attendance
/SCATTERPLOT=(*ZRESID ,*ZPRED)
/RESIDUALS HISTOGRAM(ZRESID) NORMPROB(ZRESID).

Run it from a syntax window with Run, then All, and you will get the same tables and plots as the menu route. Saving the syntax file alongside your data file is good practice: it makes your analysis reproducible and saves you from re-clicking through five dialog boxes when a supervisor asks you to rerun it.

How to interpret the output, in order

Work through the output in this order and you will not miss anything. First, the Model Summary: note R-squared as the share of variance explained, and the Durbin-Watson value for independence. Second, the ANOVA table: this tests whether the model as a whole predicts better than no predictors at all. Third, the Coefficients table: read each predictor’s unstandardized coefficient (B) as the expected change in the outcome for a one-unit increase in that predictor, holding the others constant, and check its significance and its VIF. Fourth, the plots: histogram, P-P plot, and the ZPRED-versus-ZRESID scatterplot. Only when the plots look acceptable should you lean on the p-values from the Coefficients table. If a plot looks wrong, fix the problem first and rerun — that rerun, and why you did it, belongs in your methods note.

What to do when an assumption fails: a summary

Linearity fails: consider a transformation or a squared term, and recheck the scatterplot.
Normality of residuals fails: check outliers, consider a transformation, and be cautious with small samples.
Homoscedasticity fails: check outliers and skew, then consider robust standard errors or weighted least squares.
Independence fails: reconsider the design — time-series or multilevel methods may be needed.
Multicollinearity is high: drop or combine an overlapping predictor and explain why.
In every case, report what you checked and what you found, including when you found a problem and fixed it. Reviewers trust a methods section that shows its working far more than one that claims everything was perfect.

Reporting in APA style (7th ed.)

APA style does not require a separate assumptions paragraph for every analysis, but a good results section shows that the checks were done. A concise template you can adapt:

“Prior to analysis, regression assumptions were examined. Scatterplots supported linear relationships between each predictor and the outcome. Standardized residuals were approximately normally distributed, as indicated by the histogram and Normal P-P plot, and the scatterplot of standardized residuals against standardized predicted values showed no clear funnel pattern, supporting homoscedasticity. The Durbin-Watson statistic was close to 2, indicating independence of errors, and variance inflation factors were below the conventional threshold, indicating no problematic multicollinearity.”

Then report the model itself in the usual form: the overall test, R-squared, and each predictor’s unstandardized coefficient with its confidence interval and p-value. Replace the general wording above with your actual values and, where a check was borderline, say so — for example, that the residual plot showed mild widening at higher predicted values and that robust standard errors were therefore used. Honest qualification strengthens a results section; it does not weaken it.

Key Takeaways

– Always check assumptions before interpreting p-values: a significant coefficient from a badly fitting model is not a finding.
– Linearity is checked with simple scatterplots of each predictor against the outcome, before or alongside the regression.
– Normality in regression concerns the residuals, not the raw variables — use the histogram and Normal P-P plot of standardized residuals.
– Homoscedasticity is judged from the ZPRED-versus-ZRESID scatterplot: you want a formless cloud, not a funnel.
– Durbin-Watson near 2 supports independence of errors; values far from 2 warn of autocorrelation.
– Save your syntax: it makes the whole checking routine reproducible in one click next time.

Frequently asked questions

Do I need to test the normality of my variables before regression?
Not exactly. The normality assumption applies to the residuals, not to each variable on its own. A skewed predictor does not by itself violate the assumption, though strong skew can contribute to non-normal residuals or a curved relationship, which is why the residual plots matter more than a normality test on the raw scores.

Is there a formal test for homoscedasticity in SPSS?
SPSS does not offer a single built-in heteroscedasticity test in the standard Linear Regression dialog, which is why the residual scatterplot is the expected check in student and thesis work. Formal tests exist in other software, but a clearly funnel-shaped plot is evidence enough to act on, and a formless cloud is evidence enough to proceed.

What if my sample is small and the P-P plot looks a bit wobbly?
Small samples produce wobbly plots even when nothing is seriously wrong, because each point carries more weight. Look for systematic drift — a clear curve away from the diagonal — rather than small local wiggles, and be more cautious about strong claims when the sample is small.

Can I still report my regression if one assumption is mildly violated?
Often, yes, with qualification. Regression is reasonably robust to mild violations, especially with larger samples. Report what you found, what you did about it, and how it limits the conclusion. What you should not do is hide a clear violation and report the p-values as if the model fit perfectly.

Should I check assumptions for every regression I run, or just the final one?
Check them for the model you report, and recheck them whenever you change the model — adding or dropping a predictor, transforming a variable, or removing cases can all change the residual pattern. A check done on an earlier version of the model does not automatically carry over.

Need this done on your data? Get a free quote → /services/