It is the most searched SPSS output and the most misread, and not by accident: it has <strong>two columns that look like the same thing</strong> and one that does not mean what it appears to.
Coefficients^a
Unstandardized Standardized
Coefficients Coefficients
Model B Std. Error Beta t Sig.
--------------------------------------------------------------------
1 (Constant) 12.430 2.117 5.872 .000
Hours of sleep 1.842 .461 .318 3.996 .000
Perceived stress -.274 .088 -.241 -3.114 .002
Age .031 .046 .048 .674 .502
a. Dependent Variable: Academic performanceHow much the dependent variable changes per one-unit rise in the predictor, in the original units of both. Here, each extra hour of sleep goes with 1.84 more points of performance.
It cannot be compared across predictors. That sleep’s B (1.842) exceeds stress’s (−.274) does not mean sleep matters more: it means an hour and a point on a stress scale are not the same amount of anything.
The imprecision of that B. It is what turns the coefficient into a test: t = B / standard error, and the p value follows. With 1.842 and .461 you get exactly the 3.996 in the next column.
It is the figure needed to pool your study in a meta-analysis, and the one hardly anyone reports. If you are publishing only one table, make sure it carries the standard error or the confidence interval.
The same effect with every variable put in standard deviations, so they can be compared with each other. Here sleep (.318) weighs slightly more than stress (−.241) within this model.
Comparing Betas across MODELS or across SAMPLES does not work: standardization uses the standard deviations of that sample, so the same real effect yields different Betas in two studies with different variability. To compare across studies, use B with its standard error.
The coefficient divided by its standard error. It is the test statistic for the hypothesis that this predictor adds nothing.
Its degrees of freedom are NOT in this table: they are in the ANOVA table, and they are N − k − 1, with k the number of predictors. That −1 for the intercept is the one most often forgotten.
The two-tailed p value. Here, sleep and stress depart from zero and age does not (.502).
The ".000" is not zero: it is a three-decimal rounding of something below .0005. A p value is never zero. It is written p < .001, and copying it as "p = .000" is the cheapest reviewer comment there is.
The predicted value when every predictor is zero. Here, the performance of someone with zero hours of sleep, zero stress and zero years of age.
It almost never means anything, and neither does its significance. Interpreting the constant as a finding is a sign that nobody centered the predictors. If you want it to mean something, center the variables on their means and the constant becomes the performance of an average person.
How much variance the model explains. That is in the Model Summary table, in R squared and adjusted R squared. Without it, the coefficients table is a list of effects with no context. And if what you need is f² for a power calculation, it comes from there: see R² to Cohen’s f².
Whether the model holds. The assumptions — normal residuals, homoscedasticity, independence — appear in no table: they appear in residual plots, which have to be requested separately and which hardly anyone looks at.
Whether there is collinearity. With correlated predictors the Bs become unstable and the standard errors blow up. SPSS gives it on request (Statistics → Collinearity diagnostics): a VIF above 5 already deserves a sentence in the text.
A regression is not reported by pasting the SPSS table: it is reported with a table of your own carrying B, its standard error, Beta, the 95 % confidence interval and the p value for each predictor, plus the model R² and the overall F with its two degrees of freedom.
In the text, one sentence per relevant predictor: "hours of sleep predicted performance, B = 1.84, SE = 0.46, β = .32, p < .001". The interval matters more than the asterisk.
The full format, with the table laid out, is in how to report linear regression in APA 7.
That every t adds up. Divide B by its standard error: it has to give the t on that row. If it does not, one of the three numbers is mistyped.
That p matches t. Given the residual degrees of freedom, the p value is determined. Check it in the p value against its degrees of freedom.
That the degrees of freedom match your N. They are N − k − 1. If they imply fewer cases than you report, cases were lost and never counted: degrees of freedom against N.
It should not be in simple regression, where Beta equals the correlation. In multiple regression it can happen, and it almost always indicates strong collinearity between predictors: check the VIFs before interpreting anything.
Yes, if both variables are in the same units (years, points on the same scale, euros). It is more comparable than Beta across studies, because it does not depend on each sample’s variability.
Report it. Dropping it because it did not come out is a decision made after seeing the data, and it changes every other coefficient. If you drop it, say so and show both models.
Yes: a block appears per step and the coefficients move at each one. Stepwise regression capitalizes on chance and has been discouraged for decades; if you use it, justify it and validate the model in another sample.