Hierarchical Regression Explained: How to Run, Read and Report It
Hierarchical regression enters predictors in a sequence of blocks that you decide in advance, usually on theoretical grounds, and tests what each block adds. It is the standard way to ask whether a set of predictors explains additional variance in an outcome beyond variables you already know matter. It differs fundamentally from stepwise regression, in which the software chooses the order from the data.
What is hierarchical regression?
You fit a sequence of nested linear models, each adding a block of predictors: Model 1 (controls), Model 2 (controls plus focal predictors), and so on. At each step you record R² and test the change. The key statistic is the change in R²:
ΔR² = R²full − R²reduced Fchange = (ΔR² / m) / [(1 − R²full) / (n − k − 1)]
Here m is the number of predictors added in the block, k the total number of predictors in the larger model, with df = (m, n − k − 1). The order of entry is a substantive claim: it asks what the later block explains after the earlier block has been accounted for, so reversing the order can change ΔR².
When should you use it?
Use it to control for demographic or baseline variables before testing focal predictors, to test an interaction after main effects (see moderation analysis), or to compare theory-based sets of predictors. Specify the blocks before looking at the results (Cohen et al., 2003).
Assumptions
The assumptions are those of multiple linear regression: linearity, independent errors, constant variance, approximately normal residuals, limited multicollinearity and no influential outliers. Models must be nested and fitted to the same cases, so a variable with missing values can change the sample between steps and invalidate ΔR².
Original example (illustrative)
Question. After accounting for prior achievement, do study hours and motivation explain additional variance in final exam score? Data for n = 20 students (invented):
| # | Prior | GPA | Hours | Motiv. | Final | # | Prior | GPA | Hours | Motiv. | Final |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 63.0 | 3.0 | 4.0 | 5.1 | 69.0 | 11 | 60.0 | 3.8 | 9.0 | 6.8 | 79.0 |
| 2 | 67.0 | 2.9 | 6.0 | 5.4 | 71.0 | 12 | 64.0 | 2.6 | 6.0 | 4.6 | 67.0 |
| 3 | 63.0 | 3.5 | 8.0 | 3.5 | 63.0 | 13 | 54.0 | 2.8 | 10.0 | 3.8 | 65.0 |
| 4 | 50.0 | 3.4 | 5.0 | 3.3 | 54.0 | 14 | 59.0 | 3.8 | 6.0 | 6.1 | 69.0 |
| 5 | 67.0 | 2.0 | 8.0 | 5.3 | 72.0 | 15 | 56.0 | 3.3 | 11.0 | 5.4 | 76.0 |
| 6 | 64.0 | 2.2 | 8.0 | 4.3 | 69.0 | 16 | 65.0 | 3.3 | 5.0 | 6.3 | 62.0 |
| 7 | 56.0 | 2.9 | 8.0 | 5.4 | 64.0 | 17 | 60.0 | 2.8 | 11.0 | 4.5 | 69.0 |
| 8 | 65.0 | 2.8 | 6.0 | 6.1 | 69.0 | 18 | 58.0 | 2.3 | 8.0 | 2.8 | 61.0 |
| 9 | 63.0 | 3.1 | 10.0 | 2.5 | 70.0 | 19 | 54.0 | 3.1 | 4.0 | 4.8 | 55.0 |
| 10 | 62.0 | 3.1 | 11.0 | 5.4 | 75.0 | 20 | 58.0 | 3.0 | 7.0 | 4.3 | 66.0 |
Block 1 (controls): prior exam score and prior GPA. Block 2 (focal): weekly study hours and motivation.
| Predictor | b | SE | p |
|---|---|---|---|
| Intercept | 13.77 | 21.99 | .539 |
| Prior exam score | 0.74 | 0.29 | .021 |
| Prior GPA | 2.85 | 2.88 | .337 |
| Predictor | b | SE | p |
|---|---|---|---|
| Intercept | 4.03 | 13.95 | .777 |
| Prior exam score | 0.57 | 0.19 | .009 |
| Prior GPA | 1.32 | 1.89 | .493 |
| Study hours | 1.82 | 0.34 | < .001 |
| Motivation | 2.32 | 0.77 | .008 |
| Model | R² | Adj. R² | ΔR² | F change | p |
|---|---|---|---|---|---|
| 1 | 0.277 | 0.192 | 0.277 | F(2, 17) = 3.26 | .063 |
| 2 | 0.776 | 0.716 | 0.498 | F(2, 15) = 16.64 | < .001 |
Check: Fchange = (0.498/2) / ((1 − 0.776)/15) = 16.64. Model 2 as a whole: F(4, 15) = 12.96, p < .001.
Interpretation. Prior achievement and GPA explain 27.7% of the variance in final scores. Adding study hours and motivation raises this to 77.6%, an increase of 49.8 percentage points, which is statistically significant. So the focal block explains meaningful additional variance beyond the controls. To judge practical importance, look at the coefficients in original units and their intervals, not only ΔR². With only 20 cases and four predictors the estimates are imprecise, and the example data are invented.
How to run hierarchical regression with StatFAS
Open the Hierarchical Regression Calculator. Arrange one outcome column and one column per predictor, then assign predictors to blocks in order.
Understanding the output
- R² and adjusted R² per step: the cumulative explained variance.
- ΔR² and F change: the main evidence that a block adds explanatory power.
- Coefficients per model: interpret those in the final model as adjusted effects, and note how coefficients change between steps.
- Collinearity statistics: large changes in coefficients between steps can signal correlated predictors.
APA 7 reporting example
Hierarchical multiple regression was used to test whether study hours and motivation predicted final exam score beyond prior achievement. Prior exam score and GPA in Step 1 accounted for 27.7% of the variance, F(2, 17) = 3.26, p = .063. Adding study hours and motivation in Step 2 explained an additional 49.8% of variance, ΔR² = .498, F(2, 15) = 16.64, p = < .001.
Common mistakes and troubleshooting
- Choosing the order after seeing the data, which turns a theory test into data dredging.
- Reporting only the final model. The sequence of ΔR² values is the point of the method.
- Different samples across steps because of missing data. Use the same cases for every model.
- Equating ΔR² with importance. A small ΔR² can be practically important and a statistically significant one trivial in a large sample.
- Confusing it with stepwise regression. Hierarchical order is theory-driven; stepwise selection is algorithm-driven and generally not recommended for inference.
Limitations and alternatives
Hierarchical regression shows incremental prediction, not causation, and results depend on the entry order when predictors are correlated. For many predictors with collinearity, consider ridge regression. For nested outcomes or repeated measures, mixed models are more appropriate. For comparing models that are not nested, use information criteria such as AIC.
Conclusion
Hierarchical regression answers a focused question: what does a set of predictors add after accounting for others? Define your blocks from theory, report ΔR² with its test, and interpret the coefficients and their intervals for practical importance.
1.
statfas-hierarchical-regression-interface.png, alt: "StatFAS Hierarchical Regression Calculator interface"2.
statfas-hierarchical-regression-example-data.png, alt: "Example data entered in the StatFAS Hierarchical Regression Calculator"3.
statfas-hierarchical-regression-settings.png, alt: "Hierarchical Regression Calculator settings in StatFAS"4.
statfas-hierarchical-regression-result.png, alt: "Hierarchical Regression Calculator results generated in StatFAS"Frequently asked questions
What is the difference between hierarchical and stepwise regression?
In hierarchical regression the researcher chooses the order of entry on theoretical grounds. In stepwise regression the software selects predictors based on statistical criteria, which tends to capitalise on chance and produce unstable models.
What does ΔR² tell me?
It is the additional proportion of variance in the outcome explained by the predictors added in a step, over and above those already in the model. It is tested with an F-change statistic.
Can the order of entry change the results?
Yes, when predictors are correlated. The incremental variance explained by a block depends on what has already been entered, so the order should reflect your research question.
References
- Cohen, J., Cohen, P., West, S. G., & Aiken, L. S. (2003). Applied multiple regression/correlation analysis for the behavioral sciences (3rd ed.). Erlbaum.
- Draper, N. R., & Smith, H. (1998). Applied regression analysis (3rd ed.). Wiley.