Written and maintained by CASRAI Editorial Board
Last updated
Hierarchical regression is a way of building a multiple regression model in stages: you enter predictors in blocks, in an order you fix before looking at the results, and test how much each new block adds to the variance explained. It is sometimes called sequential or block-entry regression, and it answers a specific kind of question — “does this variable (or set of variables) explain additional outcome variance after accounting for these other variables?” — that a single simultaneous regression cannot answer directly. This guide covers how block order is set, how to run and read the R² change test with a full worked example, how to report it, and how it differs from stepwise regression, the method it is most often confused with.
What makes a regression “hierarchical”
In an ordinary (simultaneous) multiple regression, every predictor enters the model at once, and the model is estimated once. Hierarchical regression instead splits the predictor set into two or more blocks and estimates the model once per block, adding predictors cumulatively:
- Block 1 enters a first set of predictors (commonly demographic or control variables) and produces R²1.
- Block 2 adds a second set of predictors on top of Block 1 and produces R²2.
- Each additional block repeats the pattern, and the model can have as many blocks as the research question needs, though two or three is typical.
The quantity of interest at each step is not R² itself but the change in R² from one block to the next (ΔR²) — the additional variance explained by the new block, over and above what the prior blocks already explained. Everything else about the underlying model (ordinary least squares estimation, the assumptions it rests on) is identical to any other multiple regression; hierarchical entry changes how the model is built and reported, not the estimator itself.
The defining rule: block order is set by theory, before you see the results
The entire defensibility of hierarchical regression rests on one condition: the number of blocks, which predictors go in which block, and the order the blocks are entered in are all decided before the analysis is run, based on a substantive rationale — not chosen or adjusted after seeing which ordering produces the largest or most significant ΔR². Three ordering rationales cover most real designs:
- Temporal or causal precedence. Variables that are logically or causally prior go in earlier blocks. Demographic characteristics (age, sex) are fixed before an intervention occurs, so they enter Block 1; the intervention or exposure of interest enters a later block.
- Control-then-focal-predictor logic. Nuisance or confound variables the researcher wants to partial out enter first; the variable the study actually exists to test enters last, so its ΔR² represents its contribution net of everything already controlled for.
- Theoretical hierarchy. A theory specifies that one class of predictors (e.g., individual-level factors) should be considered before another (e.g., organizational-level factors), independent of any temporal argument.
This is the property that separates hierarchical regression from stepwise regression, which selects and orders variables algorithmically, based on which one improves fit the most in this sample. Hierarchical entry answers a pre-specified, theory-driven question about incremental contribution; stepwise selection searches the data for whichever variables and ordering fit best, which is exactly the practice that draws a multiple-testing objection from reviewers. If the block order in a hierarchical analysis was actually chosen after inspecting which ordering “worked,” it is not meaningfully different from stepwise selection dressed in hierarchical language — the defensibility depends entirely on the ordering being fixed in advance, not on the procedure’s name.
The ΔR² significance test, worked through
The test asks whether the R² gained by adding a block is larger than would be expected by chance, given the number of new predictors and the sample size. It is an F-test built directly from the two models’ R² values, their predictor counts, and the sample size:
F = [(R²2 − R²1) / (k2 − k1)] ÷ [(1 − R²2) / (N − k2 − 1)]
where R²1 and R²2 are the R² values before and after adding the block, k1 and k2 are the total number of predictors in the model before and after, and N is the sample size. The resulting F statistic is evaluated against an F-distribution with (k2 − k1) and (N − k2 − 1) degrees of freedom — see degrees of freedom for how those two numbers are derived generally. Most statistical software (SPSS’s “R Squared Change” column, R’s anova() applied to two nested lm() models, Stata’s lrtest for the OLS case) computes this automatically, but working through it by hand once is the fastest way to understand what the output table is actually testing.
Illustrative worked example
This example uses illustrative, generic numbers to demonstrate the calculation — it is not drawn from any real study or dataset. Suppose a researcher is predicting a continuous outcome (say, a 100-point wellbeing scale) from N = 150 respondents, using a two-block design: Block 1 enters two demographic controls (age, income); Block 2 adds a single focal predictor (a social-support score) that the study exists to test, net of those controls.
| Step | Predictors in model | k | R² |
|---|---|---|---|
| Block 1 | Age, income | 2 | 0.090 |
| Block 2 | Age, income, social support | 3 | 0.160 |
ΔR² = 0.160 − 0.090 = 0.070. Plugging into the formula, with k1 = 2, k2 = 3, and N = 150:
F = [0.070 / (3 − 2)] ÷ [(1 − 0.160) / (150 − 3 − 1)] = 0.070 ÷ (0.840 / 146) = 0.070 ÷ 0.00575 ≈ 12.17
With df = (1, 146), an F of roughly 12.17 clears the critical value for α = 0.05 by a wide margin (the critical value is approximately 3.91), so the researcher would report that adding social support significantly improved the model, ΔR² = 0.070, F(1, 146) = 12.17, p < .001, after controlling for age and income. Note that the test degrees of freedom here (1) come directly from the block containing exactly one new predictor; a block that adds three predictors at once would use df = 3 for that step, and the F-test would then be a joint test of whether that whole block adds explanatory power, not a per-variable test.
Effect size for the increment: f²
ΔR² itself is a useful effect-size metric on its own scale (0 to 1, same units as R²), but Cohen’s f² expresses the same increment on a standardized scale that supports the conventional small/medium/large benchmarks used elsewhere in behavioral-science reporting (see effect size):
f² = (R²2 − R²1) / (1 − R²2)
In the worked example, f² = 0.070 / 0.840 ≈ 0.083, which sits in Cohen’s small-to-medium range (conventional cutoffs: 0.02 small, 0.15 medium, 0.35 large) — useful context beyond the significance test alone, since a large sample can make a substantively small ΔR² statistically significant, and f² helps a reader judge whether the increment is also practically meaningful.
Running it: what changes across statistical packages
The underlying logic is identical everywhere; only the mechanics of specifying blocks differ:
- SPSS supports hierarchical entry natively in Analyze > Regression > Linear by moving predictors into successive numbered blocks and selecting “Enter” as the method for each block; the Model Summary table’s “R Square Change,” “F Change,” and “Sig. F Change” columns report exactly the test worked through above, one row per block. See Multiple Regression in SPSS for the full click-path and output walkthrough.
- R has no dedicated hierarchical-entry menu; instead, fit two (or more) nested
lm()models directly and compare them withanova(model1, model2), which reports the same F-test on the R² increment (framed as a residual sum-of-squares comparison, mathematically equivalent). - Stata fits each block as a separate
regresscommand and compares nested models withlrtest(after re-fitting withmlogit-style ML estimation) or by computing the F-test on the stored R² values fromereturn listdirectly, sinceregressis fit by OLS rather than maximum likelihood.
Whichever package produces the numbers, report the same quantities: R² and ΔR² at each block, the F-change statistic and its degrees of freedom, and the significance of the change — not just the final model’s R² and coefficients, which on their own hide exactly the incremental-contribution question hierarchical entry was designed to answer.
Hierarchical regression vs. the alternatives
| Method | How predictors enter | What it answers | Multiplicity risk |
|---|---|---|---|
| Simultaneous (standard) entry | All predictors enter one model at once | Each predictor’s unique contribution controlling for all others, in a single model | Low — one model, one test per coefficient |
| Hierarchical (block) entry | Predictor blocks enter in a fixed, theory-specified order | Does this block add explanatory power beyond the prior blocks? | Low, provided the order was genuinely fixed before the analysis |
| Stepwise (forward/backward/bidirectional) | Algorithm adds/removes individual predictors based on fit in this sample | Which subset of candidates best fits this specific dataset | High — see why reviewers reject it |
Hierarchical entry and stepwise selection are frequently conflated because both involve entering predictors in stages, but they answer different questions and carry very different inferential risk. A hierarchical design can still be criticized on its own terms — reviewers may reasonably ask whether the block order was truly pre-specified, or whether the researcher tried several orderings and reported the one that looked best — but that criticism targets a violation of the method’s own rule, not a structural flaw in the method the way the stepwise objection does.
When hierarchical entry is the right choice
- You have a specific, pre-registered or theory-grounded question about whether one variable (or block of variables) adds explanatory power after accounting for others — e.g., “does perceived organizational support predict turnover intention after controlling for tenure and salary?”
- You want to control for known confounds before testing a focal predictor, and you want the contribution of that focal predictor reported net of the controls, not just its own bivariate association.
- You are comparing competing theoretical models nested within each other (a smaller model is a special case of a larger one), and you want a formal test of whether the larger model’s added complexity is justified by the data.
It is the wrong tool when the actual goal is exploratory variable selection from a large candidate pool with no prior theoretical ordering — that is a different problem, better served by a pre-specified full model, by LASSO or another regularized approach if the candidate set is genuinely large, or by explicitly framing the analysis as exploratory and reporting it as such.
Reporting checklist
- State the block structure and the rationale for the order, before presenting results — a reader should be able to see that the order was theoretically motivated, not chosen after the fact.
- Report R² for every block, not just the final model.
- Report ΔR², F-change, its degrees of freedom, and its p-value for every block after the first.
- Report the final model’s coefficients (unstandardized B, standardized β, standard errors, and p-values) alongside the block-level R² changes — the two are complementary, not interchangeable: coefficients describe each predictor’s role in the final model, while ΔR² describes each block’s incremental contribution.
- If more than two blocks are used, report the R² change test for every step, not only the last one.
Frequently asked questions
Is hierarchical regression the same as multilevel (hierarchical linear) modeling?
No, despite the shared word. Hierarchical (block-entry) regression is an ordinary multiple regression run in stages by entering predictor blocks sequentially, with one clear outcome variable and one level of observation. Multilevel modeling (also called hierarchical linear modeling or mixed-effects modeling) is a different technique entirely, used when observations are nested within groups (students within schools, patients within hospitals) and the nesting itself needs to be modeled. The naming overlap is a common source of confusion in the methods literature, not a sign the two techniques are related.
How many blocks can a hierarchical regression have?
There is no fixed limit, but two or three blocks covers the large majority of real designs, and each additional block adds another R²-change test to interpret and report. More blocks are justified when the theoretical model genuinely specifies more than two or three conceptually distinct stages (e.g., demographic controls, then situational variables, then the focal predictor of interest), not simply to break up a longer predictor list.
What sample size does hierarchical regression need?
The same general sample-size considerations that apply to any multiple regression apply here, applied to the full model (all blocks combined) rather than to each block separately, since the final block’s R² and the overall model’s degrees of freedom are what determine estimation stability. A model with more blocks and more total predictors needs a larger sample than a simpler one to detect a given ΔR² with adequate power; a formal a priori power analysis for the R² increment is preferable to a rule of thumb when the design allows for one.
Can I use hierarchical entry with a categorical or binary outcome?
The R²-change F-test described here is specific to ordinary least squares regression with a continuous outcome. For a binary or categorical outcome modeled with logistic regression, the equivalent procedure is nested-model comparison using a likelihood-ratio test rather than an F-test on R² — see the likelihood ratio test for nested models for that version of the same underlying logic.
Does a non-significant ΔR² mean the added block’s variables don’t matter?
It means the block, taken as a whole, did not explain a significant amount of additional variance beyond the prior blocks in this sample and at this sample size — not that no individual variable in the block has any real relationship with the outcome. A block can contain one strong predictor and several weak or redundant ones whose combined contribution is diluted; check the final model’s individual coefficients, and consider whether the sample size gave the block-level test adequate power, before concluding a whole block is irrelevant.
Sources
- Cohen, J., Cohen, P., West, S.G., and Aiken, L.S. Applied Multiple Regression/Correlation Analysis for the Behavioral Sciences (3rd ed.). Routledge. The standard methodological reference for hierarchical (sequential) multiple regression and the R²-change significance test.
- Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates. Source of the f² effect-size benchmarks (0.02 / 0.15 / 0.35).
