Hierarchical (Block-wise) Regression

Hierarchical regression examines the order in which a model grows rather than the coefficients of one finished model. The researcher groups predictors into blocks on theoretical grounds, fits the first block alone, then adds the second and refits, and asks whether the rise in explained variance between the two steps could plausibly be chance. The question is therefore not whether a coefficient differs from zero but whether a new block adds anything on top of what the earlier ones already accounted for. A typical setup puts variables needing to be controlled in the first block and the variables of theoretical interest in later ones, so that the contribution of the interesting set can be read net of the controls. Block order is a decision for the researcher rather than for the statistics, and it should never be settled by looking at the data. YouReply Analyze fits each step separately through the OLS class of statsmodels; two blocks are required and a third is optional, and no predictor may appear in more than one block. Every step returns the variables added, the variables then in the model, R squared, adjusted R squared, the model F with its p value, and the change in R squared with its F change, the p value of that change and a significance marker. The coefficients of the final model accompany them with standard errors, t values and p values.

• Once demographic variables are controlled, does job satisfaction still help explain organisational commitment?
• With brand awareness and price perception already in the model, how much does advertising recall add to purchase intention?
• Taking age and income as a base, do personality variables meaningfully improve the explanation of consumer spending?
• Once prior attainment is controlled, what does perceived teacher support add to the prediction of exam scores?

• When the contribution of one set of predictors has to be shown net of the effect of another set.
• When control variables and variables of theoretical interest need to be reported separately.
• When the order of the blocks follows from theory or temporal precedence, rather than being chosen by inspecting the data.
• When the dependent variable is continuous and the membership of each block was fixed before the analysis began.

• Dependent variable: a continuous column. Scale totals, durations and spending amounts all serve.
• The first and second blocks are required, each holding at least one numeric predictor, and a third block can be added optionally.
• A predictor belongs to exactly one block; placing the same column in two blocks is refused.
• Categorical predictors must be dummy-coded, and all dummies derived from one categorical variable belong in the same block.

Block order grounded in theory

The sequence of steps determines the answer: the same predictors entered in a different order take different incremental shares. The order has to rest on theory, temporal precedence or the design of the study.

Linear relationships with the outcome

Each predictor is taken to relate linearly to the dependent variable. Where the relationship curves, the R squared increments come out smaller than they should and the block's contribution disappears from view.

Normally distributed residuals with constant variance

The model F tests and the coefficient p values rest on residuals that are normally distributed and spread evenly across the fitted values. When that fails, the p values can be more optimistic than they deserve to be.

No severe overlap among predictors

The block design is especially sensitive to collinearity: when a variable in the new block largely duplicates earlier ones, the incremental R squared comes out small without that variable being inert.

Sample size adequate for the number of predictors

Every variable added lifts R squared somewhat, and in small samples much of that lift is luck. Adjusted R squared partly offsets the inflation.

• Block order grounded in theory: The engine accepts the block order as given and never asks for a justification; nothing checks whether the sequence is defensible. Reshuffling blocks until the largest increment appears is something the panel will not stop you doing, and something no publication will accept.
• Linear relationships with the outcome: Linearity is not tested and no warning is raised for curvature. To see the shape of a relationship you have to examine the predictors against the outcome separately in the correlation methods.
• Normally distributed residuals with constant variance: Nothing at all is computed about residuals in this method: no normality test, no heteroscedasticity test, no residual tables and no residual plots. Seeing how the residuals behave means fitting the model again under the linear regression method.
• No severe overlap among predictors: No regression diagnostic is computed here: no VIF, no tolerance, no condition index and no influence measures. Inside the panel the way to gauge the overlap between blocks is to obtain the correlations among the predictors in a separate analysis and interpret the increment in that light.
• Sample size adequate for the number of predictors: The engine makes no judgement about sample adequacy and enforces no floor on cases per predictor. The gap between R squared and its adjusted counterpart in the step table is the first signal of how much inflation is at work.

1. 1Upload your data and confirm on the Data tab that the outcome and the predictors are all being read as numeric.
2. 2Draw up the block plan before running anything: which variable enters at which step, and on what theoretical grounds that sequence rests.
3. 3Dummy-code the categorical predictors on the Variable tab, and keep all dummies from one variable together in a single block.
4. 4Select the method under the regression heading on the Analysis tab, then fill in the first and second blocks in the parameter form, leaving the third empty if it is not needed.
5. 5Make sure no column appears twice, since ticking the same predictor in a second block stops the analysis with an error.
6. 6Run it and read the step table from left to right: the change in R squared first, then the F change and its p value. The coefficients of the final model open in their own section. No chart is produced here, so export the tables to Excel and keep the run on the History tab.

Variables added and variables in the model per step

Which predictors entered at each step and which the model holds by then. These two columns let you confirm that your block plan reached the engine intact.

R squared and adjusted R squared

The share of variance explained at each step and the version penalised for the number of predictors. A widening gap between them suggests the added variables are not worth their cost.

Model F and p value

The overall significance of the model at that step. The test weighs every predictor entered so far together, not merely the new block.

Change in R squared

What the new block adds to the explained variance, and the real output of the method. At the first step there is no preceding model to compare with, so the change equals the model's own R squared.

F change and the p value of the change

The statistic testing whether the incremental share could be chance, with its p value. A decision about a block rests on this p value and not on the model F.

Significance marker for the change

A flag showing whether the increment at each step clears the 0.05 level. At the first step this flag simply reflects the significance of the model itself.

Coefficients of the final model

For the model containing every block, each predictor's coefficient, standard error, t value and p value. Coefficients from intermediate steps are not returned, only those from the last one.

Final R squared and adjusted R squared

The explanatory power of the completed model. A reporting sentence usually pairs these two figures with the incremental shares of each block.

Change in R squared

The share a block adds to the explained variance, and the effect size native to this method. An increment near 0.02 counts as small, one near 0.13 as medium and 0.26 upwards as large; those anchors come from the Cohen tradition and are conventions rather than boundaries. Turning them into percentages helps: a value of 0.05 means 5% of the variability in the outcome is accounted for by that block.

Adjusted R squared

Because it charges for each added predictor, it shows whether a block genuinely bought anything. When R squared rises while the adjusted figure falls, the block complicated the model more than it explained, and that reading often decides the matter when the p value of the change sits on the margin.

No quantity here arrives with an interval: neither the change in R squared, nor R squared itself, nor the coefficients of the final model. Standard errors are present in the coefficient table, so coefficient intervals can be built outside the panel, but an interval for the change in R squared is not obtained so easily and another tool is needed when a journal requires it. In practice this means the precision of a small increment, say something between 0.01 and 0.03, has to be judged from its p value alone.

The numbers below are a representative example, not data from a real study or a real user.

Research question

In an illustrative employee study, how much does job satisfaction add to the explanation of organisational commitment once demographic controls are accounted for?

Variables

Dependent variable: organisational commitment scale score (continuous) · Block 1: age and years of service · Block 2: job satisfaction score · Block 3: perceived managerial support score

Example result

n = 326. Step one gave R squared 0.06, adjusted R squared 0.05, model F(2, 323) = 10.31, p < .001, with the change equal to the model itself. Step two gave R squared 0.29, adjusted R squared 0.28, a change in R squared of 0.23, F change(1, 322) = 104.40, p < .001. Step three gave R squared 0.32, adjusted R squared 0.31, a change of 0.03, F change(1, 321) = 14.16, p < .001. In the final model the coefficient for job satisfaction was 0.48 (standard error 0.05, t = 9.60, p < .001), for managerial support 0.17 (standard error 0.05, t = 3.40, p = .001), for age 0.04 (standard error 0.03, t = 1.33, p = .184) and for years of service 0.09 (standard error 0.04, t = 2.25, p = .025).

Interpretation

The demographic controls account for 6% of the variability in commitment, a modest but significant floor. Adding job satisfaction lifts the explained share by 0.23, more than quadrupling the explanatory power of the model, and that block is the central finding of the study. Managerial support contributes a smaller 0.03, yet that small share clears the .001 level and the adjusted figure keeps rising, so retaining the block is defensible. One explanation for the small third increment is overlap between managerial support and job satisfaction, though with no collinearity measure returned that reading stays an untested conjecture. How the residuals behave is equally unknown from this output, since no normality or heteroscedasticity test is computed.

Research question: After age and income are controlled for, what do satisfaction and quality perception add to explaining loyalty?

Dependent variable

loyalty

First block predictors

age,income

Second block predictors

satisfaction,quality_score

Computation credits: scipy 1.18.0 · statsmodels 0.14.6 · scikit-learn 1.9.0 · numpy 2.5.1 · pandas 3.0.5 · semopy 2.3.11 · 89fc29a · Data seed: 20260914

In a hierarchical regression, demographic controls explained 6% of commitment; adding job satisfaction produced a significant gain in explained variance (change in R squared = .23, F(1, 322) = 104.40, p < .001), and managerial support added a small but significant share (change in R squared = .03, p < .001; final R squared = .32, n = 326).

An example sentence close to APA style; the numbers are representative.

• Not one regression diagnostic is computed: this method offers no collinearity measure, no residual normality test, no heteroscedasticity test and no influence statistics.
• Neither the change in R squared nor the coefficients arrive with intervals, so the precision of a small increment has to be inferred from its p value.
• Coefficients from intermediate steps are not returned, so how a coefficient shrinks as blocks accumulate cannot be traced and mediation claims cannot be built from this output.
• At most three blocks can be fitted, which forces designs needing a finer sequence to be split across several runs.
• The incremental share is sensitive to collinearity: in overlapping blocks a genuine effect can surface as a tiny change in R squared, and the panel never quantifies that overlap.
• Because the justification for the block order goes unchecked, nothing technical prevents the error of rearranging blocks after looking at the data.

• Linear Regression (OLS)Why: When every predictor can enter in one step, it offers considerably more output on residuals and model diagnostics.
• ANCOVA (Analysis of Covariance)Why: When the variable of interest is group membership rather than a continuous predictor, and the control enters the model as a covariate.
• Pearson CorrelationWhy: When the question concerns the strength of the pairwise links between predictors and outcome rather than what blocks contribute.
• Mediation AnalysisWhy: When the claim is that one variable carries the effect of another, since mediation estimates the indirect effect directly and an R squared change does not.

• Linear Regression (OLS)
• ANCOVA (Analysis of Covariance)
• Pearson Correlation

Is a significant change in R squared enough to keep a block in the model?

It is not, and leaning on the p value alone misleads in large samples: with over a thousand cases even an increase of 0.004 can reach significance while meaning nothing for the research. Weigh three things together. First the size of the increment: convert it to a percentage and ask whether that share matters for your question. Second the direction of adjusted R squared, because a block that lifts R squared while the adjusted figure falls is buying complexity rather than explanation. Third the theoretical warrant: a block entered because the theory demands it gets reported even with a small share, provided you say plainly how small that share was.

Why does the change in R squared equal the model R squared at the first step?

Because there is no previous model to compare against. An increment is by definition the difference between the R squared values of two models, and at the first step the preceding model is the one with no predictors, whose R squared is zero, so the difference comes out equal to the model's own figure. By the same logic the F change equals the model F and the p value of the change equals the model p value. This is neither a bug nor a duplicated cell; it simply means the first row should be read as a starting level rather than as a gain. Genuinely incremental information begins on the second row.

How do I decide the order of the blocks, and does the order change the result?

The order changes the result directly, because the share allocated to each block depends on what entered before it: between two overlapping blocks, whichever goes first takes the larger share. The sequence therefore cannot be derived from the data and must be fixed beforehand. The defensible grounds are familiar ones: variables that need controlling go first, variables that precede others in time enter before them, and the theoretically interesting set is left for last so that its contribution appears net of the rest. Nothing in the panel interrogates the ordering, so hunting for the arrangement that yields the biggest increment is technically possible; it cannot be defended in print, where you are expected to state the theoretical basis of the block plan.

Can I see the coefficients from intermediate steps as well as the final model?

You cannot; the coefficient table belongs solely to the final model containing every block. To follow how one predictor's coefficient shifts as blocks accumulate, you have to set the steps up as separate linear regression runs and compare the coefficients yourself. That route also gets you the residual diagnostics, which the block-wise method never computes. Presenting a coefficient that shrinks markedly once a block is added as evidence of mediation is a common overreach, and moving to a method designed to test indirect effects is the better course.

• Aiken, L. S., & West, S. G. (1991). Multiple Regression: Testing and Interpreting Interactions
• Field, A. (2018). Discovering Statistics Using IBM SPSS Statistics
• Tabachnick, B. G., & Fidell, L. S. (2019). Using Multivariate Statistics
• statsmodels OLS documentation