Relative-importance networks

Written by Mohammed Saqr and Sonsoles López-Pernas

When the variables of a questionnaire are strongly correlated, a regression coefficient or a partial correlation credits each predictor only with its unique contribution, and the variance that two predictors explain together is credited to neither. For a researcher who asks how much self-efficacy contributes to explaining strategy use, that unique contribution understates the answer. Relative importance analysis divides the explained variance of an outcome, its \(R^2\), among all predictors, giving each a share of both its unique and its jointly explained variance. The LMG decomposition (Lindeman et al., 1980; Groemping, 2006) does so by averaging the increase in \(R^2\) that a predictor brings over every order in which the predictors could be entered.

A relative-importance network applies this decomposition to every variable in turn, with all the others as predictors. The shares become directed edges: an edge from A to B is the part of B’s explained variance attributed to A, and the edges arriving at a node sum to that node’s \(R^2\). This article builds the network for five self-regulated learning constructs, reads the directed shares, checks that they sum to each node’s explained variance, contrasts incoming and outgoing totals, and repeats the analysis on rank correlations.

library(psychnets)

Data

SRL_GPT holds 300 sets of responses to the Motivated Strategies for Learning Questionnaire, each a construct score on a scale from 1 to 7. The responses were generated by a large language model (GPT) in a study of how such models answer psychological surveys (Saqr, 2025); they serve as a well-structured example and are not evidence about students.

head(SRL_GPT)
#>        CSU       IV       SE       SR   TA
#> 1 5.307692 5.666667 5.777778 5.333333 4.00
#> 2 5.846154 6.444444 6.000000 5.777778 4.00
#> 3 6.615385 6.666667 6.222222 6.333333 3.25
#> 4 5.692308 6.555556 6.333333 5.555556 4.50
#> 5 4.384615 5.555556 4.888889 4.777778 4.00
#> 6 4.846154 5.444444 5.666667 5.111111 3.50

The model

For outcome \(j\) with predictor set \(P\), the LMG share of predictor \(k\) is the average increase in \(R^2\) when \(k\) joins a subset \(S\) of the other predictors, weighted so that every entry order counts equally:

\[\varphi_{kj} = \sum_{S \subseteq P \setminus \{k\}} \frac{|S|!\,(|P| - |S| - 1)!}{|P|!} \left[R^2(S \cup \{k\}) - R^2(S)\right].\]

Each \(R^2(A)\) follows from the correlation matrix \(\mathbf{S}\) as \(R^2(A) = \mathbf{S}_{jA}\mathbf{S}_{AA}^{-1}\mathbf{S}_{Aj}\). The shares are non-negative and add up to the outcome’s full-model explained variance,

\[\sum_{k} \varphi_{kj} = R_j^2 .\]

The directed network has edge weight \(\varphi_{kj}\) from \(k\) to \(j\). The incoming edges of a node sum to its \(R^2\); the outgoing edges of a node sum to its total contribution to the other regressions. With \(p\) nodes, each outcome requires \(2^{p-1}\) subset regressions, so the computation is limited to 21 nodes by default.

How is each construct’s explained variance divided?

psychnet() with method = "relimp" fits the five regressions and decomposes their \(R^2\).

importance_net <- psychnet(data = SRL_GPT, method = "relimp")
summary(importance_net)
#> <psychnet> relimp network
#>   nodes: 5   edges: 20   (directed)
#>   structural certificate (decomposition residual): 2.22e-16
#>   edge weight: range [0.006, 0.299], mean 0.165

The network has 20 directed edges, one for every ordered pair, because every predictor receives some share. The shares range from 0.006 to 0.299. The edge table lists them with from as the predictor and to as the outcome:

as.data.frame(importance_net)
#>    from  to      weight
#> 1    IV CSU 0.288505266
#> 2    SE CSU 0.263574096
#> 3    SR CSU 0.274082666
#> 4    TA CSU 0.007326148
#> 5   CSU  IV 0.299473895
#> 6    SE  IV 0.216286692
#> 7    SR  IV 0.258608009
#> 8    TA  IV 0.007032583
#> 9   CSU  SE 0.280588934
#> 10   IV  SE 0.219009605
#> 11   SR  SE 0.220573037
#> 12   TA  SE 0.006045547
#> 13  CSU  SR 0.277393999
#> 14   IV  SR 0.253052645
#> 15   SE  SR 0.212824511
#> 16   TA  SR 0.052422918
#> 17  CSU  TA 0.014154964
#> 18   IV  TA 0.015416846
#> 19   SE  TA 0.012055817
#> 20   SR  TA 0.127772080

The largest share runs from cognitive strategy use to intrinsic value (0.299): of the variance in intrinsic value explained by the other four constructs, a share of 0.299 is attributed to strategy use. The reverse edge, from intrinsic value to strategy use, is 0.289. The two directions are close here but come from different regressions, and they need not agree. Among the learning constructs the shares lie between 0.213 and 0.299, so each is explained by the other three in roughly equal parts.

Test anxiety is the exception. Of its explained variance, self-regulation accounts for 0.128, while cognitive strategy use (0.014), intrinsic value (0.015) and self-efficacy (0.012) contribute little. In the other direction, test anxiety contributes almost nothing to the learning constructs (0.006 to 0.007), except to self-regulation (0.052). The association between test anxiety and the learning constructs passes through self-regulation, in both directions.

Do the shares add up to each construct’s explained variance?

By construction, the shares arriving at a node must sum to that node’s full-model \(R^2\). certificate() reports the largest difference between the two across nodes; a value near zero shows that the decomposition is complete and exact.

certificate(importance_net)
#>   method  certificate       kind certified
#> 1 relimp 2.220446e-16 structural      TRUE

The largest difference is at the level of machine precision (below \(10^{-15}\)), so every edge in the table is part of an exact division of the explained variance. The check concerns the arithmetic of the decomposition, not the adequacy of the linear regressions.

How much of each construct is explained?

net_predict() reports the full-model \(R^2\) of each node, the total of its incoming shares.

net_predict(importance_net)
#>   node     type metric predictability accuracy
#> 1  CSU gaussian     R2      0.8334882       NA
#> 2   IV gaussian     R2      0.7814012       NA
#> 3   SE gaussian     R2      0.7262171       NA
#> 4   SR gaussian     R2      0.7956941       NA
#> 5   TA gaussian     R2      0.1693997       NA

The other four constructs explain 83% of the variance in cognitive strategy use (\(R^2 = 0.833\)), 80% in self-regulation, 78% in intrinsic value and 73% in self-efficacy. They explain 17% of the variance in test anxiety (\(R^2 = 0.169\)), and the incoming shares above show that three quarters of that (0.128) comes from self-regulation. Most of the variation in test anxiety lies outside this five-construct model.

Which constructs contribute most to explaining the others?

net_centralities() sums the outgoing edges of each node. Because all shares are non-negative, strength and expected influence are equal.

net_centralities(importance_net)
#>   node  strength expected_influence
#> 1  CSU 0.8716118          0.8716118
#> 2   IV 0.7759844          0.7759844
#> 3   SE 0.7047411          0.7047411
#> 4   SR 0.8810358          0.8810358
#> 5   TA 0.0728272          0.0728272

Self-regulation has the largest outgoing total (0.881), closely followed by cognitive strategy use (0.872); test anxiety has the smallest (0.073). Self-regulation is thus the construct that contributes most to explaining the others, partly because it is the only construct with a substantial share in test anxiety. Incoming totals (\(R^2\)) describe how well a construct is explained; outgoing totals describe how much it explains. The two answer different questions and are reported separately.

What does the network look like?

cograph::splot() draws the directed network, with arrows from predictor to outcome and the share printed on each edge.

cograph::splot(importance_net, directed = TRUE, edge_labels = TRUE)

The figure is dense because every ordered pair has an edge. The thick reciprocal arrows among the learning constructs and the thin arrows to and from test anxiety are visible, but exact values are read from the edge table.

Do the conclusions depend on the correlation estimator?

Refitting with cor_method = "spearman" bases the regressions on rank correlations.

importance_spearman <- psychnet(data = SRL_GPT, method = "relimp",
                                cor_method = "spearman")
as.data.frame(importance_spearman)
#>    from  to      weight
#> 1    IV CSU 0.286725129
#> 2    SE CSU 0.270159694
#> 3    SR CSU 0.265419989
#> 4    TA CSU 0.005393971
#> 5   CSU  IV 0.299066051
#> 6    SE  IV 0.203370231
#> 7    SR  IV 0.263749828
#> 8    TA  IV 0.005279123
#> 9   CSU  SE 0.292490014
#> 10   IV  SE 0.206036816
#> 11   SR  SE 0.209495026
#> 12   TA  SE 0.004346020
#> 13  CSU  SR 0.271471270
#> 14   IV  SR 0.260383389
#> 15   SE  SR 0.203334385
#> 16   TA  SR 0.037126693
#> 17  CSU  TA 0.009240756
#> 18   IV  TA 0.011339759
#> 19   SE  TA 0.008842802
#> 20   SR  TA 0.086215285
net_predict(importance_spearman)
#>   node     type metric predictability accuracy
#> 1  CSU gaussian     R2      0.8276988       NA
#> 2   IV gaussian     R2      0.7714652       NA
#> 3   SE gaussian     R2      0.7123679       NA
#> 4   SR gaussian     R2      0.7723157       NA
#> 5   TA gaussian     R2      0.1156386       NA

The largest share remains cognitive strategy use to intrinsic value (0.299), and the learning constructs keep \(R^2\) values from 0.712 to 0.828. The explained variance of test anxiety falls from 0.169 to 0.116, and the share of self-regulation in it from 0.128 to 0.086. The structure among the learning constructs does not depend on the estimator; the explanation of test anxiety is weaker on ranks and should be reported with that caveat.

Reporting the analysis

A report states the variables, the sample size, the correlation estimator, the treatment of missing data, and whether raw or normalized shares are used. Each edge is described with its direction, from predictor to outcome, and incoming \(R^2\) and outgoing contribution are reported as separate node summaries.

In this analysis, raw LMG shares from Pearson correlations of 300 complete responses formed 20 directed edges. The largest ran from cognitive strategy use to intrinsic value (0.299). The constructs’ \(R^2\) ranged from 0.169 (test anxiety) to 0.833 (cognitive strategy use), and self-regulation carried most of the explained variance of test anxiety (0.128). A rank-based analysis kept the structure among the learning constructs and reduced the explained variance of test anxiety.

Assumptions and when to use relative importance

References

Groemping, U. (2006). Relative importance for linear regression in R: The package relaimpo. Journal of Statistical Software, 17(1), 1-27. doi:10.18637/jss.v017.i01

Lindeman, R. H., Merenda, P. F., & Gold, R. Z. (1980). Introduction to bivariate and multivariate analysis. Scott, Foresman.

Saqr, M. (2025). Delving into the psychology of machines: Exploring the structure of self-regulated learning via LLM-generated survey responses. Computers in Human Behavior, 173, 108769. doi:10.1016/j.chb.2025.108769