pith. sign in

arxiv: 2605.18206 · v1 · pith:SZJEROZ2new · submitted 2026-05-18 · 📊 stat.ME

A tool to determine the degrees of freedom in tree-structured varying coefficient models

Nikolai Spuck , Moritz Berger This is my paper

Pith reviewed 2026-05-20 00:27 UTC · model grok-4.3

classification 📊 stat.ME
keywords tree-structured varying coefficient modelsdegrees of freedomrecursive partitioningBayesian information criterionmodel selectiongeneralized regressionpredictive performance
0
0 comments X

The pith

A formula approximates the degrees of freedom in tree-structured varying coefficient models by accounting for recursive partitioning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Tree-structured varying coefficient models let covariate effects vary according to effect modifiers found by recursive partitioning. Proper use of the Bayesian information criterion for selecting among these models requires adjusting the degrees of freedom to reflect the cost of that data-driven search. The authors supply an easy-to-apply formula that approximates this effective degrees of freedom. Simulations demonstrate that the formula produces more accurate selections and stronger out-of-sample predictions than the naive choice of counting only the final free parameters. The approach is demonstrated on survey data from older European adults.

Core claim

The paper develops an easy-to-apply formula to approximate the degrees of freedom of a TSVC model. This formula is employed for model selection based on the Bayesian information criterion and compared to the naive solution, setting the DoF to the number of free model parameters, in a simulation study. Results indicated that calculation of the DoF using the proposed formula resulted in more accurate selection results with improved predictive ability.

What carries the argument

The easy-to-apply formula that approximates the effective degrees of freedom introduced by recursive partitioning when building a tree-structured varying coefficient model.

Load-bearing premise

The proposed easy-to-apply formula accurately approximates the effective degrees of freedom introduced by recursive partitioning in TSVC models.

What would settle it

A simulation study that compares held-out predictive performance of TSVC models chosen by BIC using the proposed DoF formula against models chosen by BIC with naive parameter count, or that checks whether the formula's values match cross-validated estimates of effective degrees of freedom.

Figures

Figures reproduced from arXiv: 2605.18206 by Moritz Berger, Nikolai Spuck.

Figure 1
Figure 1. Figure 1: Results of the MPF model. The figure shows the approximated DoF [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Results of the simulation study: Predictive performance (scenario 1). [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results of the simulation study: Predictive performance (scenario [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results of the simulation study: Predictive performance (scenario 3). [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results of the simulation study: Predictive performance (application [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
read the original abstract

The tree-structured varying coefficient (TSVC) model is a flexible approach for generalized regression, where the linear effects of the covariates are allowed to vary with the values of effect modifiers. Relevant effect modifiers and interactions are identified using recursive partitioning. In TSVC models, analogously to other semi- and nonparametric regression approaches, one needs to account for the cost of data-driven model building when deriving the model degrees of freedom (DoF). To address this issue, we develop an easy-to-apply formula to approximate the DoF of a TSVC model. This formula is employed for model selection based on the Bayesian information criterion (BIC) and compared to the naive solution, setting the DoF to the number of free model parameters, in a simulation study. To illustrate the proposed DoF method, TSVC models using BIC-based selection were fitted to data from the Survey of Health, Ageing, and Retirement in Europe. Results indicated that calculation of the DoF using the proposed formula resulted in more accurate selection results with improved predictive ability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper develops an easy-to-apply approximation formula for the effective degrees of freedom (DoF) in tree-structured varying coefficient (TSVC) models to account for the complexity introduced by recursive partitioning when performing BIC-based model selection. The formula is compared to the naive DoF (equal to the number of free parameters) in a simulation study, and TSVC models selected via the proposed DoF are fitted to data from the Survey of Health, Ageing, and Retirement in Europe (SHARE), with results indicating improved selection accuracy and predictive performance.

Significance. If the DoF approximation holds, the method would offer a practical correction for model complexity in semi-parametric regressions that incorporate tree-based effect modification, improving BIC-based selection and out-of-sample prediction in applications such as health and social science data analysis. The simulation results and real-data illustration provide concrete evidence of gains over the naive approach.

major comments (2)
  1. [§3.2, Eq. (8)] §3.2, Eq. (8): the DoF formula conditions on the realized tree structure after partitioning; it is unclear whether (or how) the expression adjusts for the data-driven selection of splits when the number of candidate effect modifiers is moderate to large, which is the regime where the skeptic's concern about understated DoF would be most relevant.
  2. [Table 3] Table 3, high-dimensional modifier rows: the reported gains in selection accuracy and predictive ability are shown only for the simulation settings described; without additional runs that increase tree depth or the size of the modifier candidate set, it is difficult to confirm that the approximation remains accurate outside the low-dimensional regime where the derivation is most likely to hold.
minor comments (2)
  1. [Abstract] The abstract states that the proposed DoF yields 'more accurate selection results' but does not define the accuracy metric (e.g., true-positive rate for relevant modifiers or out-of-sample MSE).
  2. [§2] Notation for the varying-coefficient functions and the partitioning operator could be introduced with a small numerical example in §2 to improve readability for readers unfamiliar with TSVC models.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments on our manuscript. We address each major comment below in a point-by-point manner and indicate the changes we will make to strengthen the paper.

read point-by-point responses

  1. Referee: [§3.2, Eq. (8)] §3.2, Eq. (8): the DoF formula conditions on the realized tree structure after partitioning; it is unclear whether (or how) the expression adjusts for the data-driven selection of splits when the number of candidate effect modifiers is moderate to large, which is the regime where the skeptic's concern about understated DoF would be most relevant.

    Authors: We agree that Equation (8) provides an approximation conditional on the realized tree structure obtained after recursive partitioning. The formula incorporates the complexity of the fitted tree by adjusting for the number and nature of the selected splits and nodes, rather than relying solely on the count of free parameters. This conditional approach is intentional, as it reflects the effective degrees of freedom of the model once the partitioning has been performed. However, we acknowledge that when the pool of candidate effect modifiers is moderate to large, the preceding data-driven search over possible splits may introduce additional selection effects not fully captured by conditioning on the final tree alone. In the revised manuscript, we will clarify this distinction in Section 3.2, add a discussion of the approximation's scope, and note that the BIC penalty using this DoF still yields improved selection performance in the regimes we examined. revision: partial

  2. Referee: [Table 3] Table 3, high-dimensional modifier rows: the reported gains in selection accuracy and predictive ability are shown only for the simulation settings described; without additional runs that increase tree depth or the size of the modifier candidate set, it is difficult to confirm that the approximation remains accurate outside the low-dimensional regime where the derivation is most likely to hold.

    Authors: The referee is correct that the current simulation results in Table 3 are limited to the settings described in the paper. To provide stronger evidence of the approximation's behavior, we will conduct additional simulation experiments that increase both tree depth and the number of candidate effect modifiers. The outcomes of these runs will be reported in a revised or supplementary version of Table 3, allowing readers to assess performance in higher-dimensional regimes. revision: yes

Circularity Check

0 steps flagged

No circularity: DoF approximation presented as independent formula

full rationale

The paper develops an easy-to-apply formula to approximate the effective degrees of freedom in tree-structured varying coefficient models that arise from recursive partitioning. This formula is then used for BIC-based model selection and compared against the naive count of free parameters in simulations and real data. No quoted derivation step reduces the proposed formula to a fitted quantity on the same data, a self-referential definition, or a load-bearing self-citation whose validity depends on the present result. The central claim therefore remains an independent approximation whose accuracy can be checked externally against simulation benchmarks and predictive performance, satisfying the criteria for a self-contained derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the validity of the DoF approximation formula for capturing the complexity of recursive partitioning; full details of any parameters or assumptions in the formula are not available from the abstract.

axioms (1)
  • domain assumption Recursive partitioning in TSVC models adds effective degrees of freedom beyond the final parameter count that can be approximated by a simple formula.
    This assumption underpins the development and use of the proposed DoF formula for BIC selection.

pith-pipeline@v0.9.0 · 5705 in / 1131 out tokens · 47084 ms · 2026-05-20T00:27:09.876645+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

147 extracted references · 147 canonical work pages · 1 internal anchor

  1. [1]

    Hothorn and K

    T. Hothorn and K. Hornik and A. Zeileis , journal=. Unbiased recursive partitioning:. 2006 , doi =

    work page 2006

  2. [2]

    Stat Probab Lett , volume=

    Mixed effects regression trees for clustered data , author=. Stat Probab Lett , volume=. 2011 , doi =

    work page 2011

  3. [3]

    1990 , address =

    Generalized Additive Models , author =. 1990 , address =

    work page 1990

  4. [4]

    1999 , journal =

    Building multivariable fractional polynomial models with selection of variables and transformations: a boostrap investigation , author =. 1999 , journal =

    work page 1999

  5. [5]

    1986 , author =

    How Biased is the Apparent Error Rate of a Prediction Rule? , journal =. 1986 , author =

    work page 1986

  6. [6]

    1998 , journal =

    On Measuring and Correcting the Effects of Data Mining and Model Selection , author =. 1998 , journal =

    work page 1998

  7. [7]

    Wang and H

    L. Wang and H. Zhao and X. Fan , journal =. Degrees of Freedom: Search Cost and Self-Consistency , year =

    work page

  8. [8]

    2020 , journal =

    On Degrees of Freedom of Projection Estimators with Applications to Multivariate Nonparametric Regression , author =. 2020 , journal =

    work page 2020

  9. [9]

    Efron , title =

    B. Efron , title =. 2004 , pages =. doi:10.1198/016214504000000692 , journal =

    work page doi:10.1198/016214504000000692 2004

  10. [10]

    Degrees of freedom for nonlinear least squares estimation

    N. Hansen and A. Sokol , title =. arXiv: 1402.2997 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Annals of Statistics , year =

    Degrees of freedom in lasso problems , author =. Annals of Statistics , year =

    work page

  12. [12]

    Zhou and T

    H. Zhou and T. Hastie and R. Tibshirani , title =. Annals of Statistics , year =

    work page

  13. [13]

    Statistica Sinica , year =

    Degrees of Freedom and Model Search , author =. Statistica Sinica , year =

    work page

  14. [14]

    Adler and S

    W. Adler and S. Potapov and B. Lausen , title =. 2011 , journal =. doi:10.1007/s00180-011-0249-1 , pages =

    work page doi:10.1007/s00180-011-0249-1 2011

  15. [15]

    Lin and C

    J. Lin and C. Zhong and D. Hu and C. Rudin and M. Seltzer , title =. Proceedings of the 37th International Conference on Machine Learning , year =

    work page

  16. [16]

    C. L. Niedzwiedz and S. V. Katikireddi and J. P. Pell and R. Mitchell , year =. Socioeconomic inequalities in thequality of life of older Europeans in different welfare regimes , journal =

    work page

  17. [17]

    I. T. M. Heyworth and M. L. Hazell and M. F. Linehan and T. L. Frank , title =. 2009 , journal =

    work page 2009

  18. [18]

    N. E. Rothrock and R. D. Hays and K. Spritzer and S. E. Yount and W. Riley and D. Cella , title =. 2010 , journal =

    work page 2010

  19. [19]

    Marengoni and S

    A. Marengoni and S. Angleman and R. Melis and F. Mangialasche and A. Karp and A. Garmen aand B. Meinow and L. Fratiglioni , title =. 2011 , journal =

    work page 2011

  20. [20]

    T. T. Makovski and S. Schmitz and M. P. Zeegers and S. Stranges and M. van der Akker , title =. Ageing Res Rev , doi =. 2019 , volume =

    work page 2019

  21. [21]

    T. T. Makovski and G. Le Coroller and P. Putrik and Y. H. Choi and M. P. Zeegers and S. Stranges and M. Ruiz Castell and L. Huiart and M. van den Akker , title =. 2020 , jorunal =

    work page 2020

  22. [22]

    von dem Knesebeck and M

    O. von dem Knesebeck and M. Wahrendorf and M. Hyde and J. Siegrist , title =. 2007 , journal =

    work page 2007

  23. [23]

    M. A. Killingsworth , year =. Experienced well-being rises with income, even above \ 75,000 per year , journal =

    work page

  24. [24]

    Rodr\'iguez-Bl\'aquez and O

    C. Rodr\'iguez-Bl\'aquez and O. Ribeiro and A. Ayala and L. Teixeira and L. Araj\'uo and M. J. Forjaz , title =. Int J Environ Res Public Health , year =

    work page

  25. [25]

    Diener and C

    E. Diener and C. Diener , title =. 1995 , journal =

    work page 1995

  26. [26]

    Demography of Europe: 2024 edition , publisher =

    work page 2024

  27. [27]

    Hajjem and F

    A. Hajjem and F. Bellavance and D. Larocque , title =. 2012 , journal =

    work page 2012

  28. [28]

    Adv Data Anal Classif , volume=

    A comparison of five recursive partitioning methods to find person subgroups involved in meaningful treatment -- subgroup interactions , author=. Adv Data Anal Classif , volume=. 2014 , doi =

    work page 2014

  29. [29]

    Tutz and M.-R

    G. Tutz and M.-R. Oelker , journal=. Modelling clustered heterogeneity:. 2017 , doi =

    work page 2017

  30. [30]

    2020 , journal =

    Robustness of linear mixed-effects models to violations of distributional assumptions , author =. 2020 , journal =

    work page 2020

  31. [31]

    Stat Model , volume=

    Clustering in linear mixed models with approximate Dirichlet process mixtures using EM algorithm , author=. Stat Model , volume=. 2013 , doi =

    work page 2013

  32. [32]

    Gr\"un and F

    B. Gr\"un and F. Leisch , journal=. 2007 , doi =

    work page 2007

  33. [33]

    2007 , publisher =

    Data Analysis Using Regression and Multilevel/Hierarchical Models , author =. 2007 , publisher =

    work page 2007

  34. [34]

    2014 , volume =

    Variable selection for generalized linear mixed models by L_1 -penalized estimation , author =. 2014 , volume =. doi:10.1007/s11222-012-9359-z , journal =

    work page doi:10.1007/s11222-012-9359-z 2014

  35. [35]

    Kim and G

    G.R. Kim and G. Netuveli and D. Blane and A. Peasey and S. Malyutina and G. Simonova and R. Kubinova and A. Pajak and S. Croezen and M. Bobak, Martin and H. Pikhart , journal=. Psychometric properties and confirmatory factor analysis of the. 2015 , publisher=

    work page 2015

  36. [36]

    Howel , journal=

    D. Howel , journal=. Interpreting and evaluating the. 2012 , publisher=

    work page 2012

  37. [37]

    Sim and B

    J. Sim and B. Bartlam and M. Bernard, Miriam , journal=. The. 2011 , publisher=

    work page 2011

  38. [38]

    International Conference on Machine Learning , pages=

    Selective inference for sparse high-order interaction models , author=. International Conference on Machine Learning , pages=. 2017 , organization=

    work page 2017

  39. [39]

    Stat Probab Lett , volume=

    Selective inference after likelihood-or test-based model selection in linear models , author=. Stat Probab Lett , volume=. 2018 , publisher=

    work page 2018

  40. [40]

    2023 , journal =

    Antiviral treatment of COVID-19: which role can clinical parameters play in therapy evaluation? , author =. 2023 , journal =

    work page 2023

  41. [41]

    and Schmid, M

    Spuck, N. and Schmid, M. and Heim, N. and Klarmann-Schulz, U. and Hoerauf, A. and Berger, M. , year =. Statistics & Computing , volume =

    work page

  42. [42]

    Release version: 9.0.0

    2024 , author =. doi:10.6103/SHARE.w9.900 , journal =

    work page doi:10.6103/share.w9.900 2024

  43. [43]

    and Wagner, M

    Bergmann, M. and Wagner, M. and B\"orsch-Supan, A. , year =

    work page

  44. [44]

    and Brandt, M

    B\"orsch-Supan, A. and Brandt, M. and Hunkler, C. and Kneip, T. and Korbmacher, J. and Malter, F. and Schaan, B. and Stuck, S. and Zuber, S. , year =. International Journal of Epidemiology , doi =

    work page

  45. [45]

    and Wiggins, R.D

    Hyde, M. and Wiggins, R.D. and Higgs, P. and Blane, D.B. , year =. Aging & Mental Health , volume =

    work page

  46. [46]

    and Stenner, P

    Bowling, A. and Stenner, P. , title =. 2011 , journal =

    work page 2011

  47. [47]

    and Ryser, V

    Borrat-Besson, C. and Ryser, V. and Goncalves, J. , year =. FORS Working Papers , doi =

    work page

  48. [48]

    and Berger, M

    Faber, J. and Berger, M. and Wilke, C. and Huebner-Schmid, J. and Schaprian, T. and Santana, M. M. and Grobe-Einsler, M. and et al. , year =. doi:10.1002/ana.26824 , journal =

    work page doi:10.1002/ana.26824

  49. [49]

    2018 , journal =

    Fixed and random effects models: making an informed choice , author =. 2018 , journal =

    work page 2018

  50. [50]

    2000 , publisher =

    Linear Mixed Models for Longitudinal Data , author =. 2000 , publisher =

    work page 2000

  51. [51]

    M. N. Wright and I. K\"onig , title =. 2019 , journal =

    work page 2019

  52. [52]

    Behav Res Methods , doi =

    Detecting treatment-subgroup interactions in clustered data with generalized linear mixed-effects model trees , author =. Behav Res Methods , doi =. 2018 , volume =

    work page 2018

  53. [53]

    and Verbeke, G

    Molenberghs, G. and Verbeke, G. , year =

    work page

  54. [54]

    and Kurland, B.F

    Heagerty, P. and Kurland, B.F. , journal =. Misspecified maximum likelihood estimates and generalised linear mixed models , year =

    work page

  55. [55]

    and Rampichini, C

    Grilli, L. and Rampichini, C. , journal =. The role of sample cluster means in multi-level models: A view on endogeneity and measurment error issues , year =

    work page

  56. [56]

    2015 , doi =

    Unbiased regression trees for longitudinal and clustered data , journal =. 2015 , doi =

    work page 2015

  57. [57]

    J Eur Surv Res Assoc , doi =

    Tree-based Machine Learning Methods for Survey Research , author =. J Eur Surv Res Assoc , doi =. 2019 , volume =

    work page 2019

  58. [58]

    1999 , publisher =

    Recursive Partitioning in Health Sciences , author =. 1999 , publisher =

    work page 1999

  59. [59]

    2009 , journal =

    An Introduction to Recursive Partitioning: Rationale, Application and Characteristics of Classification and Regression Trees, Bagging and Random Forests , author =. 2009 , journal =

    work page 2009

  60. [60]

    2007 , author =

    Biom , volume =. 2007 , author =

    work page 2007

  61. [61]

    Frias-Goytia, G. L. and Lojo-Seoane, C. and Mallo, S. C. and Nieto-Vieites, A. and Juncos-Rabad\`an, O. and Pereiro, A.X. , title =. 2024 , journal =

    work page 2024

  62. [62]

    W. D. Fisher , title =. J Am Stat Soc , year =

    work page

  63. [63]

    1996 , journal =

    Pattern Recognition and Neural Networks , author =. 1996 , journal =

    work page 1996

  64. [64]

    Sela, R. J. and Simonoff, J. S. , title =. 2012 , journal =. doi:10.1007/s10994-011-5258-3 , volume =

    work page doi:10.1007/s10994-011-5258-3 2012

  65. [65]

    J Am Stat Assoc , volume=

    Exact post-selection inference for sequential regression procedures , author=. J Am Stat Assoc , volume=. 2016 , publisher=

    work page 2016

  66. [66]

    Data Min Knowl Discov , volume=

    Decision tree boosted varying coefficient models , author=. Data Min Knowl Discov , volume=. 2022 , publisher=

    work page 2022

  67. [67]

    Int Stat Rev , volume=

    Varying coefficient regression models: a review and new developments , author=. Int Stat Rev , volume=. 2015 , doi =

    work page 2015

  68. [68]

    Stats , volume=

    The one standard error rule for model selection: does it work? , author=. Stats , volume=. 2021 , doi =

    work page 2021

  69. [69]

    Tutz and M

    G. Tutz and M. Schmid , publisher =. 2016 , address =

    work page 2016

  70. [70]

    and Tibshirani, R

    Taylor, J. and Tibshirani, R. , year =. Post-selection inference for. Can J Stat , volume =

    work page

  71. [71]

    Hastie and R

    T. Hastie and R. Tibshirani and J. Friedman , publisher =. 2009 , doi =

    work page 2009

  72. [72]

    Ann Stat , author =

    Valid post-selection inference , volume =. Ann Stat , author =. doi:10.1214/12-aos1077 , number =

    work page doi:10.1214/12-aos1077

  73. [73]

    , title =

    Akaike, H. , title =. The Statistician , year =

    work page

  74. [74]

    and Sanders, N

    Kim, S. and Sanders, N. , title =. 2017 , journal =

    work page 2017

  75. [75]

    , title =

    Kolassa, S. , title =. 2011 , journal =

    work page 2011

  76. [76]

    Burnham, K. P. and Anderson, D. R. , title =. 2002 , adress =

    work page 2002

  77. [77]

    2022 , journal =

    Tree-Values: Selective Inference for Regression Trees , author =. 2022 , journal =

    work page 2022

  78. [78]

    2022 , journal =

    Selective inference for effect modification via the lasso , author =. 2022 , journal =

    work page 2022

  79. [79]

    2008 , journal =

    Statistical methods with varying coefficient models , author =. 2008 , journal =

    work page 2008

  80. [80]

    J Comput Graph Stat , author =

    Boosted Varying-Coefficient Regression Models for Product Demand Prediction , volume =. J Comput Graph Stat , author =. 2014 , pages =. doi:10.1080/10618600.2013.778777 , number =

    work page doi:10.1080/10618600.2013.778777 2014

Showing first 80 references.