pith. sign in

arxiv: 2605.24847 · v2 · pith:E54T6H4Cnew · submitted 2026-05-24 · 📊 stat.AP

Logistic regression is not enough: The need for Bayesian nonparametric modelling for causal inference using observational data, exemplified by the 'gateway' effect

Floe Foxon , Raymond Niaura This is my paper

Pith reviewed 2026-06-30 00:10 UTC · model grok-4.3

classification 📊 stat.AP
keywords causal inferenceBayesian additive regression treese-cigarette useadolescent smokinggateway effectobservational datanonparametric modelinglongitudinal cohort
0
0 comments X

The pith

Bayesian nonparametric models applied to youth survey data find no gateway from e-cigarettes to smoking and instead a small diversionary effect among those who already smoke.

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

The paper demonstrates that logistic regression models, which have been used to claim that e-cigarette use quadruples the odds of later smoking, are limited for causal inference because they assume linear relationships and cannot easily capture complex interactions among many covariates. Applying Bayesian Additive Regression Trees to waves of the Population Assessment of Tobacco and Health youth data yields different results: e-cigarette use at baseline is linked to two fewer days of smoking among adolescents who had already tried smoking and to less than one additional day among those who had not. This pattern removes evidence for a gateway effect and aligns with the continued population-level decline in adolescent smoking even as e-cigarette use rose. The work therefore shows how model choice in observational studies can change substantive conclusions about behavioral substitution or progression between products.

Core claim

When baseline e-cigarette use is modeled as the treatment and change in days smoked is the numerical outcome in the PATH youth cohort, Bayesian Additive Regression Trees produce an average causal effect of minus two days smoked among ever-smokers (a diversionary effect) and an absolute change below one day among never-smokers (a null effect), after adjustment for socio-demographic, intra-individual, behavioral, peer, and family factors; these estimates eliminate the gateway signal previously reported from logistic regression and match the direction of observed population smoking trends.

What carries the argument

Bayesian Additive Regression Trees (BART), a nonparametric ensemble method that sums regression trees to estimate flexible treatment effects and counterfactual outcomes while supplying uncertainty intervals.

If this is right

  • Logistic regression models of the same data produce estimates inconsistent with both the BART results and population smoking trends.
  • Any projected reversal of adolescent smoking declines due to e-cigarette use rests on an artifact of model choice rather than on the observed data.
  • Causal effect estimates for product substitution or progression should be obtained with methods that allow nonlinear and interactive relationships among covariates.
  • The apparent paradox between logistic-regression predictions and real-world smoking declines is resolved once flexible modeling is used.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Reanalyses of other longitudinal cohorts on substance-use transitions with BART or similar nonparametric causal tools could test whether logistic-based gateway claims hold under more flexible specifications.
  • Epidemiologic studies that rely on binarized outcomes and logistic models for observational causal questions may systematically misestimate effects when many covariates interact.
  • Policy models that forecast youth smoking under different e-cigarette scenarios should incorporate uncertainty from model form rather than treat logistic results as definitive.

Load-bearing premise

That the measured socio-demographic, intra-individual, behavioral, peer, and family factors, when fit with BART, are sufficient to remove all confounding between baseline e-cigarette use and later changes in smoking.

What would settle it

A reanalysis of the same PATH waves or a comparable longitudinal cohort that recovers a clinically large positive causal effect of e-cigarette use on smoking days using an alternative flexible causal method such as targeted maximum likelihood estimation or causal forests would falsify the claim that BART yields unbiased estimates here.

read the original abstract

Introduction: Logistic regression (LR)-type model limitations for causal inference are explained theoretically and empirically through the lens of the purported gateway effect from e-cigarette use to smoking. Previous studies have reported that baseline e-cigarette use quadruples odds of follow-up smoking (binarized) in LR-type models of adolescent longitudinal cohorts (LCs), such that increased e-cigarette use would counteract smoking declines. However, US population-level trends show accelerated smoking declines to record-lows when e-cigarette use increased, presenting an apparent paradox. Methods: Population Assessment of Tobacco and Health (USA) Youth Waves 3 to 4 were analyzed with Bayesian Additive Regression Trees (BART) to model baseline e-cigarette use (treatment) and change in number of days smoking from baseline to follow-up (numerical response) among never- and ever-smoking respondents (group effects), adjusting for confounding risk factors (socio-demographic, intra-individual, behavioural, peer influence, and family background). Unlike LR-type models, BART provides nonlinear, nonparametric modelling with counterfactuals and provides causal effect estimates with principled uncertainty estimation. Results: The average effect of e-cigarette use on smoking was both clinically and statistically significant among ever-smoking adolescents (-2 days smoking [diversionary effect; opposite to gateway]) and was not clinically significant among never-smoking adolescents (<1-day absolute change in days smoking [null effect]). Conclusions: When LC data are analyzed with causal inference techniques, the gateway effect disappears, consistent with population-level trends. This likely explains why gateway effects predicted in previous LR-type studies have not materialized in a population-level reversal/unexpected slowing of the US adolescent smoking decline, resolving the paradox.

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 / 1 minor

Summary. The manuscript claims that logistic regression-type models are limited for causal inference on the purported 'gateway' effect from e-cigarette use to smoking in adolescents. Using Bayesian Additive Regression Trees (BART) on PATH Youth Waves 3-4 data with a numerical outcome (change in days smoked), it reports a clinically and statistically significant diversionary effect of -2 days among ever-smokers and a null effect (<1-day absolute change) among never-smokers after adjusting for socio-demographic, intra-individual, behavioural, peer, and family confounders. This is presented as resolving the paradox between prior LR-based predictions and observed population-level smoking declines.

Significance. If the estimates prove robust, the work would illustrate the advantages of flexible nonparametric Bayesian methods like BART over parametric LR for observational causal inference in tobacco epidemiology, using a continuous outcome and providing uncertainty quantification. It supplies direct comparison to external population trends, which strengthens the applied relevance if the internal validity holds.

major comments (2)
  1. [Methods (abstract)] Methods paragraph of the abstract: The headline causal claims (diversionary effect of -2 days among ever-smokers; null among never-smokers) rest on the assumption that the listed covariates plus BART suffice for conditional ignorability, but the manuscript supplies no sensitivity analyses for unmeasured confounding (e.g., e-value or Rosenbaum bounds), overlap diagnostics, or post-BART balance checks; BART addresses flexibility but does not relax the no-unmeasured-confounding assumption.
  2. [Results (abstract)] Results (abstract): The reported point estimates and clinical significance claims lack accompanying model diagnostics, sensitivity checks, full BART hyperparameter specification, or confounder coding details, which are load-bearing for evaluating whether the -2-day and <1-day effects are reliable causal estimates rather than artifacts of implementation.
minor comments (1)
  1. [Abstract] The abstract could clarify how 'clinical significance' is defined relative to the numerical outcome scale and distinguish it more explicitly from statistical significance in the reported effects.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for these constructive comments on the abstract and the supporting causal assumptions. We address each major comment below, indicating where we will revise the manuscript to strengthen the presentation of diagnostics and sensitivity checks while preserving the core methodological comparison between BART and logistic regression.

read point-by-point responses

  1. Referee: [Methods (abstract)] Methods paragraph of the abstract: The headline causal claims (diversionary effect of -2 days among ever-smokers; null among never-smokers) rest on the assumption that the listed covariates plus BART suffice for conditional ignorability, but the manuscript supplies no sensitivity analyses for unmeasured confounding (e.g., e-value or Rosenbaum bounds), overlap diagnostics, or post-BART balance checks; BART addresses flexibility but does not relax the no-unmeasured-confounding assumption.

    Authors: We agree that BART improves flexibility in modeling observed confounders but does not relax the no-unmeasured-confounding assumption required for causal identification. The manuscript already notes this standard limitation of observational data in the discussion section. To address the referee's point directly, we will add e-value calculations for the reported effects and propensity-score overlap diagnostics (including visual checks) to the methods and results. Post-BART balance checks will also be reported. These additions will be incorporated into both the main text and a revised abstract methods paragraph. We maintain that the consistency between our estimates and external population-level smoking trends provides supplementary support, but we accept that formal sensitivity analyses are needed for a complete causal claim. revision: yes

  2. Referee: [Results (abstract)] Results (abstract): The reported point estimates and clinical significance claims lack accompanying model diagnostics, sensitivity checks, full BART hyperparameter specification, or confounder coding details, which are load-bearing for evaluating whether the -2-day and <1-day effects are reliable causal estimates rather than artifacts of implementation.

    Authors: The full manuscript methods section already specifies the BART hyperparameters (default priors with 200 trees, alpha=0.95, beta=2, and 1000 posterior draws after 100 burn-in) and provides detailed coding for all confounders in the supplementary materials. However, we acknowledge that the abstract is too concise to convey these details or the requested diagnostics. In revision we will expand the abstract results paragraph to include a one-sentence summary of key diagnostics and will ensure the main text explicitly cross-references the hyperparameter settings and confounder coding. Sensitivity checks will be added as noted in the response to the first comment. These changes will be made without altering the reported point estimates. revision: partial

Circularity Check

0 steps flagged

No circularity: results are direct BART outputs on external survey data

full rationale

The paper applies BART to PATH Youth Waves 3-4 data to estimate average treatment effects of baseline e-cigarette use on change in smoking days, stratified by smoking status and adjusted for listed covariates. The reported effects (-2 days among ever-smokers; <1 day among never-smokers) are model outputs compared against independent population trends; no equation reduces a claimed prediction to a fitted parameter by construction, no self-citation chain bears the central causal claim, and no ansatz or uniqueness theorem is imported from prior author work. The derivation is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The analysis rests on the standard causal assumption that the observed covariates capture all confounding and that BART correctly models the response surface; no free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption No unmeasured confounding remains after adjustment for the listed socio-demographic, intra-individual, behavioural, peer influence, and family background factors
    Invoked when the abstract states that BART provides causal effect estimates after adjusting for these factors.

pith-pipeline@v0.9.1-grok · 5849 in / 1254 out tokens · 31087 ms · 2026-06-30T00:10:22.085461+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [2]

    Selya AS, Foxon F

    doi:10.1111/add.15838 8. Selya AS, Foxon F. Trends in electronic cigarette use and conventional smoking: quantifying a possible ‘diversion’ effect among US adolescents. Addiction . 2021;116(7):1848 –

    work page doi:10.1111/add.15838 2021

  2. [3]

    Wagner LM, Clifton SM

    doi:10.1111/add.15385 9. Wagner LM, Clifton SM. Modeling the public health impact of e - cigarettes on adolescents and adults. Chaos: An Interdisciplinary Journal of Nonlinear Science . 2021;31(11) . doi:10.1063/5.0063593 10. Shahab L, Brown J, Boelen L, Beard E, West R, Munafò MR. Unpacking the Gateway Hypothesis of E - Cigarette Use: The Need for Triang...

    work page doi:10.1111/add.15385 2021

  3. [4]

    Meza R, Jimenez - Mendoza E, Levy DT

    doi:10.1136/tobaccocontrol - 2019 - 055310 12. Meza R, Jimenez - Mendoza E, Levy DT. Trends in Tobacco Use Among Adolescents by Grade, Sex, and Race, 1991 - 2019. JAMA Network Open . 2020;3(12):e2027465 – e2027465. doi:10.1001/jamanetworkopen.2020.27465 13. Delnevo CD, Villanti AC. Dramatic Reductions in Cigarette Smoking Prevalence among High School Yout...

    work page doi:10.1136/tobaccocontrol 2019

  4. [5]

    Jackson SE, Farrow E, Brown J, Shahab L

    doi:10.1111/add.70294 21. Jackson SE, Farrow E, Brown J, Shahab L. Is dual use of nicotine products and cigarettes associated with smoking reduction and cessation behaviours? A prospective study in England. BMJ Open . 2020;10(3):e036055. doi:10.1136/bmjopen - 2019 - 036055 22. Carpenter MJ, Smith TT, Walters KJ, Wahlquist AE, Koval KR, Klemperer EM. Evalu...

    work page doi:10.1111/add.70294 2020

  5. [6]

    SI 2020 Health Economics

    doi:10.1016/j.drugpo.2015.01.006 24. Cotti CD, Courtemanche CJ, Maclean JC, Nesson ET, Pesko MF, Tefft N. The Effects of E - Cigarette Taxes on E - Cigarette Prices and Tobacco Product Sales: Evidence from Retail Panel Data. presented at "SI 2020 Health Economics", July 20 - 21, 2020. National Bureau of Economic Research Working Paper Series . 2020;No. 26...

    work page doi:10.1016/j.drugpo.2015.01.006 2015

  6. [7]

    Carlin JB, Moreno - Betancur M

    doi:10.4236/jamp.2019.77103 26. Carlin JB, Moreno - Betancur M. On the Uses and Abuses of Regression Models: A Call for Reform of Statistical Practice and Teaching. Statistics in Medicine . 2025;44(13 - 14):e10244. doi:10.1002/sim.10244 27. Keller - Hamilton B, Lu B, Roberts ME, Berman ML, Root ED, Ferketich AK. Electronic cigarette use and risk of cigare...

    work page doi:10.4236/jamp.2019.77103 2019

  7. [8]

    Hill J, Perrett G, Dorie V

    doi:10.1017/pan.2019.11 29. Hill J, Perrett G, Dorie V. Machine Learning for Causal Inference. Handbook of Matching and Weighting Adjustments for Causal Inference . Chapman and Hall/CRC; 2023. 30. Tan YV, Roy J. Bayesian additive regression trees and the General BART model. Statistics in Medicine . 2019;38(25):5048 –

    work page doi:10.1017/pan.2019.11 2019

  8. [9]

    Hoffman SR, Gangan N, Chen X, et al

    doi:10.1002/sim.8347 31. Hoffman SR, Gangan N, Chen X, et al. A step - by - step guide to causal study design using real - world data. Health Services and Outcomes Research Methodology . 2025/06/01 2025;25(2):182 –

    work page doi:10.1002/sim.8347 2025

  9. [10]

    doi:10.1007/s10742 - 024 - 00333 - 6 32. Hill JL. Bayesian Nonparametric Modeling for Causal Inference. Journal of Computational and Graphical Statistics . 2011/01/01 2011;20(1):217 – 240. doi:10.1198/jcgs.2010.08162 33. Hill J. 2016 Atlantic Causal Inference Conference Competition. Accessed 21 March, 2026. https://jenniferhill7.wixsite.com/acic - 2016/co...

    work page doi:10.1007/s10742 2011

  10. [11]

    doi:10.1214/19 - BA1195 39. Dorie V. Causal Inference using Bayesian Additive Regression Trees. CRAN. Accessed 11 May, 2026. https://cran.r - project.org/web/packages/bartCause/bartCause.pdf 40. Lumley T. Analysis of Complex Survey Samples. Journal of Statistical Software . 04/15 2004;9(8):1 –

    work page doi:10.1214/19 2026

  11. [12]

    Impact of the e - cigarette era on cigarette smoking among youth in the United States: A population - level study

    doi:10.18637/jss.v009.i08 41. Xu S, Hill J, Cui L, Feng Y, Niaura R. Can Electronic Nicotine Delivery Systems (ENDS) Help Adults Reduce Cigarette smoking?: A Bart Analysis for Causal inference. presented at: Society for Research on Nicotine and Tobacco Annual Meeting; 2026; Baltimore, MD. https://web.archive.org/web/20260215002004/https://trg.joynadmin.or...

    work page doi:10.18637/jss.v009.i08 2026

  12. [13]

    Foxon F, Selya AS

    doi:10.5694/mja2.70000 3. Foxon F, Selya AS. Electronic cigarettes, nicotine use trends and use initiation ages among US adolescents from 1999 to 2018. Addiction . 2020;115(12):2369 – 2378. doi:10.1111/add.15099 4. King G. Advanced Quantitative Research Methodology, Lecture Notes: Model Dependence in Counterfactual Inference. https://web.archive.org/web/2...

    work page doi:10.5694/mja2.70000 1999

  13. [14]

    Lee PN, Fry JS

    doi:10.1111/add.15246 9. Lee PN, Fry JS. Investigating gateway effects using the PATH study [version 2; peer review: 2 approved]. F1000Research 2019;8(264)doi:10.12688/f1000research.18354.2 10. Delnevo CD. e - Cigarette and Cigarette Use Among Youth: Gateway or Common Liability? JAMA Network Open . 2023;6(3):e234890 – e234890. doi:10.1001/jamanetworkopen....

    work page doi:10.1111/add.15246 2019

  14. [15]

    Grant SW, Hickey GL, Head SJ

    doi:10.4103/ijpsym.Ijpsym_193_19 17. Grant SW, Hickey GL, Head SJ. Statistical primer: multivariable regression considerations and pitfalls. European Journal of Cardio - Thoracic Surgery . 2018;55(2):179 –

    work page doi:10.4103/ijpsym.ijpsym_193_19 2018

  15. [16]

    Carlin JB, Moreno - Betancur M

    doi:10.1093/ejcts/ezy403 18. Carlin JB, Moreno - Betancur M. On the Uses and Abuses of Regression Models: A Call for Reform of Statistical Practice and Teaching. Statistics in Medicine . 2025;44(13 - 14):e10244. doi:10.1002/sim.10244 19. Hill J, Perrett G, Dorie V. Machine Learning for Causal Inference. Handbook of Matching and Weighting Adjustments for C...

    work page doi:10.1093/ejcts/ezy403 2025

  16. [17]

    Rutter M

    doi:10.1093/aje/kws412 21. Rutter M. Proceeding From Observed Correlation to Causal Inference: The Use of Natural Experiments. Perspectives on Psychological Science . 2007;2(4):377 – 395. doi:10.1111/j.1745 - 6916.2007.00050.x

    work page doi:10.1093/aje/kws412 2007