arxiv: 2603.00041 · v2 · submitted 2026-02-09 · 💻 cs.LG · cs.AI· econ.EM· stat.ME

Recognition: 2 theorem links

· Lean Theorem

Econometric vs. Causal Structure-Learning for Time-Series Policy Decisions: Evidence from the UK COVID-19 Policies

Bruno Petrungaro , Anthony C. Constantinou

Authors on Pith no claims yet

Pith reviewed 2026-05-16 06:04 UTC · model grok-4.3

classification 💻 cs.LG cs.AIecon.EMstat.ME

keywords causal discoverytime serieseconometricsmachine learningpolicy decisionsCOVID-19Bayesian networksgraphical models

0 comments

The pith

Econometric methods enforce strict temporal rules in time-series causal graphs while causal machine learning recovers denser structures with more identifiable relationships, shown on UK COVID-19 policy data.

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

The paper compares traditional econometric techniques for establishing causality in sequential data with causal machine learning algorithms that learn graphical structures. It tests both families on real UK COVID-19 policy time series to determine how each supports policy decisions by recovering cause-and-effect links. Econometric approaches supply explicit rules that respect time order in the resulting graphs. Causal machine learning explores a wider range of possible structures and tends to produce denser graphs that identify additional causal relationships. The work also supplies translation code so econometric outputs can be used in standard Bayesian network software.

Core claim

Four econometric methods are evaluated against eleven causal machine learning algorithms on their recovery of graphical structures from UK COVID-19 policy time-series data. Econometric methods supply clear rules for temporal ordering within the graphs they produce. Causal machine learning algorithms search a larger space of graph structures, resulting in denser networks that capture more identifiable causal relationships. These differences are examined for their value in supporting policy decision-making.

What carries the argument

Direct comparison of four econometric time-series causality methods against eleven causal machine learning algorithms applied to the same UK policy intervention data, measuring differences in graph structure, dimensionality, and number of recoverable causal effects.

If this is right

Econometric outputs can be converted into standard Bayesian network formats for wider use in policy modeling.
Causal machine learning may be chosen when the goal is to identify the largest number of potential causal effects for exploratory policy analysis.
Strict temporal constraints from econometric methods reduce ambiguity when models must respect the order of policy interventions.
Denser graphs from machine learning can surface additional policy interactions that stricter econometric rules would omit.

Where Pith is reading between the lines

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

Policy analysts could run both families in parallel on the same data and retain only the causal links on which they agree to increase robustness.
The observed difference in graph density suggests that future time-series causal work may benefit from hybrid algorithms that combine econometric temporal rules with machine learning search.
Testing these methods on simulated time series with known ground-truth causal graphs would quantify how often each family recovers the correct structure under controlled noise levels.

Load-bearing premise

The UK COVID-19 policy time series contains identifiable causal effects that the algorithms can recover without hidden confounders or measurement error that would invalidate the graphs.

What would settle it

If the graphs recovered by the econometric methods and the causal ML algorithms disagree on the direction of key causal links between policy variables and outcomes, or if neither set of graphs aligns with documented timelines of actual policy impacts, the claim that both families recover valid causal structures would be challenged.

read the original abstract

Causal machine learning (ML) recovers graphical structures that inform us about potential cause-and-effect relationships. Most progress has focused on cross-sectional data with no explicit time order, whereas recovering causal structures from time series data remains the subject of ongoing research in causal ML. In addition to traditional causal ML, this study assesses econometric methods that some argue can recover causal structures from time series data. The use of these methods can be explained by the significant attention the field of econometrics has given to causality, and specifically to time series, over the years. This presents the possibility of comparing the causal discovery performance between econometric and traditional causal ML algorithms. We seek to understand if there are lessons to be incorporated into causal ML from econometrics, and provide code to translate the results of these econometric methods to the most widely used Bayesian Network R library, bnlearn. We investigate the benefits and challenges that these algorithms present in supporting policy decision-making, using the real-world case of COVID-19 in the UK as an example. Four econometric methods are evaluated in terms of graphical structure, model dimensionality, and their ability to recover causal effects, and these results are compared with those of eleven causal ML algorithms. Amongst our main results, we see that econometric methods provide clear rules for temporal structures, whereas causal-ML algorithms offer broader discovery by exploring a larger space of graph structures that tends to lead to denser graphs that capture more identifiable causal relationships.

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.

Desk Editor's Note private letter to a colleague

Head-to-head benchmark of econometric and causal ML methods on COVID policy time series, but validation of the graphs is missing.

read the letter

Colleague, This paper compares four econometric methods for recovering causal structures in time series against eleven causal machine learning algorithms, using UK COVID-19 policy data as the case study. They also supply code to translate the econometric results into the bnlearn library. The main observation is that econometric methods enforce clear temporal ordering rules, whereas the causal ML approaches explore a wider space and produce denser graphs that identify more causal relationships. What works here is the direct empirical comparison on the same dataset. It highlights practical differences that could help someone decide which family of methods to use for similar policy time series problems. The code contribution makes it easier to combine or compare outputs across tools. The soft spots center on evaluation. The claim about causal ML capturing more identifiable relationships assumes we can distinguish true edges from spurious ones, but the paper provides no external validation, such as known policy impacts from epidemiology, to score the graphs. The abstract mentions results on structure and causal recovery without numbers or error bars, so the strength of the evidence is unclear from the summary alone. This is a common issue with observational causal discovery but needs addressing for the conclusions to land firmly. This is relevant for applied researchers working on causal discovery in time series, particularly those dealing with policy decisions. A reader looking for benchmark-style guidance would get value from it. The work engages honestly with the literature from both econometrics and causal ML. I think it deserves peer review so that the full paper can be assessed for methods details and any additional validation steps.

Referee Report

1 major / 1 minor

Summary. The manuscript compares four econometric methods against eleven causal ML algorithms for recovering causal graph structures from UK COVID-19 policy time-series data. It supplies code to export econometric outputs into the bnlearn library and reports that econometric approaches impose clear temporal ordering rules while causal-ML methods explore a larger space of graphs, producing denser structures that the authors interpret as capturing more identifiable causal relationships. The evaluation focuses on graphical structure, model dimensionality, and ability to recover causal effects.

Significance. If an objective validation benchmark were supplied, the work could usefully highlight complementary strengths of the two families for time-series policy analysis. In its current form the comparison remains difficult to interpret because no external ground truth (e.g., known policy effects from epidemiology or randomized evidence) is used to score edge correctness, so the claim that denser graphs are preferable cannot be assessed quantitatively.

major comments (1)

[Abstract] Abstract: the assertion that causal-ML algorithms 'capture more identifiable causal relationships' is unsupported by any precision/recall metric, edge-recovery score, or external benchmark against documented UK policy effects; without such validation the preference for higher edge density is untestable and load-bearing for the headline comparison.

minor comments (1)

The manuscript should report the exact preprocessing steps applied to the UK COVID-19 policy series and any explicit assumptions regarding hidden confounders or measurement error.

Simulated Author's Rebuttal

1 responses · 1 unresolved

We thank the referee for the detailed and constructive report. The observation that our comparison lacks an external ground-truth benchmark for edge correctness is correct and highlights a genuine limitation of the study. We have revised the abstract and added explicit discussion of this limitation, softening all claims about denser graphs capturing 'more identifiable causal relationships' to focus instead on observable differences in graph structure and temporal constraints. We cannot, however, supply a quantitative validation benchmark because no verified external ground truth exists for the UK COVID-19 policy time series.

read point-by-point responses

Referee: [Abstract] Abstract: the assertion that causal-ML algorithms 'capture more identifiable causal relationships' is unsupported by any precision/recall metric, edge-recovery score, or external benchmark against documented UK policy effects; without such validation the preference for higher edge density is untestable and load-bearing for the headline comparison.

Authors: We agree that the original wording was too strong. The manuscript's empirical observation is that causal-ML methods return denser graphs than the econometric approaches, which impose stricter temporal ordering. Without an external benchmark we cannot demonstrate that these additional edges correspond to true causal effects. We have therefore revised the abstract to remove the phrase 'capture more identifiable causal relationships' and replaced it with language that reports the structural difference (greater edge density and broader exploration of the graph space) while explicitly noting the absence of ground-truth validation. A new limitations paragraph has been added that discusses the difficulty of obtaining precision/recall scores for real-world policy time series and the consequent interpretive caution required. revision: yes

standing simulated objections not resolved

No verified external ground truth (e.g., known policy effects from epidemiology or randomized evidence) is available for the UK COVID-19 time-series data, so quantitative edge-recovery metrics cannot be computed.

Circularity Check

0 steps flagged

Empirical benchmark comparison with no derivation chain

full rationale

The paper applies existing econometric methods and causal ML algorithms to the same UK COVID-19 policy time-series dataset, then compares their recovered graphs on structure, dimensionality, and identifiable effects. No equations or steps derive a new result from fitted parameters that are then relabeled as predictions. No self-citation chain is invoked to justify uniqueness or force a modeling choice. The central claim (econometric methods give clearer temporal rules while causal-ML yields denser graphs) is an empirical observation from running standard algorithms, not a reduction to the paper's own inputs by construction. Absence of external ground-truth validation is a validity concern, not circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The comparison rests on standard causal discovery assumptions such as causal sufficiency and the ability of the chosen algorithms to recover the true graph from observational time series; no new free parameters or invented entities are introduced in the abstract.

axioms (2)

domain assumption Causal sufficiency (no unobserved confounders)
Implicit in all causal discovery algorithms evaluated; required for the recovered graphs to be interpreted as causal.
domain assumption Time order reflects causal order
Used by the econometric methods to constrain possible edges.

pith-pipeline@v0.9.0 · 5570 in / 1330 out tokens · 42040 ms · 2026-05-16T06:04:30.738294+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Four econometric methods are evaluated... compared with those of eleven causal ML algorithms... econometric methods provide clear rules for temporal structures, whereas causal-ML algorithms offer broader discovery by exploring a larger space of graph structures
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We use routinely collected publicly available aggregated data... VAR process of order p... LASSO... Hill-Climbing (HC)... Structural Hamming Distance (SHD)

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.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Time series causal discovery with variable lags
cs.LG 2026-04 unverdicted novelty 7.0

A Tabu-based algorithm learns time-ordered causal graphs from time series by optimizing per-edge lags with a decomposable BIC score and explicit lag penalty.
Time series causal discovery with variable lags
cs.LG 2026-04 unverdicted novelty 5.0

A Tabu-based algorithm learns time-ordered causal graphs from time series with variable per-edge lags using a decomposable BIC score and explicit lag penalty.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages · cited by 1 Pith paper · 2 internal anchors

[1]

Journal of Economic Perspectives31(2), 3–32 (2017) https://doi.org/ 10.1257/jep.31.2.3

Athey, S., Imbens, G.W.: The state of applied econometrics: Causality and policy evaluation. Journal of Economic Perspectives31(2), 3–32 (2017) https://doi.org/ 10.1257/jep.31.2.3

work page doi:10.1257/jep.31.2.3 2017
[2]

American Economic Review107(1), 138–68 (2017) https://doi.org/10

Pinotti, P.: Clicking on heaven’s door: The effect of immigrant legalization on crime. American Economic Review107(1), 138–68 (2017) https://doi.org/10. 1257/aer.20150355

work page 2017
[3]

2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), 392–395 (2019) https://doi.org/10.1109/ICMLA

Kocacoban, D., Cussens, J.: Online causal structure learning in the presence of latent variables. 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), 392–395 (2019) https://doi.org/10.1109/ICMLA. 2019.00073 30

work page doi:10.1109/icmla 2019
[4]

In: UAI Workshop on Causal Structure Learning (2012)

Kummerfeld, E., Danks, D.: Online learning of time-varying causal structures. In: UAI Workshop on Causal Structure Learning (2012)

work page 2012
[5]

Advances in neural information processing systems26(2013)

Kummerfeld, E., Danks, D.: Tracking time-varying graphical structure. Advances in neural information processing systems26(2013)

work page 2013
[6]

Proceedings of the Royal Society of London A115(772), 700–721 (1927)

Kermack, W., McKendrick, A.: Contributions to the mathematical theory of epidemics. Proceedings of the Royal Society of London A115(772), 700–721 (1927)

work page 1927
[7]

Biometrika82, 669–710 (1995)

Pearl, J.: Causal diagrams for empirical research. Biometrika82, 669–710 (1995)

work page 1995
[8]

Expert Systems with Applications234(2023) https: //doi.org/10.1016/j.eswa.2023.121069

Constantinou, A., et al.: Open problems in causal structure learning: A case study of covid-19 in the uk. Expert Systems with Applications234(2023) https: //doi.org/10.1016/j.eswa.2023.121069

work page doi:10.1016/j.eswa.2023.121069 2023
[9]

The Official UK Govern- ment Website for Data and Insights on Coronavirus (COVID-19)

Coronavirus (COVID-19): Testing in United Kingdom. The Official UK Govern- ment Website for Data and Insights on Coronavirus (COVID-19)

work page
[10]

Coronavirus (COVID-19): Healthcare in United Kingdom

GOV.UK (2022d). Coronavirus (COVID-19): Healthcare in United Kingdom. The Official UK Government Website for Data and Insights on Coronavirus (COVID- 19)

work page
[11]

doi: 10.1115/1.3662552

Kalman, R.E.: A new approach to linear filtering and prediction problems. Journal of Basic Engineering82(1) (1960) https://doi.org/10.1115/1.3662552

work page doi:10.1115/1.3662552 1960
[12]

The R Journal9(1), 207–218 (2017) https://doi.org/10.32614/RJ-2017-009

Moritz, S., Bartz-Beielstein, T.: imputeTS: Time Series Missing Value Imputation in R. The R Journal9(1), 207–218 (2017) https://doi.org/10.32614/RJ-2017-009

work page doi:10.32614/rj-2017-009 2017
[13]

Journal of the royal statistical society

Hartigan, J.A., Wong, M.A.: Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics)28(1), 100–108 (1979)

work page 1979
[14]

Springer (2013)

Radhakrishnan, N., et al.: Bayesian Networks in R with Applications in Systems Biology. Springer (2013)

work page 2013
[15]

Annals of Statistics34, 1436–1462 (2006)

Meinshausen, N., B¨ uhlmann, P.: Variable selection and high-dimensional graphs with the lasso. Annals of Statistics34, 1436–1462 (2006)

work page 2006
[16]

Journal of the Royal Statistical Society Series B: Statistical Methodology58(1), 267–288 (1996)

Tibshirani, R.: Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology58(1), 267–288 (1996)

work page 1996
[17]

Hastie, T., Efron, B.: lars: Least angle regression, lasso and forward stagewise (Retrieved December, 2023)

work page 2023
[18]

The Annals of Statistics32(2), 407–451 (2004) 31

Efron, B.,et al.: Least angle regression. The Annals of Statistics32(2), 407–451 (2004) 31

work page 2004
[19]

BMC bioinformatics8(2), 1–8 (2007)

Opgen-Rhein, R., Strimmer, K.: Learning causal networks from systems biol- ogy time course data: an effective model selection procedure for the vector autoregressive process. BMC bioinformatics8(2), 1–8 (2007)

work page 2007
[20]

Schaefer, J., et al.: Genenet: Modeling and inferring gene networks (Retrieved December, 2023)

work page 2023
[21]

Bioinfor- matics25(3), 417–418 (2009)

Chiquet, J.,et al.: Simone: Statistical inference for modular networks. Bioinfor- matics25(3), 417–418 (2009)

work page 2009
[22]

Applied Intelligence55(6), 496 (2025)

Petrungaro, B., Kitson, N.K., Constantinou, A.C.: Investigating potential causes of sepsis with bayesian network structure learning. Applied Intelligence55(6), 496 (2025)

work page 2025
[23]

Journal of Statistical Software35(3), 1–22 (2010) https://doi.org/10.18637/jss.v035.i03

Scutari, M.: Learning bayesian networks with the bnlearn R package. Journal of Statistical Software35(3), 1–22 (2010) https://doi.org/10.18637/jss.v035.i03

work page doi:10.18637/jss.v035.i03 2010
[24]

Colombo, D.,et al.: Order-independent constraint-based causal structure learn- ing. J. Mach. Learn. Res.15(1), 3741–3782 (2014)

work page 2014
[25]

Social science computer review9(1), 62–72 (1991)

Spirtes, P.,et al.: An algorithm for fast recovery of sparse causal graphs. Social science computer review9(1), 62–72 (1991)

work page 1991
[26]

PhD thesis, School of Computer Science, Carnegie Mellon University Pittsburgh, PA, USA (2003)

Margaritis, D.: Learning bayesian network model structure from data. PhD thesis, School of Computer Science, Carnegie Mellon University Pittsburgh, PA, USA (2003)

work page 2003
[27]

In: FLAIRS Conference, vol

Tsamardinos, I.,et al.: Algorithms for large scale markov blanket discovery. In: FLAIRS Conference, vol. 2, pp. 376–380 (2003). St. Augustine, FL

work page 2003
[28]

In: Fifth IEEE International Conference on Data Mining (ICDM’05), p

Yaramakala, S.,et al.: Speculative markov blanket discovery for optimal feature selection. In: Fifth IEEE International Conference on Data Mining (ICDM’05), p. 4 (2005). IEEE

work page 2005
[29]

In: European Conference on Evolutionary Computation, Machine Learning and Data Mining in Bioinformatics, pp

Pena, J.M.: Learning gaussian graphical models of gene networks with false discovery rate control. In: European Conference on Evolutionary Computation, Machine Learning and Data Mining in Bioinformatics, pp. 165–176 (2008). Springer

work page 2008
[30]

Expert Systems with Applications41(15), 6755–6772 (2014)

Gasse, M.,et al.: A hybrid algorithm for bayesian network structure learning with application to multi-label learning. Expert Systems with Applications41(15), 6755–6772 (2014)

work page 2014
[31]

Learning Bayesian Network Structure from Massive Datasets: The "Sparse Candidate" Algorithm

Friedman, N., et al.: Learning bayesian network structure from massive datasets: The ”sparse candidate” algorithm. arXiv preprint arXiv:1301.6696 (2013)

work page internal anchor Pith review Pith/arXiv arXiv 2013
[32]

PhD thesis (1995) 32

Bouckaert, R.R.: Bayesian belief networks: from construction to inference. PhD thesis (1995) 32

work page 1995
[33]

Machine learning65, 31–78 (2006)

Tsamardinos, I.,et al.: The max-min hill-climbing bayesian network structure learning algorithm. Machine learning65, 31–78 (2006)

work page 2006
[34]

Chapman and Hall, Boca Raton (2021)

Scutari, M., Denis, J.-B.: Bayesian Networks with Examples in R, 2nd edn. Chapman and Hall, Boca Raton (2021). ISBN 978-0367366513

work page 2021
[35]

Journal of Statistical Software27(4) (2008)

Pfaff, B.: Var, svar and svec models: Implementation within R package vars. Journal of Statistical Software27(4) (2008)

work page 2008
[36]

IEEE transactions on automatic control19(6), 716–723 (1974)

Akaike, H.: A new look at the statistical model identification. IEEE transactions on automatic control19(6), 716–723 (1974)

work page 1974
[37]

The annals of statistics, 461– 464 (1978)

Schwarz, G.: Estimating the dimension of a model. The annals of statistics, 461– 464 (1978)

work page 1978
[38]

Journal of the Royal Statistical Society: Series B (Methodological)41(2), 190–195 (1979)

Hannan, E.J., Quinn, B.G.: The determination of the order of an autoregression. Journal of the Royal Statistical Society: Series B (Methodological)41(2), 190–195 (1979)

work page 1979
[39]

Ann Inst Stat Math21, 243–247 (1969) https://doi.org/10.1007/BF02532251

Akaike, H.: Fitting autoregressive models for prediction. Ann Inst Stat Math21, 243–247 (1969) https://doi.org/10.1007/BF02532251

work page doi:10.1007/bf02532251 1969
[40]

International Statistical Review/Revue Internationale de Statistique, 163–172 (1987)

Jarque, C.M., Bera, A.K.: A test for normality of observations and regression residuals. International Statistical Review/Revue Internationale de Statistique, 163–172 (1987)

work page 1987
[41]

Komsta, L., Novomestky, F.: Moments: Moments, Cumulants, Skewness, Kurtosis and Related Tests. (2022). R package version 0.14.1. https://CRAN.R-project. org/package=moments

work page 2022
[42]

Econometrica: Journal of the Econometric Society, 1293–1301 (1978)

Godfrey, L.G.: Testing against general autoregressive and moving average error models when the regressors include lagged dependent variables. Econometrica: Journal of the Econometric Society, 1293–1301 (1978)

work page 1978
[43]

Australian economic papers17(31), 334–355 (1978)

Breusch, T.S.: Testing for autocorrelation in dynamic linear models. Australian economic papers17(31), 334–355 (1978)

work page 1978
[44]

R News 2(3), 7–10 (2002)

Zeileis, A., Hothorn, T.: Diagnostic checking in regression relationships. R News 2(3), 7–10 (2002)

work page 2002
[45]

Springer, New York, USA (2016)

Wickham, H.: Ggplot2: Elegant Graphics for Data Analysis. Springer, New York, USA (2016). https://ggplot2.tidyverse.org

work page 2016
[46]

Journal of Statistical Software76(12), 1–30 (2017) https://doi.org/10.18637/jss

Tikka, S., Karvanen, J.: Identifying causal effects with the R package causaleffect. Journal of Statistical Software76(12), 1–30 (2017) https://doi.org/10.18637/jss. v076.i12

work page doi:10.18637/jss 2017
[47]

Identification of Conditional Interventional Distributions

Shpitser, I., Pearl, J.: Identification of conditional interventional distributions. 33 arXiv preprint arXiv:1206.6876 (2012)

work page internal anchor Pith review Pith/arXiv arXiv 2012
[48]

In: Proceedings of the National Conference on Artificial Intelligence, vol

Shpitser, I., Pearl, J.: Identification of joint interventional distributions in recur- sive semi-markovian causal models. In: Proceedings of the National Conference on Artificial Intelligence, vol. 21, p. 1219 (2006). Menlo Park, CA; Cambridge, MA; London; AAAI Press; MIT Press; 1999

work page 2006
[49]

Nature (2023) 34

Ferretti, L., et al.: Digital measurement of sars-cov-2 transmission risk from 7 million contacts. Nature (2023) 34

work page 2023