Identifying Elasticities in Autocorrelated Time Series Using Causal Mechanism Graphs

arXiv:2409.15530 · econ.EM, stat.ME, stat.ML · Submitted 2024-09-23 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Identifying Elasticities in Autocorrelated Time Series Using Causal Mechanism Graphs".

Jane: The gist: Causal time graphs can simplify instrumental variable identification and help select consistent estimators for price elasticity in autocorrelated time series <ref:2409.15530#pg3>.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: Okay, so this paper is "Identifying Elasticities in Autocorrelated Time Series Using Causal Mechanism Graphs". It's tackling the problem of estimating price elasticity from observational data using instrumental variables, but it points out that simple instrumental variables can be inconsistent when you have time series that are autocorrelated.

Jane: So the main idea here seems to be using causal time graphs to make identifying these instrumental variables easier and more consistent, especially in settings where the data has autocorrelation. What's the core claim they are making about this approach?

Lu: They propose modeling the equilibrium condition using an unobserved confounder, which leads them to a directed acyclic graph, or DAG, while keeping that assumption of simultaneous determination of prices and quantities intact. Then they exploit graphical inference to derive valid instrumental variables that can achieve consistency by simultaneously estimating nuisance effects.

Meng: That sounds complicated on the setup side. So, what's the practical payoff for someone trying to actually run an analysis? Does this mean we can finally trust some IV estimates in these complex time series?

Lalam: It does. The paper argues that observing significant differences between estimates of presumably valid instrumental variables can help us reject false model assumptions, which improves our understanding of the underlying economic dynamics. This is about testing the model itself, not just getting a single number.

Tom: So it's not just about finding one better estimator, it's about having a systematic way to test if your underlying structure—your demand equation—is actually right when you have these autocorrelated time series. What kind of structures are they testing?

Jane: They explore three different demand equations: Model I for inertial demand, Model II for partially responsive demand, and Model III for demand shifting. These models all create autocorrelated data but they each have different implications for how the conditional instrumental variable estimators behave.

Lu: They formalize the equilibrium relationship as a system with time-instantaneous hidden confounding between price and demand, which is represented by that DAG structure they derive. They then use graphical criteria like d-separation to check if an estimator meets three specific requirements for being a valid conditional IV estimator.

Paper summary: Meng: Three requirements? I need those concrete rules if I'm going to know what's actually working versus what's just theoretical. What are these conditions for a conditional IV to be valid?

Tom: Requirement one says that the variables you are estimating and the instrument must be d-separated given some other set of variables in the graph, which is denoted as G X not→Y, and requirement two says that the confounder B isn't a descendant of X or Y in the graph.

Jane: And requirement three is about rank, saying that the matrix E

cov(XX, XIXB): has full row rank with respect to X. If you meet those three things using an estimator based on I, X, and B—that's when it becomes a valid conditional IV estimator.

Lu: The paper then discusses two main ways to build these estimators: one is blocking paths via conditioning, like the lag-augmented local projection IV estimator which relies on blocking paths for validity in Model I.

Meng: And the other way is estimating nuisance effects simultaneously, like CIV(Wt - s L s=zero (Pt, Pt-one Dt-one) → Dt ∅). They say that this second approach can be more powerful when you have small sample sizes.

Tom: The empirical performance analysis on the German electricity market data shows that discrepancies between estimates of presumably valid estimators actually help reject false model assumptions. For example, in their simulations based on Model I, the confidence intervals of all valid estimators overlap.

Jane: But if you had assumed Model II instead, you would be forced to reject that assumption because certain estimators like one and six deviate from the others <ref:2409.15530#pg2>. That's a strong signal for testing assumptions.

Lu: And they also found that observing the autocorrelation of demand isn't enough on its own to correct for the bias of a naive instrumental variable estimator; the divergence between conditional IV estimates in German electricity demand suggests that ignoring time dynamics leads to significant bias.

Meng: So, if I were deploying a model, what does this mean for my pipeline? Does this mean I have to explicitly build these causal graphs into my data prep stage before running any estimation?

Paper summary: Tom: It means you're moving beyond just fitting a simple equation and you need to be explicit about the structural assumptions in your time series setup. The paper shows that simply ignoring time dynamics overestimates the elasticity by up to an order of magnitude, which is a big number.

Jane: So, what's the actual result on the numbers? Are these new causal graph methods yielding specific elasticity values for electricity demand?

Tom: In linear specification, their CIV estimators and even their benchmark estimator two converge to an approximate value of negative two hundred MW per Euro per Megawatt-hour <ref:2409.15530#pg1>. In log-log specification, that estimate moves to about negative zero point one.

Lu: Those results confirm that the dynamics of electricity markets are more complex than what the simple Models I, II, and III suggest because when taken in isolation, none of those models are sufficient to explain the disagreement between the different estimates.

Meng: So for production use, if we only look at on-peak demand data—which is where their confidence intervals overlap—we can expect high demand response there. But for off-peak periods, it looks like a totally different dynamic that needs investigation.

Jane: That seems to be the point about the off-peak periods needing further study, because the structural dynamics there look different from what the simple models predict. This paper gives us a way to investigate that difference using these causal graphs.

Tom: The paper concludes by saying that causal graphs provide a transparent expression of structural assumptions, which helps in understanding biasing dynamics and testing model validity. It's a powerful addition to the toolkit for anyone working with time series data structured this way.

Lu: They suggest extending causal time graphs to estimate impulse response functions as future work, because those can strengthen the intuition behind recent results about local projection IV estimators, which is really interesting for understanding dynamic systems.

Jane: So what does this mean for the broader field of econometrics? It seems like a way to systematically identify which IV estimators are actually consistent based on the data's underlying structure. This paper is "Identifying Elasticities in Autocorrelated Time Series Using Causal Mechanism Graphs".

Conclusion: Tom: So we’ve been looking at how to estimate price elasticity when you have time series data that has autocorrelation, and this paper is called "Identifying Elasticities in Autocorrelated Time Series Using Causal Mechanism Graphs".

Jane: Right, so essentially they are using these causal graphs to make sure the instrumental variables they use are actually consistent, even with all that time structure messing things up.

Lu: They take the equilibrium conditions and model them as a directed acyclic graph, or DAG. This structure lets them systematically analyze why some estimators work and others fail based on graphical criteria like d-separation.

Meng: So they’re not just picking an IV; they’re using the graph to prove it’s valid before you even run the regression. That sounds like a lot of setup for production, though.

Tom: It is, but that setup is what lets them test if their model assumptions are right. They show that observing differences between estimates of these presumably valid estimators can actually help you reject false assumptions about the underlying demand structure.

Jane: That’s interesting because it moves beyond just getting one number and makes the whole process about testing your economic theory against the data's structure itself.

Lalam: I’m pretty sure this direction is solid. The paper shows that ignoring time dynamics alone isn't enough to fix the bias in a naive instrumental variable estimator; you need that structural insight from the causal graph.

Lu: They test three different demand models, Model I, Model II, and Model III. Each one creates different kinds of autocorrelation, and they show how those dependencies change what’s valid for an estimator.

Tom: And their empirical results on the German electricity market data are actually pretty telling. They found that while all their valid estimators gave overlapping confidence intervals under one model assumption, switching to another model assumption forced you to reject the first one because the estimates just didn't align anymore.

Jane: That means we can use these results as a diagnostic tool for the real world, not just a theoretical exercise on clean data. It shows us where our simple models are breaking down in practice.

Meng: So if I’m shipping a model to production, this suggests I need to think about what kind of structural dependencies my instrument is actually capturing over time, because autocorrelation alone isn't the whole story.

Tom: Exactly. The point of the paper is that causal graphs give you transparency into those structural assumptions, helping you understand why some IV methods are biased and others aren't.

Jane: And it opens up a way to test your economic intuition directly against the patterns hidden in time series data, which is really powerful for anyone trying to model complex systems.

Centre for Sustainability, Hertie School · Department of Mathematics, ETH Zurich · Department of Economics and Institute of Energy Economics, University of Cologne

econ.EM, stat.ME, stat.ML

Submitted: 2024-09-23

Updated: 2026-10-19

Comments: Substantially revised; author order changed to alphabetical by agreement

Code: https://github.com/jscanales/elasticities_with_graphs

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: The gist: Causal time graphs can simplify instrumental variable identification and help select consistent estimators for price elasticity in autocorrelated time series <ref:2409.15530#pg3>.

Key concepts

Causal Mechanism Time Graphs: These are acyclic graphs used to model the equilibrium conditions where prices and quantities are determined simultaneously. They allow researchers to maintain the assumption of simultaneous determination while providing an interpretable structure for analyzing causal relationships between variables over time.
d-separation Criteria (CIV1, CIV2, CIV3): These are three specific requirements derived from the causal graph that determine if a conditional IV estimator is valid. They ensure that the instrument and the endogenous variable are appropriately separated in the graph structure, which is necessary for obtaining a consistent estimate.
Conditional Instrumental Variable (CIV) Estimator: This is an estimator based on instrumental variables that satisfies specific graphical criteria to be considered valid. The paper discusses two main ways to construct these, including blocking paths using lagged terms or estimating nuisance effects simultaneously, depending on the model structure.

Terminology

Consistency — 1× in this paper · explained in 3 episodes
Directed Acyclic Graph (DAG) — 1× in this paper · explained in 2 episodes

Summary

The gist: Causal time graphs can simplify instrumental variable identification and help select consistent estimators for price elasticity in autocorrelated time series <ref:2409.15530#pg3>.

How it works

The paper proposes using causal time graphs to model the equilibrium condition of a system with unobserved confounding between prices and quantities, which leads to a directed acyclic graph (DAG) <ref:2409.15530#pg7>. This approach allows for the systematic analysis of estimator validity by exploiting graphical criteria such as d-separation <ref:2409.15530#pg8>. The authors argue that observing significant differences between estimates of presumably valid estimators can help reject false model assumptions, thereby improving our understanding of underlying economic dynamics <ref:2409.15530#pg9>.

The paper introduces three alternative demand equations (Model I: Inertial Demand, Model II: Partially responsive demand, and Model III: Demand shifting) to explore different structural dependencies in time <ref:2409.15530#pg10>. These models generate autocorrelated time series but have different implications for the validity of conditional instrumental variable estimators. The paper formalizes the equilibrium relationship as a system with time-instantaneous hidden confounding between price and demand, represented by a DAG.

Graphical Criteria for Validity

The validity of conditional IV (CIV) estimators is determined by three requirements derived from d-separation in the causal graph G. These criteria are:

(CIV1) I and Y are d-separated given B in the graph GX not→Y, which denotes the graph obtained when removing all direct edges from X to Y from G.
(CIV2) B is not a descendant of X ∪ Y in G.
(CIV3) the matrix E[cov(XX, XIXB)] has rank X, that is, full row rank.

If an estimator is based on I, X, B such that these requirements are met, it is called a valid CIV estimator and the resulting estimate βˆ:= arg min b ˆcov(XY − b T XX, XIXB) is a consistent estimator for β if T → ∞.

Types of Valid Estimators

The paper presents two main concepts for constructing valid estimators:

Blocking paths via conditioning: This approach includes lagged terms in the conditioning set, such as estimator CIV(WtPt → Dt Dt−1) which is valid for Model I. The lag-augmented local projection IV estimator (estimator 4) is a benchmark estimator that relies on blocking paths to achieve validity.
Estimating nuisance effects: This approach involves simultaneously estimating effects that are not of primary interest and whose estimates are afterward disregarded, such as CIV(Wt−s L s=0 (Pt, Pt−1, Dt−1) → Dt ∅). This approach is noted to be more powerful at small sample sizes.

Empirical Performance and Model Validation

The paper applies these CIV estimators to simulated and real-world data from the German electricity market. The findings show that discrepancies between estimates of presumably valid estimators can help reject false model assumptions. For instance, in simulations based on Model I, the confidence intervals of all valid estimators overlap, but if one had assumed Model II, one would be forced to reject this assumption since estimators like 1 and 6 deviate from others. The empirical performance analysis shows that for Model I, the nuisance IV estimator (estimator 8) tends to have a smaller average absolute percentage error and yields shorter confidence intervals at smaller sample sizes than the conditional IV estimators.

The study also demonstrates that knowing the observed autocorrelation of demand is not sufficient to correct for the bias of the naive IV estimator. The divergence between CIV estimates when applied to German electricity demand suggests that the widely used IV estimator, which ignores time dynamics, cannot identify the own-price elasticity without significant bias.

Conclusion and Future Directions

The paper demonstrates the advantages of using directed acyclic graphs to estimate price elasticity with autocorrelated instruments. Causal graphs allow for a transparent expression of structural assumptions, understanding biasing dynamics, and testing model validity. Future work is suggested to extend causal time graphs to estimate impulse response functions. Causal full time graphs can help strengthen the intuition behind recent results in this literature, such as the robustness of local projection IV estimators. The graphical approach provides a powerful addition to the toolkit of every economist working with data structured as a time series.

Data and Application

The application uses hourly electricity demand, price, and wind generation data from ENTSO-E spanning November 1st, 2017, to December 31st, 2020. Wind power generation is used as an exogenous time series instrument because it is relevant and credibly exogenous. The analysis considers structural autocorrelation in both the instrument and the demand equation, showing that observed autocorrelation alone cannot predict potential bias. The results indicate that nighttime consumption has different dynamics than daytime consumption, suggesting that simple models are insufficient to explain real-world dynamics.

Quantitative Results

The CIV estimators, including the benchmark estimator 2, have overlapping confidence intervals across different structural assumptions for the own-price elasticity of German electricity demand. The estimates converge to a value of approximately −200 MW/(EUR/MWh) in linear specification and −0.1 in log-log specification. The results confirm that the dynamics of electricity markets are more complex than suggested by the simple Models I, II, and III because taken in isolation, none of the models is sufficient to explain the disagreement between the estimators. The analysis shows that on-peak demand exhibits a pattern where confidence intervals overlap, suggesting that this period is where one would expect the highest demand response. The off-peak periods suggest a different dynamic which could be investigated further.

Robustness Analysis

The robustness analysis shows that the overlap pattern holds primarily for on-peak demand. The results indicate that the structural dynamics during off-peak periods suggest a different dynamic, which could be investigated further with our proposed method. The authors conclude that the similarity between the CIV estimators derived from causal graphs and the lag-augmented local projection IV estimator calls for an extension of causal time graphs to estimate impulse response functions. The graphical approach is a powerful addition to the toolkit of every economist working with data structured as a time series.

Data Details

The data used spans from November 1st, 2017, to December 31st, 2020. Key variables include electricity load (MW), electricity price (€/MWh), and wind generation (MW) from ENTSO-E. Control series incorporated include solar PV power generation, surface temperature data for Germany from NASA’s MERRA-2, and various commodity prices. The analysis includes seasonal dummies and weather controls to increase precision.

Improvements for AI systems

Strengthen Instrument Validity Check for Time Series Data: The improved system will systematically test structural assumptions by comparing the overlap of several presumably valid estimators to reject incorrect structural assumptions, as the paper argues that If the researcher’s assumptions encoded in the structural model are correct, two valid estimators should yield comparable results (up to finite sample errors).
Identify Model Misspecification via Estimator Divergence: The system will utilize discrepancies between them [valid estimators] can inform researchers about the validity of the underlying models, allowing it to reject assumptions, such as rejecting Model II if the confidence intervals do not overlap when assuming a different structure.
Detect Hidden Confounding in Autocorrelated Systems: The system will use causal time graphs to determine consistency by checking d-separation criteria, specifically identifying when an estimator is invalid because a path is unblocked given S, which directly addresses the issue where naive IV estimators may be inconsistent in settings with autocorrelated time series.
Evaluate Dynamic Response Heterogeneity: The system will analyze the divergence between CIV estimates across different demand models (I, II, and III) to determine if the dynamics of electricity markets are more complex than suggested by the simple Models I, II, and III because, taken in isolation, none of the models is sufficient to explain the disagreement between the estimators.
Select Optimal Estimator Strategy: The system will dynamically choose between path-blocking approaches and nuisance effect estimation based on sample size; the nuisance IV estimator [estimator] outperforms them [path-blocking estimators], thereby providing a useful alternative for smaller sample sizes.

Abstract

Instrumental variable (IV) methods are widely used to estimate the price elasticity of demand from observational data. However, naive IV estimators may be inconsistent in settings with autocorrelated time series. We argue that causal graphs taking into account time structure can simplify IV identification and help select consistent estimators. More precisely, we propose to model equilibrium conditions using causal mechanism time graphs. These (acyclic) graphs allow us to maintain the assumption of a simultaneous determination of prices and quantities and come with an interpretable notion of intervention. They further allow us to exploit recent advances in graphical inference to derive valid IV estimators that are consistent under weak assumptions. Finally, we argue that observing significant differences between the estimates of presumably valid estimators can help to reject false model assumptions, thereby improving our understanding of underlying economic dynamics. We apply our approach to the German electricity market, estimating the price elasticity of demand from simulated and real-world data. The findings underscore the importance of accounting for structural autocorrelation in IV-based analysis.

Sources

Identifying Causal Effects using Instrumental Time Series: Nuisance IV and Correcting for the Past