Pith. sign in

REVIEW 5 major objections 8 minor 300 references

Causal Inference with Video Features as Treatments

T0 review · 5 major / 8 minor · reviewed 2026-07-08 · glm-5.2

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Video Features as Treatments: Causal Effects from Generative AI

desk verdict First methodology for causal inference with video features as treatments; Mario benchmark with known ground truth is independently valuable; central theorem proof omitted and estimation procedure has a gap between identification and estimation. read the letter →

arxiv 2607.06126 v1 pith:3N4ACMRH submitted 2026-07-07 stat.AP

Kentaro Nakamura , Adam Breuer , Michael H. Crespin , Bryce J. Dietrich , Kosuke Imai This is my paper
classification stat.AP
keywords causalinferencevideodatagenerativeAIdeconfoundermarginalstructuralmodelsdynamicstochasticinterventionsreal-timeresponsemeasurementcampaignadvertising
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces a method for causal inference where the treatment is a specific visual feature appearing at specific moments in a video, and the outcome is a real-time response trajectory from a viewer. The core problem is that video features are confounded by other latent, high-dimensional visual characteristics that also drive viewer responses. The authors solve this by reproducing each video with a deep generative model and extracting the model's internal representations as a learned, low-dimensional summary of video content. They show that adjusting for a deconfounder derived from these internal representations nonparametrically identifies the average potential-outcome trajectory under dynamic stochastic interventions on video features. They build a consistent and asymptotically normal estimator using a longitudinal neural network architecture, validate it on a benchmark dataset of 10,000 Super Mario Bros. levels where the ground-truth causal effect is known to be zero, and apply it to 849 ads from the 2020 U.S. presidential campaign, finding that increasing the probability of a candidate appearing over time leads to higher average viewer evaluations.

What carries the argument

Deep generative model internal representations; learned deconfounder via longitudinal neural network; dynamic stochastic interventions (incremental propensity score interventions adapted to use observed treatment probabilities); marginal structural models for longitudinal data; influence functions with Neyman orthogonality and cross-fitting; multiplier bootstrap for uniform confidence bands over the outcome trajectory

What would settle it

If, in a real-world video, a treatment feature structurally determines a confounding feature within the same segment (e.g., a candidate's appearance mechanically triggers specific background visuals), the deconfounder cannot separate them, and the causal estimate would be biased.

Watch

Extended reading notes

Core claim

The central object is the learned deconfounder: a low-dimensional latent representation extracted from the internal representations of a deep generative model used to reproduce video segments. The paper proves (Theorem 1) that adjusting for this deconfounder nonparametrically identifies the average counterfactual outcome trajectory under dynamic stochastic interventions on time-varying video features. The key insight is that because the entire video is observed, latent confounding features can be recovered from the generative model's internal representations, relaxing the standard sequential ignorability assumption that would otherwise require all confounders to be directly observed.

Load-bearing premise

The method requires that, within each video segment, the confounding features (like background imagery or tone) are not deterministically caused by the treatment feature (like a candidate's appearance). If the treatment mechanically produces the confounder, the identification result fails.

Editorial extensions

If this is right

  • Researchers can now ask which specific visual features (e.g., candidate appearance, emotional imagery, text overlays), appearing at which points in a video, causally influence audience responses over time, rather than only estimating effects of whole-video exposure.
  • The Mario benchmark dataset of 10,000 gameplay videos with known zero ground-truth effects provides a reusable testbed for evaluating future causal inference methods on dynamic, high-dimensional, confounded video data.
  • The framework extends naturally to other modalities (text, audio, images) where treatments and confounders are embedded in unstructured data, provided a generative model can reproduce the content and yield internal representations.
  • The finding that coarse hand-crafted adjustments (e.g., pipe counts) leave residual bias while the learned deconfounder recovers the null effect suggests that common practices in unstructured-data causal inference may systematically under-adjust for confounding.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the sequential separability assumption fails in practice (e.g., a candidate's appearance mechanically causes specific background imagery), the identification result breaks. A natural extension would be to develop sensitivity analyses quantifying how large violations of separability would need to be to overturn a given causal conclusion.
  • The framework currently operates on discrete video segments. A continuous-time extension could avoid the segment-length tradeoff the authors note between temporal resolution and computational cost, potentially via continuous-time neural processes or neural ODEs for the deconfounder and outcome model.
  • The reliance on a specific generative model (here, the NVIDIA Cosmos tokenizer) means the deconfounder quality depends on that model's reconstruction fidelity. As generative models improve, the method's power should improve with them, but a systematic study of how model choice affects identification and estimation would be valuable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 8 minor

Summary. This paper introduces a methodology for causal inference with video features as treatments, using deep generative models to extract low-dimensional internal representations that serve as learned deconfounders. The framework extends the GenAI-Powered Inference (GPI) methodology to the longitudinal setting, establishing nonparametric identification of average potential-outcome trajectories under dynamic stochastic interventions (Theorem 1) and deriving an influence-function-based estimator with cross-fitting (Theorem 2, Proposition 2). The authors validate their approach on a novel Super Mario Bros. benchmark with known zero ground-truth causal effect, showing that their method recovers the null while naive adjustments do not. They then apply the methodology to 2020 U.S. presidential campaign advertisements, finding that increasing candidate appearance probability raises viewer evaluations among co-partisans. The paper addresses an important and timely problem at the intersection of causal inference, machine learning, and computational social science.

Significance. The paper makes several valuable contributions. First, the Mario benchmark dataset with known ground-truth causal effects is a genuinely useful resource for the community and is of independent interest. Second, the extension of GPI to the dynamic/longitudinal video-as-treatment setting is a non-trivial methodological contribution. Third, the influence function derivation (Theorem 2) and the full estimation algorithm (Appendix S2.2) provide a complete inference pipeline. The empirical application to campaign advertisements is well-motivated and the real-time response data collection is carefully designed. The uniform confidence band construction via multiplier bootstrap is appropriate. The paper ships a falsifiable benchmark and a reproducible estimation procedure.

major comments (5)
  1. Section 3.5, Equation (5): The estimation objective is plain squared loss with no visible mechanism enforcing the mean-independence condition (Equation 3) that links identification to consistent estimation. The paper states that 'in cross-sectional settings, the deconfounder can be learned with a neural network architecture that encodes this mean-independence relationship' (Section 3.5), but does not describe how this encoding extends to the longitudinal setting. The architecture in Figure 4 shows a standard encoder-decoder with no adversarial term, constraint, or regularization that would prevent the learned f̂(R) from capturing aspects of R correlated with Y through pathways not mediated by U, or from implicitly encoding W through shared video structure. If the learned f̂ violates Equation (3), the identification formula in Theorem 1 does not apply to the estimator. This gap between an
  2. identification result (existence of f) and an estimation procedure (finding the right f) is load-bearing for the central claim of consistent estimation. The paper should either (a) specify the architectural or algorithmic mechanism that enforces mean-independence in the longitudinal setting, or (b) provide an argument for why the squared-loss solution satisfies Equation (3) under the stated assumptions, or (c) acknowledge this as a gap and discuss its implications for the empirical results.
  3. Theorem 1 (Section 3.4): The proof is omitted with the statement that it is 'similar to that of Theorem 1 of Nakamura and Imai (2026),' which is an unpublished working paper (arXiv:2605.07834). While the high-level argument is plausible — if U is a deterministic function of R (Assumptions 4, 6) and Y depends on X only through W and U (Assumption 5), then f = g_U ∘ h should satisfy mean-independence — the extension to the longitudinal setting with temporally aggregated outcomes involves subtleties that warrant at least a sketch. Specifically, the mean-independence condition in Equation (3) conditions on the full history of deconfounders {f(R_{is'})}_{s'=1}^s, and it is not immediately obvious why a single learned f should satisfy this for all s simultaneously. A proof sketch or at least a precise statement of how the longitudinal extension differs from the cross-sectional case would allow
  4. independent verification of this load-bearing result.
  5. Section 3.1, Assumption 5 (Sequential Separability): The second component requires that confounding features U are not deterministic functions of the treatment feature W within the same segment. The paper acknowledges this is 'not guaranteed by the study design and may therefore be restrictive.' In the campaign advertisement application (Section 5), this assumption requires that candidate appearance does not mechanically determine other visual features (e.g., background imagery, text overlays). This is a substantive restriction: campaigns likely design ads where candidate appearance co-occurs with specific visual contexts. The paper should discuss the plausibility of this assumption in the campaign ad setting specifically, and whether the empirical diagnostic mentioned (inspecting estimated propensity scores, Section 3.4) was actually applied and what it showed.
minor comments (8)
  1. Section 3.1, Assumption 5: The formal statement involves functions g' and g̃_U^(s') whose roles are somewhat hard to parse on first reading. A brief verbal gloss after the formal statement (in addition to the existing explanation) would improve readability.
  2. Section 4.2: The learning rate is reported as 3.912×10^{-4}. The apparent precision suggests hyperparameter optimization output; noting this explicitly would help readers understand the usage.
  3. Section 5: The learning rate 7.304×10^{-5} is similarly precise. Consistent labeling of these as Optuna-selected values would help.
  4. Figure 6: The y-axis label 'Estimated Difference in Jumps' could be more precise, e.g., 'Ψ̂(5.0) − Ψ̂(0.5): Difference in Expected Jumps.'
  5. Section 2.2: The survey dates (June 22 and June 26, 2026) are in the future relative to the manuscript date (July 2026). If these are correct, no change needed; if they are typos, please correct.
  6. Appendix S2.1, Theorem 2: The proof is also omitted ('follows directly from Theorem 2 of Nakamura and Imai (2026)'). While the influence function expression is given in full (Equations S1–S3), a brief derivation sketch would strengthen the paper's self-containedness.
  7. Section 3.3, Equation (1): The notation q_s(δ_s; w_{s-1}) uses a semicolon to separate δ_s from w_{s-1}, but the text refers to 'the observed treatment probability conditional on treatment history.' Clarifying that p_s(w_{s-1}) is estimated (not the propensity score) earlier in the paragraph would help readers unfamiliar with this distinction.
  8. Section 4.1: The Mario AI agent is described as 'Glenn-Hartmann' in one place and 'Glenn–Hartmann' (en-dash) in another. Consistent usage would help.

Simulated Author's Rebuttal

5 responses · 0 unresolved

We thank the referee for a careful and constructive report. The referee correctly identifies that the gap between the identification result (existence of a deconfounder f satisfying mean-independence) and the estimation procedure (finding such an f via squared loss) is the most important methodological issue in the paper. We address each major comment below and describe revisions we will make.

read point-by-point responses

  1. Referee: Section 3.5, Equation (5): The estimation objective is plain squared loss with no visible mechanism enforcing the mean-independence condition (Equation 3). The architecture in Figure 4 shows a standard encoder-decoder with no adversarial term, constraint, or regularization that would prevent the learned f-hat from capturing aspects of R correlated with Y through pathways not mediated by U, or from implicitly encoding W through shared video structure. If the learned f-hat violates Equation (3), the identification formula in Theorem 1 does not apply to the estimator.

    Authors: The referee is correct that the current manuscript does not adequately explain how the squared-loss estimation procedure connects to the mean-independence condition in Equation (3). This is a genuine gap between identification and estimation, and we agree it is load-bearing. We will address it through option (b): providing an argument for why the squared-loss solution satisfies Equation (3) under the stated assumptions, while also being explicit about the conditions under which the argument holds. The key observation is as follows. Under Assumptions 4 and 6, the confounding features U are deterministic functions of the internal representation R. The mean-independence condition in Equation (3) requires that f(R) be sufficiently informative that conditioning on it renders Y mean-independent of R given W and f(R). If f(R) recovers the confounding features U (or a sufficient statistic for them), then mean-independence follows from Assumption 5 (Sequential Separability), because Y depends on X only through W and U. The squared-loss objective in Equation (5) trains f and the outcome model mu_s jointly. The outcome model mu_s takes W and {f(R_{s'})} as inputs and predicts Y. If f is too low-dimensional to capture U, the squared loss will be high because mu_s cannot predict Y well without the confounders. If f is high-dimensional enough to capture U, the squared loss is minimized. The concern the referee raises is that f might capture aspects of R correlated with Y through pathways not mediated by U. However, under Assumption 5, there are no such pathways: Y depends on X only through W and U. So any information in R about Y is either through W (which is explicitly conditioned on in mu_s) or through U. The remaining concern is that f might implicitly encode W itself, violating a revision: no

  2. Referee: The paper should either (a) specify the architectural or algorithmic mechanism that enforces mean-independence in the longitudinal setting, or (b) provide an argument for why the squared-loss solution satisfies Equation (3) under the stated assumptions, or (c) acknowledge this as a gap and discuss its implications for the empirical results.

    Authors: We will pursue option (b) in the revision, providing the argument sketched above. However, we also acknowledge that the argument relies on the outcome model mu_s being sufficiently flexible to extract the confounding information from f(R), and on f being sufficiently expressive. These are empirical conditions that cannot be verified purely from the architecture. We will therefore also add a discussion of this as a potential limitation and note that the Mario benchmark, where ground truth is known, serves as an empirical check: if the squared-loss solution violated mean-independence in practice, we would expect the Mario results to show bias, which they do not. We will also add discussion of the architectural mechanism from the cross-sectional GPI framework (Imai and Nakamura, 2026) and explain how it extends to the longitudinal setting. Specifically, in the cross-sectional case, the deconfounder is learned with an architecture that prevents it from being a function of W by construction — the network takes R as input and produces f(R) before W is observed by the outcome model. In the longitudinal setting, the same principle applies segment-by-segment: f^{(s)}(R_{is'}, s') is computed from R_{is'} alone, before W_{is'} enters the outcome model. The segment indicator allows the deconfounder to vary across segments but does not introduce dependence on W. We will make this architectural detail explicit in the revised manuscript and add a paragraph to Section 3.5 explaining the connection between the squared-loss objective and the mean-independence condition. revision: partial

  3. Referee: Theorem 1 (Section 3.4): The proof is omitted with the statement that it is 'similar to that of Theorem 1 of Nakamura and Imai (2026),' which is an unpublished working paper. The extension to the longitudinal setting with temporally aggregated outcomes involves subtleties that warrant at least a sketch. Specifically, the mean-independence condition in Equation (3) conditions on the full history of deconfounders, and it is not immediately obvious why a single learned f should satisfy this for all s simultaneously.

    Authors: The referee is right that omitting the proof is insufficient, particularly given that the cited paper is unpublished. We will add a proof sketch in the appendix. The key steps are: (1) Under Assumptions 4 and 6, U^{(s)}_{is} is a deterministic function of R_{is}, so there exists a function f^{(s)}_s such that f^{(s)}_s(R_{is}) = U^{(s)}_{is} (or a sufficient statistic thereof). (2) Under Assumption 5, Y_{is} depends on X_{is} only through W_{is} and U^{(s)}_{is}, so conditioning on W_{is} and the full history {f^{(s)}_{s'}(R_{is'})}_{s'=1}^s renders Y_{is} mean-independent of R_{is} given these variables. This is because R_{is} contains no information about Y_{is} beyond what is in W_{is} and U^{(s)}_{is}, and f captures U. (3) The reason a single f (parameterized with a segment indicator) can satisfy the condition for all s is that f^{(s)} is defined separately for each target outcome s, as indicated by the superscript. The network f^{(s)}(R_{is'}, s') takes the segment index as an input, so it can represent different mappings for different s'. For a fixed target outcome s, the deconfounder history {f^{(s)}(R_{is'}, s')}_{s'=1}^s is computed using the same network, and the mean-independence condition holds for that specific s. The temporally aggregated estimand then marginalizes over s. (4) The identification formula follows by applying the backdoor adjustment with the learned deconfounder and then marginalizing over the intervention distribution. We will write this up formally in Appendix S3. revision: yes

  4. Referee: Section 3.1, Assumption 5 (Sequential Separability): The second component requires that confounding features U are not deterministic functions of the treatment feature W within the same segment. In the campaign advertisement application, this requires that candidate appearance does not mechanically determine other visual features. Campaigns likely design ads where candidate appearance co-occurs with specific visual contexts. The paper should discuss the plausibility of this assumption in the campaign ad setting specifically, and whether the empirical diagnostic mentioned (inspecting estimated propensity scores, Section 3.4) was actually applied and what it showed.

    Authors: The referee raises a valid concern. Assumption 5's second component is indeed potentially restrictive in the campaign advertising setting. We will add a substantive discussion of its plausibility. The key consideration is that the assumption does not require candidate appearance to be uncorrelated with other visual features — it requires that candidate appearance does not deterministically determine them. In practice, candidate appearance co-occurs with certain visual contexts (e.g., positive imagery, patriotic backgrounds), but the co-occurrence is stochastic rather than deterministic: there exist segments where the candidate appears without those contexts and segments with those contexts but without the candidate. The empirical diagnostic we mentioned in Section 3.4 — inspecting the distribution of estimated propensity scores — was applied in both the Mario and campaign ad analyses. In the Mario benchmark, the diagnostic confirms that the estimated propensity scores are bounded away from 0 and 1, consistent with Proposition 1. In the campaign ad application, the estimated propensity scores are also bounded away from 0 and 1 across all segments, which is consistent with the support condition implied by sequential separability. We will report these diagnostic results explicitly in the revised manuscript. We acknowledge that the diagnostic is not a definitive test of Assumption 5 itself but rather of its observable implication (Proposition 1). We will also discuss the specific concern about campaigns designing ads where candidate appearance co-occurs with specific visual contexts, and note that while this is plausible, the stochastic nature of the co-occurrence is what the assumption requires. If certain visual features were perfectly determined by candidate appearance, revision: no

  5. Referee: The paper should discuss the plausibility of this assumption in the campaign ad setting specifically, and whether the empirical diagnostic mentioned (inspecting estimated propensity scores, Section 3.4) was actually applied and what it showed.

    Authors: As noted above, we will add this discussion and report the diagnostic results. We will also acknowledge the limitation honestly: the assumption is not guaranteed by the study design, and there may be specific visual features that are mechanically tied to candidate appearance in ways that violate the assumption. The Mario benchmark provides some reassurance about the methodology's robustness, but the campaign ad setting involves richer confounding structures. We will add this as a limitation in the conclusion. revision: yes

Circularity Check

2 steps flagged · score 2.0 of 10

Load-bearing self-citation to unpublished Nakamura & Imai (2026) for both main theorems, but paper retains independent content through longitudinal extension, included proofs, and ground-truth benchmark validation.

  1. self citation load bearing [Section 3.4, Theorem 1 (Identification)]
    "This result is a natural extension of Theorem 1 of Nakamura and Imai (2026), which identifies the causal effect on a single outcome at a given segment under the same dynamic intervention. [...] The proof is omitted as it is similar to that of Theorem 1 of Nakamura and Imai (2026)."

    The central identification result (Theorem 1) has its proof omitted and defers to Nakamura and Imai (2026, arXiv:2605.07834), an unpublished arXiv preprint co-authored by two of the current paper's five authors. The existence of the deconfounder satisfying the mean-independence condition (Eq. 3) and the identification formula for the aggregated outcome both rest on this unverified self-citation. However, the paper is extending from single-outcome to temporally aggregated outcomes, and Proposition 2's proof (included) shows the aggregation step is a standard product-rule application. The self-citation is load-bearing for the base identification argument but the longitudinal extension has independent content.

  2. self citation load bearing [Appendix S2.1, Theorem 2 (Influence Function)]
    "The proof is omitted since it follows directly from Theorem 2 of Nakamura and Imai (2026)."

    The influence function for the single-outcome estimand (Theorem 2) is stated as following 'directly' from Theorem 2 of the same unpublished self-citation (Nakamura and Imai, 2026). The phrase 'follows directly' suggests the influence function is essentially imported rather than independently derived. However, Proposition 2 (proof included) derives the aggregated influence function from the single-outcome one via a standard product rule, which is self-contained. The concern is limited to the base influence function, not the paper's novel aggregation contribution.

full rationale

The paper's two main theorems (Theorem 1: identification; Theorem 2: influence function) both have omitted proofs deferring to Nakamura and Imai (2026, arXiv:2605.07834), an unpublished preprint by two of the current authors. This is a load-bearing self-citation: without accessing that reference, the core identification existence argument and the single-outcome influence function cannot be independently verified. However, the paper is not circular in the strict sense. It provides self-contained proofs for Proposition 1 (support independence under sequential separability) and Proposition 2 (aggregated influence function via product rule). The longitudinal extension from single-outcome to temporally aggregated outcomes is a genuine methodological step. The Mario benchmark (Section 4) provides independent empirical validation with ground-truth causal effects known by construction (zero effect of Princess Peach), where the proposed method recovers the null while naive approaches fail. The dynamic stochastic intervention (Eq. 1) traces to Kennedy (2019), an external published reference. The estimation gap (Eq. 5 not visibly enforcing mean-independence Eq. 3) is a correctness concern, not circularity. Overall, the self-citation is load-bearing but the paper retains substantial independent content, warranting a score of 2.

Assumptions & free parameters 4 free parameters · 7 assumptions · 1 invented entities

The axiom ledger reveals that the framework relies on seven assumptions, of which Sequential Separability (Assumption 5) is the most restrictive and ad hoc. The free parameters are primarily hyperparameters of the neural network architecture, selected via Optuna, and the intervention parameter delta chosen by the researcher. The invented entity (deconfounder function) has partial independent evidence from the synthetic benchmark.

free parameters (4)
  • delta (incremental parameter) = 0.5 and 5.0
    User-specified intervention parameters determining the shift in treatment probability. Not fitted to data but chosen by researchers.
  • Neural network architecture parameters = channel sizes (8,16,32), (16,32,64); hidden layers (256,32), (64), (256,128)
    Architecture choices for the deconfounder and outcome networks, selected via Optuna hyperparameter tuning.
  • Learning rate = 3.912e-4 (Mario), 7.304e-5 (ads)
    Selected via Optuna with 100 trials.
  • alpha, beta, epsilon (Mario DGP) = alpha=-1.5, beta=1, epsilon=0.05
    Parameters controlling treatment assignment confounding in the synthetic benchmark. Chosen by design, not fitted to real data.
assumptions (7)
  • domain assumption Consistency (Assumption 1)
    Observed outcome equals potential outcome under realized video. Rules out spillover between respondents. Standard in causal inference.
  • domain assumption Random Assignment of Video (Assumption 2)
    Videos are randomly assigned to respondents. Guaranteed by study design in the empirical application.
  • domain assumption Treatment feature exists as deterministic function (Assumption 3)
    There exists a mapping g_W from video segment to binary treatment. Invoked in Section 3.1.
  • domain assumption Confounding features are low-dimensional (Assumption 4)
    Confounders lie in a much lower-dimensional space than the video itself. Required for tractability of the deconfounder.
  • ad hoc to paper Sequential Separability (Assumption 5)
    Confounding features are not deterministic functions of the treatment feature within a segment. Section 3.1 acknowledges this is 'not guaranteed by the study design and may therefore be restrictive.' This is the key load-bearing assumption.
  • domain assumption Factorized Deterministic Decoding (Assumption 6)
    The generative model's output layer is deterministic given the latent representation. Required to ensure features are deterministic functions of R_is. Section 3.2.
  • domain assumption Bounded relative overlap (Assumption 7)
    Propensity score is bounded relative to observed treatment probability. Required for identification under stochastic interventions.
invented entities (1)
  • Deconfounder function f(s) independent evidence
    purpose: A learned low-dimensional representation of the generative model's internal representation that satisfies the mean-independence condition (Equation 3) and captures confounding information
    The paper provides a falsifiable validation: on the Mario benchmark, the deconfounder recovers the known null effect. In the real-world application, the face validity of results (partisan alignment effects) provides partial evidence, though ground truth is absent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Inference with Video Features as Treatments." pith.science (2026). https://pith.science/paper/3N4ACMRH

@misc{pith2026260706126,
  author       = {Pith},
  title        = {Pith review of: Causal Inference with Video Features as Treatments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3N4ACMRH}},
  note         = {Machine review of arXiv:2607.06126}
}
abstract

We develop the first statistical methodology for causal inference with video features as treatments. Video is the most engaging content modality on the internet. A central causal question is how audience reactions change in response to treatment features that unfold over the course of a video. Unfortunately, standard causal inference methods are not applicable because confounding features are latent, high-dimensional, and dynamically related to both the treatment sequence and the outcome trajectory. To address these challenges, we first reproduce each video using a deep generative model and leverage the model's internal representations as learned, low-dimensional summaries of video content for causal estimation. We then establish that the average potential-outcome trajectory under dynamic stochastic interventions is nonparametrically identified. Lastly, we propose a consistent and asymptotically normal estimator based on a longitudinal neural network architecture. We empirically validate our approach by constructing a new causal inference benchmark consisting of $10{,}000$ Super Mario Bros. levels played by fixed Mario AI agents, where ground-truth causal effects are known by construction. Finally, we apply our method to television advertisements from the 2020 U.S. presidential campaign and find that increasing the probability of a candidate appearing over time leads to higher average viewer evaluations. With the proposed methodology, researchers can ask which visual features, appearing at which points in a video, influence audience responses, while benchmarking new methods against datasets with known ground-truth causal effects.

Figures

Figures reproduced from arXiv: 2607.06126 by the authors.

Figure 1
Figure 1. Candidate Identification and Rating Tasks. Panel [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Schematic illustration of our experimental setup. We segment each video into short clips. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the assumed data generating process when we are interested in the outcome [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Diagram Illustrating the Proposed Neural Network Architecture for the case with four [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: A Screenshot from the Mario environment. The treatment is the appearance of Princess [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Estimated difference in the average number of jumps under two stochastic interventions: [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: The original video frame (left) and the reconstructed video frame using the NVIDIA [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: Estimated average potential outcome trajectory of candidate dial ratings for each re [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

300 extracted references · 300 canonical work pages

  1. [1]

    The Llama 3 Herd of Models

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    work page Pith review arXiv

  2. [2]

    IEEE Transactions on Computational Intelligence and AI in Games , volume=

    The mario ai benchmark and competitions , author=. IEEE Transactions on Computational Intelligence and AI in Games , volume=. 2012 , publisher=

    work page 2012

  3. [3]

    Human communication research , volume=

    Reliability in content analysis: Some common misconceptions and recommendations , author=. Human communication research , volume=. 2004 , publisher=

    work page 2004

  4. [4]

    and Ridout, Travis N

    Fowler, Erika Franklin and Franz, Michael M. and Ridout, Travis N. and Baum, Laura and Bogucki, Colleen and Floyd, Breeze , title =. 2025 , howpublished =

    work page 2025

  5. [5]

    American Political Science Review , volume=

    How experiments help campaigns persuade voters: Evidence from a large archive of campaigns’ own experiments , author=. American Political Science Review , volume=. 2024 , publisher=

    work page 2024

  6. [6]

    Nature Human Behaviour , volume=

    A 2 million-person, campaign-wide field experiment shows how digital advertising affects voter turnout , author=. Nature Human Behaviour , volume=. 2023 , publisher=

    work page 2023

  7. [7]

    2008 , publisher=

    In defense of negativity: Attack ads in presidential campaigns , author=. 2008 , publisher=

    work page 2008

  8. [8]

    The Journal of Politics , volume=

    The effects of negative political campaigns: A meta-analytic reassessment , author=. The Journal of Politics , volume=. 2007 , publisher=

    work page 2007

Show all 300 references
  1. [9]

    1995 , publisher =

    Going Negative: How Political Advertisements Shrink and Polarize the Electorate , author =. 1995 , publisher =

    1995

  2. [10]

    American political science review , volume=

    Does attack advertising demobilize the electorate? , author=. American political science review , volume=. 1994 , publisher=

    1994

  3. [11]

    Proceedings of the National Academy of Sciences , volume=

    Quantifying the potential persuasive returns to political microtargeting , author=. Proceedings of the National Academy of Sciences , volume=. 2023 , publisher=

    2023

  4. [12]

    Journal of Quantitative Description: Digital Media , volume=

    Digital advertising in us federal elections, 2004-2020 , author=. Journal of Quantitative Description: Digital Media , volume=

    2004

  5. [13]

    The Forum , volume=

    Understanding the Message (s): Spending and Content of Political Advertising on Television in 2024 , author=. The Forum , volume=. 2025 , organization=

    2024

  6. [14]

    The Forum , volume=

    Election Advertising on Meta, Google, and Snapchat in 2024 , author=. The Forum , volume=. 2025 , organization=

    2024

  7. [15]

    The International Journal of Press/Politics , volume=

    The intersection of candidate gender and ethnicity: How voters respond to campaign messages from Latinas , author=. The International Journal of Press/Politics , volume=. 2023 , publisher=

    2023

  8. [16]

    The Journal of Politics , volume=

    Gendered targeting: Do parties tailor their campaign ads to women? , author=. The Journal of Politics , volume=. 2026 , publisher=

    2026

  9. [17]

    Electoral Studies , volume=

    Candidate appearance in campaign advertisements , author=. Electoral Studies , volume=. 2021 , publisher=

    2021

  10. [18]

    Political Research Quarterly , volume=

    Harnessing the Power of Emotions: How Women of Color Use Emotional Appeals in Their Campaign Messages , author=. Political Research Quarterly , volume=. 2025 , publisher=

    2025

  11. [19]

    Political Behavior , volume=

    Negative advertising and the dynamics of candidate support , author=. Political Behavior , volume=. 2016 , publisher=

    2016

  12. [20]

    American Political Science Review , volume=

    How large and long-lasting are the persuasive effects of televised campaign ads? Results from a randomized field experiment , author=. American Political Science Review , volume=. 2011 , publisher=

    2011

  13. [21]

    arXiv preprint arXiv:2605.07834 , year=

    GenAI Powered Dynamic Causal Inference with Unstructured Data , author=. arXiv preprint arXiv:2605.07834 , year=

    arXiv

  14. [22]

    Humanities and Social Sciences Communications , volume=

    How to convince in a televised debate: the application of machine learning to analyze why viewers changed their winner perception during the 2021 German chancellor discussion , author=. Humanities and Social Sciences Communications , volume=. 2023 , publisher=

    2021

  15. [23]

    International Journal of Public Opinion Research , volume=

    Reliability and validity of real-time response measurement: A comparison of two studies of a televised debate in Germany , author=. International Journal of Public Opinion Research , volume=. 2007 , publisher=

    2007

  16. [24]

    Proceedings of the 2023 CHI conference on human factors in computing systems , pages=

    Short-form videos degrade our capacity to retain intentions: Effect of context switching on prospective memory , author=. Proceedings of the 2023 CHI conference on human factors in computing systems , pages=

    2023

  17. [25]

    Management Science , volume=

    Choosing to discover the unknown: The effects of choice on user attention to online video advertising , author=. Management Science , volume=. 2024 , publisher=

    2024

  18. [26]

    Journal of Marketing Research , volume=

    Neural signals of video advertisement liking: insights into psychological processes and their temporal dynamics , author=. Journal of Marketing Research , volume=. 2024 , publisher=

    2024

  19. [27]

    Statistics in medicine , volume=

    Estimating the causal effect of zidovudine on CD4 count with a marginal structural model for repeated measures , author=. Statistics in medicine , volume=. 2002 , publisher=

    2002

  20. [28]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  21. [29]

    Advances in neural information processing systems , volume=

    Distributed representations of words and phrases and their compositionality , author=. Advances in neural information processing systems , volume=

  22. [30]

    Journal of marketing research , volume=

    Emotion-induced engagement in internet video advertisements , author=. Journal of marketing research , volume=. 2012 , publisher=

    2012

  23. [31]

    International Journal of Public Opinion Research , volume=

    Measuring real-time response in real-life settings , author=. International Journal of Public Opinion Research , volume=. 2020 , publisher=

    2020

  24. [32]

    The Econometrics Journal , volume=

    Debiased machine learning of conditional average treatment effects and other causal functions , author=. The Econometrics Journal , volume=. 2021 , publisher=

    2021

  25. [33]

    International conference on machine learning , pages=

    Robust speech recognition via large-scale weak supervision , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  26. [34]

    Public opinion quarterly , volume=

    Breaking out of the lab: Measuring real-time responses to televised political content in real-world settings , author=. Public opinion quarterly , volume=. 2016 , publisher=

    2016

  27. [35]

    arXiv preprint arXiv:2201.10005 , year=

    Text and code embeddings by contrastive pre-training , author=. arXiv preprint arXiv:2201.10005 , year=

    arXiv

  28. [36]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    Fine-tuning llama for multi-stage text retrieval , author=. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  29. [37]

    PLoS One , volume=

    Continuous ratings of movie watching reveal idiosyncratic dynamics of aesthetic enjoyment , author=. PLoS One , volume=. 2019 , publisher=

    2019

  30. [38]

    Advances in Neural Information Processing Systems , volume=

    Using imperfect surrogates for downstream inference: Design-based supervised learning for social science applications of large language models , author=. Advances in Neural Information Processing Systems , volume=

  31. [39]

    Science , volume=

    Prediction-powered inference , author=. Science , volume=. 2023 , publisher=

    2023

  32. [40]

    Mathematical modelling , volume=

    A new approach to causal inference in mortality studies with a sustained exposure period—application to control of the healthy worker survivor effect , author=. Mathematical modelling , volume=. 1986 , publisher=

    1986

  33. [41]

    arXiv preprint arXiv:2507.03897 , year=

    Genai-powered inference , author=. arXiv preprint arXiv:2507.03897 , year=

    arXiv

  34. [42]

    Proceedings of the National Academy of Sciences , volume=

    The (minimal) persuasive advantage of political video over text , author=. Proceedings of the National Academy of Sciences , volume=. 2021 , publisher=

    2021

  35. [43]

    British Journal of Political Science , volume=

    Public opinion and emphatic legislative speech: Evidence from an automated video analysis , author=. British Journal of Political Science , volume=. 2025 , publisher=

    2025

  36. [44]

    outside lobbying

    “outside lobbying” over the airwaves: A randomized field experiment on televised issue ads , author=. American Political Science Review , volume=. 2022 , publisher=

    2022

  37. [45]

    Science advances , volume=

    Psychological inoculation improves resilience against misinformation on social media , author=. Science advances , volume=. 2022 , publisher=

    2022

  38. [46]

    Science advances , volume=

    The small effects of political advertising are small regardless of context, message, sender, or receiver: Evidence from 59 real-time randomized experiments , author=. Science advances , volume=. 2020 , publisher=

    2020

  39. [47]

    American Political Science Review , volume=

    Gender, candidate emotional expression, and voter reactions during televised debates , author=. American Political Science Review , volume=. 2021 , publisher=

    2021

  40. [48]

    Journal of the American Statistical Association , year =

    Causal Inference with Generative Artificial Intelligence: Application to Texts as Treatments , author =. Journal of the American Statistical Association , year =. 2410.00903 , archivePrefix =

    arXiv

  41. [49]

    arXiv preprint arXiv:2501.03575 , year=

    Cosmos world foundation model platform for physical ai , author=. arXiv preprint arXiv:2501.03575 , year=

    arXiv

  42. [50]

    American Political Science Review , volume=

    The effect of television advertising in United States elections , author=. American Political Science Review , volume=. 2022 , publisher=

    2022

  43. [51]

    American Political Science Review , volume=

    The effect of TV ads and candidate appearances on statewide presidential votes, 1988--96 , author=. American Political Science Review , volume=. 1999 , publisher=

    1988

  44. [52]

    Journal of the Royal Statistical Society, Series B (Statistical Methodology) , year =

    Papadogeorgou, Georgia and Imai, Kosuke and Lyall, Jason and Li, Fan , title =. Journal of the Royal Statistical Society, Series B (Statistical Methodology) , year =

  45. [53]

    Journal of the American Statistical Association , year =

    Imai, Kosuke and Jiang, Zhichao , title =. Journal of the American Statistical Association , year =

  46. [54]

    2018 , publisher=

    Improving language understanding by generative pre-training , author=. 2018 , publisher=

    2018

  47. [55]

    arXiv preprint arXiv:1705.08582 , year=

    On the multiply robust estimation of the mean of the g-functional , author=. arXiv preprint arXiv:1705.08582 , year=

    arXiv

  48. [56]

    Biometrics , volume=

    Population intervention causal effects based on stochastic interventions , author=. Biometrics , volume=. 2012 , publisher=

    2012

  49. [57]

    Working Paper , year=

    Replication for Language Models Problems, Principles, and Best Practice for Political Science , author=. Working Paper , year=

  50. [58]

    2024 , eprint=

    Gemma: Open Models Based on Gemini Research and Technology , author=. 2024 , eprint=

    2024

  51. [59]

    Journal of Business & Economic Statistics , number=

    Double debiased machine learning nonparametric inference with continuous treatments , author=. Journal of Business & Economic Statistics , number=. 2025 , publisher=

    2025

  52. [60]

    American Political Science Review , year =

    Egami, Naoki and Hartman, Erin , title =. American Political Science Review , year =

  53. [61]

    2024 , url =

    Llama 3 Model Card , author=. 2024 , url =

    2024

  54. [62]

    Large language models in medicine , url =

    Thirunavukarasu, Arun James and Ting, Darren Shu Jeng and Elangovan, Kabilan and Gutierrez, Laura and Tan, Ting Fang and Ting, Daniel Shu Wei , date =. Large language models in medicine , url =. Nature Medicine , number =. 2023 , bdsk-url-1 =. doi:10.1038/s41591-023-02448-8 , id =

    2023 doi

  55. [63]

    ChatGPT for good? On opportunities and challenges of large language models for education , journal =

    Enkelejda Kasneci and Kathrin Sessler and Stefan Küchemann and Maria Bannert and Daryna Dementieva and Frank Fischer and Urs Gasser and Georg Groh and Stephan Günnemann and Eyke Hüllermeier and Stephan Krusche and Gitta Kutyniok and Tilman Michaeli and Claudia Nerdel and Jürge...

    2023 doi

  56. [64]

    Liu , title =

    Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu , title =. Journal of Machine Learning Research , year =

  57. [65]

    Dwivedi and Thomas H

    Nir Kshetri and Yogesh K. Dwivedi and Thomas H. Davenport and Niki Panteli , keywords =. Generative artificial intelligence in marketing: Applications, opportunities, challenges, and research agenda , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.ijinfomgt.2023.10271...

    2024 doi

  58. [66]

    Political Analysis , author=

    Multiple Hypothesis Testing in Conjoint Analysis , volume=. Political Analysis , author=. 2023 , pages=. doi:10.1017/pan.2022.30 , number=

    2023 doi

  59. [67]

    Journal of Economic Literature , year =

    Abadie, Alberto , title =. Journal of Economic Literature , year =

  60. [68]

    2017 , OPTkey =

    An Evalution of 2016 Election Polls in the. 2017 , OPTkey =

    2016

  61. [69]

    2017 , OPTkey =

    User Guide and Codebook for the. 2017 , OPTkey =

    2017

  62. [70]

    2017 , OPTkey =

    Ansolabehere, Stephen and Schaffner, Brian and Luks, Sam , title =. 2017 , OPTkey =

    2017

  63. [71]

    R: A Language and Environment for Statistical Computing , Url =

  64. [72]

    2004 , OPTkey =

    The Health Consequences of Smoking: A Report of the Surgeon General , institution =. 2004 , OPTkey =

    2004

  65. [73]

    Abadie, Alberto , Journal =

  66. [74]

    Abadie, Alberto and Angrist, Joshua and Imbens, Guido , Journal =

  67. [75]

    Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of

    Abadie, Alberto and Diamond, Alexis and Hainmueller, Jens , Journal =. Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of

  68. [76]

    American Economic Review , year =

    Abadie, Alberto and Gardeazabal, Javier , title =. American Economic Review , year =

  69. [77]

    , Journal =

    Abadie, Alberto and Imbens, Guido W. , Journal =

  70. [78]

    , Journal =

    Abadie, Alberto and Imbens, Guido W. , Journal =. On the Failure of the Bootstrap for Matching Estimators , Volume =

  71. [79]

    , Journal =

    Abadie, Alberto and Imbens, Guido W. , Journal =. Bias-Corrected Matching Estimators for Average Treatment Effects , Volume =

  72. [80]

    , Journal =

    Abadie, Alberto and Imbens, Guido W. , Journal =. A Martingale Representation for Matching Estimators , Year =

  73. [81]

    , title =

    Abramowitz, Alan I. , title =. Journal of Politics , year =

  74. [82]

    2018 , OPTkey =

    Abraham, Sarah and Sun, Liyang , title =. 2018 , OPTkey =

    2018

  75. [83]

    , Journal =

    Acemoglu, Daron and Johnson, Simon and Robinson, James A. , Journal =

  76. [84]

    , Journal =

    Achen, Christopher H. , Journal =

  77. [85]

    , Publisher =

    Achen, Christopher H. , Publisher =

  78. [86]

    and Bartels, Larry , Journal =

    Achen, Christopher H. and Bartels, Larry , Journal =

  79. [87]

    2010 , OPTkey =

    Achen, Christopher and Blais, Andre , title =. 2010 , OPTkey =

    2010

  80. [88]

    and Shively, W

    Achen, Christopher H. and Shively, W. Phillips , Publisher =

  81. [89]

    and Smith, Dennis J

    Adams, Williams C. and Smith, Dennis J. , Journal =

  82. [90]

    Adolph, Christopher and King, Gary , Journal =

  83. [91]

    and Shotts, Kenneth W

    Adolph, Christopher and King, Gary and Herron, Michael C. and Shotts, Kenneth W. , Journal =

  84. [92]

    Statistical Methods for the Social Sciences , publisher =

    Agresti, Alan and Finlay, Barbara , ALTeditor =. Statistical Methods for the Social Sciences , publisher =. 2008 , OPTkey =

    2008

  85. [93]

    , title =

    Ahlquist, John S. , title =. Political Analysis , year =

  86. [94]

    2017 , note =

    Ahlquist, John , title =. 2017 , note =

    2017

  87. [95]

    2013 , OPTkey =

    Ahlquist, John and Mayer, Kenneth and Jackman, Simon , title =. 2013 , OPTkey =

    2013

  88. [96]

    and Mayer, Kenneth R

    Ahlquist, John S. and Mayer, Kenneth R. and Jackman, Simon , title =. Election Law Journal , year =

  89. [97]

    , Journal =

    Aigner, Dennis J. , Journal =

  90. [98]

    Mixed membership stochastic blockmodels , Volume =

    Airoldi, Edoardo M and Blei, David M and Fienberg, Stephen E and Xing, Eric P , Journal =. Mixed membership stochastic blockmodels , Volume =

  91. [99]

    , Publisher =

    Aitchison, J. , Publisher =

  92. [100]

    , Journal =

    Aitchison, J. , Journal =

  93. [101]

    and Anderson, D

    Aitkin, M. and Anderson, D. and Hinde, J. , Journal =

  94. [102]

    and Rubin, Donald B

    Aitkin, M. and Rubin, Donald B. , Journal =

  95. [103]

    Akaike, Hirotugu , Chapter =

  96. [104]

    Akaike, Hirotugu , Journal =

  97. [105]

    Akaike, Hirotugu , Journal =. A

  98. [106]

    , Journal =

    Albert, Jeffrey M. , Journal =

  99. [107]

    , Journal =

    Albert, James H. , Journal =. Bayesian Estimation of Normal Ogive Item Response Curves Using

  100. [108]

    and Chib, Siddhartha , Journal =

    Albert, James H. and Chib, Siddhartha , Journal =

  101. [109]

    and Chib, Siddhartha , Journal =

    Albert, James H. and Chib, Siddhartha , Journal =. Bayes Inference Via

  102. [110]

    and Nelson, Suchitra , Journal =

    Albert, Jeffrey M. and Nelson, Suchitra , Journal =

  103. [111]

    2019 , OPTkey =

    Albright, Alex , title =. 2019 , OPTkey =

    2019

  104. [112]

    , Journal =

    Aldrich, John H. , Journal =

  105. [113]

    Alesina, Alberto and Rosenthal, Howard , Publisher =

  106. [114]

    2014 , publisher=

    Our Declaration: A Reading of the Declaration of Independence in Defense of Equality , author=. 2014 , publisher=

    2014

  107. [115]

    Difference without Domination: Pursuing Justice in Diverse Democracies , year =

  108. [116]

    , Journal =

    Allison, Paul D. , Journal =

  109. [117]

    The Annals of Statistics , volume=

    Identifiability of parameters in latent structure models with many observed variables , author=. The Annals of Statistics , volume=

  110. [118]

    and King, Gary , Journal =

    Alt, James E. and King, Gary , Journal =

  111. [119]

    Rutgers Computer & Technology Law Journal , year =

    Altman, Micah , title =. Rutgers Computer & Technology Law Journal , year =

  112. [120]

    From Crayons to Computers: The Evolution of Computer Use in Redistricting , journal =

    Altman, Micah and. From Crayons to Computers: The Evolution of Computer Use in Redistricting , journal =. 2005 , OPTkey =

    2005

  113. [121]

    Journal of Statistical Software , year =

    Altman, Micah and. Journal of Statistical Software , year =

  114. [122]

    On Rates of Convergence of Scholastic Relaxation for

    Amit, Yali , Journal =. On Rates of Convergence of Scholastic Relaxation for

  115. [123]

    Anderson, T. W. , Publisher =

  116. [124]

    , ALTeditor =

    Anderson, Margo and Fienberg, Stephen E. , ALTeditor =. Who Counts?: The Politics of Census-Taking in Contemporary America , publisher =. 1999 , OPTkey =

    1999

  117. [125]

    Anderson, T. W. and Hsiao, Cheng , title =. Journal of the American Statistical Association , year =

  118. [126]

    Consistent moment selection procedures for generalized method of moments estimation , Volume =

    Andrews, Donald , Journal =. Consistent moment selection procedures for generalized method of moments estimation , Volume =

  119. [127]

    2020 , OPTkey =

    Andrews, Isaiah and Kitagawa, Toru and McCloskey, Adam , title =. 2020 , OPTkey =

    2020

  120. [128]

    Andrews, Donald W. K. and Soares, Gustavo , Journal =

  121. [129]

    and Di Maro, V

    Angelucci, M. and Di Maro, V. , title =. Journal of Development Effectiveness , year =

  122. [130]

    , Journal =

    Angrist, Joshua D. , Journal =

  123. [131]

    , Journal =

    Angrist, Joshua D. , Journal =. Lifetime Earnings and the

  124. [132]

    , Journal =

    Angrist, Joshua D. , Journal =. The draft lottery and voluntary enlistment in the

  125. [133]

    Angrist, Joshua and Fernandez-. Extrapo

  126. [134]

    and Imbens, Guido W

    Angrist, Joshua D. and Imbens, Guido W. , Journal =

  127. [135]

    Angrist, J. D. and Imbens, G. W. and Krueger, A. B. , Journal =

  128. [136]

    and Imbens, Guido W

    Angrist, Joshua D. and Imbens, Guido W. and Rubin, Donald B. , Journal =

  129. [137]

    and Krueger, Alan B

    Angrist, Joshua D. and Krueger, Alan B. , Journal =

  130. [138]

    and Krueger, Alan B

    Angrist, Joshua D. and Krueger, Alan B. , Journal =. The effect of age at school entry on educational attainment:

  131. [139]

    and Lavy, V

    Angrist, J. and Lavy, V. , Institution =. The Effect of High School Matriculation Awards: Evidence from Randomized Trials , Type =

  132. [140]

    and Lavy, Victor , Journal =

    Angrist, Joshua D. and Lavy, Victor , Journal =. Using

  133. [141]

    and Pischke, J\"

    Angrist, Joshua D. and Pischke, J\"

  134. [142]

    2016 , OPTnote =

    Angwin, Julia and Larson, Jeff and Mattu, Surya and Kirchner, Lauren , title =. 2016 , OPTnote =

    2016

  135. [143]

    Anselin, Luc , Journal =

  136. [144]

    Tam , Journal =

    Anselin, Luc and Cho, Wendy K. Tam , Journal =

  137. [145]

    Jr , Journal =

    Ansolabehere, Stephen and Hansen, John Mark and Hirano, Shigeo and Snyder, James M. Jr , Journal =. The incumbency advantage in

  138. [146]

    Political Analysis , year =

    Ansolabehere, Stephen and Hersh, Eitan , title =. Political Analysis , year =

  139. [147]

    Ansolabehere, Stephen and Hirano, Shigeo and Snyder, Jim and Ueda, Michiko , Journal =

  140. [148]

    Political Communication , year =

    Ansolabehere, Stephen and Iyengar, Shanto , title =. Political Communication , year =

  141. [149]

    Ansolabehere, Stephen and Iyengar, Shanto , Journal =

  142. [150]

    Ansolabehere, Stephen and Iyengar, Shanto , Publisher =

  143. [151]

    and Snyder, James M

    Ansolabehere, Stephen and Snowberg, Erik C. and Snyder, James M. , Journal =. Television and the Incumbency Advantage in

  144. [152]

    Why Is There so Little Money in

    Ansolabehere, Stephen, and Snyder, James and. Why Is There so Little Money in. Journal of Economic Perspectives , Pages =

  145. [153]

    and Stewart, Charles , Journal =

    Ansolabehere, Stephen and Snyder, James M. and Stewart, Charles , Journal =

  146. [154]

    , Journal =

    Antoniak, Charles E. , Journal =. Mixtures of

  147. [155]

    Arceneaux, Kevin , Journal =

  148. [156]

    Computing robust standard errors for within-groups estimators , Volume =

    Arellano, Manuel , Journal =. Computing robust standard errors for within-groups estimators , Volume =

  149. [157]

    Review of Economic Studies , year =

    Arellano, Manuel and Bond, Stephen , title =. Review of Economic Studies , year =

  150. [158]

    Ariely, Dan and Loewenstein, George and Prelec, Drazen , Journal =

  151. [159]

    and Imbens, Guido W

    Arkhangelsky, Dmitry and Athey, Susan and Hirshberg, David A. and Imbens, Guido W. and Wager, Stefan , title =. 2018 , OPTkey =

    2018

  152. [160]

    2018 , OPTkey =

    Arkhangelsky, Dmitry and Imbens, Guido , title =. 2018 , OPTkey =

    2018

  153. [161]

    , title =

    Aronow, Peter M. , title =. , year =

  154. [162]

    and Samii, Cyrus , title =

    Aronow, Peter M. and Samii, Cyrus , title =. American Journal of Political Science , year =

  155. [163]

    Annals of Applied Statistics , year =

    Aronow, Peter and Samii, Cyrus , title =. Annals of Applied Statistics , year =

  156. [164]

    and Sovey, Allison J

    Aronow, Peter M. and Sovey, Allison J. , Institution =. Beyond

  157. [165]

    , Edition =

    Arrow, Kenneth J. , Edition =

  158. [166]

    and Kent, John T

    Arslan, Olcay and Constable, Peter D.L. and Kent, John T. , Journal =. Domains of Convergence for the

  159. [167]

    and Kent, John T

    Arslan, Olcay and Constable, Peter D.L. and Kent, John T. , Journal =. Convergence Behavior of the

  160. [168]

    Journal of Marketing Research , year =

    Ascarza, Eva , title =. Journal of Marketing Research , year =

  161. [169]

    , Journal =

    Asher, Herbert B. , Journal =

  162. [170]

    Review of Economics and Statistics , year =

    Ashenfelter, Orley and Card, David , title =. Review of Economics and Statistics , year =

  163. [171]

    Ashenfelter, Orley and Krueger, Alan , Journal =

  164. [172]

    , Journal =

    Ashworth, Scott and Clinton, Joshua and Meirowitz, Adam and Ramsay, Kristopher W. , Journal =

  165. [173]

    and Pocock, Stuart J

    Assmann, Susan F. and Pocock, Stuart J. and Enos, Laura E. and Kasten, Linda E. , Journal =

  166. [174]

    , title =

    Athey, Susan and Eckles, Dean and Imbens, Guido W. , title =. Journal of the American Statistical Association , year =. doi:10.1080/01621459.2016.1241178 , OPTannote =

    2016 doi

  167. [175]

    , Journal =

    Athey, Susan and Imbens, Guido W. , Journal =

  168. [176]

    Proceedings of the National Academy of Sciences , year =

    Athey, Susan and Imbens, Guido , title =. Proceedings of the National Academy of Sciences , year =

  169. [177]

    2018 , OPTkey =

    Athey, Susan and Imbens, Guido , title =. 2018 , OPTkey =

    2018

  170. [178]

    and Wager, Stefan , title =

    Athey, Susan and Imbens, Guido W. and Wager, Stefan , title =. Journal of the Royal Statistical Society, Series. 2018 , OPTkey =

    2018

  171. [179]

    , Journal =

    Austin, Peter C. , Journal =. A critical appraisal of propensity-score matching in the medical literature between 1996 and 2003 , Volume =

    1996

  172. [180]

    Avin, Chen and Shpitser, Ilya and Pearl, Judea , Booktitle =

  173. [181]

    and Kaplan, Noah , Journal =

    Bafumi, Joseph and Gelman, Andrew and Park, David K. and Kaplan, Noah , Journal =. Practical Issues in Implementing and Understanding

  174. [182]

    2010 , Journal =

    Bafumi, Joseph and Herron, Michael , Title =. 2010 , Journal =

    2010

  175. [183]

    , Journal =

    Bagley, C.R. , Journal =

  176. [184]

    and Bergstrand, Jeffrey H

    Baier, Scott L. and Bergstrand, Jeffrey H. , Journal =. Do free trade agreements actually increase members' international trade? , Volume =

  177. [185]

    , title =

    Bailey, Michael A. , title =. Journal of Politics , year =

  178. [186]

    The Institutional Roots of

    Bailey, Michael and Goldstein, Judith and Weingast, Barry , Journal =. The Institutional Roots of

  179. [187]

    and Strezhnev, Anton and Voeten, Erik , title =

    Bailey, Michael A. and Strezhnev, Anton and Voeten, Erik , title =. Journal of Conflict Resolution , year =

  180. [188]

    and Hecock, Donald S

    Bain, Henry M. and Hecock, Donald S. , Publisher =

  181. [189]

    and Small, D

    Baiocchi, M. and Small, D. S. and Lorch, S. and Rosenbaum, P. R. , Journal =. Building a stronger instrument in an observational study of prenatal care for premature infants , Volume =

  182. [190]

    Aislinn and McIntosh, Craig and Ozler, Berk , title =

    Baird, Sarah and Bohren, J. Aislinn and McIntosh, Craig and Ozler, Berk , title =. Review of Economics and Statistics , year =

  183. [191]

    Baker, Judy , Institution =

  184. [192]

    Kate Baldwin , Date-Added =

  185. [193]

    Balke, Alexander and Pearl, Judea , Journal =

  186. [194]

    and Bogatz, G

    Ball, S. and Bogatz, G. A. , Institution =

  187. [195]

    and Ford, L.A

    Band, D.L. and Ford, L.A. and Matteson, J.L. and Briggs, M.S. and Paciesas, W.S. and Pendleton, G.N. and Preece, R.D. and Palmer, D.M. and Teegarden, B.J. and Schaefer, B.E. , Journal =. Batse Gamma-Ray Burst Line Search

  188. [196]

    and Ryder, S

    Band, D.L. and Ryder, S. and Ford, L.A. and Matteson, J.L. and Palmer, D.M. and Teegarden, B.J. and Briggs, M.S. and Paciesas, W.S. and Pendleton, G.N. and Preece, R.D. , Journal =. Batse Gamma-Ray Burst Line Search

  189. [197]

    and Ford, L.A

    Band, D.L. and Ford, L.A. and Matteson, J.L. and Briggs, M.S. and Paciesas, W.S. and Pendleton, G.N. and Preece, R.D. , Journal =. Batse Gamma-Ray Burst Line Search

  190. [198]

    and Finley, Andrew O

    Banerjee, Sudipto and Gelfand, Alan E. and Finley, Andrew O. and Sang, Huiyan , title =. Journal of the Royal Statistical Society, Series. 2008 , OPTkey =

    2008

  191. [199]

    Birds of the Same Feather Tweet Together: Bayesian Ideal Point Estimation Using Twitter Data , journal =

    Barber\'. Birds of the Same Feather Tweet Together: Bayesian Ideal Point Estimation Using Twitter Data , journal =. 2015 , OPTkey =

    2015

  192. [200]

    Barber, Benjamin , Publisher =

  193. [201]

    and Hill, Jennifer L

    Barnard, John and Frangakis, Constantine E. and Hill, Jennifer L. and Rubin, Donald B. , Journal =. Principal Stratification Approach to Broken Randomized Experiments: A Case Study of School Choice Vouchers in

  194. [202]

    Fairness and Machine Learning , author =

  195. [203]

    and Kenny, David A

    Baron, Reuben M. and Kenny, David A. , Journal =

  196. [204]

    , Journal =

    Barreto, Matt A. , Journal =. Si

  197. [205]

    and Donald, Stephen N

    Barrett, Garry F. and Donald, Stephen N. , Journal =

  198. [206]

    and Segura, Gary M

    Barreto, Matt A. and Segura, Gary M. and Woods, Nathan D. , Journal =. Mobilizing Effect of Majority-Minority Districts , Volume =

  199. [207]

    , Journal =

    Bartels, Larry M. , Journal =

  200. [208]

    Bas, Mihammet and Signorino, Curtis and Walker, Robert , Journal =

  201. [209]

    , title =

    Basse, Guillaume and Airoldi, Edoardo M. , title =. Sociological Methodology , year =

  202. [210]

    Journal of the American Statistical Association , year =

    Basse, Guillaume and Feller, Avi , title =. Journal of the American Statistical Association , year =

  203. [211]

    State Politics & Policy Quarterly , year =

    Battista, James Coleman and Peress, Michael and Richman, Jesse , title =. State Politics & Policy Quarterly , year =

  204. [212]

    and Preacher, Kristopher J

    Bauer, Daniel J. and Preacher, Kristopher J. and Gil, Karen M. , Journal =

  205. [213]

    Baum, L. E. and Petrie, T. and Soules, G. and Weiss, N. , Journal =. A Maximization Technique Occurring in the Statistical Analysis of Probabilistic Functions of

  206. [214]

    , Chapter =

    Bayarri, Susie and Berger, James O. , Chapter =

  207. [215]

    Beale, E. M. L. , Journal =

  208. [216]

    and Ghahramani, Zoubin , title =

    Beal, Matthew J. and Ghahramani, Zoubin , title =. 2001 , OPTkey =

    2001

  209. [217]

    Annual Review Political Science , year =

    Beck, Nathaniel , title =. Annual Review Political Science , year =

  210. [218]

    , title =

    Beck, Natheniel and Katz, Jonathan N. , title =. American Political Science Review , year =

  211. [219]

    American Political Science Review , year =

    Beck, Nathaniel and King, Gary and Zeng, Langche , title =. American Political Science Review , year =

  212. [220]

    Progressing through

    Behrman, Jere and Sengupta, Piyali and Todd, Petra , Journal =. Progressing through

  213. [221]

    and Rubin, Donald B

    Belin, Thomas R. and Rubin, Donald B. , title =. Journal of the American Statistical Association , year =

  214. [222]

    A Probability Model for the Measurement of Ecological Segregation , Volume =

    Bell, Wendell , Journal =. A Probability Model for the Measurement of Ecological Segregation , Volume =

  215. [223]

    Political Science Research and Methods , year =

    Bell, Andrew and Jones, Kelvyn , title =. Political Science Research and Methods , year =

  216. [224]

    Bell, Robert M. and. Bias reduction in standard errors for linear regression with multi-stage samples , journal =. 2002 , OPTkey =

    2002

  217. [225]

    and Lerman, Steven R

    Ben-Akiva, Moshe E. and Lerman, Steven R. , Publisher =

  218. [226]

    and Berger, David and Luby, Stephen P

    Benjamin-Chung, Jade and Arnold, Benjamin F. and Berger, David and Luby, Stephen P. and Miguel, Edward and Colford Jr, John M. and Hubbard, Alan E. , title =. International Journal of Epidemiology , year =

  219. [227]

    Benjamini, Yoav and Hochberg, Yosef , Journal =

  220. [228]

    Benjamini, Yoav and Yekutieli, Daniel , Journal =

  221. [229]

    2019 , OPTkey =

    Ben-Michael, Eli and Feller, Avi and Rothstein, Jesse , title =. 2019 , OPTkey =

    2019

  222. [230]

    Beran, Rudolf , Journal =

  223. [231]

    and Krosnick, Jon A

    Berent, Matthew K. and Krosnick, Jon A. and Lupia, Arthur , title =. 2011 , OPTkey =

    2011

  224. [232]

    and Krosnick, Jon A

    Berent, Matthew K. and Krosnick, Jon A. and Lupia, Arthur , title =. Public Opinion Quarterly , year =

  225. [233]

    , Edition =

    Berger, James O. , Edition =

  226. [234]

    , Publisher =

    Berinsky, Adam J. , Publisher =. Silent Voices: Opinion Polls and Political Representation in

  227. [235]

    Journal of Experimental Criminology , year =

    Berk, Richard , title =. Journal of Experimental Criminology , year =

  228. [236]

    Annals of Statistics , year =

    Berk, Richard and Brown, Lawrence and Buja, Andreas and Zhang, Kai and Zhao, Linda , title =. Annals of Statistics , year =

  229. [237]

    Sociological Methods & Research , year =

    Berk, Richard and Heidari, Hoda and Jabbari, Shahin and Kearns, Michael and Roth, Aaron , title =. Sociological Methods & Research , year =

  230. [238]

    Public Opinion Quarterly , year =

    Bernstein, Robert and Chadha, Anita and Montjoy, Robert , title =. Public Opinion Quarterly , year =

  231. [239]

    Bertrand, Marianne and Duflo, Esther and Mullainathan, Sendhil , Journal =

  232. [240]

    Bertrand, Marianne and Mullainathan, Sendhil , Journal =. Are

  233. [241]

    , Publisher =

    Bertsimas, Dimitris and Tsitsiklis, John N. , Publisher =

  234. [242]

    Spatial Interaction and the Statistical Analysis of

    Besag, Julian , Journal =. Spatial Interaction and the Statistical Analysis of

  235. [243]

    , Journal =

    Besag, Julian and Green, Peter J. , Journal =. Spatial Statistics and

  236. [244]

    Besag, Julian and Green, Peter and Higdon, David and Mengersen, Kerrie , Journal =

  237. [245]

    and Moran, Patrick A.P

    Besag, Julian E. and Moran, Patrick A.P. , Journal =. On the Estimation and Testing of Spatial Interaction in

  238. [246]

    The Economics of Preferential Trade Agreements , Year =

  239. [247]

    and Levina, Elizaveta , title =

    Bickel, Peter J. and Levina, Elizaveta , title =. Annals of Statistics , year =

  240. [248]

    Perspectives on Politics , year =

    Biddle, Stephen , title =. Perspectives on Politics , year =

  241. [249]

    Biemer, Paul and Brown, Gordon , Journal =

  242. [250]

    Annals of Statistics , year =

    Bien, Jacob and Taylor, Jonathan and Tibshirani, Robert , title =. Annals of Statistics , year =

  243. [251]

    Billewicz, W. Z. , Journal =

  244. [252]

    Billheimer, Dean and Guttorp, Peter and Fagan, William F , Journal =

  245. [253]

    , Publisher =

    Bishop, Christopher M. , Publisher =

  246. [254]

    and Fienberg, Stephen E

    Bishop, Yvonne M.M. and Fienberg, Stephen E. and Holland, Paul W. , Pages = 557, Publisher =. Discrete Multivariate Analyses:

  247. [255]

    A Framework for Dynamic Causal Inference in Political Science , Volume =

    Blackwell, Matthew , Journal =. A Framework for Dynamic Causal Inference in Political Science , Volume =

  248. [256]

    and Berger, Mark C

    Black, Dan A. and Berger, Mark C. and Scott, Frank A. , Journal =

  249. [257]

    American Political Science Review , year =

    Blackwell, Matthew and Glynn, Adam , title =. American Political Science Review , year =

  250. [258]

    Blais, Andr\'e , Publisher =

  251. [259]

    , author=

    Why Do Civilians Hold Bargaining Power in State Revenue Conflicts? Evidence from Nigeria. , author=

  252. [260]

    Poverty and Support for Militant Politics: Evidence from

    Blair, Graeme and Fair, Christine and Malhotra, Neil and Shapiro, Jacob , Journal =. Poverty and Support for Militant Politics: Evidence from

  253. [261]

    Blattman, Christopher , Journal =

  254. [262]

    and Ng, Andrew Y

    Blei, David M. and Ng, Andrew Y. and Jordan, Michael I. , journal =

  255. [263]

    and Ng, Andrew Y

    Blei, David M. and Ng, Andrew Y. and Jordan, Michael I. , title =. Journal of Machine Learning Research , year =

  256. [264]

    and Hwang, Jessica , ALTeditor =

    Blitzstein, Joseph K. and Hwang, Jessica , ALTeditor =. Introduction to Probability , publisher =. 2015 , OPTkey =

    2015

  257. [265]

    , Journal =

    Bloom, Howard S. , Journal =

  258. [266]

    and Orr, Larry L

    Bloom, Howard S. and Orr, Larry L. and Bell, Stephen H. and Cave, George and Doolittle, Fred and Lin, Winston and Bos, Johannes M. , Journal =. The Benefits and Costs of

  259. [267]

    Bobo, Lawrence and Gilliam, Frank D. Jr. , Journal =. How Black Candidates Affect Voter Turnout , Volume =

  260. [268]

    Darrell and Aitkin, Murray , title =

    Bock, R. Darrell and Aitkin, Murray , title =. Psychometrika , year =

  261. [269]

    and Goldberger, Arthur S

    Bohrnstedt, George W. and Goldberger, Arthur S. , Journal =

  262. [270]

    Journal of the American Statistical Association , year =

    Bojinov, Iavor and Shephard, Neil , title =. Journal of the American Statistical Association , year =

  263. [271]

    , Journal =

    Bollen, Kenneth A. , Journal =

  264. [272]

    , Journal =

    Bollinger, Christopher R. , Journal =

  265. [273]

    American Political Science Review , year =

    Bond, Robert and Messing, Solomon , title =. American Political Science Review , year =

  266. [274]

    Teoria statistica delle classi e calcolo delle probabilit\`

    Bonferroni, Carlo Emilio , Journal =. Teoria statistica delle classi e calcolo delle probabilit\`

  267. [275]

    2012 , OPTkey =

    Bonhomme, Stephane and Manresa, Elena , title =. 2012 , OPTkey =

    2012

  268. [276]

    American Journal of Political Science , year =

    Bonica, Adam , title =. American Journal of Political Science , year =

  269. [277]

    and Hobart, James G

    Booth, James P. and Hobart, James G. , Journal =. Maximizing Generalized Linear Mixed Model Likelihoods With an Automated

  270. [278]

    2016 , note =

    RecordLinkage: Record Linkage in R , author =. 2016 , note =

    2016

  271. [279]

    2017 , OPTkey =

    Borusyak, Kirill and Jaravel, Xavier , title =. 2017 , OPTkey =

    2017

  272. [280]

    and Eavey, Cheryl L

    Bottom, William P. and Eavey, Cheryl L. and Miller, Gary J. and Victor, Jennifer Nicoll , Journal =

  273. [281]

    and Baker, Regina M

    Bound, John and Jaeger, David A. and Baker, Regina M. , Journal =

  274. [282]

    and Hunter, J

    Box, George E.P. and Hunter, J. Stuart and Hunter, William G. , Edition =. Statistics for Experimenters: Design, Innovation, and Discovery , Year =

  275. [283]

    Box, George E. P. and Tiao, George C. , Publisher =

  276. [284]

    , Journal =

    Boyles, R.A. , Journal =. On the Convergence of the

  277. [285]

    European Journal of Operational Research , year =

    Bozkaya, Burcin and Erkut, Erhan and Laporte, Gilbert , title =. European Journal of Operational Research , year =

  278. [286]

    , Journal =

    Brady, Henry E. , Journal =

  279. [287]

    , Date-Added =

    Bradburn, N., Sudman, S., Blair, E., & Stocking, C. , Date-Added =. Public Opinion Quarterly , Pages =

  280. [288]

    and Collier, David , Publisher =

    Brady, Henry E. and Collier, David , Publisher =

  281. [289]

    and Herron, Michael and Mebane, Walter R

    Brady, Henry E. and Herron, Michael and Mebane, Walter R. and Sekhon, Jasjeet S. and Shotts, Kenneth and Wand, Jonathan , Journal =

  282. [290]

    Brader, Ted and Valentino, Nicholas and Suhay, Elizabeth , Journal =

  283. [291]

    and Verba, Sidney and Schlozman, Kay Lehman , Journal =

    Brady, Henry E. and Verba, Sidney and Schlozman, Kay Lehman , Journal =. Beyond

  284. [292]

    Political Analysis , year =

    Brambor, Thomas and Clark, William Roberts and Golder, Matt , title =. Political Analysis , year =

  285. [293]

    and Freeman, John R

    Brandt, Patrick T. and Freeman, John R. , Journal =. Advances in

  286. [294]

    and Feng, Ziding , Journal =

    Braun, Thomas M. and Feng, Ziding , Journal =

  287. [295]

    Breiman, Leo , Journal =

  288. [296]

    and Olshen, R.A

    Breiman, Leo and Friedman, Jerome and Stone, Charles J. and Olshen, R.A. , Publisher =

  289. [297]

    and Clayton, D.G

    Breslow, Norman E. and Clayton, D.G. , Journal =

  290. [298]

    and Lin, X

    Breslow, Norman E. and Lin, X. , Journal =

  291. [299]

    Proceedings of the 18th Conference of Uncertainty in Artificial Intelligence , year =

    Brito, Carlos and Pearl, Judea , title =. Proceedings of the 18th Conference of Uncertainty in Artificial Intelligence , year =

  292. [300]

    and Grogger, Jeff , Journal =

    Bronars, Stephen G. and Grogger, Jeff , Journal =

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.