PaperlayerView exact revision on arXiv

Source-linked AI summary

Observational-Interventional Priors for Dose-Response Learning

Ricardo Silva

arXiv:1605.01573v1stat.ML

TL;DR

Estimating dose-response curves is difficult when interventions are costly and observational data may be confounded. The paper combines observationally informed Gaussian process priors with affine transformations learned from interventions, and reports advantages over using either data source alone. Its experiments also identify computational cost and overlap as practical boundaries.

  • Problem

    Interventional samples can be limited by cost, while observational associations may differ from causal effects because of confounding.

    Source: §1 · p. 1, §2 · p. 3

  • Method

    A hierarchical Gaussian process learns a distribution over dose-response curves from observational data and reshapes it with an affine transformation informed by interventions.

    Source: §3 · p. 3, §3.1 · p. 4

  • Results

    The combined approach outperformed estimates based on observational data alone or interventional data alone across scenarios, with reported robustness to varying unmeasured confounding.

    Source: §1 · p. 1, §4.1 · p. 7

  • Takeaways & Limitations

    The case study found improved average normalized absolute differences over an interventional-only Gaussian process with squared exponential covariance.

    Source: §4.2 · p. 8

    Computing the observational covariance matrix can be expensive at O(T^2N^2); approximation is left for future research.

    Source: §3.1 · p. 4

Abstract

from arXiv · show

Controlled interventions provide the most direct source of information for learning causal effects. In particular, a dose-response curve can be learned by varying the treatment level and observing the corresponding outcomes. However, interventions can be expensive and time-consuming. Observational data, where the treatment is not controlled by a known mechanism, is sometimes available. Under some strong assumptions, observational data allows for the estimation of dose-response curves. Estimating such curves nonparametrically is hard: sample sizes for controlled interventions may be small, while in the observational case a large number of measured confounders may need to be marginalized. In this paper, we introduce a hierarchical Gaussian process prior that constructs a distribution over the dose-response curve by learning from observational data, and reshapes the distribution with a nonparametric affine transform learned from controlled interventions. This function composition from different sources is shown to speed-up learning, which we demonstrate with a thorough sensitivity analysis and an application to modeling the effect of therapy on cognitive skills of premature infants.

1 Contribution

The paper studies how to estimate a causal dose-response curve across controlled treatment levels, where observational associations may differ from causal effects because of confounding. It combines observational and interventional data through coupled Gaussian process priors.

1 Contribution

  • Coupled Gaussian process priors combine observational and interventional data, and outperform estimates using either source alone across a variety of scenarios.Source: §1 · p. 1
  • Observational data can reflect confounded associations rather than causal effects, while controlled trials provide an unconfounded treatment-outcome relationship.Source: §1 · p. 1
  • Under the paper’s invariance assumptions, measuring covariates Z that block back-door paths allows observational data to identify the interventional distribution.Source: §1 · p. 1, Figure 1 · p. 2
  • Dose-response learning estimates a treatment’s causal effect across quantitative levels, assuming the causal structure is known.Source: §1 · p. 2

2 Background

The background distinguishes the interventional target from observational association and describes adjustment over covariates as one way to estimate that target. It also notes limits of adjustment, especially when treatment overlap is weak.

2 Background

  • The interventional outcome distribution is obtained by integrating conditional outcomes over Z, rather than weighting covariates by their treatment-conditioned distribution.Source: §2 · p. 2
  • A back-door-adjusted estimator averages predictions g(x, z) over observed covariates; under conditions, estimating the resulting curve is a one-dimensional nonparametric regression problem.Source: §2 · pp. 2–3
  • Estimating g(x, z) can be daunting when Z is high-dimensional, even though the adjusted curve can be easier to estimate.Source: §2 · pp. 2–3
  • Weak overlap can require extrapolating from distant treatment observations, producing unreliable estimates and poorly calibrated credible intervals, especially for continuous nonlinear responses.Source: §2 · p. 3
  • The paper uses adjusted observational data to inform priors, then applies pointwise affine transformations to address possible oversmoothing with interventional data.Source: §2 · p. 3
  • Conditioning on selected covariates can yield stratum-specific average causal effects, but fine-grained randomized effects may require stronger smoothing and extrapolation assumptions.Source: §2 · p. 3

3 Hierarchical Priors via Inherited Smoothing and Local Affine Changes

The model uses observational data to form a Gaussian-process prior over dose-response curves, then reshapes it with a nonparametric affine transformation informed by interventions. Its nonstationary covariance inherits information from the observational adjustment, while inference estimates the transformed curve and its uncertainty.

3.1 Two-layered Priors for Dose-responses

  • The prior begins with a nonstationary Gaussian process over observationally adjusted dose-response values, with covariance computation costing O(T^2N^2).The paper notes this cost can be expensive, while intervention-point sets are often not particularly large; approximating the covariance is left for future work.Source: §3.1 · pp. 3–4
  • The curve is modeled as f(X) = a(X) ⊙ fobs(X) + b(X), where a and b are distortion and translation functions that allow uncertainty to vary with fobs.The elementwise multiplicative term also induces a heavy-tail prior on f.Source: §3.1 · p. 4

3.2 Hyperpriors

  • The distortion covariance borrows estimated smoothness and scale from observational data, linking correlation to distance between scaled dose-and-mean points and amplitude to Kobs's diagonal.Hyperparameters control the distance weighting and global amplitude adjustment.Source: §3.2 · p. 4
  • In the synthetic example, combining observational and interventional data improves the fit despite misspecified adjustment, but can still miss the strongest-dose behavior when intervention draws are unlucky.The adjustment uses 15 of 25 confounders; the example warns that a small interventional dataset may not overcome a prior pointing in the wrong direction.Source: §3.2 · p. 5, Figure 2 · p. 5

3.3 Inference, Stratified Learning and Active Learning

  • Inference uses Gibbs sampling over f(X), a(X), b(X), and transformation hyperparameters, with slice sampling for the hyperparameters.The observational-process hyperparameters are fit separately because updating Kobs within each iteration would be costly.Source: §3.3 · p. 6
  • For a small set of discrete covariates, the implementation fits independent priors and affine transformations for each stratum rather than borrowing strength across strata.The paper also considers simple active-learning schemes that use the prior's varying uncertainty across treatment levels.Source: §3.3 · p. 6

4 Experiments

The experiments use simulated dose-response data, including curves generated from a real infant-development study, to assess performance when intervention samples are limited. The method is robust to varying unmeasured confounding and improves over an interventional-only Gaussian process in the reported settings.

4.1 Synthetic Data Studies

  • The synthetic benchmark varied quadratic or cubic ground truth, 50% or 80% signal range, and interventional sample sizes M = 40, 100, and 200, with 50 simulations per condition.The observational sample had 1,000 observations and 25 confounders, while algorithms received only 15 confounders.Source: §4.1 · p. 6, Table 1 · p. 7
  • The experiments assess estimates with normalized absolute difference and posterior log density, comparing the method with fixed affine terms, a generic Gaussian-process correction, and an interventional-only Gaussian process.Positive table differences indicate the proposed method is better under the corresponding criterion.Source: §4.1 · p. 7, Table 1 · p. 7
  • The method was robust to varying unmeasured confounding; its advantage over interventional-only Gaussian processes was strongest at M = 40 and diminished as M increased.The study varied intervention sample size, polynomial curve form, signal range, and whether 10 confounders were randomly or adversarially omitted.Source: §4.1 · pp. 6–7

4.2 Case Study

  • The infant-development case study used 347 eligible infants and 21 covariates to fit an adjusted curve, then generated synthetic intervention datasets because the dose level was not randomized and its true response curve was unknown.The simulated dose was days assigned to a development program, and the outcome was IQ at age three.Source: §4.2 · p. 8
  • 0.06, 0.07, and 0.08 average normalized absolute-difference improvements were reported for the high-school, college, and combined strata, respectively, over an interventional-only Gaussian process.Error was reduced in 82%, 89%, and 91% of runs, respectively, with 10 interventional samples per treatment level per stratum.Source: §4.2 · p. 8, Figure 3 · p. 8

5 Conclusion

The paper combines observational and interventional measurements and assesses the method’s accuracy and robustness. It also identifies unresolved limits involving confounding, sample selection, and active learning.

5 Conclusion

  • The paper presents a principled way to combine observational and interventional measurements and evaluates its accuracy and robustness.Its evaluation emphasizes robustness to model misspecification and sensitivity to individual prior components.Source: §5 · p. 8
  • The infant-study curves are not claimed to be true dose-response curves because treatment levels may have changed in response to infants’ outcomes.The authors say measured covariates may not reliably account for this feedback effect.Source: §5 · p. 8
  • The analysis leaves sample-selection bias unaddressed, and strong unmeasured confounding limits expected performance.The authors also call for a more thorough analysis of active learning under possible model misspecification.Source: §5 · p. 9

Appendix

The appendix documents synthetic-data generation and IHDP preprocessing, illustrates active learning, and compares the method with deep Gaussian-process approaches.

Appendix

  • The appendix covers the synthetic-data protocol, IHDP preprocessing, an active-learning illustration, and a comparison with deep Gaussian-process methods.The authors describe the active-learning and deep-GP comparisons as illustrative.Source: Appendix · p. 10

A Synthetic Data Generator

The synthetic generator creates treatment and outcome data with shared covariate causes and controlled levels of confounding. Its sensitivity comparisons indicate that the affine component chiefly changes posterior uncertainty, especially in the Adversarial regime.

A Synthetic Data Generator

  • The generator makes covariates Z common causes of treatment X and outcome Y, with off-diagonal entries of ΣZ set to 0.5.Causes of X that do not cause Y contribute approximately 20% to 40% of the variance contributed by common causes.Source: A Synthetic Data Generator · p. 10
  • The outcome contribution from X is a polynomial dose-response function of degree 2 or 3, with its scale controlled relative to other causes of Y.The generator rejects and resamples data when the absolute empirical rank correlation between X and f_yz is below 0.2, enforcing a minimum degree of confounding.Source: A Synthetic Data Generator · p. 11
  • The polynomial maps standardized X to a range approximately matching the span between selected empirical quantiles of the other-cause contribution.The experiments use α = 0.1 or α = 0.25.Source: A Synthetic Data Generator · p. 11
  • Table 2 adds an unadjusted observational-only method and a translation-only variant to the comparisons.Method IV omits back-door adjustment; Method V drops the multiplicative component and uses a non-stationary covariance for b.Source: A Synthetic Data Generator · p. 11, Table 2 · p. 12
  • The affine component does not appear to change absolute error, but it significantly changes posterior uncertainty, with differences more prominent in the Adversarial regime.The authors partly attribute this pattern to the heavy-tailed, non-Gaussian nature of a ⊙ fobs; a 0.05 difference in average LV corresponds to a density ratio of 2.7 for |X| = 20.Source: A Synthetic Data Generator · pp. 11–12

B Preprocessing of the Infant Health and Development Program Data

The IHDP preprocessing produces a 347-person sample and fits a Gaussian-process regression to model the observational response. The resulting intervention simulator uses a normal approximation for convenience, although the implied outcomes may be heteroscedastic or non-Gaussian.

B Preprocessing of the Infant Health and Development Program Data

  • After removing 30 individuals with missing outcomes from 985 individuals, including 377 treated, the authors retained 347 and further filtered or binarized variables.The preprocessing began from the version provided by.Source: B Preprocessing of the Infant Health… · p. 12
  • The processed dataset has 21 columns, and the observational regression uses a Gaussian-process prior with Gaussian likelihood.The prior uses a Matérn 3/2 covariance with automatic relevance determination, and hyperparameters are fitted by marginal maximum likelihood.Source: B Preprocessing of the Infant Health… · pp. 12–13
  • The treated subgroup’s treatment durations span 0–468 days, and the interventional levels are set at 0, 25, 50, …, 450.The outcome is IQ score at age 3, standardized using the observational data’s empirical mean and standard deviation.Source: B Preprocessing of the Infant Health… · p. 13
  • Normality is used as a convenient approximation even though the implied do(x) outcome may be heteroscedastic or non-Gaussian.The authors say heteroscedastic regression could be adopted without conceptual changes but is omitted for convenience.Source: B Preprocessing of the Infant Health… · p. 13

C Active Learning Illustration

The D-optimal scheme selects doses by posterior uncertainty, approximated with variance from MCMC samples. In the IHDP illustration, active selection did at least as well as uniform allocation, with a clearer advantage where the prior missed the true curve.

C Active Learning Illustration

  • The D-optimal scheme chooses the dose with the highest target-function entropy, approximated by its variance across current MCMC samples.The authors note that variance can rank uncertainty using standard second-order approximations.Source: C Active Learning Illustration · p. 13
  • The IHDP illustration initialized one outcome per dose level and allocated 95 further trials across two strata, updating hyperparameters after every five new points.The trial budget was 5 × |X| = 95.Source: C Active Learning Illustration · p. 13
  • Active selection did at least as well as uniform allocation, with a clear advantage at low high-school doses where the prior missed the true curve.The advantage also appeared in regions with no more than two datapoints.Source: C Active Learning Illustration · p. 13
  • The 114-experiment allocation histogram covers 19 initial trials and 95 adaptively selected trials; most doses were assigned at the endpoints of X.The high-school stratum received 31 further trials and the college stratum received 64, compared with uniform allocation of six trials per dose level.Source: C Active Learning Illustration · p. 13, Figure 5 · p. 14

D A Note on Generic Deep Gaussian Processes

The comparison illustrates the factorized affine prior against a generic deep Gaussian process, which the authors say offers no evident advantage for scalar-dose learning. In this IHDP illustration, the affine approach performed better and ran substantially faster, but the authors avoid generalizing that comparison.

D A Note on Generic Deep Gaussian Processes

  • The affine transformation parameters a and b are not identifiable; their value is as a useful prior for f, not as separately identified quantities.The authors assess the prior through comparisons and sensitivity analyses.Source: D A Note on Generic Deep Gaussian… · p. 14
  • The generic alternative transforms fobs through a Gaussian-process function whose mean is the identity and whose covariance uses smooth functions.This prior is biased toward the unknown observational curve.Source: D A Note on Generic Deep Gaussian… · p. 14
  • For a scalar treatment dose, the authors see no clear reason to use a more complicated model motivated by combining high-dimensional signals.They find no evident advantage for the generic construction over the pointwise affine prior.Source: D A Note on Generic Deep Gaussian… · p. 14
  • Sampling the generic deep setup is difficult because fobs appears in Ku; the comparison used a squared-exponential covariance and no parameter sharing across strata.The authors describe that prior as a convenient sampling setup, not an optimal dose-response prior.Source: D A Note on Generic Deep Gaussian… · p. 14
  • The authors frame the comparison as an illustration, not a general claim of inferential superiority over standard deep Gaussian processes.They report no evidence that the standard deep Gaussian process offers an advantage.Source: D A Note on Generic Deep Gaussian… · p. 14
  • 1200 seconds for HMC versus 54 seconds for factorized-prior inference, with the affine prior performing substantially better in this IHDP illustration.Figure 7 compares posterior dose-response samples from the non-factorized and factorized priors.Source: D A Note on Generic Deep Gaussian… · pp. 14–15, Figure 7 · p. 16
Loading 1605.01573v1…