Processing math: 0%
+ - 0:00:00
Notes for current slide
Notes for next slide

Methods for Estimating Causal Effects for Multivariate Continuous Exposures

Justin Williams

2020-11-05

1/37

Introduction

2/37

What If...

  • Q: "What if the batter had swung?"

  • "What-if" questions require us to think causally and imagine unobserved potential outcomes

  • Inability to observe all potential outcomes

    • Fundamental problem of causal inference (Holland, 1986)
    • Third rung of ladder of causation (Pearl and Mackenzie, 2018)
3/37

Background

  • Gold standard for answering causal questions is the randomized clinical trial
    • Balance on known and unknown confounders
  • However, using the gold standard RCT may not be possible in real world settings

    • Unrealistic, e.g., surgery vs. placebo
    • Unethical, e.g., forcing patients to smoke
  • How to estimate causal effects without randomization?

    • Exposure may be systematically related to outcome
  • Utilize propensity score method (Rosenbaum and Rubin, 1983)

Propensity score = Probability of receiving exposure given set of confounders

Pr

4/37

Brief History of Propensity Score

Estimation

  1. Originally proposed by Rosenbaum and Rubin (1983)
    • Handled binary exposure only, i.e., exposed vs. control
  2. Methods extended to categorical, or multiple, treatments by Imbens (2000)
    • Term "generalized propensity score" (GPS) is introduced
  3. Adaptions for continuous treatments introduced by Hirano and Imbens (2004); Imai and Van Dyk (2004)
    • Use Gaussian densities for estimating probability of exposure level

Bias Removal

  1. Covariate adjustment
  2. Subclassification/stratification
  3. Matching
  4. Inverse probability of treatment weighting (IPTW)
5/37
  • Throughout we use GPS to refer to continuous rather than categorical treatment

Propensity Score with Continuous Exposure

  • Popular applications
    • Economics: difference in employment outcomes due to length of time in job training program (Flores, Flores-Lagunes, Gonzalez, and Neumann, 2012; Kluve, Schneider, Uhlendorff, and Zhao, 2012)
    • Health outcomes: childhood obesity and duration of breastfeeding (Jiang and Foster, 2013)
    • Education policy: transfer rate and number of credits taken (Doyle, 2011)
  • Methodology development
    • SuperLearner (Kreif, Grieve, Díaz, and Harrison, 2015)
    • Kernel density estimation (Flores, Flores-Lagunes, Gonzalez, and Neumann, 2012)
    • Gradient boosting (Zhu, Coffman, and Ghosh, 2015)
    • Covariate balancing generalized propensity score (CBGPS) (Fong, Hazlett, and Imai, 2018)
    • Entropy balancing (Vegetabile, Griffin, Coffman, Cefalu, and McCaffrey, 2020; Tübbicke, 2020)
6/37

Multivariate Continuous Exposures

  • All methodological development has focused on single univariate continuous exposure
    • Multiple exposures noted only briefly in Imai and Van Dyk (2004)
  • Need for methods to handle multivariate exposures
    • Identify potential combination therapies from observational studies
      • COVID-19 treatments (Sanders, Monogue, Jodlowski, and Cutrell, 2020)
    • Estimate combined effects of childhood obesity intervention programs in the community
  • Challenges:
    • Define multivariate densities for generalized propensity score
    • Multiple sets of confounders
    • Correlation of exposures
    • Assessing high dimensional balance
    • Properly defining estimable multivariate region for inference
7/37

Motivating Example

8/37

Project Goal

  • High obesity rates of low-income 2-5 year olds in Los Angeles County (PHFE WIC, 2010)

    • 20% classified as obese during period from 2003-2009
  • Estimate causal effects of non-randomly assigned obesity interventions on WIC preschool-aged child obesity rates

    • Early Childhood Obesity Systems Science (ECOSyS) study funded by NIH R01 HD072296
  • Intervention Period: 2010-2016

  • Exposure:

    • WIC intervention programs
  • Outcome:

    • Change in obesity prevalence from post- to pre-intervention measured at census tracts
9/37
  • In 2018 WIC served approximately half of all children under age 5 in Los Angeles County

Outcome Data

  • Source:
  • Quantification:
    • Obesity prevalence at census tracts
    • Include only census tracts with at least 30 WIC children for all years
      • Total of 1079 included
    • Pre-intervention=2007-2009
    • Post-intervention=2012-2016
      • Allow for lagged treatment effect Y=\bar{p}_{post}-\bar{p}_{pre}

10/37

Exposure Data

  • Source:
    • Interviews with WIC staff
    • Total of 32 programs identified across 7 WIC agencies in LAC
    • Targeted 8 regions during interview process
  • Quantification:
    1. Categorize intervention strategies used by program
    2. Construct adapted "intervention dose index"
    3. Identify clinics that implemented intervention
    4. Map dose from clinics to census tracts

11/37

Intervention Strategies

  • Classified interventions into 9 potential strategies
    • Grouped as micro (targeted individuals) or macro (general population)
# Name Group n (%)
1 Government Policies Macro 1 (3%)
2 Institutional Polices Macro 4 (12%)
3 Infrastructure Investments Macro 3 (9%)
4 Business Practices Macro 4 (12%)
5 Group Education Micro 21 (66%)
6 Counseling Micro 14 (44%)
7 Health Communication Micro 17 (53%)
8 Home Visitation Micro 8 (25%)
9 Screening and Referral Micro 14 (44%)
12/37
  • Micro strategies were more frequently used than macro
  • Program could use multiple strategies

Intervention Dose Index

  • Developed to quantify community exposure (Wang, Crespi, Jiang, Nobari, Roper-Fingerhut, Rauzon, Robles, Blocklin, Davoudi, Kuo, MacLeod, Seto, Whaley, and Prelip, 2018) \begin{align} IDI = RS \times FS \times SS \end{align}
  • RS\in[0,1] = reach score:
    • probability of receiving the intervention program
  • FS\in[0,1] = fidelity score:
    • degree to which program was followed during implementation
  • SS\in[1,9] = strength score:

    • effectiveness of strategy as rated by subject matter experts
  • Estimated for each strategy and then summed by group, macro and micro




Adapted IDI construction example

Adapted IDI construction example

13/37
  • We assume FS=1 for all programs and omit SS in our adapted version as we want to estimate this from the data

Catchment Areas

  • Inherent geographic misalignment
    • Exposure data on intervention programs at clinic locations
    • Outcome data on obesity rates at census tracts
  • Construct mapping using catchment areas
  • Catchment area construction:
    1. Based on client participation patterns of attendance at each clinic
    2. Form a circle encompassing X% of clients who attend the clinic
      • Macro strategies: 80%
      • Micro strategies: 50%
    3. Census tracts that overlap catchment area receive program IDI

Micro

Macro

14/37
  • Key feature is that macro catchment areas are larger than micro reflecting belief that they have broader impact than micro
  • Outlined we also see the 8 regions of interest where interviews were targeted

Exposure Dose

15/37
  • In red is the trimmed convex hull region at the 95th quantile
  • Each point represents a census tract

Multivariate Generalized Propensity Score

mvGPS

16/37

Notation

  • Y: outcome of interest
  • \mathbf{D}: multivariate exposure of dimension m
    • In simulation study and application we fix m=2, i.e., bivariate exposure
  • \mathcal{C}=\{\mathbf{C}_{1}, \dots, \mathbf{C}_{m}\}: set of random variables of length m
    • Each \mathbf{C}_{j} for j=1,\dots,m, is a p_{j} vector of baseline confounders associated with the j^{th} exposure and outcome
  • Y_{i}(\mathbf{d}): potential outcome of i^{th} subject assigned exposure vector \mathbf{d}=(d_{1},\dots,d_{m})
  • \mu(\mathbf{d})=\mathbb{E}[Y(\mathbf{d})]: average dose-response function
  • n: total number of units
  • C_{ijk}: k^{th} confounder of the j^{th} exposure for the i^{th} individual

  • Observed data

    • (Y_{i}, D_{i1},\dots,D_{im}, C_{i11}, \dots, C_{i1p_{1}}, \dots, C_{im1}, \dots, C_{imp_{m}})
17/37

Identifying Assumptions

  • Three key assumptions for valid causal quantities
  1. Weak ignorability
  2. Positivity
  3. Stable-unit treatment value (SUTVA)
  • Exposure is conditionally independent of the potential outcomes given the appropriate set of confounders.

\begin{align} Y_{i}(\mathbf{d})\perp \!\!\! \perp\mathbf{D}_{i}\mid \mathcal{C} \quad \forall \quad \mathbf{d}\in\mathcal{D}. \end{align}

  • Replace \mathcal{C} with f_{\mathbf{D}\mid\mathcal{C}}

    • Especially with high dimensional set and/or continuous confounders
  • Difficult to rationalize

    • Requires perfect knowledge and collection of confounders
    • We assume it holds, i.e., no unmeasured confounding
  • "Weak" refers to conditional independence for each level of exposure not joint distribution of exposures

  • All units have the potential to receive a particular level of exposure given any value of the confounders

\begin{align} 0<f_{\mathbf{D}\mid\mathcal{C}}(\mathbf{D}=\mathbf{d}\mid \mathcal{C})<1 \quad \forall \quad \mathbf{d}\in\mathcal{D} \end{align}

  • Grid or convex hull

  • Trimmed or untrimmed

  • \text{Cov}(D_{1}, D_{2})=0.5
  • Convex hull and grid
  • The potential outcome of each unit does not depend on the exposure that other units receive and that there exists only one version of each exposure

  • Assumes there is no interference between units

    • No herd immunity
  • No errors in defining the potential outcomes caused by multiple versions of the exposure

  • Y_{i} = Y_{i}(\mathbf{d})

18/37

mvGPS

  • Conditional probability of receiving multivariate exposure given set of confounders

\begin{align} f(\mathbf{D}|\mathbf{C}_{1},\dots,\mathbf{C}_{m}) \end{align}

  • For bias removal construct inverse probability of treatment weights

\begin{align} w=\frac{f(\mathbf{D})}{f(\mathbf{D}|\mathbf{C}_{1},\dots,\mathbf{C}_{m})}, \end{align}

  • Propose multivariate normal models \begin{align} \mathbf{D}\sim \text{N}_{m}(\boldsymbol{\mu}, \boldsymbol{\Sigma}) \quad \mathbf{D}\mid \mathbf{C}_{1},\dots,\mathbf{C}_{m}\sim \text{N}_{m}\Bigg(\begin{bmatrix}\boldsymbol{\beta}_{1}^{T}\mathbf{C}_{1}\\ \vdots \\ \boldsymbol{\beta}_{m}^{T}\mathbf{C}_{m}\end{bmatrix}, \boldsymbol{\Omega}\Bigg), \end{align}
    • Tractable, well-behaved asymptotics, full univariate conditionals
    • Estimate parameters using ordinary least squares
19/37

Properties of mvGPS weights

  • Returns consistent estimate of average dose-response function

    • \mathbb{E}[wY\mid\mathbf{D}]=\mathbb{E}[Y(\mathbf{d})]
  • Covariance between each exposure, D_{j} and confounder, C_{jk}, is zero

    • \mathbb{E}[w(D_{j}-\mu_{D_j})(C_{jk}-\mu_{C_{jk}})]=0
  • Weights are normalized and maintain marginal moments

    • \mathbb{E}[w]=1
    • \mathbb{E}[w D_{j}]=\mathbb{E}[D_{j}]
    • \mathbb{E}[w\mathbf{C}_{j}]=\mathbb{E}[\mathbf{C}_{j}]
20/37

Simulation Study

21/37

Design

  • Sample size : n=200

  • Repetitions: B=1000

  • Compare mvGPS to other univariate alternative methods

    • Covariate balancing generalized propensity score, CBGPS, (Fong, Hazlett, and Imai, 2018)
    • Entropy balancing (Tübbicke, 2020)
    • Generalized linear propensity score, PS
  • Simulate bivariate normal exposure

    • marginal correlation
    • overlap of confounding sets
  • Effect of trimming weights (Kang and Schafer, 2007)

    • None vs. 99th quantile vs 95th quantile
22/37
  • Univariate methods may perform well when exposures are highly correlated and have the same confounders
    • Perform poorly with weakly correlated exposures and/or separate sets of confounders

Scenarios

  • M1: No common confounding

    • Each exposure has separate set of confounders
    • \mathbf{C}_{1} and \mathbf{C}_{2}
  • M2: Partially common confounding

    • Each exposure have some unique confounders but also share set of common confounders
    • \mathbf{C}_{1}, \mathbf{C}_{2}, and \mathbf{C}_{12}
  • M3: Common confounding

    • Both exposures share exact same set of common confounders
    • \mathbf{C}_{12}

23/37

Balance Results

  • Three balance metrics were assessed
    1. Average exposure-covariate correlation
    2. Maximum exposure-covariate correlation
    3. Effective sample size, (\Sigma_i w_i)^{2}/\Sigma_i w_i^2

24/37
  • Average summarizes total balance on all confounders
  • Max describes the most imbalanced confounder
  • ESS is equivalent to the relative power for inference

Dose Response Results

  • Two performance metrics were assessed
    1. Absolute total bias
    2. Root mean squared error (RMSE)

25/37

Takeaways

  • Multivariate methods are critical for achieving proper balance across multiple exposures
    • Protects against any single confounder being strongly imbalanced
    • Slight penalty of lower average balance in certain scenarios
  • Univariate methods achieved best balance with high overlap of confounders
    • Despite balance, high total bias of treatment effects
  • Using multivariate weights decreased ESS resulting in lower power and higher RMSE
    • Weight trimming reduced these effects, q=0.99
26/37

Application Results

Estimating Causal Effect of WIC Intervention

27/37

Potential Confounders

  • Three data sources used: American Community Survey, WIC Admin, and NETS
  • All values for propensity equation are average of three years prior to intervention, i.e., 2007-2009

American Community Survey

  1. % Spanish Speaking
  2. Median Household Income
  3. % with at least High School Diploma
  4. % of population less than 5 years old

WIC Admin

  1. Obesity Prevalence
  2. Overweight Prevalence

NETS (National Establishment Time-Series)

  1. Healthy Food Outlets per sq. mile
  2. Unhealthy Food Outlets per sq. mile
28/37

Covariate Balance Total

29/37

Covariate Balance Summary

Max Abs. Corr. Avg. Abs. Corr. ESS Method
0.12 0.04 637 mvGPS
0.18 0.08 580 CBGPS (Macro)
0.22 0.07 659 PS (Macro)
0.25 0.08 541 Entropy (Macro)
0.35 0.12 779 Entropy (Micro)
0.37 0.14 828 PS (Micro)
0.49 0.16 679 CBGPS (Micro)
0.41 0.19 1079 Unweighted
30/37

Dose-Response Surface

31/37

Takeaways

Balance

  • Significant imbalance prior to weighting

    • mvGPS had greater reduction in imbalance than univariate methods
  • ESS for all methods are reduced but power is still reasonably high

Dose-Response

  • Unweighted estimate was monotonic plane where effectiveness increased as you increased either exposure dose

  • mvGPS estimate was a saddle with optimal effectiveness using high micro and low macro

    • Showed that high levels of both exposures was less effective
32/37

Exploring Tail Behavior

33/37

Discussion

  • Achieved balance across multiple simultaneous exposures
  • Adapted for confounding sets and/or correlation of exposures
  • Visualized and interpreted effect of two continuous interventions with dose-response surface
  • Defined high dimensional positivity space with convex hull
  • mvGPS R package available now on CRAN

Limitations

  • Multivariate normal parametric distribution
  • Investigation of higher dimensional exposure
  • Time-varying multivariate exposure
  • Multiple versions possible for macro and micro doses
  • Geographic interference (SUTVA violations)
34/37

References I

Doyle, W. R. (2011). "Effect of increased academic momentum on transfer rates: an application of the generalized propensity score". In: Economics of Education Review 30.1, pp. 191-200.

Flores, C. A., A. Flores-Lagunes, A. Gonzalez, and T. C. Neumann (2012). "Estimating the effects of length of exposure to instruction in a training program: The case of job corps". In: Review of Economics and Statistics 94.1, pp. 153-171.

Fong, C., C. Hazlett, and K. Imai (2018). "Covariate balancing propensity score for a continuous treatment: application to the efficacy of political advertisements". In: Annals of Applied Statistics 12.1, pp. 156-177.

Hirano, K. and G. W. Imbens (2004). "The propensity score with continuous treatments". In: Applied Bayesian Modeling and Causal Inference from Incomplete-Data Perspectives. Ed. by A. Gelman and X. Meng. Hoboken, N.J.: John Wiley & Sons, pp. 73-84.

Holland, P. W. (1986). "Statistics and Causal Inference". In: Journal of the American Statistical Association 81.396, pp. 945-960.

Imai, K. and D. A. Van Dyk (2004). "Causal inference with general treatment regimes: Generalizing the propensity score". In: Journal of the American Statistical Association 99.467, pp. 854-866.

Imbens, G. W. (2000). "The role of the propensity score in estimating dose-response functions". In: Biometrika 87.3, pp. 706-710.

Jiang, M. and E. M. Foster (2013). "Duration of breastfeeding and childhood obesity: a generalized propensity score approach". In: Health Services Research 48.2.1, pp. 628-651.

Kang, J. D. and J. L. Schafer (2007). "Demystifying double robustness: A comparison of alternative strategies for estimating a population mean from incomplete data". In: Statistical Science 22.4, pp. 523-539.

Kluve, J., H. Schneider, A. Uhlendorff, and Z. Zhao (2012). "Evaluating continuous training programmes by using the generalized propensity score". In: Journal of the Royal Statistical Society: Series A 175.2, pp. 587-617.

Kreif, N., R. Grieve, I. Díaz, and D. Harrison (2015). "Evaluation of the effect of a continuous treatment: A machine learning approach with an application to treatment for traumatic brain injury". In: Health Economics 24.9, pp. 1213-1228.

35/37

References II

Pearl, J. and D. Mackenzie (2018). The Book of Why: The New Science of Cause and Effect. New York: Basic Books.

PHFE WIC (2010). WIC data 2003-2009: A report on low income families with young children in Los Angeles County. Technical Report. Special Supplemental Nutrition Assistance Program for Women, Infants and Children.

Rosenbaum, P. R. and D. B. Rubin (1983). "The central role of the propensity score in observational studies for causal effects". In: Biometrika 70.1, pp. 41-55.

Sanders, J. M., M. L. Monogue, T. Z. Jodlowski, and J. B. Cutrell (2020). "Pharmacologic Treatments for Coronavirus Disease 2019 (COVID-19): A Review". In: JAMA 323.18, pp. 1824-1836.

Tübbicke, S. (2020). "Entropy Balancing for Continuous Treatments". In: arXiv preprint arXiv:2001.06281. arXiv: 2001.06281 [econ.EM].

Vegetabile, B. G., B. A. Griffin, D. L. Coffman, M. Cefalu, et al. (2020). "Nonparametric Estimation of Population Average Dose-Response Curves using Entropy Balancing Weights for Continuous Exposures". In: arXiv preprint arXiv:2003.02938. arXiv: 2003.02938 [stat.ME].

Wang, M. C., C. M. Crespi, L. H. Jiang, T. Nobari, et al. (2018). "Developing an index of dose of exposure to early childhood obesity community interventions". In: Preventive Medicine 111, pp. 135-141.

Zhu, Y., D. L. Coffman, and D. Ghosh (2015). "A boosting algorithm for estimating generalized propensity scores with continuous treatments". In: Journal of Causal Inference 3.1, pp. 25-40.

36/37

Dose Refinements

  1. Radius truncation

  2. Micro exposure inverse distance weighting

  3. Overlapping catchment area adjustment

38/37

Dose Refinement Example

39/37

Introduction

2/37
Paused

Help

Keyboard shortcuts

, , Pg Up, k Go to previous slide
, , Pg Dn, Space, j Go to next slide
Home Go to first slide
End Go to last slide
Number + Return Go to specific slide
b / m / f Toggle blackout / mirrored / fullscreen mode
c Clone slideshow
p Toggle presenter mode
t Restart the presentation timer
?, h Toggle this help
oTile View: Overview of Slides
Esc Back to slideshow