Granger Mediation Analysis of Functional Magnetic Resonance Imaging Time Series

Similar documents
arxiv: v4 [stat.ap] 7 Jul 2017

arxiv: v1 [stat.me] 15 Sep 2017

Mediation Analysis in Neuroimaging Studies

Package gma. September 19, 2017

MIXED EFFECTS MODELS FOR TIME SERIES

Functional Causal Mediation Analysis with an Application to Brain Connectivity. Martin Lindquist Department of Biostatistics Johns Hopkins University

A hierarchical group ICA model for assessing covariate effects on brain functional networks

Functional Mediation Analysis with an Application to Functional Magnetic Resonance Imaging Data

Structural Nested Mean Models for Assessing Time-Varying Effect Moderation. Daniel Almirall

Dynamic Causal Modelling for fmri

Flexible mediation analysis in the presence of non-linear relations: beyond the mediation formula.

HST 583 FUNCTIONAL MAGNETIC RESONANCE IMAGING DATA ANALYSIS AND ACQUISITION A REVIEW OF STATISTICS FOR FMRI DATA ANALYSIS

Effective Connectivity & Dynamic Causal Modelling

Causal modeling of fmri: temporal precedence and spatial exploration

Modelling temporal structure (in noise and signal)

Causal Inference for Mediation Effects

Extracting fmri features

Modern Mediation Analysis Methods in the Social Sciences

Data Analysis I: Single Subject

High-dimensional Multivariate Mediation with Application to Neuroimaging Data

University of California, Berkeley

Causal Mechanisms Short Course Part II:

Structural Nested Mean Models for Assessing Time-Varying Effect Moderation. Daniel Almirall

Causal mediation analysis: Definition of effects and common identification assumptions

Casual Mediation Analysis

13.1 Causal effects with continuous mediator and. predictors in their equations. The definitions for the direct, total indirect,

Discussion of Papers on the Extensions of Propensity Score

An Introduction to Causal Mediation Analysis. Xu Qin University of Chicago Presented at the Central Iowa R User Group Meetup Aug 10, 2016

Conceptual overview: Techniques for establishing causal pathways in programs and policies

Statistical Methods for Causal Mediation Analysis

Computationally Efficient Estimation of Multilevel High-Dimensional Latent Variable Models

Title. Description. var intro Introduction to vector autoregressive models

Ratio of Mediator Probability Weighting for Estimating Natural Direct and Indirect Effects

Some Theories about Backfitting Algorithm for Varying Coefficient Partially Linear Model

Statistical Analysis Aspects of Resting State Functional Connectivity

The General Linear Model (GLM)

An Efficient Estimation Method for Longitudinal Surveys with Monotone Missing Data

Functional Connectivity and Network Methods

9 Graphical modelling of dynamic relationships in multivariate time series

+ + ( + ) = Linear recurrent networks. Simpler, much more amenable to analytic treatment E.g. by choosing

When Should We Use Linear Fixed Effects Regression Models for Causal Inference with Longitudinal Data?

A Longitudinal Look at Longitudinal Mediation Models

Causal inference in epidemiological practice

Extending causal inferences from a randomized trial to a target population

Statistical Analysis of Causal Mechanisms

On the Power of Tests for Regime Switching

Revision list for Pearl s THE FOUNDATIONS OF CAUSAL INFERENCE

Optimization of Designs for fmri

Help! Statistics! Mediation Analysis

SEM REX B KLINE CONCORDIA D. MODERATION, MEDIATION

For more information about how to cite these materials visit

EEG/MEG Inverse Solution Driven by fmri

New Procedures for False Discovery Control

Bootstrapping the Grainger Causality Test With Integrated Data

State-space Model. Eduardo Rossi University of Pavia. November Rossi State-space Model Fin. Econometrics / 53

Specification Errors, Measurement Errors, Confounding

The General Linear Model. Guillaume Flandin Wellcome Trust Centre for Neuroimaging University College London

Causal Inference with a Continuous Treatment and Outcome: Alternative Estimators for Parametric Dose-Response Functions

Quasi-likelihood Scan Statistics for Detection of

Using Estimating Equations for Spatially Correlated A

Ph.D. Qualifying Exam Friday Saturday, January 6 7, 2017

GARCH Models. Eduardo Rossi University of Pavia. December Rossi GARCH Financial Econometrics / 50

Experimental design of fmri studies

Specifying Latent Curve and Other Growth Models Using Mplus. (Revised )

Bayesian Inference. Chapter 4: Regression and Hierarchical Models

When Should We Use Linear Fixed Effects Regression Models for Causal Inference with Longitudinal Data?

2012 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media,

arxiv: v1 [stat.me] 15 May 2011

MFE Financial Econometrics 2018 Final Exam Model Solutions

Mediation and Interaction Analysis

Integrative Methods for Functional and Structural Connectivity

SC705: Advanced Statistics Instructor: Natasha Sarkisian Class notes: Introduction to Structural Equation Modeling (SEM)

New Machine Learning Methods for Neuroimaging

University of California, Berkeley

Exploratory Causal Analysis in Bivariate Time Series Data Abstract

Causal Effect Estimation Under Linear and Log- Linear Structural Nested Mean Models in the Presence of Unmeasured Confounding

Measurement error effects on bias and variance in two-stage regression, with application to air pollution epidemiology

F & B Approaches to a simple model

Beyond Univariate Analyses: Multivariate Modeling of Functional Neuroimaging Data

Propensity Score Weighting with Multilevel Data

Stochastic Dynamic Causal Modelling for resting-state fmri

Identification and Inference in Causal Mediation Analysis

Introduction to Simple Linear Regression

Identification, Inference, and Sensitivity Analysis for Causal Mediation Effects

The General Linear Model (GLM)

The ASL signal. Parenchy mal signal. Venous signal. Arterial signal. Input Function (Label) Dispersion: (t e -kt ) Relaxation: (e -t/t1a )

Partial factor modeling: predictor-dependent shrinkage for linear regression

Model Specification Testing in Nonparametric and Semiparametric Time Series Econometrics. Jiti Gao

Abstract Title Page. Title: Degenerate Power in Multilevel Mediation: The Non-monotonic Relationship Between Power & Effect Size

Bayesian modelling of fmri time series

Bayesian Inference. Chapter 4: Regression and Hierarchical Models

Spatial Source Filtering. Outline EEG/ERP. ERPs) Event-related Potentials (ERPs( EEG

Inference With Interference Between Units in an f MRI Experiment of Motor Inhibition

Experimental design of fmri studies & Resting-State fmri

Towards a Regression using Tensors

A Unified Estimation Framework for State- Related Changes in Effective Brain Connectivity

Approximate Bayesian Computation

Unpacking the Black-Box: Learning about Causal Mechanisms from Experimental and Observational Studies

SIMPLE EXAMPLES OF ESTIMATING CAUSAL EFFECTS USING TARGETED MAXIMUM LIKELIHOOD ESTIMATION

State-space Model. Eduardo Rossi University of Pavia. November Rossi State-space Model Financial Econometrics / 49

Transcription:

Granger Mediation Analysis of Functional Magnetic Resonance Imaging Time Series Yi Zhao and Xi Luo Department of Biostatistics Brown University June 8, 2017

Overview 1 Introduction 2 Model and Method 3 Simulation Study 4 An fmri Study 5 Discussion

Introduction

Motivation Credit: NSF Task-related functional MRI (fmri) fmri: measures brain activities task fmri: perform task under fmri scanner response conflict task GO trial: push the button STOP trial: withhold the pushing 1 / 29

Motivation Credit: NSF Task-related functional MRI (fmri) fmri: measures brain activities task fmri: perform task under fmri scanner response conflict task GO trial: push the button STOP trial: withhold the pushing 1 / 29

Motivation Task-related functional MRI (fmri) fmri: measures brain activities Credit: NSF task fmri: perform task under fmri scanner response conflict task GO trial: push the button STOP trial: withhold the pushing Objective identify causal effects of task stimulus on brain activity infer brain connectivity (effective connectivity) 1 / 29

Functional MRI Data 10 5 10 6 uniformly spaced voxels 264 putative functional regions 1 1 Power et al., Neuron, 2011 2 / 29

Functional MRI Data 10 5 10 6 uniformly spaced voxels 264 putative functional regions 1 1 Power et al., Neuron, 2011 2 / 29

Functional MRI Data 105 106 uniformly spaced voxels 264 putative functional regions1 time series or functional data 1 Power et al., Neuron, 2011 2 / 29

Response conflict task Brain regions of interest primary motor cortex (PMC): responsible for movement presupplementary motor area (presma): primary region for motor response prohibition Objective Quantify the causal effects stimulus presma, stimulus PMC presma PMC 2 2 Obeso et al., Brain Stimulation, 2013 3 / 29

Response conflict task Brain regions of interest primary motor cortex (PMC): responsible for movement presupplementary motor area (presma): primary region for motor response prohibition Objective Quantify the causal effects stimulus presma, stimulus PMC presma PMC 2 2 Obeso et al., Brain Stimulation, 2013 3 / 29

Brain activities can be modeled as a linear superposition of task and spontaneous random fluctuations 3 task-related signal stimulus presma random fluctuation stimulus PMC 3 Cole et al., Neuron, 2014. 4 / 29

Brain activities can be modeled as a linear superposition of task and spontaneous random fluctuations 3 task-related signal stimulus presma random fluctuation stimulus PMC 3 Cole et al., Neuron, 2014. 4 / 29

Brain activities can be modeled as a linear superposition of task and spontaneous random fluctuations 3 task-related signal stimulus presma random fluctuation stimulus PMC 3 Cole et al., Neuron, 2014. 4 / 29

A E 1 presma (M) B Unmeasured confounding (U) E 2 Stimulus (Z) C PMC (R) M = ZA + E 1 R = ZC + MB + E 2 Task-related signal indirect effect (AB): stimulus presma PMC direct effect (C): stimulus PMC Random fluctuation spatiotemporal dependency Unmeasured confounding random fluctuation of other regions 4 ; system errors: head motion 5 4 Fox et al., Nature Neuroscience (2006), Mason et al., Science (2007) 5 Sobel and Lindquist, JASA, 2014 5 / 29

A E 1 presma (M) B Unmeasured confounding (U) E 2 Stimulus (Z) C PMC (R) M = ZA + E 1 R = ZC + MB + E 2 Task-related signal indirect effect (AB): stimulus presma PMC direct effect (C): stimulus PMC Random fluctuation spatiotemporal dependency Unmeasured confounding random fluctuation of other regions 4 ; system errors: head motion 5 4 Fox et al., Nature Neuroscience (2006), Mason et al., Science (2007) 5 Sobel and Lindquist, JASA, 2014 5 / 29

A E 1 presma (M) B Unmeasured confounding (U) E 2 Stimulus (Z) C PMC (R) M = ZA + E 1 R = ZC + MB + E 2 Task-related signal indirect effect (AB): stimulus presma PMC direct effect (C): stimulus PMC Random fluctuation spatiotemporal dependency Unmeasured confounding random fluctuation of other regions 4 ; system errors: head motion 5 4 Fox et al., Nature Neuroscience (2006), Mason et al., Science (2007) 5 Sobel and Lindquist, JASA, 2014 5 / 29

A E 1 presma (M) B Unmeasured confounding (U) E 2 Stimulus (Z) C PMC (R) M = ZA + E 1 R = ZC + MB + E 2 Task-related signal indirect effect (AB): stimulus presma PMC direct effect (C): stimulus PMC Random fluctuation spatiotemporal dependency Unmeasured confounding random fluctuation of other regions 4 ; system errors: head motion 5 4 Fox et al., Nature Neuroscience (2006), Mason et al., Science (2007) 5 Sobel and Lindquist, JASA, 2014 5 / 29

Data structure Randomized stimulus Participant 1 presma time series Participant 2 PMC time series. Causal effects for Participant 1 Participant N 6 / 29

Data structure Randomized stimulus Participant 1 presma time series Participant 2 PMC time series. Causal effects for Participant 1 Participant N 6 / 29

Major challenges Unmeasured confounding Iterregional and temporal dependency Individual variation 7 / 29

Existing Methods Mediation Baron and Kenny (1996), Imai et al. (2010) Kenny et al. (2003): multilevel mediation Ten Have et al. (2007), Small (2011): instrumental variables Lindquist (2012): IV and functional mediation Robins et al. (2000), van der Laan and Petersen (2008), VanderWeele (2009): longitudinal mediation Granger causality (time series) Granger (1969) Granger causality of two time series time series y is said to Granger cause time series x if the current value of x can be predicted by the past values of x and y Goebel et al. (2003), Harrison et al. (2003), Valdes-Sosa (2004) vector autoregression (VAR) of fmri time series 8 / 29

Multilevel Granger mediation model Lower-level model fmri time series of one participant causal mediation + spatiotemporal dependency unmeasured confounding nonidentifiability issue Higher-level model individual variation identifiability and consistency 9 / 29

Model and Method

Lower-level model Z 1 Z 2 Z T 1 Z T M 1 ẽ 11 M 2 ẽ 12 M T 1 ẽ 1,T 1 M T ẽ 1T E 11 E 12 E 1,T 1 E 1T R 1 R 2 R T 1 R T E 21 E 22 E 2,T 1 E 2T ẽ 21 ẽ 22 ẽ 2,T 1 ẽ 2T Time 1 Time 2 Time T 1 Time T task-related signal: Z t M t, Z t R t, Z t M t R t spontaneous random fluctuations: E 1t M t, E 2t R t unmeasured confounding: U t 10 / 29

Lower-level model Z 1 Z 2 Z T 1 Z T M 1 ẽ 11 M 2 ẽ 12 M T 1 ẽ 1,T 1 M T ẽ 1T E 11 E 12 E 1,T 1 E 1T R 1 R 2 R T 1 R T E 21 E 22 E 2,T 1 E 2T ẽ 21 ẽ 22 ẽ 2,T 1 ẽ 2T Time 1 Time 2 Time T 1 Time T task-related signal: Z t M t, Z t R t, Z t M t R t spontaneous random fluctuations: E 1t M t, E 2t R t unmeasured confounding: U t 10 / 29

Lower-level model Z 1 Z 2 Z T 1 Z T M 1 ẽ 11 M 2 ẽ 12 M T 1 ẽ 1,T 1 M T ẽ 1T E 11 E 12 E 1,T 1 E 1T R 1 R 2 R T 1 R T E 21 E 22 E 2,T 1 E 2T ẽ 21 U 1 ẽ 22 U 2 ẽ 2,T 1 U T 1 ẽ 2T U T Time 1 Time 2 Time T 1 Time T task-related signal: Z t M t, Z t R t, Z t M t R t spontaneous random fluctuations: E 1t M t, E 2t R t unmeasured confounding: U t 10 / 29

Causal mediation with VAR(p) autoregressive error Regression models: task-induced brain activations Structural equation modeling for causal mediation M t = Z t A + E 1t R t = Z t C + M t B + E 2t Remaining residuals: resting-state random fluctuations 6 VAR(p) model E 1t = p j=1 ω 11 j E 1,t j + p j=1 ω 21 j E 2,t j + ɛ 1t E 2t = p j=1 ω 12 j E 1,t j + p j=1 ω 22 j E 2,t j + ɛ 2t ( ) ( ) ɛ1t σ 2 N (0, Σ), Σ = 1 δσ 1 δ 2 ɛ 2t δσ 1 σ 2 σ2 2 (ɛ 1t, ɛ 2t) Gaussian white noise, (ɛ 1t, ɛ 2t) 6 Fair et al., NeuroImage,2007 = (ɛ 1s, ɛ 2t) for s t 11 / 29

Causal mediation with VAR(p) autoregressive error Regression models: task-induced brain activations Structural equation modeling for causal mediation M t = Z t A + E 1t R t = Z t C + M t B + E 2t Remaining residuals: resting-state random fluctuations 6 VAR(p) model E 1t = p j=1 ω 11 j E 1,t j + p j=1 ω 21 j E 2,t j + ɛ 1t E 2t = p j=1 ω 12 j E 1,t j + p j=1 ω 22 j E 2,t j + ɛ 2t ( ) ( ) ɛ1t σ 2 N (0, Σ), Σ = 1 δσ 1 δ 2 ɛ 2t δσ 1 σ 2 σ2 2 (ɛ 1t, ɛ 2t) Gaussian white noise, (ɛ 1t, ɛ 2t) 6 Fair et al., NeuroImage,2007 = (ɛ 1s, ɛ 2t) for s t 11 / 29

Z 1 Z 2 Z T 1 Z T M 1 ẽ 11 M 2 ẽ 12 M T 1 ẽ 1,T 1 M T ẽ 1T E 11 E 12 E 1,T 1 E 1T R 1 R 2 R T 1 R T E 21 E 22 E 2,T 1 E 2T ẽ 21 U 1 ẽ 22 U 2 ẽ 2,T 1 U T 1 ẽ 2T U T E 1t = (ω 11j E 1,t j + ω 21j E 2,t j) + ν 1U t + ẽ 1t = (ω 11j E 1,t j + ω 21j E 2,t j) + ɛ 1t j E 2t = (ω 12j E 1,t j + ω 22j E 2,t j) + ν 2U t + ẽ 2t = (ω 12j E 1,t j + ω 22j E 2,t j) + ɛ 2t j j j Assumption: additive unmeasured confounding ν 1 ν 2 0 ɛ 1t, ɛ 2t correlated 12 / 29

Z 1 Z 2 Z T 1 Z T M 1 ẽ 11 M 2 ẽ 12 M T 1 ẽ 1,T 1 M T ẽ 1T E 11 E 12 E 1,T 1 E 1T R 1 R 2 R T 1 R T E 21 E 22 E 2,T 1 E 2T ẽ 21 U 1 ẽ 22 U 2 ẽ 2,T 1 U T 1 ẽ 2T U T E 1t = (ω 11j E 1,t j + ω 21j E 2,t j) + ν 1U t + ẽ 1t = (ω 11j E 1,t j + ω 21j E 2,t j) + ɛ 1t j E 2t = (ω 12j E 1,t j + ω 22j E 2,t j) + ν 2U t + ẽ 2t = (ω 12j E 1,t j + ω 22j E 2,t j) + ɛ 2t j j j Assumption: additive unmeasured confounding ν 1 ν 2 0 ɛ 1t, ɛ 2t correlated 12 / 29

Assumptions Mediation model (A1) no interference the treatment assignment regime is the same (A2) models are correctly specified (A3) Z t is randomized with positive probabilities (A4) no unmeasured confounding effect on mediator-outcome relationship unmeasured confounding U t only affects current time mediator and outcome, and the effect is additive Granger causality (A5) {(E 1t, E 2t )} t stationary 7 7 Chang and Glover, NeuroImage, 2010. 13 / 29

Granger mediation analysis (GMA): causal mediation SEM + VAR(p) M t = Z t A + E 1t E 1t = p j=1, ω 11 j E 1,t j + p j=1 ω 21 j E 2,t j + ɛ 1t R t = Z t C + M t B + E 2t E 2t = p j=1 ω 12 j E 1,t j + p j=1 ω 22 j E 2,t j + ɛ 2t (E 1t, E 2t ) s not independent difficult to derive estimators in explicit form (ɛ 1t, ɛ 2t ) Gaussian white noise: (ɛ 1t, ɛ 2t ) = (ɛ 1s, ɛ 2s ) for t s reparametrize the model M t = Z t A + p j=1 φ 1 j Z t j + p j=1 ψ 11 j M t j + p j=1 ψ 21 j R t j + ɛ 1t R t = Z t C + M t B + p j=1 φ 2 j Z t j + p j=1 ψ 12 j M t j + p j=1 ψ 22 j R t j + ɛ 2t 14 / 29

GMA { M t = Z ta + E 1t R t = Z tc + M tb + E 2t { E 1t = j ω 11 j E 1,t j + j ω 21 j E 2,t j + ɛ 1t E 2t = j ω 12 j E 1,t j + j ω 22 j E 2,t j + ɛ 2t Reparametrized model M t = Z ta + j φ 1 j Z t j + j ψ 11 j M t j + j ψ 21 j R t j + ɛ 1t R t = Z tc + M tb + j φ 2 j Z t j + j ψ 12 j M t j + j ψ 22 j R t j + ɛ 2t Z t Z t+1 Z t Z t+1 ẽ 1t ẽ 1,t+1 M t ẽ 1t M t+1 ẽ 1,t+1 M t M t+1 E 1t E 1,t+1 R t R t+1 R t R t+1 E 2t E 2,t+1 ẽ 2t U t ẽ 2,t+1 U t+1 ẽ 2t U t ẽ 2,t+1 U t+1 Time t Time (t + 1) Time t Time (t + 1) 15 / 29

GMA { M t = Z ta + E 1t R t = Z tc + M tb + E 2t { E 1t = j ω 11 j E 1,t j + j ω 21 j E 2,t j + ɛ 1t E 2t = j ω 12 j E 1,t j + j ω 22 j E 2,t j + ɛ 2t Reparametrized model M t = Z ta + j φ 1 j Z t j + j ψ 11 j M t j + j ψ 21 j R t j + ɛ 1t R t = Z tc + M tb + j φ 2 j Z t j + j ψ 12 j M t j + j ψ 22 j R t j + ɛ 2t Z t Z t+1 Z t Z t+1 ẽ 1t ẽ 1,t+1 M t ẽ 1t M t+1 ẽ 1,t+1 M t M t+1 E 1t E 1,t+1 R t R t+1 R t R t+1 E 2t E 2,t+1 ẽ 2t U t ẽ 2,t+1 U t+1 ẽ 2t U t ẽ 2,t+1 U t+1 Time t Time (t + 1) Time t Time (t + 1) 15 / 29

GMA (causal mediation SEM + VAR(p)) { { M t = Z ta + E 1t E 1t = j, ω11 j E 1,t j + j ω21 j E 2,t j + ɛ 1t R t = Z tc + M tb + E 2t E 2t = j ω12 j E 1,t j + j ω22 j E 2,t j + ɛ 2t Lemma Reparametrized model { M t = Z ta + p j=1 φ1 j Z t j + p j=1 ψ11 j M t j + p j=1 ψ21 j R t j + ɛ 1t R t = Z tc + M tb + p j=1 φ2 j Z t j + p j=1 ψ12 j M t j + p j=1 ψ22 j R t j + ɛ 2t Given (φ j, Ψ j ) and (A, B, C), Ω j can be uniquely determined. 16 / 29

M t = Z t A + p j=1 φ 1 j Z t j + p j=1 ψ 11 j M t j + p j=1 ψ 21 j R t j + ɛ 1t R t = Z t C + M t B + p j=1 φ 2 j Z t j + p j=1 ψ 12 j M t j + p j=1 ψ 22 j R t j + ɛ 2t Given δ = Cor(ɛ 1t, ɛ 2t ), estimate the rest by conditional likelihood Let θ 1 = {A, {φ 1j }, {ψ 11j }, {ψ 21j }}, θ 2 = {C, {φ 2j }, {ψ 12j }, {ψ 22j }}, ˆθ 1 = (X X) 1 X M, ˆσ 1 2 = M (I P X )M/(T p) ˆθ 2 = (X (I P M )X) 1 X (I P M )R + δˆσ 2/ˆσ 1(X X) 1 X M ˆB = (M M) 1 M(I X(X (I P M )X) 1 X (I P M ))R δˆσ 2/ˆσ 1 ˆσ 2 2 = R (I P MX P M )R/(T p)(1 δ 2 ) 17 / 29

M t = Z t A + p j=1 φ 1 j Z t j + p j=1 ψ 11 j M t j + p j=1 ψ 21 j R t j + ɛ 1t R t = Z t C + M t B + p j=1 φ 2 j Z t j + p j=1 ψ 12 j M t j + p j=1 ψ 22 j R t j + ɛ 2t Given δ = Cor(ɛ 1t, ɛ 2t ), estimate the rest by conditional likelihood Let θ 1 = {A, {φ 1j }, {ψ 11j }, {ψ 21j }}, θ 2 = {C, {φ 2j }, {ψ 12j }, {ψ 22j }}, ˆθ 1 = (X X) 1 X M, ˆσ 2 1 = M (I P X )M/(T p) ˆθ 2 = (X (I P M )X) 1 X (I P M )R + δˆσ 2/ˆσ 1(X X) 1 X M ˆB = (M M) 1 M(I X(X (I P M )X) 1 X (I P M ))R δˆσ 2/ˆσ 1 ˆσ 2 2 = R (I P MX P M )R/(T p)(1 δ 2 ) Theorem For fixed δ ( 1, 1), given the initial p observations, the conditional likelihood of the reparametrized model achieves the same maximum. 17 / 29

(δ, θ 2, B, σ 2 ) not identifiable under lower-level model time series of one participant fmri study with hierarchically nested data structure participants time series individual variation in the causal effects and overparametrization in the lower-level model 18 / 29

Lower-level model: for participant i M it = Z it A i + E i1t E i1t = j, ω i 11j E i1,t j + j ω i 21j E i2,t j + ɛ i1t R it = Z it C i + M it B i + E i2t E i2t = j ω i 12j E i1,t j + j ω i 22j E i2,t j + ɛ i2t Higher-level model: linear model for the coefficients A i A ɛ A i b i = B i = B + ɛ B i = b + η i C i C ɛ C i A, B, and C are the population-level causal effects of interest η i N (0, Λ) (δ i, θ i2, B i, σ 2i ) not identifiable Assumption (A6): δ i = δ across participants 19 / 29

l(υ) = N T N log P(R it, M it Z it, Z (p) i, M (p) t 1 i, R (p) t 1 i, θ i1, θ i2, B t 1 i, δ, σ 1i, σ 2i ) + log P(b i b, Λ) i=1 } t=1 {{ } l 1 : lower-level time series i=1 }{{} l 2 : causal coefficients Υ = {δ, b, Λ, (θ i1, θ i2, B i ), (σ 1i, σ 2i )} optimization problem max Υ:{Λ,(σ 1i,σ 2i )} S l(υ) S: constraint set for variance components 20 / 29

l(υ) = N T N log P(R it, M it Z it, Z (p) i, M (p) t 1 i, R (p) t 1 i, θ i1, θ i2, B t 1 i, δ, σ 1i, σ 2i ) + log P(b i b, Λ) i=1 } t=1 {{ } l 1 : lower-level time series i=1 }{{} l 2 : causal coefficients Υ = {δ, b, Λ, (θ i1, θ i2, B i ), (σ 1i, σ 2i )} optimization problem max Υ:{Λ,(σ 1i,σ 2i )} S l(υ) S: constraint set for variance components Theorem Given δ, negative l is conditional convex. 20 / 29

l(υ) = N T N log P(R it, M it Z it, Z (p) i, M (p) t 1 i, R (p) t 1 i, θ i1, θ i2, B t 1 i, δ, σ 1i, σ 2i ) + log P(b i b, Λ) i=1 } t=1 {{ } l 1 : lower-level time series i=1 }{{} l 2 : causal coefficients Block coordinate-descent optimize over l(υ) joint optimization of a large number of parameters fast block search for conditional convex function Two-stage optimize over l 1 first, and then over l 2 easy to compute b i s latent parameters, not actually observed 21 / 29

l(υ) = N T N log P(R it, M it Z it, Z (p) i, M (p) t 1 i, R (p) t 1 i, θ i1, θ i2, B t 1 i, δ, σ 1i, σ 2i ) + log P(b i b, Λ) i=1 } t=1 {{ } l 1 : lower-level time series i=1 }{{} l 2 : causal coefficients Block coordinate-descent Two-stage log likelihood (l 1 + l2) 34400 34350 34300 δ^ = 0.492 higher level log likelihood (l 2) 480 460 440 420 400 380 360 δ^ = 0.466 0.5 0.0 0.5 δ 0.5 0.0 0.5 δ 21 / 29

Simulation Study

Simulation Study N = 50 subjects, T i Poisson(100), Z it Bern(0.5) Each time point is a randomized trial A = 0.5, B = 1, C = 0.5 Λ = 0.5I 3, σ i1 = 1, σ i2 = 2 (E 1t, E 2t ) stationary: p = 1, Ω transition matrix, (cov (E 0 ) = Ξ) ( ) ( ) E1t = Ω E1,t 1 + ɛ t, ɛ t N (0, Σ), Ξ = Ω ΞΩ + Σ E 2t E 2,t 1 Methods Granger mediation using block coordinate descent (GMA-h); Granger mediation with two-stage approach (GMA-ts); CMA-h (and CMA-ts) by Zhao and Luo (2014) (R macc package); multilevel SEM (KKB) 8 ; Baron and Kenny (BK) method. 8 Kenny et al., Psychological methods, 2003. 22 / 29

δ estimate ˆδ δ δ^ 0.5 0.0 0.5 GMA h GMA ts CMA h CMA ts True δ^ δ 0.04 0.02 0.00 0.02 GMA h GMA ts CMA h CMA ts 0.5 0.0 0.5 δ 0.5 0.0 0.5 δ 23 / 29

Direct effect Indirect effect 0.5 0.0 0.5 1.0 C^ GMA h GMA ts CMA h KKB (δ = 0) BK (δ = 0) True (C = 0.5) AB^ p 1.0 0.5 0.0 GMA h GMA ts CMA h KKB (δ = 0) BK (δ = 0) True (AB = 0.5) 0.5 0.0 0.5 0.5 0.0 0.5 δ δ 24 / 29

Ω 2 1 0 1 ω^ 11 ω^ 21 ω^ 12 ω^ 22 True 0.5 0.0 0.5 δ GMA-h estimates 25 / 29

Asymptotic property ˆδ ) MSE (ˆδ δ^ 0.44 0.46 0.48 0.50 50 100 200 500 1000 2000 5000 N GMA h GMA ts True (δ = 0.5) MSE(δ^) 0.000 0.002 0.004 0.006 0.008 0.010 50 100 200 500 1000 2000 5000 N GMA h GMA ts N: number of participants; T i = N: number of time points 26 / 29

An fmri Study

fmri study N = 96 with T i = 295 time points randomized STOP/GO trials outcome (R): time series of the primary motor cortex (PMC) mediator (M): time series of the presupplementary motor area (presma) Methods GMA-h, GMA-ts CMA-h KKB BK participant-level bootstrap repeated 500 times 27 / 29

Causal effects Transition matrix Ω 1.0 0.5 0.0 0.5 1.0 GMA h GMA ts CMA h KKB (δ = 0) BK (δ = 0) Ω^ 0.4 0.2 0.0 0.2 0.4 ω^ 11 = 0.401 ω^ 12 = 0.020 ω^ 21 = 0.231 ω^ 22 = 0.344 Direct Indirect (product) δ Indirect (difference) GMA-h 0.095 (-0.012, 0.209) GMA-ts 0.065 (-0.041, 0.178) CMA-h 0.073 (-0.045, 0.201) 0 100 200 300 400 500 ω 21 0: feedback effect of PMC (R) on presma (M) p = 1, estimate of Ω 1 under VAR(1) p = 2, Ω 2 0 28 / 29

Discussion

Discussion 9 Propose a two-level mediation model for fmri time series. remove unmeasured confounding capture both interregional and temporal dependencies address individual variation issue Jointly optimize a large number of parameters Improve estimation accuracy and consistency Identify feedback effect of (outcome) PMC on (mediator) presma Future work theory about δ identifiability and consistency (convergence rate) covariates more general form of the spatio/temporal correlation structure spatial variation in the haemodynamic response function (HRF) relax the concurrent influence assumption of unmeasured confounding 9 2017 JSM Mental Health Section one of two highest rate papers 29 / 29

Acknowledgements Advisor Xi (Rossi) Luo, PhD, Department of Biostatistics, Brown University Dissertation Committee Reader Joseph Hogan, ScD, Department of Biostatistics, Brown University Jerome Sanes, PhD, Department of Neuroscience, Brown University Eli Upfal, PhD, Department of Computer Science, Brown University Yen-Tsung Huang, MD & ScD, Departments of Epidemiology and Biostatistics, Brown University Brian Caffo, PhD, Department of Biostatistics, Johns Hopkins Bloomberg School of Public Health

Thank you!