Report

Functional Connectivity: PPI and beta Series With thanks to Derek Nee & Bob Spunt 1 Localization vs integration • Localization • Integration – What areas of the brain respond to experimental manipulation? – How do regions of the brain influence each other? – How is this influence affected by experimental manipulation? – Localize functions to distinct regions of the brain – Mechanize functions to brain interactions 2 Basic Idea 3 Some common approaches • Between subjects functional connectivity • Time series correlations • Beta Series – Look for changes in correlation as a function of condition • Are X and Y more tightly coupled in condition A compared to condition B? • Psychophysiological Interaction (PPI) – Look for changes in the regression slope as a function of condition • Does more X activation produce more Y activation in condition A compared to condition B? 4 Between subjects correlation • Do participants who tend to show increased brain activity in region X also tend to show increased brain activity in region Y for a specific contrast? 5 Example 6 Why/ How Example • Do people who show more activity in DMPFC also show more activity in other regions associated with mentalizing? – As opposed to the appearance of a “network” coming from multiple different people activating sub regions • Slightly stronger evidence 7 How to do it • Method 1 (ROI method): – Extract parameter estimates at group level from a priori hypothesized ROIs – Examine their correlations with one another • Method 2 (whole brain search) – Extract PEs at group level from an a priori hypothesized ROI or peak voxel in a theoretically relevant cluster – Regress onto brain activity in whole brain analysis at group level • Variants (see yesterday’s lecture) 8 Between subjects connectivity Strengths • Very simple to run • Very simple to understand • Easy to combine with other individual difference measures Limits • Throws out a lot of temporal information • Does not actually get at whether regions are coactive during the task (only individual differences across people) • No ability to make causal inference 9 Within subject approaches • For a given seed region – Find areas that show changes in their relationship with the seed region • Within conditions • As a function of different task conditions • Beta series– takes advantage of within trial variation • PPI– treats within trail variability as noise in a more traditional interaction analysis 10 Example • What brain regions is DMPFC working with during attribution? – i.e., “why” in the how/why task 11 Standard GLM • Typical GLM for our experiment: Y = β0 + Xwhyβwhy + Xhowβhow + ε Xwhy is predictor for ‘why’ condition Xhow is predictor for ‘how’ condition • Trials are combined into a single predictor – Individual trial variation considered noise Trial1 Trial2 Trial3 TrialN 12 Beta series GLM • Beta series method assumes that individual trial variation is meaningful • For a given seed region, what other regions show similar trial-bytrial variability? – i.e. simple correlation • To examine between-trial variability, need a separate predictor for each trial Y = β0 + Xwhy1βwhy1 + Xwhy2βwhy2 + Xwhy3βwhy3 + … + XwhyNβwhyN + Xhow1βhow1 + Xhow2βhow2 + Xhow3βhow3 + … + XhowNβhowN + ε 13 Beta series Each predictor is now replaced with a series of predictors When fit to the GLM, this will yield a series of betas why1 = 1.1 why2 = 1.3 why2 = 0.7 14 Beta series correlation Take beta series from seed region = 1.1 why2 = 1.3 why3 = 0.7 … whyN = 1.8 why1 Yields a correlation map Correlate the seed beta series with the beta series at every other voxel of the brain Map of correlations during why events (note: not real data for this task) 15 Beta series correlation Repeat process for a different event = 0.6 2 = 1.1 3 = 0.3 … N = 1.4 1 Yields a correlation map Correlate the seed beta series with the beta series at every other voxel of the brain Map of correlations during attend how events (note: not real data for this task) 16 Note • Can learn descriptive information about what regions co-vary during specific task conditions • But, to figure out what is specific to our condition of interest… – Logic similar to subtraction analysis in standard GLM analysis 17 Beta series Comparison Examine changes in correlations as a function of condition through simple subtraction _ Note: important to first normalize the correlation maps, so that t-statistics can be performed = Send normalized correlation diff maps (1 per subject) to 2nd level for simple one-sample ttest 18 Selected other examples • Persistence of emotional memories (Ritchey et al., Cerebral Cortex 2008) – Increased connectivity between amygdala and hippocampus during encoding predicts increased temporal durability of emotional memories • Emotional regulation in depression (Heller et al., PNAS 2010) – Decreased NAcc activity in depressed individuals is related to diminished connectivity between NAcc and PFC • Individual differences in financial risk-taking (Samanez-Larkin et al., J Neurosci 2010) – Individuals with reduced connectivity between the NAcc and PFC made more risk-seeking mistakes 19 Beta series evaluation • Pro’s – Allows flexible modeling • Good for multi-event per trial designs • Tease apart sub parts of psychological process – After 1st level GLM is estimated, can repeat correlations on any number of seeds and conditions – Relatively more powerful for event related designs • Con’s – No directionality of inference – Individual beta estimates are noisy – Massive 1st level model • All the beta images take a lot of harddrive space – No precooked SPM implementation – Relatively less powerful for block designs 20 Within subject approaches • Beta Series – Look for changes in correlation as a function of condition • Are X and Y more tightly coupled in condition A compared to condition B? • Psychophysiological Interaction (PPI) – Look for changes in the regression slope as a function of condition • Does more X activation produce more Y activation in condition A compared to condition B? 21 PsychoPhysiological Interaction (PPI) • Specifies the GLM with 3 predictors of interest – 1) Psychological term • Contrast of interest – E.g. why – how – 2) Physiological term • Time series from seed region – E.g. DMPFC – 3) Interaction term • Psych X Phys • Interaction of the seed time series with the psychological contrast of interest 22 PPI GLM Y = β0 + (why – how)β1 + DMPFCβ2 + (why - how)*DMPFCβ3 + ε • Hypothesis: Physiological variable TPJ Activation Psychological variable – H0: β3 = 0, there is no interaction – Ha: β3 > 0, positive interaction Interaction why how DMPFC Activation 23 PPI deconvolution DMPFC BOLD • Accomplished by Deconvolve – Taking BOLD signal – Deconvolving to putative neuronal inputs – Computing interactions at neural input level – Convolving with HRF to predict BOLD signal why – how DMPFC Neural X Reconvolve Gitelman et al., 2003, NeuroImage 24 Interpreting PPIs (do not make causal claims) Attribution • 2 possible interpretations: – 1) Contribution of the source area to the seed area response (or vice versa) depends upon experimental context • E.g. DMPFC input to TPJ is modulated by attribution DMP FC Attributi on – 2) Seed response to experimental context depends on activation in the source area (or vice versa) • E.g. Effect of attribution on TPJ is modulated by DMPFC TPJ DMP FC TPJ 25 PPI in SPM • First, must perform standard GLM analysis • 1) Create a volume of interest (VOI) – Examine results, go to seed and click “Eigenvariate” • • • • • Will need to name the VOI (e.g. DMPFC_1) Specify session (e.g. 1) Define VOI shape (e.g. sphere, box, cluster) Repeat for each session Each VOI will be saved (e.g. “VOI_DMPFC_1.mat”) 26 PPI In SPM • 2) click “PPIs” in the main menu – Select the standard GLM’s “SPM.mat” – Select “psychophysiological interaction” – SPM will go through each predictor in the standard GLM and ask if you want to include it as part of the Psychological variable • If included, set a weight (i.e. 1 for why, -1 for how) – Name the PPI (e.g. DMPFC_why-how1) – Repeat for each session – Each PPI will be saved (e.g. “PPI_ DMPFC_whyhow1.mat”) 27 PPI IN SPM • 3) Specify a new GLM: a GLM for PPI – Each of the saved PPI_.mat files contains the 3 regressors of interest – PPI.ppi – the interaction – PPI.P – the psychological term – PPI.Y – the physiological term – For each session, load the appropriate PPI_.mat file in MATLAB and type the above variables in as regressors – Include any other nuisance regressors you normally would (e.g. motion regressors) 28 PPI In SPM • 4) After estimating, the contrast is simply a 1 for the interaction term (e.g. [1 0 0 0] for the design to the right) • 5) Submit the interaction contrasts from each subject to second-level one-sample t-test • For more precise details on each step and a tutorial data set, consult the SPM8 manual 29 PPI Pros and Cons • Pro’s – Model-based with an approximated neuronal input structure – Implemented in SPM • Con’s – New model for each seed – New model for each psychological contrast – Optimized for simple (e.g. 2-condition) designs, but may not be suitable for more complex designs • See http://www.nitrc.org/projects/gppi/ for a potential solution to this – Claims to be “effective connectivity”, but still is not much more than a simple correlation 30 Comparison of PPI and beta series • gPPI and beta series produced bigger effects than sPPI – Modeling each condition separately may produce better effects that treating the contrast in one step • A comparison of statistical methods for detecting context-modulated functional connectivity in fMRI • Cisler, Bush & Steele, 2014, Neuroimage 31 Selected shortcomings • Both beta series and PPI require a task – Scott will talk about task-free/resting-state connectivity • Both beta series and PPI requires specification of seeds – Places strong constraints on revealed networks – May prefer a data driven approach • Neither beta series nor PPI specify direction of influence – May want methods to examine effective connectivity • Scott and Luis will cover methods that are well-suited to address these shortcomings 32 Questions? 33