Home

Effective Connectivity Modeling with the euSEM and GIMME

image

Contents

1. Standardized RMR values around or below 0 05 and NNFI and CFI values should be above 0 95 It is possible to continue improving a model as indicated by significant chi squared scores that satisfies these other diagnostic Effective Connectivity Modeling with the euSEM and GIMME 8 criteria but it remains in the judgment of the researcher to what extent to pursue increased fit versus parsimony of the model Using GIMME however this process of model evaluation is automated The script will automatically search the parameters freeing one at a time until the model is no longer significantly improved GIMME performs this analysis first at the group level searching for parameters that when freed will improve the fit of the majority of participants models and then refine each participant s model to account for individual differences The procedure performing this search it outlined in the next section A brief GIMME user s manual written by the software author Prof Kathleen Gates kmgates unc edu is also available Section II Data Analysis Connectivity analysis of functional neuroimaging data is conducted in several steps Here we will outline the procedure for MRI data using the SPM software package Friston amp Holmes 1995 This procedure is conceptually applicable across different software packages but given the ubiquity of SPM we will focus on this example The three main steps which will compose this analysis are 1 Ext
2. and stimulus data coded into matrices and saved in the proper locations GIMME makes estimation of a connectivity map very easy Before enumerating the steps for doing this estimation though it 1s valuable to have some understanding of the underlying process GIMME 1s a set of Matlab functions which assist the user by composing LISREL syntax to describe the time series data and guide the iterative search process for the best fitting model over the group data To complete this process GIMME will use Matlab s operating Effective Connectivity Modeling with the euSEM and GIMME 10 system interface to call the LISREL program and execute the commands that GIMME has composed for LISREL It is therefore necessary to have LISREL installed in the location where GIMME expects to find it In the following steps the procedure for estimating the connectivity map is outlined For additional technical details about the design and use of GIMME see the GIMME users manual Gates 2012 l Add GIMME to Matlab s executable path addpath C path to GIMME scripts Start the top level script by typing GIMME into the Matlab command line A window will appear with several parameters that need to be defined Number of ROIs Enter the number of regions that you have specified in the time series data for each participant If you participant time series data have 3 columns you have 3 ROIs one time series for each region In that case you would ent
3. connectivity map e final subj mat files for each individual participant describing that participant s individual connectivity map Within each output file you will find data structures containing the regression weights described for Extended Unified SEM in Section I Section III References Friston K J amp Holmes A 1995 Statistical parametric maps in functional imaging a general linear approach Human Brain Mapping 2 189 210 Gates K M 2012 Group Iterative Multiple Model Estimation GIMME User s Manual Retrieved from www personal psu edu kmg3 11 GIMME Using 20GIMME pdf Gates K M amp Molenaar P C M 2012 Group search algorithm recovers effective connectivity maps for individuals in homogeneous and heterogeneous samples Neurolmage 63 310 319 Gates K M Molenaar P C M Hillary F G amp Slobounov S 2011 Extended unified SEM approach for modeling event related MRI data Neurolmage 54 2 1151 8 Kim J Zhu W Chang L Bentler P M Ernst T 2007 Unified structural equation modeling approach for the analysis of multisubject multivariate functional MRI data Human Brain Mapping 28 85 93
4. Effective Connectivity Modeling with the euSEM and GIMME Benjamin Zinszer PhD candidate Department of Psychology MAS candidate Department of Statistics Center for Language Science Pennsylvania State University Supported by Guangwai Brain amp Language Lab Director Prof Yangping Dong 276 Associate Director Dr Jing Yang 4 At www gwball com Version 1 0 18 Nov 2013 Effective Connectivity Modeling with the euSEM and GIMME 2 Section I Connectivity Models The present tutorial will describe the extended unified SEM euSEM Gates Molenaar Hillary amp Slobounov 2011 a model for estimating relationships between multiple regions of interest in neuroimaging data Before addressing the practical steps for preparing and analyzing datasets it is important to have some background on the statistical principles underlying the euSEM First I introduce structural equation modeling SEM and vector auto regression VAR two models commonly used to estimate connectivity and integrated under the euSEM Next I outline the euSEM model and statistical assumptions necessary for applying it to neuroimaging data Finally I describe some of the statistics used in evaluating and comparing euSEM models The practical steps for implementing these statistical models are available in the next section Structural equation modeling or SEM is a statistical tool for comparing causal models between multiple variables by estimating a set of linear reg
5. archer has explicitly assumed that activity in X causes the activity in Y and uses SEM to estimate the magnitude of that connection If in fact activity in B is the cause of activity in A SEM may still yield a significant connection weight coefficient in the wrong X gt Y direction Two approaches for reducing these errors are searching a wide variety of models to compare which direction best explains the observed data comparing the model X Y to the model YX or using time sequence as a causal constraint by regressing Y over X since activity in Y could not cause a change in X backwards in time This second approach is known as vector auto regression Vector autoregression VAR like SEM uses a system of linear regression equations to estimate the connections between ROIs However unlike SEM VAR compares the lagged or delayed in time relationships between ROIs To illustrate this method statistically consider only one ROI X Because X s activation has been measured repeatedly through time it could be described by a time series Many time series have a property called autocorrelation which means that the value of X at time point is correlated with the value of X at a previous time point In the following discussion we will look only at the immediately preceding time point t denoted as a VAR 1 model Autoregression reveals the systematic changes that occur in a variable over time and this relationship is represented by the f
6. er 3 Indicate directory where your data are This is the directory which contains the time series data files for each participant may be csv xls or txt format Make sure that every participant s data file is in that directory no additional directories and make sure there are no other files in that directory as GIMME will attempt to open each of the files it finds Indicate directory where your results should go This directory will be populated with mat files for each participant Critically if you are estimated an euSEM your onsets mat file must also be stored in this directory before you start GIMME TR The value you enter here should be the length of a single scan one TR measured in seconds What type of files are your data in Here you will select the integer value that corresponds to the data type you saved your participants time series data files in The correct choice here is the file type that you used in Preparation of the Data Set when you saved the time series matrices Put 0 to conduct a uSEM 1 for a euSEM If you have a stimulus with specified onsets in the onsets mat file that you would like to include in the estimation of the connectivity map select the euSEM by entering a 1 If you do not want to include Stimulus data enter 0 If this is an euSEM how many inputs are there If you selected 1 for the previous question you must now enter the number of inputs stimuli
7. er When a new model ceases to yield a significant improvement by the addition of new open parameters we conclude that the most parsimonious model has been generated An example model may look like this Bey O Bez 0 0 Byes P 0 By Byz Tu 0 0 0 0 0 LO 60 0 0 0 0 A Py 0 J 4 0 Bay 0 C p 0 J Rar r 10 0 O 0 0 Diagnostic statistics for euSEM model provide a quantitative evaluation of the model s 7 improvement over a baseline model such as the model immediately prior to freeing an additional parameter The chi squared test compares the likelihood of two models to test the null hypothesis that two models are equally fit to the data If we reject this null hypothesis we can conclude that the model with the new parameter significantly improves the model The a test 1s performed by the LISREL software package each time a model 1s estimated If you want to estimate a connectivity model without using the GIMME package you will have to manually update your model by freeing individual parameters re running LISREL and checking the x test for the significance of the improvement Other diagnostics computed by LISREL such as the Root Mean Square Error of Approximation RMSEA Standardized Root Mean square Residual Standardized RMR Non Normed Fit Index NNFI and Comparative Fit Index CFI also provide information on the fit of the model In general a well fit model should produce RMSEA and
8. graphical representation of this model shows causal relationships an arrow pointing from causal ROI to affected ROI in every possible direction Effective Connectivity Modeling with the euSEM and GIMME 3 A more reasonable model might make use of some causal relationships that the researcher hypothesizes For instance perhaps activation is predicted to flow from X to Y to Z This structural equation model would be described by the following regression equations and graphic Xt Bo TE Y B3 B4X xX Z Bo BY Matrix notation X Box TO 0 OXJ ex Y gt Ba t Pyx 9 J J er LZ Faz 0 fy OILZ Ez Estimating the B coefficients in these regression equations yields the estimated connectivity between each region Specifically B4 is the influence of X on Y and fs is the influence of Y on Z Real connectivity studies however are unlikely to begin with such specific models In this case model comparison is necessary to determine which model from either a set of competing hypotheses or through a search process best fits the observed data There are a number of statistics useful for comparing models which will be discussed in a subsequent section One further benefit of SEM is the possibility of estimating exogenous influences on the ROIs Exogenous variables are defined as those which may cause changes in other variables but are not themselves influenced by any othe
9. ill describe the lagged relationships between the endogenous and exogenous variables and Io will describe the contemporaneous relationships e The bilinear interaction terms between the exogenous and endogenous variables will be described as To and qt for time points f and t respectively Finally the matrix will be described as matrices for each n u and bilinear term t i 0 0 O QO 0 gt 1 M1 0 0 0 0 OFF u G a 1 Te 14e i 0 0 O O 0 117 ity 1 Kutt 4 Notice however that the number of free parameters in the B matrix is relatively small compared to the size of the matrix Since we hold S as exogenous a stimulus unaffected by brain activity all regressors of S are set to zero Similarly since we maintain that causality cannot take place backwards in time all regressors of an ROI at time point t e g X over ROIs at time point t Effective Connectivity Modeling with the euSEM and GIMME 7 e g X are also set to zero That is activity at X could not possibly influence activity at X 1 What remains is only predictors of n the vector of observations at the ROIs for time point f As in the previous models solving this regression equation is unlikely to yield an interesting result if we attempt to estimate a parameter at every allowable connection Instead we can search the model iteratively by opening a single parameter at a time and comparing the resulting model to a preceding model without that paramet
10. le formats accepted by GIMME csv comma separated variable txt text or xls xIsx Excel Marsbar outputs time series data in the t by r matrix format correct dimensions for use in GIMME However Marsbar starts every column with a label for the ROI from which measurements are taken For that reason it is necessary to open the Marsbar output in Excel remove the first row containing labels and save the data again in an acceptable format one file for each participant If you are performing an extended unified SEM analysis including stimulus events in the connectivity analysis you will also need to build an onset matrix for these events This matrix must be saved in mat format Matlab workspace in a filed called onsets mat containing a variables onsets1 This variable is a t by n matrix where the vertical dimension f is the set of time points in either units of either scans or seconds and the horizontal dimension n is the number of participants Each column of this matrix should be filled with zeros except for ones at the time points when a stimulus was presented for that particular participant If a second stimulus type is included in the analysis a second variable named onsets2 maybe be included in the onsets mat file following the same convention as onsets1 This onsets mat file should be stored in another directory outside the source directory titled output Estimation of the Connectivity Map With all of the time series data
11. nous and endogenous variables such that stimuli may modulate the connections between ROIs as opposed to modulating activity within an ROI As in the preceding time series models euSEM relies on the assumption of stationarity for each of the time series measured at each ROI Stationarity entails a mean value for the time series that is constant over time In graphical terms the time series should not drift up or down but should center around a single mean value Variance must also be constant for the time series to maintain stationarity Graphically the width of the distribution of observations around the mean should be equally wide over time If either of these conditions is not met the assumptions for performing the regression Over a time series are invalid and may produce spurious results The matrix form of the regression equation for euSEM becomes quite large and some shorthand will be necessary to present it on the page Let the endogenous variables ROIs such as X Y and Z be represented by the vector n at time points t and t Exogenous variables like S will be represented by the vector u also at time points t and t The B matrix will be composed of several sub matricies e The lagged relationships previously described in VAR for endogenous variables will be described in the phi matrix e The contemporaneous relationships previously described in SEM for endogenous variables will be described in the A alpha matrix e I w
12. ollowing regression equation X Bo Bi Xei Vector autoregression generalizes this equation to allow the regression of several time series over themselves and each other Extending the regression equation for X we could describe X in relation to all three ROIs Xt Bo BrXe1 Bo Yer B3Ze1t E The same autoregression equations can also be estimated for each ROI and vector autoregression provides a tool for solving them all simultaneously Effective Connectivity Modeling with the euSEM and GIMME 5 x Boa Byy Byy Bxz Xe 1 r Pos Erx Prr Br er Le Poc Bzx Bzy Bzzil4e 1 Ez The graphical representation of this vector autoregression depicted above highlights that each ROI has a time lagged relationship denoted by the dashed lines with every ROI including on itself As was the case in SEM the B matrix in the foregoing VAR equation is not very useful because it allows each ROI to have a lagged influence over every ROI including itself making practical interpretation difficult Also like SEM VAR is suitable for comparing competing hypotheses or may be searched iteratively to add only connections which significantly improve the model as will be described in detail for the euSEM model In this way the B matrix may be sparsely populated with only the hypothetically relevant or statistically best fitting connections The following regression equation and graphic represent one
13. possible model i Boa Bo a B Oc xx 0 0 x t 1 Ey Byx 0 0 1 Er 0 Bzy Pzzi Ey 2 Pz One considerable advantage to examining lagged relationships is the additional causal evidence provided Autoregression utilizes a form of causal inference known as Granger Causality which simply states that if two events happen sequentially in time the causal relationship cannot occur backwards in time and must occur forwards in time For example if activity in ROI Y is correlated with activity in ROI X at a previous time point the activity in Y could not have caused the past activity in X Consequently one may infer that activity in X is the cause of activity in Y This inference ignores the third variable problem that X and Y may both be caused by something else and do not influence each other at all but it remains a stronger inference than correlation of simultaneous measurements as inferred in SEM Effective Connectivity Modeling with the euSEM and GIMME 6 Extended Unified SEM euSEM is a functional connectivity modeling technique designed to integrate the properties of both structural equation modeling SEM and vector autoregression VAR To this end euSEM estimates both contemporaneous connections as in SEM and time lagged connections as in VAR for both endogenous variables other ROIs and exogenous variables stimuli Finally euSEM also provides for interactions between exoge
14. r variable in the system In the previous model the ROI labeled A is technically an exogenous variable but it is unconventional to think of brain measurements as exogenous because we want to maintain the possibility that A is influenced by some other region or by a truly exogenous variable In event related neuroimaging studies a typical exogenous variable would be a stimulus such as a checkerboard image projected into the participant s visual field In a block design study the exogenous variable is not necessary as each block type may be independently modeled and compared In this case the variable would have a value 1 when the checkerboard is present and 0 when the checkerboard is absent If the checkerboard has a continuous property of some sort such as contrast or luminance its value might be anywhere between one and zero The regression equations and graphical representation might then be as follows Effective Connectivity Modeling with the euSEM and GIMME 4 Xt Bo BoS Y B3 BaXt BroS Matrix notation J _ Y t 4 FY o 0 0 0 Bxys i E a it Pyg 0 0 Bys r 2 Boz 0 Bay 0 0 zy L5 0 0 0 0 ae Eg m try bt While SEM is an extremely valuable tool for estimating connectivity for known or predicted models it is critical to understand that connectivity estimates are based on causal assumptions of the estimated model For example in the third model the rese
15. raction of the time series data 2 Preparation of the data set for connectivity analysis 3 Estimation of the connectivity map with the GIMME package Extraction of Time Series Data using Marsbar Time series data that is repeated measurements of the same region of interests ROI one or more voxels in an anatomically circumscribed region of the brain through time are the basic observations used in the connectivity analysis The locations of these ROIS are arbitrary defined by the researcher depending on her interests In order to extract the time series data it will be necessary to define these ROIs in SPM ROIs can be defined in SPM by selecting anatomical regions using the Pick Atlas tool and saving the resulting mask file Alternately ROIs can be defined manually by inputting MNI coordinates and defining the radius of a sphere surrounding the voxel at those coordinates In either case the Matlab package Marsbar is used to extract these time series data from your SPM files Marsbar is launched by selecting Marsbar from SPMS8 s Toolbox menu At the initial screen open ROI definition and select Build At this stage you will decide whether to define the ROI anatomically based on the a mask you defined using the Pick Atlas or manually based on a sphere around MNI coordinates In the third menu Type of ROD select Image for an anatomical mask or select Sphere for a manual ROI For anatomically defined ROI
16. ression equations In the case of neuroimaging data these variables represent regions of interest ROIs that are each measured as a time series SEM allows researchers to propose a set of connections between the ROIs and estimate the regression equations describing these connections The proposed set of connections constitutes the causal model that is a hypothesis about what regions cause changes in activity for other regions and this causal model is compared to the observed data The best fitting set of coefficients describing the weights of the proposed connections is estimated by solving the matrix form of these regression equations The resulting covariance is compared to the covariance of the observed data to obtain a y chi squared statistic a test for the significance of the model covered in more detail later To illustrate SEM conceptually consider the hemodynamic activity measured in three ROIs X Y and Z repeatedly measured at time point t The relationship between these three regions can be described as a set of three linear regression equations Xt Bo Br Y B24 Y B3 BaXt BsZ Li Bo T B7X Bg Y yr yez This model is the weakest least specific possible hypothesis It assumes that every ROI has Matrix notation x Box 0 By Byz e x LaS B oy Pyx 0 Byz Y y LZ LBoz lWPzx Bry 0 112 Lez some causal influence on all of the other ROIs A
17. s Select the saved mask for the contrast you have defined in SPM After selecting the mask respond Yes to Maintain binary image and No to Apply function to image Effective Connectivity Modeling with the euSEM and GIMME 9 For manually defined ROIs Enter the MNI coordinates for the center of your sphere and the radius of the sphere in millimeters mm In both cases save the file by entering a name in Label for ROI field opening the Save dialogue and clicking Save To extract the time series data select Extract ROI Data default from the Data menu Select the ROI mask that you just saved and then select the SPM mat design file for the participant whose time series you are trying to extract Once the data have been extracted you can select Export data from the Data menu choose Summary times series and save the time series to an MS Excel format file Preparation of the Data Set Having acquired the time series from Marsbar some work will be required to insure these files follow a format compatible with the GIMME software In brief GIMME requires each participant s time series data to be saved in a separate file as a t by r matrix where the vertical dimension is the set of time points the length of a single time series and the horizontal dimension r is the number of ROIs It s most convenient to edit these files in Microsoft Excel or LibreOffice Calc and save them to one of several fi
18. which you wish to estimate Effective Connectivity Modeling with the euSEM and GIMME 1 At this time GIMME only supports or 2 stimuli so these are the only possible stimuli If you selected 0 for the previous question leave this question blank If euSEM are you input vectors in seconds 1 or TRs 2 If you are including stimuli inputs in the connectivity map you must specify whether the variables in onsets mat were defined by seconds one observation per second or TRs one observation per TR Another way to ask this question is whether the input matrix s vertical dimension f is the number of seconds in the study or the number of scans If you are estimating a uSEM and therefore have no stimuli to input enter O for this value Would you like to start with the autoregressive terms estimated The connectivity analysis is expedited by first estimating the effect of each ROI on itself over time by performing an independent autoregression at each ROI This speeds up the later estimation of the connectivity map between ROIs but the effects of doing so are not fully understood 1 e whether this will introduce error into the connectivity analysis After selecting all of the appropriate options for your analysis click GIMME 4 GIMME may take several minutes to complete the individual and group level analyses When it is finished the specified output folder will be populated with e finalALL mat describing the best fit group

Download Pdf Manuals

image

Related Search

Related Contents

802.11b/g/n Draft 2.0 Wireless LAN USB Adapter  SALAiSE iNfo N°107  MANUEL D`UTILISATION SIMPLIFIÉE  OWNERS MANUAL - Wall to Wall Communications  イヤープロテクター 取扱説明書 - 西華デジタルイメージ株式会社    Triarch 23175 User's Manual  2008 JK Jeep Wrangler Owner Manual  Here - Radiology Business Management Association  Tilburg University A User's Guide to the Brave  

Copyright © All rights reserved.
DMCA: DMCA_mwitty#outlook.com.