Statistical Inference Guillaume Flandin Wellcome Trust Centre for Neuroimaging University College London SPM Course Zurich, February 2008

Statistical Inference Image time-series Spatial filter Design matrix Statistical Parametric Map Realignment Smoothing General Linear Model Statistical Inference RFT Normalisation p <0.05 Anatomical reference Parameter estimates

Voxel-wise time series analysis Model specification Parameter estimation Hypothesis Statistic Time Time BOLD signal single voxel time series SPM

Overview A recap of model specification and parameters estimation Hypothesis testing Contrasts and estimability T-tests F-tests Design orthogonality A first example A second example Design efficiency

Model Specification: The General Linear Model = + y Sphericity assumption: Independent and identically distributed (i.i.d.) error terms N: number of scans, p: number of regressors The General Linear Model is an equation that expresses the observed response variable in terms of a linear combination of explanatory variables X plus a well behaved error term. Each column of the design matrix corresponds to an effect one has built into the experiment or that may confound the results.

Parameter Estimation: Ordinary Least Squares Find β that minimize The Ordinary Least Estimates are: Under i.i.d. assumptions, the Ordinary Least Squares estimates are Maximum Likelihood.

Hypothesis Testing To test an hypothesis, we construct “test statistics”. The Null Hypothesis H0 Typically what we want to disprove (no effect).  The Alternative Hypothesis HA expresses outcome of interest. The Test Statistic T The test statistic summarises evidence about H0. Typically, test statistic is small in magnitude when the hypothesis H0 is true and large when false.  We need to know the distribution of T under the null hypothesis. Null Distribution of T

Hypothesis Testing u t Type I Error α: Acceptable false positive rate α. Level  threshold uα Threshold uα controls the false positive rate  Observation of test statistic t, a realisation of T Null Distribution of T P-value: A p-value summarises evidence against H0. This is the change of observing value more extreme than t under the null hypothesis. t P-val The conclusion about the hypothesis: We reject the null hypothesis in favour of the alternative hypothesis if t > uα Null Distribution of T

One cannot accept the null hypothesis (one can just fail to reject it) Absence of evidence is not evidence of absence. If we do not reject H0, then all that we are saying is that there is no strong evidence in the data to reject that H0 might be a reasonable approximation to the truth. This certainly does not mean that there is a strong evidence to accept H0.

cTβ = 0xb1 + -1xb2 + 1xb3 + 0xb4 + 0xb5 + . . . Contrasts We are usually not interested in the whole β vector. A contrast select a specific effect of interest:  a contrast c is a vector of length p.  cTβ is a linear combination of regression coefficients β. cT = [1 0 0 0 0 …] cTβ = 1xb1 + 0xb2 + 0xb3 + 0xb4 + 0xb5 + . . . cT = [0 -1 1 0 0 …] cTβ = 0xb1 + -1xb2 + 1xb3 + 0xb4 + 0xb5 + . . . Under i.i.d assumptions:

Estimability of a contrast parameters images Factor 1 2 Mean parameter estimability (gray ® b not uniquely specified) If X is not of full rank then we can have Xb1 = Xb2 with b1≠ b2 (different parameters). The parameters are not therefore ‘unique’, ‘identifiable’ or ‘estimable’. For such models, XTX is not invertible so we must resort to generalised inverses (SPM uses the pseudo-inverse). One-way ANOVA (unpaired two-sample t-test) 1 0 1 0 1 1 Rank(X)=2 Example: [1 0 0], [0 1 0], [0 0 1] are not estimable. [1 0 1], [0 1 1], [1 -1 0], [0.5 0.5 1] are estimable.

T-test - one dimensional contrasts – SPM{t} Question: box-car amplitude > 0 ? = b1 = cTb> 0 ? cT = 1 0 0 0 0 0 0 0 b1 b2 b3 b4 b5 ... Null hypothesis: H0: cTb=0 T = contrast of estimated parameters variance estimate Test statistic:

T-contrast in SPM For a given contrast c: ResMS image beta_???? images con_???? image spmT_???? image SPM{t}

T-test: a simple example Passive word listening versus rest cT = [ 1 0 ] Q: activation during listening ? Design matrix 0.5 1 1.5 2 2.5 10 20 30 40 50 60 70 80 Null hypothesis: SPMresults: Height threshold T = 3.2057 {p<0.001} voxel-level p uncorrected T ( Z º ) mm mm mm 13.94 Inf 0.000 -63 -27 15 12.04 -48 -33 12 11.82 -66 -21 6 13.72 57 -21 12 12.29 63 -12 -3 9.89 7.83 57 -39 6 7.39 6.36 36 -30 -15 6.84 5.99 51 0 48 5.65 -63 -54 -3 6.19 5.53 -30 -33 -18 5.96 5.36 36 -27 9 5.84 5.27 -45 42 9 5.44 4.97 48 27 24 5.32 4.87 36 -27 42 Statistics: p-values adjusted for search volume set-level c p cluster-level corrected uncorrected k E voxel-level FWE-corr FDR-corr T ( Z º ) mm mm mm 0.000 10 520 13.94 Inf -63 -27 15 12.04 -48 -33 12 11.82 -66 -21 6 426 13.72 57 -21 12 12.29 63 -12 -3 9.89 7.83 57 -39 6 35 7.39 6.36 36 -30 -15 9 6.84 5.99 51 0 48 0.002 3 0.024 0.001 5.65 -63 -54 -3 8 6.19 5.53 -30 -33 -18 0.003 5.96 5.36 36 -27 9 0.005 2 0.058 0.004 5.84 5.27 -45 42 9 0.015 1 0.166 0.022 5.44 4.97 48 27 24 0.036 5.32 4.87 36 -27 42

T-test: a few remarks T-test is a signal-to-noise measure (ratio of estimate to standard deviation of estimate). T-contrasts are simple combinations of the betas; the T-statistic does not depend on the scaling of the regressors or the scaling of the contrast. H0: vs HA: Unilateral test: -5 -4 -3 -2 -1 1 2 3 4 5 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 n =1 =2 =5 =10 = ¥ Probability Density Function of Student’s t distribution

F-test - the extra-sum-of-squares principle Model comparison: Full vs. Reduced model? Null Hypothesis H0: True model is X0 (reduced model) X1 X0 RSS Or Reduced model? X0 RSS0 Test statistic: ratio of explained variability and unexplained variability (error) 1 = rank(X) – rank(X0) 2 = N – rank(X) Full model ?

F-test - multidimensional contrasts – SPM{F} Tests multiple linear hypotheses: H0: True model is X0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 cT = H0: b3 = b4 = ... = b9 = 0 test H0 : cTb = 0 ? SPM{F6,322} X0 X1 (b3-9) X0 Full model? Reduced model?

F-contrast in SPM ( RSS0 - RSS ) ResMS image beta_???? images ess_???? images spmF_???? images ( RSS0 - RSS ) SPM{F}

F-test example: movement related effects To assess movement-related activation: There is a lot of residual movement-related artifact in the data (despite spatial realignment), which tends to be concentrated near the boundaries of tissue types. Even though we are not interested in such artifact, by including the realignment parameters in our design matrix, we “covary out” linear components of subject movement, reducing the residual error, and hence improve our statistics for the effects of interest.

Slightly more complicated… Think of it as constructing 3 regressors from the 3 differences and complement this new design matrix such that data can be fitted in the same exact way (same error, same fitted data).

F-test: a few remarks F-tests can be viewed as testing for the additional variance explained by a larger model wrt a simpler (nested) model  Model comparison. F tests a weighted sum of squares of one or several combinations of the regression coefficients b. In practice, we don’t have to explicitly separate X into [X1X2] thanks to multidimensional contrasts. Hypotheses: In testing uni-dimensional contrast with an F-test, for example b1 – b2, the result will be the same as testing b2 – b1. It will be exactly the square of the t-test, testing for both positive and negative effects.

Design orthogonality For each pair of columns of the design matrix, the orthogonality matrix depicts the magnitude of the cosine of the angle between them, with the range 0 to 1 mapped from white to black. The cosine of the angle between two vectors a and b is obtained by: If both vectors have zero mean then the cosine of the angle between the vectors is the same as the correlation between the two variates.

Shared variance Independent contrasts

Shared variance Testing for the green: Correlated regressors, for example: green: subject age yellow: subject score

Shared variance Testing for the red: correlated contrasts

Shared variance Testing for the green: Entirely correlated contrasts ?  Non estimable !

Shared variance Testing for the green and yellow If significant ? Could be G or Y ! Entirely correlated contrasts ? Non estimable !

A “not so good” model True signal and observed signal (--) Model (green, pic at 6sec) TRUE signal (blue, pic at 3sec) Fitting (b1 = 0.2, mean = 0.11) Residual (still contains some signal)  Test for the green regressor not significant

A “not so good” model Residual Var.= 0.3 p(Y| b1 = 0)  p-value = 0.1 (t-test) p(Y| b1 = 0)  p-value = 0.2 (F-test) = + Y X b e

A “better” model True signal + observed signal Model (green and red) and true signal (blue ---) Red regressor : temporal derivative of the green regressor Global fit (blue) and partial fit (green & red) Adjusted and fitted signal Residual (a smaller variance)  t-test of the green regressor significant  F-test very significant  t-test of the red regressor very significant

A “better” model Residual Var. = 0.2 p(Y| b1 = 0)  p-value = 0.07 (t-test) p(Y| b1 = 0, b2 = 0)  p-value = 0.000001 (F-test) = + Y X b e

Correlation between regressors True signal Model (green and red) Fit (blue : global fit) Residual

Correlation between regressors Residual var. = 0.3 p(Y| b1 = 0)  p-value = 0.08 (t-test) P(Y| b2 = 0)  p-value = 0.07 p(Y| b1 = 0, b2 = 0)  p-value = 0.002 (F-test) = + Y X b e 1 2 1 2

Correlation between regressors True signal Model (green and red) red regressor has been orthogonalised with respect to the green one  remove everything that correlates with the green regressor Fit Residual

Correlation between regressors (0.79) (0.85) (0.06) Residual var. = 0.3 p(Y| b1 = 0) p-value = 0.0003 (t-test) p(Y| b2 = 0) p-value = 0.07 p(Y| b1 = 0, b2 = 0) p-value = 0.002 (F-test) = + Y X b e 1 2 1 2

Design efficiency The aim is to minimize the standard error of a t-contrast (i.e. the denominator of a t-statistic). This is equivalent to maximizing the efficiency e: Noise variance Design variance If we assume that the noise variance is independent of the specific design: This is a relative measure: all we can really say is that one design is more efficient than another (for a given contrast).

Design efficiency The efficiency of an estimator is a measure of how reliable it is and depends on error variance (the variance not modeled by explanatory variables in the design matrix) and the design variance (a function of the explanatory variables and the contrast tested). XTX represents covariance of regressors in design matrix; high covariance increases elements of (XTX)-1. High correlation between regressors leads to low sensitivity to each regressor alone. cT=[1 0]: 5.26 cT=[1 1]: 20 cT=[1 -1]: 1.05

Example: working memory B C Stimulus Response Stimulus Response Stimulus Response Time (s) Time (s) Time (s) Correlation = -.65 Effciency ([1 0]) = 29 Correlation = +.33 Effciency ([1 0]) = 40 Correlation = -.24 Effciency ([1 0]) = 47 B: Jittering time between stimuli and response. C: Requiring a response on a randomly half of trials.

Bibliography: Statistical Parametric Mapping: The Analysis of Functional Brain Images. Elsevier, 2007. Plane Answers to Complex Questions: The Theory of Linear Models. R. Christensen, Springer, 1996. Statistical parametric maps in functional imaging: a general linear approach. K.J. Friston et al, Human Brain Mapping, 1995. Ambiguous results in functional neuroimaging data analysis due to covariate correlation. A. Andrade et al., NeuroImage, 1999. Estimating efficiency a priori: a comparison of blocked and randomized designs. A. Mechelli et al., NeuroImage, 2003. With many thanks to J.-B. Poline, Tom Nichols, S. Kiebel, R. Henson for slides!