Circulating inflammatory markers are associated with MR visible Perivascular Spaces but not directly with White Matter Hyperintensities
Online-Only Supplementary Material
Measurement of Inflammatory Markers
Bivariate Statistics
Multivariate Statistics using Structural Equation Modelling
Measurement of Inflammatory Markers
C-reactive protein (CRP) and interleukin 6 (IL6) were analysed at the University of Glasgow using high sensitivity ELISA from R&D Systems (Oxford, UK). For IL-6, the minimum detectable dose (MDD) ranged from 0.016-0.110 pg/mL (mean=0.039 pg/mL). The intra-assay coefficient of variation (CV) ranged from 6.9% to 7.8%, while the inter-assay CV ranged from 6.6% to 9.6%. For CRP, the MDD ranged from 0.005-0.022 ng/mL (mean=0.010 ng/mL). The intra-assay CV ranged from 3.8% to 8.3%, while the inter-assay CV ranged from 6.0% to 7.0%. Fibrinogen was measured using an automated Clauss assay (TOPS coagulometer, Instu-mentation Laboratory, Warrington, UK) with a CV of 3.69% [1].
Bivariate Statistics
Supplementary Table 1 shows the descriptive statistics. The total median (and interquartile range )Fazekas score was, 2.0 (1.0), range 0 to 6 (periventricular median score was 1.0 (1.0), range 0 to 3, and deep median score was 1.0 (0), range 0 to 3). The median and interquartile range PVS scores were 0 (0), range 0 to 2 in the hippocampus, 1.0 (1.0), range 0 to 4 in the BG, and 2.0 (1.0), range 0 to 4 in the CS. Inflammatory marker levels are given in Supplementary Table 1.
Supplementary Table 2 presents the bivariate association between markers of Inflammation, WMH and PVS measures. The main results have been presented in the main text. As expected, Fibrinogen, CRP and IL-6 were all significantly positively correlated with each other as were all measures of WMH with each other and all measures of PVS with each other.
Correlations between the covariates and inflammation markers, WMH and PVS, are presented in Supplementary Table 3. History of hypertension and of stroke correlated consistently with WMH (r range 0.08 to 0.19, P<0.01). History of smoking correlated consistently with inflammation (r range 0.08 to 0.16, P<0.001). Advancing age (even within the narrow range included) was associated with declining fibrinogen (r = -0.13, P<0.001), but not CRP or IL-6, and increase in WMH (r=0.14, P<0.001). Sex was significantly associated with Fazekas deep score (r=0.11, P=0.003) with women having higher Fazekas scores than men. Other health covariates (diabetes, hypercholesterolaemia and cardiovascular disease histories) showed no significant correlation with markers of inflammation, WMH or PVS.
Multivariate Statistics using Structural Equation Modelling
In statistical modelling, latent variables are employed when the parameter of interest is inadequately measured by a single variable. Latent variables are estimated based on the shared common variance between a set of measured indicator variables, and this allows the exclusion of measurement errors when estimating the contribution that each measured variable makes to the latent variable [2, 3]. Structural equation models (SEM), used in this study, simultaneously provides estimates of such latent variables, and the correlations and uni-directed paths (regression parameters, which may be thought of like standardised, partial beta weights) between latent variables. The SEM is conventionally represented by a diagram as in the figures presented here where the rectangular boxes each represent measured observations, eg CRP or WMH volume. Each circle or ellipse represents a latent construct, eg‘inflammation’ or ‘PVS’, which is a weighted combination of the correlated individual measured variables. Paths in the diagrams that are single-headed arrows represent hypothesised causal pathways from a predictor variable to the outcome variable, and double headed arrows represent correlations.Solid arrows represent associations that are statistically significant at p <0.05 while dashed arrows represent associations that are not significant at p <0.05. Numbers adjacent to paths may be squared to obtain the shared variance between adjacent variables. The numbers on the arrows pointing from the latent to the measured variables, called the factor loading or correlation coefficient, represent the strength of the association between the measured and latent variables, +/- 1 indicate the strongest association. Thick arrows pointing to the measured variables mean that that error term was accounted for in the modelling and the numbers adjacent represent the r2 which is the square of the factor loading.
The three measurement models of ‘inflammation’, ‘WMH’ and ‘PVS’ all fitted well (model fit statistics are presented in figure legends). The factor loadings for the latent variables of ‘inflammation’, ‘WMH’ and ‘PVS’ were moderate to large, indicating that the measurement models were good for the three latent traits. The standardised path values were all large (shown in Figures 1-3 and supplementary Figures 1 to 3) indicating that the latent variables are robust and account for between approximately 20% and 88% of variance in the measured variables in all models.
Model fit parameters are shown in the legends: the maximum modification index (Max MI, which indicates the degree of greatest local strain within the model in terms of a potential reduction in model chi square); the comparative fit index, (CFI, 0.90 indicates acceptable fit); and the root mean square error of approximation (RMSEA, <0.06 indicates acceptable fit).
All models were estimated using full information maximum likelihood estimation. Model fit was evaluated based on commonly adopted cut-off points for the Root Mean Square Error Approximation (RMSEA), Tucker-Lewis Index (TLI), Comparative Fit Index (CFI), incremental Fit index (IFI) and the Standardized Root Mean Square of <0.06, >0.90, >0.90, >0.90 and <0.06 respectively [4]. Model modifications were based on the standardized residual covariance < 1.96. Associations were modelled using a stepwise approach. The first set of models included only the measures of inflammation markers & PVS measures, inflammation markers & WMH measures, PVS measures & WMH measures or inflammation markers & PVS measures & WMH measures (Supplementary Figure 1, Supplementary Table 4). Then, demographic variables were introduced (Supplementary Figure 2). All health variables were then introduced (Supplementary Figure 3). Model fit measures and p-values for regression coefficients were inspected to identify covariates which did not make any significant contributions to the model fit, such covariates were removed to improve model fit. Finally, the standardized residual covariance matrices were inspected to identify significant associations which were not included in the models, such associations were then included to produce the final models (Figures 1 - 3) which produced satisfactory model fit (Supplementary Table 5). Supplementary Figures 1 to 3 show some of the intermediate results.
All imaging features were defined according to the STandards for ReportIngVascular Changes on NEuroimaging (STRIVE)v1[5].
Supplementary Table 1: Descriptive statistics for markers of inflammation, measures of WMH, PVS, demographic and health conditions
N=634 / Measures / Mean(SD)Markers of Inflammation / CRP / 2.94(5.97)
Fibrinogen / 3.34(0.60)
IL6 / 2.00(1.74)
Measures of PVS / Hippo-PVS / 0.25(0.45)
BG-PVS / 1.43(0.58)
CS-PVS / 2.07(0.74)
WMH and related measures / WMH (mm3) / 12186.18(13289.57)
ICV (mm3) / 1442352.06(142201.34)
% WMH in ICV / 0.85(0.92)
Fazekas Periventricular scores, median (IQR) / 1.00(1.00)
Fazekas Deep scores, median (IQR) / 1.00(0.00)
Demographic / Age in years, mean (SD) / 72.5(0.72)
% men / 53
YES
Health conditions / History of hypertension (%) / 48.70
History of diabetes (%) / 11
History of stroke (%) / 6.9
History of smoking (%) current or ever smoked / 56.1
History of hypercholesterolemia (%) / 41.4
Note. CRP=C-reactive Protein, IL6=Interleukin 6, hippo=Hippocampus, BG=Basal Ganglia, CS=Centrum Semiovale, PVS=Perivascular Spaces, WMH=White Matter Hyperintenities, ICV=Intracranial Volume, IQR=Interquartile Range, SD=Standard Deviation
SupplementaryTable 2: Bivariate Correlations (p values) within and between markers of Inflammation, WMH and PVS measures; Bonferroni corrected (N=634)
CRP / Fibrinogen / IL6
% WMH in ICV / 0.00(0.97) / 0.00(0.92) / 0.02(0.54)
FazekasPeri / 0.00(0.99) / 0.05 (0.21) / 0.04(0.30)
Fazekas Deep / 0.01 (0.74) / 0.00(0.97) / 0.02(0.57)
CRP / 0.41 (<0.001) / 0.41(<0.001)
Fibrinogen / 0.37(<0.001)
PVS and WMH
% WMH in ICV / FazekasPeri / Fazekas Deep
Hippo-PVS / 0.11(0.03) / 0.21(0<0.001) / 0.16(0<0.001)
BG-PVS / 0.31(0<0.001) / 0.32(0<0.001) / 0.31(0<0.001)
CS-PVS / 0.18(0<0.001) / 0.23(0<0.001) / 0.19(0<0.001)
% WMH in ICV / 0.72(<0.001) / 0.65(<0.001)
FazekasPeri / 0.52(<0.001)
Inflammation and PVS
PVS hippo / BG-PVS / CS-PVS
CRP / 0.03(0.40) / 0.01(0.86) / 0.10(0.10)
Fibrinogen / 0.06(0.15) / 0.05(0.24) / 0.03(0.42)
IL6 / 0.02(0.58) / 0.02(0.54) / 0.09(0.22)
Hippo-PVS / 0.25(<0.001) / 0.27(<0.001)
BG-PVS / 0.40(<0.001)
Abbreviations: See note Table 1.
1
Supplementary Table 3: Pearson bivariate and biserial (*) correlations (p values) between covariate risk factors, markers of inflammation, WMH and PVS measures
Sex / Age / Smoking* / Hypertension* / Diabetes* / Hypercholesterolaemia* / Cardiovascular Disease* / Stroke*% WMH in ICV / 0.01(0.85) / 0.14(<0.001) / 0.08(0.04) / 0.13(0.001) / 0.03(0.45) / 0.07(0.05) / 0.00(0.91) / 0.08(0.03)
FazekasPeri / 0.01(0.85) / 0.03(0.48) / 0.06(0.12) / 0.10(0.008) / 0.02(0.65) / 0.05(0.16) / 0.00(0.92) / 0.10(0.01)
Fazekas Deep / 0.11(<0.01) / 0.08(0.04) / 0.03(0.43) / 0.19(<0.001) / 0.00(0.95) / 0.03(0.43) / 0.01(0.73) / 0.08(0.04)
CRP / 0.02(0.55) / 0.04(0.24) / 0.12(<0.001) / 0.10(0.01) / 0.02(0.53) / 0.01(0.76) / -0.03(0.43) / 0.09(0.01)
Fibrinogen / 0.08(0.03) / -0.13(<0.001) / 0.16(<0.001) / 0.04(0.24) / 0.02(0.65) / 0.02(0.49) / 0.01(0.75) / 0.13(<0.001)
IL6 / -0.04(0.21) / 0.04(0.24) / 0.16(<0.001) / 0.06(0.08) / 0.10(<0.01) / 0.00(0.95) / 0.02(0.56) / 0.09(0.01)
Hippo-PVS / 0.03(0.38) / -0.18(<0.001) / -0.05(0.21) / 0.02(0.61) / -0.04(0.28) / 0.03(0.41) / -0.01(0.72) / 0.02(0.56)
BG-PVS / -0.02(0.68) / 0.02(0.60) / -0.03(0.51) / 0.11(0.01) / -0.03(0.39) / 0.05(0.24) / -0.01(0.88) / 0.01(0.76)
CS-PVS / 0.00(0.99) / 0.02(0.56) / -0.01(0.72) / 0.03(0.41) / -0.04(0.30) / -0.05(0.19) / -0.05(0.19) / -0.03(0.49)
Abbreviations: See note Table 1.
1
Model 1 / Model 2 / Mode 3 / Model 4WMH / / PVS / 0.47 (<0.0001) / 0.47 (<0.0001) / 0.47 (<0.0001) / 0.47 (<0.0001)
HIP / PVS / 0.53 (<0.0001) / 0.38 (<0.0001) / 0.37 (<0.0001) / 0.38 (<0.0001)
BG / PVS / 0.36 (<0.0001) / 0.73 (<0.0001) / 0.74 (<0.0001) / 0.73 (<0.0001)
CS / PVS / 0.74 (<0.0001) / 0.53 (<0.0001) / 0.53 (<0.0001) / 0.53 (<0.0001)
WMH V / WMH / 0.92 (<0.0001) / 0.92 (<0.0001) / 0.92 (<0.0001) / 0.92 (<0.0001)
FazPeri / WMH / 0.78 (<0.0001) / 0.77 (<0.0001) / 0.77 (<0.0001) / 0.77 (<0.0001)
Faz Deep / WMH / 0.72 (<0.0001) / 0.72 (<0.0001) / 0.72 (<0.0001) / 0.72 (<0.0001)
PVS / Age / -0.03 (0.531)
PVS / Sex / -0.02 (0.676)
WMH / Sex / 0.08 (0.043)
WMH / Age / 0.14(<0.0001) / 0.13 (<0.0001) / 0.13 (0.003)
PVS / High BP / 0.11 (0.038) / 0.11 (0.037)
PVS / Stroke / 0.00 (0.955) / 0.00 (0.999)
WMH / High BP / 0.09 (0.019) / 0.10 (0.019)
WMH / Stroke / 0.08 (0.051) / 0.08 (0.055)
WMH / Smoking / 0.09 (0.024) / 0.09 (0.023)
Supplementary Table 4: Association between measures of PVS and WMH using structural
Note. Models predicted ‘WMH’ from ‘PVS’. Model 1 is the basic model that contained only ‘WMH’ and ‘PVS’. In model 2, age and sex were introduced. In model 3 all health variables (history of cardiovascular disease, diabetes, hypertension, smoking, hypercholesterolemia and stroke) were introduced but only those that passed the inclusion criteria test were retained. Model 4 is the modified version of model 3 to improve the model fit. This was done using the standardized residual covariance. Arrows point from the predictors to outcome variables.
Supplementary Table 5: Model fit statistics for association between WMH and EPVS
IFI / TLI / CFI / RMSEA / SRMRModel 1 / 0.972 / 0.947 / 0.972 / 0.078 / 0.0347
Model 2 / 0.938 / 0.898 / 0.938 / 0.081 / 0.0468
Model 3 / 0.94 / 0.915 / 0.939 / 0.074 / 0.0436
Model 4 / 0.968 / 0.952 / 0.968 / 0.044 / 0.0316
Note. Models predicted ‘WMH’ from ‘PVS’. Model 1 is the basic model that contained only ‘WMH’ and ‘PVS’. In model 2, age and sex were introduced. In model 3 included health variables. Model 4, the final model (Figure 1) is the modified version of model 3 to improve the model fit. This was done using the standardized residual covariance. RMSEA = Root Mean Square Error Approximation, TLI = Tucker-Lewis Index, CFI = Comparative Fit Index, IFI = incremental Fit index (IFI) and SRMS = Standardized Root Mean Square. A model fits well if RMSEA <0.06, TLI >0.90, CFI >0.90, IFI >0.90 and SRMS<0.06 [4].
1
Supplementary Figure 1: SEM of the association between measures of PVS and WMH using the basic model that included only the measures of PVS and measures of WMH. Model fit: See Supplementary Table 5.
Note. ‘PVS’=latent variable derived from measures of PVS, ‘WMH’=latent variable derived from measures of WMH. Models included only covariates that were significantly associated with WMH load or measures of PVS. BG=measured basal ganglia PVS, HIP=measured hippocampus PVS, CS=measured centrum semiovale PVS, high BP=history of hypertension, WMH v=% of WMH in ICV, Faz=Fazekas, Peri=Perivascular. Solid or thick arrows represent associations that are statistically significant at p < 0.05 while dashed arrows represent associations that are not significant at p < 0.05.
Supplementary Figure 2: SEM of the association between measures of PVS and WMH. Model included age and Sex. Dotted lines are associations that are not significant.
Note.Abbreviations:See Figure 1.Model fit: See Supplementary Table 5
Supplementary Figure 3: SEM of the association between measures of PVS and WMH. Model included age and Sex and all health covariates (history of cardiovascular disease, diabetes, hypertension, smoking, hypercholesterolemia and stroke). Health variables that showed no significant association with WMH or PVS and that reduced model fit were removed from the model.
Note.Abbreviations:See Figure 1.Model fit: See Supplementary Table 5
1
References
1.Hart CL, Taylor MD, Smith GD, Whalley LJ, Starr JM, Hole DJ, et al. Childhood IQ and cardiovascular disease in adulthood: prospective observational study linking the Scottish Mental Survey 1932 and the Midspan studies. SocSci Med. 2004;59:2131-2138.
2.Pike RG.Using structural equation models with latent variables to study student growth and development.Research in Higher Education. 1991;32:499-524.
3.Stephenson M, Holbert R, Zimmerman R.On the use of structural equation modeling in health communication research. Health Commun.n 2006;20:159-167.
4.Hooper D, Coughlan J, Mullen M. Structural equation modelling: guidelines for determining model fit. Electronic Journal of Business Research Methods. 2008;6:53-60.
5.Wardlaw JM, Smith EE, Biessels GJ, Cordonnier C, Fazekas F, Frayne R, et al. Neuroimaging standards for research into small vessel disease and its contribution to ageing and neurodegeneration. Lancet Neurol. 2013;12:822-838.
6.Blunch N: Introduction to Structural equation modeling using IBM SPSS statistics and Amos. SAGE publications limited 2012.
7.Arbuckle LJ: IBM SPSS Amos 19 User's guide. IBM 2010.
1