Télécharger le texte intégral - ISPED-Enseignement à distance

Annexes 188170 C. Proust, H. Jacqmin-GaddaTable 2 Estimates of the homogeneous linear mixedmodel and the heterogeneous linear mixed model withtwo components of mixture obtained with HETMIXLINprogram for the MMSE evolution adjusted on age, occupation,educational level, living place and interactionwith time for age and educational levelParameterHomogeneousmodel (G = 1)Heterogeneousmodel (G = 2)Estimate S.E. Estimate S.E. 1 1 0.98 0.005 2 0.02 01 −1.92 0.32 −1.51 0.31 11 −0.57 0.16 −0.42 0.16 21 −0.042 0.018 −0.028 0.024 02 0.89 0.36 12 0.46 0.19 22 0.055 0.028var(u 0i ) 0.43 0.025 0.32 0.024cov(u 0i , u 1i ) 0.074 0.010 0.034 0.0095var(u 1i ) 0.066 0.014 0.029 0.013cov(u 0i , u 2i ) 0.042 0.0073 0.028 0.0069cov(u 1i , u 2i ) 0.052 0.010 0.038 0.010var(u 2i ) 0.073 0.016 0.060 0.016 0.45 0.009 0.45 0.009−L 4662.9 4609.5AIC 9367.7 9269.0BIC 9477.8 9400.0where X ij is the vector of covariates for subjecti at visit j including age, occupation, educationallevel, living place and interactions with timefor age and educational level; u i = (u 0i , u 1i , u 2i ) ′ ∼∑ Gg=1 g N( g , D) with g = ( 0g , 1g , 2g ) ′ and ij ∼ iid N(0, 2 ).We fitted the heterogeneous linear mixed modelwith two components of mixture. The Akaike InformationCriterion (AIC) and the Bayesian InformationCriterion (BIC) were largely improvedcompared with the homogeneous mixed model(AIC = 98.7; BIC = 77.8). Table 2 displaysthe parameter estimates for the homogeneousand the heterogeneous linear mixed models obtainedwith HETMIXLIN. The heterogeneous linearmixed model distinguished two different MMSEcourses (Fig. 1). First, a large class including 98%of the sample follows a linear evolution witha slight decline of 0.42 points ( = 0.15) peryear. The second class including 2% of the samplefollows a non-linear evolution which speedsto fell down from the second follow-up to theend.Then, we tried to evaluate if the cognitive profileshighlighted by the model were associated withdementia diagnosis. Among the 1392 subjects, 26Fig. 1 Mean curves of the MMSE between T1 and T8 forthe heterogeneous linear mixed model with two componentsof mixture for a 70-year-old worker subject with noeducation and living in Dordogne. Solid line: first componentwith a probability of 98%; dashed line: second componentwith a probability of 2%.(1.9%) are classified in the second component withthe non-linear decline (Table 3). Among them, 21have a positive dementia diagnosis at T8. Thepredictive positive value (81%) and the specificity(99.6%) of this classification are high but the sensitivityis poor (31%): 47 subjects among the 68subjects diagnosed as demented at T8 are not detectedby the model. These subjects are significantlyless disabled (p