Statistical analyses

All cross-sectional and longitudinal data were analyzed across the ɛ3/ɛ3, ɛ3/ɛ4, and ɛ4/ɛ4 genotypes. One-way analysis of variance (ANOVA) followed by Least Significant Difference (LSD) test were used to test between-group differences of clinical and demographic data presented in Table 1.

Locally estimated scatterplot smoothing (LOESS) regression, traced with 70% smoothing and uniform distribution as pre-set parameters, was used for nonparametric, graphical representation of time and the ɛ4 allele dependent trends in analyzed serial cognitive and volumetric measures. They also motivated the segmented linear mixed model (LMM) analysis [43] on the data taken before participants transitioned to AD dementia (i.e., when they carried an MCI diagnosis) and on the data taken on and after the transition to explicitly adjust for AD-dementia onset and account for the overall nonlinearity in time. All serial pre and post transition data sets were assessed for linearity (Supplementary Tables 1 and 2). The majority of volumetric and cognitive data sets revealed a linear relationship with time during each of the MCI and AD dementia segments, justifying the selection of LMM. Segmented LMM analysis exemplifies a multilevel modeling approach, and considers the data collected during repeated visits of each subject as a cluster allowing for comparison between rates of change even if subjects had different numbers of visits or were missing individual data points. Within each segment, the LMM also reduces non-random attrition bias and models random intercepts, and thus values of the dependent variable for each individual measure are predicted by the fixed effects including the intercept that varies across groups. A significant main effect of time in the LMM analysis would indicate that a given cognitive or volumetric measure changed significantly over time in all participants adjusted for demographics: sex, ADNI baseline visit age (for pre-transition analysis; MCI), transition age (for post-transition analysis; AD), and years of education. A significant main effect of the ɛ4 allele would indicate that a baseline data set for a given measure varied significantly across ɛ3/ɛ3, ɛ3/ɛ4, and ɛ4/ɛ4 genotypes. The baseline data sets used for the pre-transition analysis were the data collected during the ADNI baseline visit, while the baseline data sets used for post-transition analysis were the data collected during the visit when a participant was diagnosed with AD dementia. A significant interaction between time and the ɛ4 allele would indicate that the rate of cognitive decline or brain atrophy varied as a function of the ɛ4 allele. This interaction would determine not only the overall magnitude of an ɛ4 effect but also the allele-specific dose dependency pattern by directly comparing ɛ3/ɛ4 and ɛ4/ɛ4 genotypes. Additionally, stratified LMM analyses of cognitive measures were conducted on data collected after AD transition by stratifying the participants by median age of the transition (<76.1 years versus≥76.1 years), sex, education (<16 years versus≥16 years), and the ADNI study they were originally enrolled (ADNI-1 versus ADNI-GO/2). Race and ethnicity were excluded from the stratified analysis because of the low number of non-Whites (n < 10). These stratified LMM analyses tested interactions between the main effect of time and ɛ4 allele separately for ɛ3/ɛ4 and ɛ4/ɛ4 genotypes with ɛ3/ɛ3 as the reference group. For each LMM analysis the p value and the regression coefficient (β)±standard error (SE) were calculated.

A multiple linear regression model was used to compute yearly rates of change for all analyzed cognitive and volumetric measures in each genotype. Parameter estimates from LMM analysis were used as the dependent regression variables and time as the independent variable. Separate analyses were performed for all measures pre and post transition to AD dementia.

All statistical analyses were performed using IBM^® SPSS^® Statistics 25 (IBM Corp., Armonk, NY).