Statistical analysis
Given the skewness in the distributions of CSF biomarkers, concentrations were log10-transformed to meet the assumptions of Gaussian distribution for the relevant statistical analyses. Differences between clinical diagnostic groups were first examined. Continuous variables were compared using one-way ANOVA. Categorical variables were assessed through chi-square tests. To determine whether plasma p-tau181 differed between clinical diagnostic groups, log10-transformed plasma p-tau181 levels were analyzed using ANCOVA with diagnosis as the independent variable of interest, while age and gender were included as covariates. Receiver operating characteristic (ROC) curve analysis was performed to summarize the ability of plasma p-tau181 in differentiating diagnostic subgroups. Tukey corrected post hoc pairwise comparisons were applied to examine differences between diagnostic groups. The nonparametric Mann-Whitney U test or Kruskal-Wallis test followed by Dunn’s corrected post hoc comparisons (non-parametric) were applied to examine differences between diagnostic groups as a sensitivity analysis.
Plasma p-tau181 associations with CSF and PET biomarkers at baseline were evaluated using linear regression models and controlling for age and sex. Linear regression models were also used to examine the effects of baseline plasma p-tau181, clinical diagnosis group, and their interaction on baseline global cognition after controlling for age, sex, and education. This analysis was repeated for CDR-SB.
Linear mixed models were used to examine (1) longitudinal change in plasma p-tau181, (2) whether baseline plasma p-tau181 predicted change in global cognition (MoCA), and (3) whether baseline plasma p-tau181 predicted change in function (CDR-SB). For the assessment of longitudinal change in plasma p-tau181, we conducted a linear mixed model with fixed effects of clinical diagnosis (CU, MCI, AD) at the time of the blood draw, time from initial blood draw, and their interaction, as well as a random intercept to account for within-subject correlations. In this model, contrasts were examined to determine whether change in plasma p-tau181 was (1) significantly different than zero within each clinical diagnostic group and (2) whether change over time differed across diagnostic groups. This regression analysis was repeated with additional adjustment of age, sex, and their interactions with time since the initial blood draw. Of the 219 participants with more than one plasma measure (range = 2–5), the mean follow-up time between their first and last blood draw was 2.05 years (SD = 1.03) for CU, 1.99 years (SD = 0.90) for MCI, and 1.86 years (SD = 0.71) for AD.
Specifically, we fitted the following two mixed effects regression models:
$${Y}_{ij}={\beta}_1+{\beta}_2^{\prime }{X}_{i,\mathit{\operatorname{diag}}}+\left({\beta}_3+{\beta}_4^{\prime }{X}_{i,\mathit{\operatorname{diag}}}\right)\times {t}_{ij}+{b}_i+{e}_{ij}$$
and
$${Y}_{ij}={\beta}_1+{\beta}_2^{\prime }{X}_{i,\mathit{\operatorname{diag}}}+{\beta}_3 Ag{e}_i+{\beta}_4 Se{x}_i+\left({\beta}_5+{\beta}_6^{\prime }{X}_{i,\mathit{\operatorname{diag}}}+{\beta}_7 Ag{e}_i+{\beta}_8 Se{x}_i\right)\times {t}_{ij}+{b}_i+{e}_{ij}$$
where
Yij is the Plasma p-tau181 level at the jth visit for the ith participant,
tij is the time of the jth plasma measurement relative to the first plasma measurement (in years) for the ith participant,
Xi, diag is the baseline diagnosis group for the ith participant (CU, MCI, and AD),
bi is the random intercept for the ith participant, and
eij is the mean zero random error at the jth visit of the ith participant.
To examine the association between baseline plasma p-tau181 with change in global cognition, we first used mixed regression models including a random intercept and time from initial MoCA, clinical diagnosis (CU, MCI, and AD) at the time of blood draw, age at time of blood draw, sex, education, and their interactions with time as predictors. In this analysis, the interaction between baseline plasma p-tau181 and time indicated whether plasma p-tau181 levels were associated with change in cognition after controlling for demographics and clinical diagnosis. To determine the associations between plasma p-tau181 and change in cognition within each clinical diagnosis as well as whether the association differs between diagnoses, we conducted a second mixed effects regression analysis that additionally included a three-way interaction term between clinical diagnosis, baseline plasma p-tau181, and time. Specifically, we fitted the following two regression models:
Model 1
$${Y}_{ij}={\beta}_1+{\beta}_2 Ag{e}_i+{\beta}_3 Se{x}_i+{\beta}_4 Educatio{n}_i+{\beta}_5^{\prime }{X}_{i,\mathit{\operatorname{diag}}}+{\beta}_6{X}_{i,P-\tau 181}+\left({\beta}_7+{\beta}_8 Ag{e}_i+{\beta}_9 Se{x}_i+{\beta}_{10} Educatio{n}_i+{\beta}_{11}^{\prime }{X}_{i,\mathit{\operatorname{diag}}}+{\beta}_{12}{X}_{i,P-\tau 181}\right)\times {t}_{ij}+{b}_i+{e}_{ij}$$
and
Model 2
$${Y}_{ij}={\beta}_1+{\beta}_2 Ag{e}_i+{\beta}_3 Se{x}_i+{\beta}_4 Educatio{n}_i+{\beta}_5^{\prime}\log {X}_{i,\mathit{\operatorname{diag}}}+{\beta}_6{X}_{i,P-\tau 181}+\left({\beta}_7+{\beta}_8 Ag{e}_i+{\beta}_9 Se{x}_i+{\beta}_{10} Educatio{n}_i+{\beta}_{11}^{\prime }{X}_{i,\mathit{\operatorname{diag}}}+{\beta}_{12}{X}_{i,P-\tau 181}+{\beta}_{13}{X}_{i,\mathit{\operatorname{diag}}}\times \log \left({X}_{i,P-\tau 181}\right)\right)\times {t}_{ij}+{b}_i+{e}_{ij}$$
where
Yij is the MoCA total scores at the jth visit of the ith participant,
Agei is the centered age for the ith participant at the first blood draw,
Sexi: is the sex (male, female) of the ith participant,
Educationi is the centered education in years for the ith participant,
Xi, diag is the clinical diagnosis closest to first blood draw (CU, MCI, AD) for the ith participant,
Xi, P − τ181 is the plasma p-tau181 level for the ith participant at the first blood draw,
tij is the time in years of the jth visit since the first MoCA assessment within 1 year of the first blood draw for the ith participant,
bi is the random intercept for the ith participant, and
eij: the mean zero random error
These models were repeated with CDR sum of boxes as the outcome (replacing MoCA total score) to examine the relationship with plasma p-tau181. Statistical tests were two-sided at a significance level of α = 0.05. Tests were performed using GraphPad Prism software and R.