Sex, Amyloid, and APOEε4 and risk of cognitive decline in preclinical Alzheimer’s disease: findings from three well-characterized cohorts

Abstract

INTRODUCTION

Our objective was to investigate the effect of sex on cognitive decline within the context of β-amyloid (Aβ) burden and apolipoprotein (APOE) genotype.

METHODS

We analyzed sex-specific effects on Aβ-PET, APOE and rates of change on the Preclinical Alzheimer Cognitive Composite-5 (PACC-5) across three cohorts, ADNI, AIBL and HABS (n=755; clinical dementia rating (CDR)=0; Age(SD)=73.6(6.5); Female=55%). Mixed-effects models of cognitive change by sex, Aβ-PET and APOEε4 were examined with quadratic time-effects over a median of 4 years of follow-up.

RESULTS

APOEε4 prevalence and Aβ burden did not differ by sex. Sex did not directly influence cognitive decline. Females with higher Aβ exhibited faster decline than males. Post-hoc contrasts suggested that females who were Aβ and APOEε4 positive declined faster than their male counterparts.

DISCUSSION

Although Aβ did not differ by sex, cognitive decline was greater in females with higher Aβ. Our findings suggest sex may play a modifying role on risk of AD-related cognitive decline.

1. INTRODUCTION

Investigating the extent to which the early pathophysiology and cognitive decline vary by sex is critical for understanding the course of Alzheimer’s disease (AD) dementia [1]. Epidemiological studies suggest higher incidence rates for women relative to men for AD dementia in older ages [2], however, this is not always supported [3–5]. Studies are mixed with regard to the effect of sex on cognitive decline in clinically-normal older adults [6, 7]. Steeper decline in delayed recall [8], performance IQ and executive function [9] has been reported in female apolipoprotein ε4 (APOEε4) carriers, however, this finding is not supported by other studies [6, 10]. In addition, some studies suggest males show steeper decline than females in areas of speed, integration and visuospatial ability [11]. A recent meta-analysis found that APOEε4 carriers exhibit particular vulnerability for progression to Alzheimer’s disease (AD) dementia relative to males and female APOEε4 non-carriers between the ages 65-75 years [12]. Carrying the APOEε4 allele confers higher risk for abnormal levels of β-amyloid (Aβ) burden [13] and accumulation [14], and results in steeper cognitive decline when accompanied by high Aβ burden [15, 16]. Hence, it is possible that APOEε4 genetic risk in the presence of Aβ may impart particular susceptibility for females to AD clinical symptoms.

It is unclear whether sex differences in AD dementia risk in those with APOEε4 is exacerbated by higher Aβ burden, whether APOEε4 and Aβ play independent roles, or whether there are other “downstream” mechanisms that account for the increased vulnerability to dementia risk. Disentangling sex-specific effects with respect to Aβ and APOE is particularly relevant during the preclinical stage, given that abnormal levels of Aβ start to accrue decades before the onset of clinical symptoms [17], and that greater focus is being placed on prevention trials [18]. If Aβ burden is differentially associated with cognitive decline in males and females, this will have implications for recruitment and treatment practices in clinical trials.

Our first aim was to ascertain whether sex differences exist in relation to APOE and Aβ burden as estimated by positron emission tomography (PET). Our second aim was to examine the effect of sex on cognitive decline, and whether this was influenced by APOEε4 carriage, abnormal Aβ or both APOE and Aβ. To improve our ability to detect the extent to which females could increase risk for AD biomarkers and cognitive decline, we harmonized data from three well-characterized, longitudinal datasets: the Alzheimer’s Disease Neuroimaging Initiative (ADNI), the Australian Imaging, Biomarker and Lifestyle (AIBL) study of ageing, and the Harvard Aging Brain Study (HABS). This allows us sufficient statistical power to determine potentially small magnitude relationships between sex and AD risk.

2. METHODS

2.1. Participants

Cohort-specific inclusion criteria for recruitment have been published previously [19–21]. For the current study, the baseline was considered to be an individual’s first Aβ-PET scan. Participants were all required to be clinically-normal at baseline (Global clinical dementia rating (CDR) score = 0, MMSE ≥ 24); ADNI’s subjective cognitive decline (SCD) group was included in the current study, given that these participants attained a CDR score of 0. Participants were included if their baseline Aβ-PET scan was within 1 year of a neuropsychological testing session (either before or after the scan), and they had at least 2 follow-up neuropsychological assessments after their baseline visit. We excluded participants who carried APOEε2/ε4 and APOEε2/ε2 (total < 2.9%), given that the effect of these genotypes on AD risk are unclear. For analysis, 755 participants (ADNI, n=330; AIBL, n=161; HABS, n=268) formed the final participant group. We conducted the procedures for this study under the ethical guidelines stipulated by the Partners Human Research Committee, which is the Institutional Review Board for the Massachusetts General Hospital and Brigham and Women’s Hospital.

2.2. Cognitive outcome

We examined cognitive decline using the Preclinical Alzheimer Cognitive Composite score with an additional semantic processing component (PACC-5) [22]. This composite modifies the PACC that was developed as a sensitive measure of Aβ-related cognitive decline [23]. We used the PACC-5 as recent findings suggest the inclusion of a semantic component, specifically the Categories task, adds unique variance associated with Aβ-related cognitive decline beyond that provided by the original PACC components [22]. Each study used a version of the PACC-5 that has been previously published [23–25]. In each study, the PACC-5 includes some overlapping tests (Mini-Mental State Examination [MMSE] and Logical Memory Delayed Recall) and some non-overlapping tests (for ADNI: ADAS-Cog Word Recall, Trails B, and Categories (Animals); for AIBL: the California Verbal Learning Test (second edition), Digit Symbol substitution, and Categories (Animals/Names); for HABS: the Free and Cued Selective Reminding Test, and Categories (Animals/Vegetables/Fruit). It is important to note that all three studies had two overlapping tests in the PACC-5 (MMSE and Logical Memory), however, the other three tests were unique to each study. All test scores were standardized within their own study according to the baseline mean and standard deviation of CDR=0 participants from the respective cohort. The PACC-5 was formed by averaging these z-scores. Baseline and longitudinal slopes for the PACC-5 were compared across the three studies to determine whether means and variances were similar (see Appendix A for cross-cohort distributions). ADNI and HABS participants completed these tests approximately every year, whereas AIBL participants underwent testing every 1.5 years. All available testing sessions following the analysis-defined baseline session were used (for ADNI: 324 participants completed 3 visits, 248 completed 4 visits, 166 completed 5 visits, and 28 completed 6 visits; for AIBL: 155 completed 3 visits, 73 completed 4 visits; for HABS: 244 completed 3 visits, 234 completed 4 visits, 183 completed 5 visits, 116 completed 6 visits). In order to adjust for baseline performance in our models, our cognitive outcome variable was PACC-5 change from baseline (with baseline cognitive performance as a covariate).

2.3. Aß positron emission tomography (PET)

ADNI uses the ¹⁸F-AV45 (Florbetapir or FBP) Aβ-PET tracer, while AIBL and HABS use the ¹¹C-Pittsburgh Compound-B (PiB) Aβ-PET tracer. The PET acquisition parameters for each study have been published previously [21, 26, 27]. In brief, ADNI and AIBL’s PET acquisition time was 50–70 minutes post-injection (http://adni.loni.usc.edu/), while for HABS, PiB-PET data were collected 40–60 minutes post-injection. All raw Aβ-PET data were processed with a standard pipeline. For this pipeline, PET data underwent reconstruction and attenuation correction, were evaluated for head motion, and were co-registered/normalized to a PET template in MNI space using the SPM12 unified segmentation, normalization routine, which applies a rigid body registration, followed by an affine registration, and a nonlinear mapping that fits the image to pre-specified 6-class tissue probability map. Summary measures for regions of interest (ROIs) were computed from a probabilistic GTM-Seg atlas in MNI space (Freesurfer v6.0 [28]) as standard uptake value ratios (SUVrs). The following ROIs that have been validated as AD regions of interest in previous publications [16] were aggregated: the frontal, lateral, and retrosplenial (FLR) regions. Values were normalized against the whole cerebellum to yield an Aβ FLR SUVr for each participant. To ensure cross-tracer equivalency, we applied a novel nonlinear transformation mapping approach (NLTM; further details in Appendix B). With this equating method, we extracted equivalent FBP SUVrs (FBP_equiv) for all PiB data so that they conformed to the FBP distribution. Although the NTLM is bidirectional, we chose to conform PiB SUVrs to the FBP distribution due to its more limited dynamic range and scale (Appendix B shows the pre- and post- distributions after applying NLTM in comparison with raw values and a linear-only transformation). This equivalence allowed for combined analysis of continuous Aβ FLR SUVr. For a post-hoc analysis, we also dichotomized Aβ using a Gaussian mixture modelling procedure [16], which gave an FBP_equiv cut-off of 1.082 for Aβ+ (referred to as Aβ_status to differentiate between the continuous Aβ measure).

2.4. Statistical analysis

Analyses were performed using R version 3.3.2. To determine sex differences in Aβ burden and APOEε4 we ran group comparisons using Wilcoxon/Mann-Whitney and chi-square (χ²) tests, respectively. To investigate sex-specific effects on cognitive decline in association with Aβ and APOEε4 status, we ran a series of hierarchical linear mixed models with subject-specific random intercept nested within cohort. We also modeled cohort as a fixed effect. Covariates were age, years of education, and baseline cognitive performance. As PACC-5 change is best modeled with quadratic time [29, 30], interaction terms with this time effect and all other covariates were included in the models. Only quadratic time terms were modeled, as the global extremum (vertex of parabola) of PACC-5 performance passed through zero at baseline, and an assessment of goodness-of-fit parameters suggested adequate fit against models including the linear terms. We compared goodness-of-fit of increasingly complex models using a log-likelihood ratio test. Multiple independent comparisons (n=5) were accounted for according to a Sidak correction of α = 0.01.

The following hierarchy of models were run:

1. PACC-5_change ~ sex^*time + covariates^*time

2. PACC-5_change ~ [Aβ OR APOEε4]^*sex^*time + covariates^*time

3. PACC-5_change ~ Aβ^*APOEε4^*sex^*time + covariates^*time

4. PACC-5_change ~ Aβ^*sex^*Age^*time + covariates^*time

where PACC-5_change is the change in PACC-5 from baseline, OR indicates different terms used in models A and B, Aβ is the continuous FBP_equiv SUVr, and the covariates were age, years of education and PACC-5 performance at baseline. Note that all lower-order terms were included in each model.

To further explore interactions in Model C between the dichotomized variables Aβ_status (+/−), APOEε4 (+/−) and sex (M/F), pairwise comparisons were performed for the groups: (Aβ−/APOEε4−/M (n=191), Aβ−/APOEε4−/F (n=220), Aβ−/APOEε4+/M (n=41), Aβ−/APOEε4+/F (n=62), Aβ+/APOEε4−/M (n=46), Aβ+/APOEε4−/F (n=59), Aβ+/APOEε4+/M (n=46), and Aβ+/APOEε4+/F (n=54)).

We also ran post-hoc analyses within each study to determine whether patterns of findings were consistent with the combined-cohort results.

3. RESULTS

3.1. Cohort characteristics

We examined 755 clinically-normal individuals with a median of 4 years of follow-up in their respective study (range = 3 – 7 years across the combined group), as summarized in Table 1. ADNI had the shortest follow-up duration (6.08 years), whereas AIBL had the longest follow-up duration (6.98 years). There were no significant differences in follow-up length by sex (t = 0.25, p = 0.80), Aβ status (t = 0.05, p = 0.96), APOE (t = −0.70, p = 0.48), sex^*Aβ status (F = 1.61, p = 0.21), or by sex^*APOE (F = 0.92, p = 0.34). AIBL participants were significantly younger, and had fewer years of education than the other studies. There were no cohort-level differences in the frequencies of APOEε4 carriers, females/males, or those with Aβ+ status. Differences also did not exist in baseline cognitive performance or individual cognitive slopes between the cohorts (cognitive slopes were extracted from ordinary least squares regression models).

Table 1.

Comparison between the studies on demographics and cognitive performance

	ADNI (n = 330) Mean (SD)	AIBL (n = 161) Mean (SD)	HABS (n=268) Mean (SD)	F, χ2	Effect size (ηp2 or Cramer’s V)	p
Follow-up (yrs)	3.29 (1.2)	4.99 (1.3)	4.56 (1.3)	129.1	0.26	<0.001
Age	74.6 (6.5)	71.6 (6.8)*	73.4 (6.0)	11.3	0.03	<.001
Females (n/%)	173 (52%)	86 (53%)	159 (59%)	3.1	0.06	0.21
Years of	16.4 (2.7)	13.7 (2.2)*	15.8 (3.0)	53.8	0.12	<0.001
Education
APOEε4+ (n/%)	88 (27%)	49 (32%)	67 (27%)	1.13	0.04	0.56
Aβ SUVr (FBPequiv)	1.04 (0.2)	1.05 (0.2)	1.04 (0.2)	0.23	0.0006	0.79
Aβ + status (n/%)	96 (29%)	46 (29%)	76 (28%)	0.01	0.004	0.99
MMSEbaseline	29 (1.2)	29 (1.2)	29 (1.1)	0.37	0.001	0.69
PACCbaseline	0.03 (0.6)	0.07 (0.6)	0.02 (0.7)	0.34	0.001	0.71
PACCslope	−0.05 (0.2)	−0.03 (0.2)	−0.03 (0.2)	1.01	0.003	0.36

Note: ADNI = Alzheimer’s Disease Neuroimaging Initiative; Australian Imaging, Biomarker and Lifestyle (AIBL) study; HABS = Harvard Aging Brain Study (HABS); SD = standard deviation; FBP_equiv = Florebetapir equivalent score; APOE = apolipoprotein

3.2. Sex differences in Aβ burden and APOE carrier status

Females did not exhibit greater median Aβ burden (Diff = −0.006, CI 95% [−0.023, 0.011], p = 0.48; see Figure 1A and cohort-level findings in Appendix D). We also did not find sex differences according to Aβ status (χ² = 0.80, Cramer’s V = .04, p = 0.37). Females were not more likely be APOEε4 carriers (χ² = 0.39, Cramer’s V = .02, p = 0.53; see Figure 1B). Neither were female APOEε4 carriers more likely to exhibit greater median Aβ than male APOEε4 carriers (Diff = 0.010, CI 95% [−0.045, 0.066], p = 0.76).

Figure 1.

(A) Females and males display equal proportions of APOEε4 carrier status and (B) Aβ burden. (C) Represents decline in global cognition (as measured by the PACC-5) by Aβ and sex. These are model estimates from a continuous model of Aβ; high and low Aβ are represented by the first quartile (on the left) and third quartile (on the right) of Aβ along the continuous spectrum. Each line extends to the longest follow-up period within that group.

3.3. Longitudinal change in cognition

For the first set of mixed-effect models which looked at main effects of sex over time, a main effect of sex did not associate with cognitive decline after adjusting for covariates (p = 0.05; see Table 2 for model estimates of terms of interest, with full models to be found in Appendix D).

Table 2.

Summary of linear mixed models for PACC-5 change (only terms of interest included)

	Estimate	Std Error	t value	p value
Model A: Sex
Sex	0.11	0.04	2.73	0.006
Sex*time2	− 0.003	0.002	− 1.98	0.05
Model B1: Aβ*Sex
Aβ*time2	− 0.04	0.01	− 5.09	< 0.001
Sex*time2	0.03	0.01	2.58	0.01
Aβ*Sex*time2	− 0.03	0.01	− 2.96	0.003
Model B2: APOEε4*Sex
APOEε4*time2	− 0.01	0.003	− 3.95	< 0.001
Sex*time2	− 0.005	0.002	− 2.30	0.02
APOEε4*Sex*time2	− 0.006	0.005	1.32	0.18
Model C: Aβ*Sex*APOEε4
APOEε4*Sex*time2	− 0.03	0.02	− 1.43	0.15
Aβ*Sex*time2	− 0.04	0.02	− 2.44	0.01
Aβ*Sex*APOEε4*time2	0.04	0.02	1.37	0.17
Model D: Aβ*Sex*Age
Aβ*Sex*time2	0.33	0.12	2.70	0.007
Aβ*Age*time2	0.002	0.001	1.54	0.12
Sex*Age*time2	0.005	0.002	2.70	0.007
Age*Aβ*Sex	0.008	0.03	0.25	0.80
Aβ*Sex*Age*time2	− 0.005	0.002	− 2.80	0.005
Model: Comparisons across Aβstatus/APOEε4/Sex groups
Aβ−/APOEε4−/M*Time2|Aβ−/APOEε4−/F*Time2	0.001	0.002	0.36	0.72
Aβ−/APOEε4+/M*Time2|Aβ−/APOEε4+/F*Time2	− 0.004	0.005	− 0.78	0.43
Aβ+/APOEε4−/M*Time2|Aβ+/APOEε4−/F*Time2	− 0.009	0.005	− 1.85	0.06
Aβ+/APOEε4+/M*Time2|Aβ+/APOEε4+/F*Time2	− 0.009	0.005	− 2.04	0.04

Note: Aβ is the FBP SUVr_equiv. Aβ+/− is formed according to the Aβ SUVr cut-off = 1.082. Esimates are unstandardized and the reference group for sex is female

For the next set of models looking at Aβ^*sex and APOE^*sex interactions over time on cognitive decline, females with elevated Aβ exhibited steeper cognitive decline than males with elevated Aβ over time (p = 0.003; see Figure 1C). A comparison between the Aβ main effects model and Aβ^*sex interaction model showed that the latter model fit significantly better than the former, Log Likelihood Ratio (15, 19) = 185.57, p <0.001. By contrast, we did not find an interaction effect of sex and APOE on cognitive decline (p = 0.19).

For the final interaction model between Aβ^*sex^*APOE over time, a significant interaction did not exist on cognitive decline (p = 0.17). Post-hoc contrasts, however, suggested a weak effect of female APOEε4 carriers with high Aβ_status (Aβ+/APOEε4+/F) having steeper cognitive decline in comparison with male APOEε4 carriers with high Aβstatus (p = 0.04; see Figure 3). This effect, however, did not survive multiple comparison adjustment.

Figure 3.

Decline in global cognition (as measured by the PACC-5) by APOEε4 status, sex and Aβ status (see legend for colours). These are model estimates from a factorial model with Aβ status represented by cut-off 1.082. Each line extends to the longest follow-up period within that group.

We next examined the effect of age on the interaction between Aβ and sex on cognitive decline, and found that females with higher Aβ displayed steeper cognitive decline after approximately 80 years of age in comparison with males (see Figure 2). This model did not fit better than a simpler model of Aβ^*sex, Log Likelihood Ratio (17, 23) = 4.55, p = 0.60, and as such, we interpreted this finding with caution. We did not consider an interaction between sex, amyloid and education, as our models were not statistically powered to provide reliable estimates (given that only 15 males and 28 females with high Aβ had 12 or less years of education).

Figure 2.

Decline in global cognition (as measured by the PACC-5) by Aβ and sex across the age span (females = pink, males = blue). These are model estimates from a continuous model of Aβ; high and low Aβ is represented by the first quartile (on the top) and third quartiles (on the bottom) of Aβ along the continuous spectrum. Age is also treated continuously in the model, with this visualization showing model estimates at the following ages: 65, 70, 75, 80, 85 years. Each line extends to the longest follow-up period within that group.

When considering each of the three studies in isolation, the same pattern of effects existed for the Aβ^*sex^*time interaction on cognitive decline. Due to statistical power constraints, however, many of these models did not meet the conventional threshold for statistical significance (see Appendix D). Models of Aβ^*sex^*time interactions were also conducted on each z-scored test within the PACC-5 (see Appendix A). We found that decline on list-learning delayed recall was significantly steeper in females than males for a given level of high Aβ (estimate = −0.07, SE = 0.02, t = −4.07, p < 0.001), with performance on Logical Memory delayed recall trending (estimate = −0.03, SE = 0.02, t = −1.91, p = 0.05). As list-learning tests were different across the cohorts (ADNI = ADAS-Cog Word Recall, AIBL = CVLT, HABS = FCSRT), we argue that these exploratory analyses warrant replication.

4. DISCUSSION

In a large cross-cohort dataset of clinically-normal older adults, females with elevated Aβ were found to decline in cognition more rapidly than males with a comparable level of Aβ. This mirrors recent work in ADNI reporting an interactive effect of sex and CSF Aβ₄₂ on episodic memory decline [31]. We did not find an interactive effect of sex and APOE on cognitive decline, however, there was a weak effect of female APOEε4 carriers with higher Aβ demonstrating faster rates of cognitive decline in comparison with their male counterparts. There was no main effect of sex on cognitive decline, although females exhibited better cognitive performance at baseline, supporting the notion that females outperform males on tests of verbal memory tasks [32], which are key components of the PACC-5 [22].

The mechanism explaining the interaction between sex and Aβ remains unclear. In line with previous studies [33–35], sex was not associated with Aβ burden at baseline (as either a continuous or dichotomous variable), even in APOEε4 carriers. It is important to note that sex differences in amyloid load have been reported in relation to family history; for instance, maternal history of sporadic AD dementia influences Aβ-PET retention in normal subjects [36], and closer proximity to one’s parental estimated year of onset of sporadic AD is associated with elevated amyloid burden in female, but not male, subjects [37]. We did not measure family history in the current study, however, it is possible that co-varying for proximity to parental age at onset may highlight differences in amyloid burden between males and females.

An alternative possibility is that sex effects in relation to Aβ burden do exist, but perhaps at earlier ages, such as during menopause, which was not examined in the current study. Postmortem work shows extensive senile plaque build-up in women who are in the neurofibrillary stages I, II and III compared with men in similar stages, and particularly in those with APOEε4 [38]. Animal studies have also reported elevated risk for greater Aβ burden in females than males in a range of transgenic mouse models [39, 40]; much more so than sex differences in tauopathy [41]. Estrogen is often implicated in findings of sex differences in animal models [42], with ovariectomies increasing Aβ burden in female mice [43], and estrogen replacement reducing the risk of Aβ burden [44]. These findings mirror human studies suggesting that higher estradiol in females is associated with better memory performance [45], oophorectomies that occur prior to menopause increases risk of cognitive impairment in females [46], and epidemiological studies that show chronic (10+ years) use of hormone replacement therapy close to menopause may be a protective factor for AD risk [47]. Clinical trials of estrogen replacement, however, have been disappointing to date [48], highlighting the complex role that sex hormones may play in AD dementia risk [1]. We did not examine menopausal onset in the current study, nor did we measure use of hormone replacement therapy. It will be important for future studies to assess cognitive changes in relation to Aβ, APOE and sex prior to and during the full menopause phase.

Other interpretations should also be considered. If similar levels of high Aβ burden lead to steeper cognitive decline in women, this may imply a greater sensitivity in women to Aβ burden relative to men, which may be mediated by greater levels of tau and/or neurodegeneration in Aβ positive females. It is also possible that, despite similar levels of Aβ burden at baseline, females may accumulate Aβ at a faster rate, however, the literature has not comprehensively investigated this supposition. In one recent study, proximity to parental estimated year of onset for sporadic AD corresponded with higher Aβ burden cross-sectionally in females relative to males, but no evidence was found of sex differences in Aβ accumulation longitudinally [37].

As mentioned above, downstream mechanisms of Aβ burden, such as tauopathy and neurodegeneration, may occur to a greater extent in Aβ positive females than males, that would mirror our finding of greater cognitive decline among the same group. Mounting cross-sectional evidence suggests that females, particularly those with APOEε4, exhibit greater levels of tau pathology and neurodegeneration [33, 35, 49, 50]. Female APOEε4 carriers show greater levels of CSF total tau [33], hippocampal atrophy [51], cortical thinning [50], and lower intrinsic connectivity in the default network [49] in comparison with male carriers. However, other work, by Jack and colleagues [35] showed that in normal subjects, males have smaller hippocampal volumes in comparison with females, and Koran and colleagues [31] did not find an interaction between sex and CSF total tau on episodic memory decline in a model including normal, MCI and AD dementia patients. As such, further research should focus on the question of whether sex differences in tauopathy or neurodegeneration might drive risk for AD-related cognitive decline.

Unlike previous studies on clinical progression to mild cognitive impairment (MCI) or AD dementia [12, 33, 52], we did not find a sex-by-APOE interaction effect on cognitive decline. It is possible that previous APOE findings represent proxies for Aβ and/or neurodegeneration effects that are not feasibly measurable in epidemiological cohorts. Previous studies report that risk for clinical progression in female APOEε4 carriers exists largely within the specific 65-75 age range [12, 52], suggesting that cognitive decline would need to occur prior to this age. We were unable to accurately estimate cognitive decline in those below 65 years of age due to small sample sizes in that range, and so we may have failed to capture sex-APOE effects on cognition. A weak three-way sex-APOE-Aβ interaction was found, but this finding did not survive multiple comparison adjustment and thus requires replication.

Another departure from previous studies relates to the age at which sex effects appeared; we found cognitive decline was more likely at later decades, unlike epidemiological studies that report effects in those between 65-75 years [12]. It is possible that our finding represents a ‘survivor bias’ effect [53], such that sex-related cognitive decline in the oldest-old is a ‘second-wave’ risk. Individuals who have survived past factors that reduce the age of onset to AD dementia, such as APOE [54], and other factors that increase mortality, such as early male death from cardiovascular disease [55] may result in differential vulnerability to sex effects. As followup time did not differ in our study by sex it is unlikely that our findings are simply driven by the fact that women were followed for longer. Our age interactive model did not significantly explain more variance than simpler models, and so any interpretation of age effects should be considered with this caveat in mind.

Our study has several limitations. We investigated three cohorts of convenience: participants are primarily educated, of higher socioeconomic status, and are not very racially diverse compared to the general population. In addition, the AIBL study is enriched for APOEε4 carriers, thus increasing the proportion of carriers over the population prevalence rate [20]. As such, these cohorts may not reflect sex effects in the general population. In addition, we assumed similarity between cohorts, although demographically they were slightly different at baseline. In order to account for this, we processed Aβ-PET data with the same PET-pipeline, and scrutinized baseline and longitudinal cognitive performance to ensure similarity of cognitive performance. We also accounted for random cohort effects in our linear mixed effects models, and ran analyses within each cohort to confirm that the pattern of results was similar across the cohorts. As such, we argue that the advantage of combining cohorts allows idiosyncratic noise from within each cohort to be smoothed out. Nevertheless, these findings require replication in other large datasets, such as Mayo Clinic, the Framingham Heart Study, or even meta-analysis datasets such as the Amyloid Biomarker Study [56].

It is important to note that harmonizing across cohorts is a complex endeavor [57], particularly with regard to equating across neuropsychological test performance. As the PACC-5 composite included both overlapping and non-overlapping tests, we found similar baseline and slope variability across the cohorts. This does not negate the fact that biases may be introduced by including three non-overlapping cognitive tests to form the PACC-5 across the three cohorts. Our future aim will be to further harmonize across neuropsychological tests in order to probe within cognitive domains (e.g. episodic memory, executive function, processing speed, etc). Neuropsychological tests within each cognitive domain largely do not overlap between the cohorts, and as such, it will be necessary to implement item-response theory equating methods, similar to those applied in other cohort-harmonization studies [58]. We found that memory, in this case, performance from list-learning and story learning tests from the PACC-5, may be the domain most affected by the sex^*Aβ interaction, however, as these tests do not overlap between the studies, our analyses are highly exploratory and need to be replicated in other cohorts.

One strength of our study involves our treatment of the different Aβ-PET tracers across the studies. In order to address differences in dynamic range, sensitivity and scaling between PiB and FBP Aβ-PET tracers, we employed a non-linear transformation mapping approach. We used this method to account for non-linearities that exist in the extreme ends of the PiB and FBP SUVr distributions. An assessment of cumulative distribution functions between raw and equivalent SUVrs showed that non-linear mapping preserves the distributional properties of the original SUVr much better than linear transformations (see Appendix B). Regardless, methods to accurately translate SUVrs across Aβ tracers is complex, and will require further refinement and testing in out-of-sample cohorts.

Determining sex-specific effects on rates of cognitive decline in the context of genetic risk and AD pathology will aid in more accurate detection of individuals most vulnerable to AD pathology, and further clarify recruitment and treatment approaches for AD clinical trials with reference to sex. We found that sex differences, both as a main effect and interacted with APOEε4, were not apparent on cognitive decline, however, females with high Aβ burden did show steeper cognitive decline trajectories in comparison to males with similarly high Aβ burden. Taken together, these findings suggest that sex effects may be salient in preclinical AD, when Aβ burden is apparent [1]. The mechanism underlying sex-specific sensitivity to Aβ, however, is yet to be elucidated. Taken together, these findings imply that the influence of sex and sex-specific risk factors on AD risk should be considered with respect to the cascade of events thought to underlie the development of AD dementia, and that the impact of sex is present during the preclinical stage of the disease. The effects observed in this study do, however, highlight the benefit of harmonizing across smaller datasets of clinically normal older adults to detect complex modifying effects in preclinical AD, which will become important at the scalable-level of large clinical trials, such as Anti-Amyloid Treatment in Asymptomatic Alzheimer’s Disease (A4) [18].

Research in context.

Systematic review: The authors reviewed the literature using Google Scholar and PubMed sources. Sex effects on cognitive decline within the context of Alzheimer’s disease biomarkers are not yet as widely published, however, there have been some publications on sex differences in dementia incidence and AD biomarker risk which the authors referred to. These relevant citations are appropriately cited.

Interpretation: Although the sexes did not differ on amyloid burden or APOE carrier status, our findings suggest that females decline faster in cognition for a given level of amyloid in comparison with males.

Future directions: In order to examine sex-related amyloid effects on cognition, it will be important to measure neurodegenerative and tauopathy associations. Further, cognitive domains should be investigated; it is possible that certain cognitive domains may be more affected in females than males for a given level of amyloid, however, this remains to be investigated.

Highlights.

• Females are not more likely to exhibit steeper cognitive decline

• Females are not more likely to exhibit high amyloid or carry APOEε4 than males

• For a given level of amyloid-PET, females exhibit steeper cognitive decline than males

• For a given level of APOE risk, females do not exhibit steeper decline than males

APPENDIX A: Density plots of baseline PACC performance and longitudinal PACC slopes for each cohort

APPENDIX B: Visualisation of the Amyloid-PET SUVr distributions

Visualisation of the Amyloid-PET SUVr distributions of (A) raw SUVrs across the three cohorts (AIBL and HABS use PiB, while ADNI uses Florebetapir), (B) linearly transformed SUVrs, and (C) our method of non-linearly transformed SUVrs of PiB to be equated with the Florbetapir distribution

Non-linear method:

We utilize baseline Aβ-PET datasets of clinically-normal participants from ADNI (FBP), AIBL (PIB), and HABS (PIB). Summary measures were computed via a PET only pipeline with direct spatial normalization of PET data to MNI space. Measurements were made as SUVr using a cortical composite (FLR) with a whole-cerebellum reference region. Cross-sample mapping was performed via 10,000 bootstrapped samples of PiB and FBP matched for age, sex and apolipoprotein ε4 (APOEε4) status. A transfer function for each bootstrapped sample was generated via smoothed cumulative distribution functions. An SUVr equivalency map and confidence intervals were extracted across these 10,000 fits using sliding-window PCA.

APPENDIX C. Model estimates for each mixed-effects model in the combined group, followed by the model estimates within each cohort for the sex^*amyloid model

	Estimate	Std Error	t value	p value
Model: Sex
Time2	0.03	0.01	2.70	<0.001
Baseline PACC-5	−0.21	0.03	−6.18	<0.001
Age	−0.01	0.003	−3.07	0.01
Education	0.01	0.007	1.95	0.07
Sex	0.11	0.04	2.74	0.003
Cohort	−0.03	0.02	−1.28	0.42
Cohort*time2	0.02	0.001	9.26	<0.001
Baseline PACC-5*time2	0.003	0.002	1.68	0.06
Age*time2	−0.001	0.0001	−6.25	<0.001
Education*time2	0.0004	0.0003	1.12	0.24
Sex*time2	−0.003	0.002	−1.99	0.05
Model: Aβ*Sex
Time2	0.07	0.01	4.59	<0.001
Baseline PACC-5	−0.22	0.01	−6.62	<0.001
Age	−0.009	0.003	−2.90	0.02
Education	0.02	0.007	2.32	0.02
Aβ	−0.31	0.17	−1.79	0.07
Sex	0.02	0.24	0.06	0.95
Cohort	−0.03	0.02	−1.29	0.41
Cohort*time2	0.01	0.001	8.90	<0.001
Baseline PACC-5*time2	0.001	0.002	0.82	0.41
Age*time2	−0.0008	0.0002	−4.95	<0.001
Education*time2	0.0002	0.0004	0.62	0.54
Aβ*Sex	0.10	0.23	0.44	0.66
Aβ*time2	−0.04	0.01	−5.09	<0.001
Sex*time2	0.03	0.01	2.58	0.01
Aβ*Sex*time2	−0.03	0.01	−2.96	0.003
Model: APOEε4*Sex
Time2	0.05	0.01	3.51	<0.001
Baseline PACC-5	−0.21	0.03	−5.98	<0.001
Age	−0.01	0.003	−2.97	0.003
Education	0.02	0.01	2.11	0.04
APOEε4	−0.03	0.07	−0.48	0.63
Sex	0.11	0.05	2.24	0.03
Cohort	−0.03	0.02	−1.36	0.40
Cohort*time2	0.01	0.001	9.24	<0.001
Baseline PACC-5*time2	0.003	0.002	1.59	0.11
Age*time2	−0.001	0.0001	−6.66	<0.001
Education*time2	0.0003	0.0003	0.83	0.40
APOEε4*time2	−0.01	0.004	−3.95	0.001
Sex*time2	−0.006	0.003	−2.30	0.02
APOEε4*Sex*time2	0.006	0.005	1.32	0.19
Model: Aβ*Sex*APOEε4
Time2	0.06	0.02	3.36	<0.001
Baseline PACC-5	−0.16	0.03	−5.47	<0.001
Age	−0.009	0.003	−3.15	0.002
APOEε4	0.45	0.34	1.31	0.19
Education	0.01	0.006	2.04	0.04
Aβ	−0.03	0.21	−0.13	0.90
Sex	0.31	0.29	1.09	0.28
Cohort	−0.03	0.02	−1.23	0.43
Cohort*time2	0.01	0.001	8.33	<0.001
Baseline PACC-5*time2	0.0003	0.002	0.15	0.88
Age*time2	−0.0007	0.0002	−4.42	<0.0001
APOEε4*time2	0.02	0.02	0.99	0.32
Education*time2	0.0003	0.0003	0.91	0.36
Aβ*time2	−0.03	0.01	−2.24	0.03
Sex*time2	0.04	0.02	2.30	0.02
Aβ*Sex*time2	−0.04	0.02	−2.44	0.01
Aβ*APOEε4*time2	−0.02	0.02	−1.17	0.23
Sex*APOEε4*time2	−0.03	0.02	−1.43	0.15
APOEε4*Aβ*Sex	0.42	0.41	1.03	0.30
Aβ*Sex*APOEε4*time2	0.03	0.02	1.37	0.17
Model: Aβ*Sex*Age
Time2	0.19	0.09	2.14	0.03
Baseline PACC-5	−0.16	0.03	−5.73	<0.001
Age	−0.001	0.02	−0.04	0.96
Education	0.01	0.006	2.10	0.04
Aβ	0.42	1.79	0.24	0.81
Sex	0.51	2.49	0.20	0.84
Cohort	−0.03	0.02	−1.30	0.42
Cohort*time2	0.01	0.001	8.89	<0.001
Baseline PACC-5*time2	0.0009	0.002	0.52	0.60
Age*time2	−0.002	0.001	−1.91	0.06
Education*time2	0.0001	0.0003	0.45	0.65
Aβ*time2	−0.17	0.09	−2.41	0.02
Sex*time2	−0.31	0.13	−2.41	0.01
Aβ*Sex*time2	0.32	0.12	2.60	0.009
Aβ*Age *time2	0.002	0.001	1.54	0.12
Sex*Age*time2	0.004	0.002	2.57	0.01
Age*Aβ*Sex	0.008	0.03	0.25	0.80
Aβ*Sex*Age* time2	−0.005	0.002	−2.80	0.005
Model: Comparisons across Aβ status/APOEε4/Sex groups
Aβ−/APOEε4−/M*Time2 vs Aβ−/APOEε4−/F*Time2	0.001	0.002	0.36	0.72
Aβ−/APOEε4+/M*Time2 vs Aβ−/APOEε4+/F*Time2	−0.004	0.005	−0.78	0.43
Aβ+/APOEε4−/M*Time2 vs Aβ+/APOEε4−/F*Time2	−0.009	0.005	−1.85	0.06
Aβ+/APOEε4+/M*Time2 vs Aβ+/APOEε4+/F*Time2	−0.009	0.005	−2.04	0.04

Note: Aβ is the FBP SUVr_equiv. Aβ+/− is formed according to the Aβ SUVr cut off = 1.082

Comparison of Aβ-amyloid^*sex model within each cohort

Coefficients	Study
	ADNI			AIBL			HABS
	Estimate	Conf. Int.	p-value	Estimate	Conf. Int.	p-value	Estimate	Conf. Int.	p-value
Fixed Parts
Intercept	0.69	−0.11-1.49	.090	0.52	−0.66-1.71	.390	0.77	−0.13-1.67	.095
I(timc_yrs^2)	0.12	0.05-0.19	.001	0.04	0.00-0.08	.030	0.12	0.08-0.16	<.001
PACC_bl	−0.20	−0.29-−0.10	<.001	−0.21	−0.37-−0.06	.008	−0.12	−0.21-−0.03	.013
YrsEd	0.01	−0.01-0.03	.321	0.01	−0.04-0.05	.795	0.01	−0.01-0.03	.195
Age	−0.01	−0.02-−0.00	.005	−0.01	−0.02-0.01	.417	−0.01	−0.01-0.00	.245
SEXF	0.08	−0.52-0.68	.796	0.18	−0.91-1.27	.749	−0.03	−0.72-0.67	.940
Amyloid	0.02	−0.42-0.46	.923	−0.26	−0.93-0.41	.451	−0.66	−1.20-−0.12	.017
I(timc_yrs^2):PACC_bl	0.01	0.00-0.02	.046	−0.00	−0.01-0.00	.633	−0.00	−0.00-0.00	.994
I(timc_yrs^2):YrsEd	0.00	−0.00-0.00	.126	0.00	0.00-0.00	.015	−0.00	−0.00-0.00	.096
I(timc_yrs^2):Agc	−0.00	−0.00-−0.00	.004	−0.00	−0.00-0.00	.269	−0.00	−0.00-−0.00	<.001
I(timc_yrs^2):SEXF	0.01	−0.04-0.06	.665	0.01	−0.02-0.05	.550	0.02	−0.01-0.05	.140
I(timc_yrs^2):Amyloid	−0.08	−0.12-−0.04	<.001	−0.05	−0.07-−0.02	<.001	−0.03	−0.05-−0.01	.005
SEXF:Amyloid	0.00	−0.57-0.57	.999	−0.08	−1.11-0.95	.883	0.14	−0.52-0.80	.682
I(timc_yrs^2): SEXF: Amyloid	−0.00	−0.05-0.04	.855	−0.01	−0.04-0.02	.550	−0.03	−0.06-−0.00	.045
Random Parts
σ2		0.214			0.131			0.143
τ00,ID		0.116			0.189			0.119
NID		330			160			265
ICCID		0.351			0.590			0.455
Observations		1102			510			1164
R2/Ω02		.613/.591			.761/.749			.653/.642