We utilized longitudinal models of different classes of biomarkers to test hypotheses derived from the Jack model (Jack et al., 2010
) about the temporal sequence of biomarkers. As expected, CSF Aβ and tau were independently associated with brain structure and function, and all classes of biomarkers that we investigated (CSF Aβ, CSF tau, neuroimaging) were independently associated with the cognitive outcomes of episodic memory and executive functioning. The focus of our analysis strategy was to examine how these individual relations changed as additional explanatory variables, downstream in the Jack model, were added. Results generally supported hypotheses derived from the Jack model. CSF Aβ effects on neuroimaging and cognitive outcomes were significantly diminished by controlling for CSF tau and independent effects on cognitive outcomes were not present after accounting for neuroimaging variables. Similarly, CSF tau effects on cognitive change were attenuated substantially by effects of neuroimaging variables. Results deviated from expectations generated by the Jack model in that CSF Aβ had effects on brain structure and function that were independent of CSF tau, and CSF tau had effects on baseline cognition that were independent of neuroimaging measures. There were some minor differences in the pattern of results for neuroimaging measures as outcomes, where CSF tau was more specifically related to FDG PET.
The Jack model derives from a substantial foundation of research that identifies production of Aβ as a critical step in the causal sequence leading to brain degeneration and cognitive impairment. Aβ is posited to have a direct neurotoxic effect, leading to impaired tau processing, impaired cell function presumably reflected at an aggregate level by FDG PET, to cell death reflected at an aggregate level by brain atrophy, and eventually to impaired cognition resulting from the structural and functional changes occurring in the brain. This is a causal model in which tau mediates effects of Aβ, and neuroimaging variables mediate effects of tau (and Aβ through tau). The sequence of analyses in this study was designed to examine relations among CSF Aβ and tau and longitudinal trajectories of brain structure, brain function, and cognition to test mediation effects predicted by the Jack model.
Longitudinal studies have major advantages for studying causal mediation. Cross-sectional studies are limited by correlations obtained at one point in time, and can provide misleading information about true longitudinal relationships in causal pathways (Maxwell & Cole, 2007
). An important advantage of longitudinal studies is that they can utilize information about temporal precedence, an important criterion for causal inference (MacKinnon et al., 2007
). Studies in which proposed mediator variables are measured longitudinally are especially valuable because they can be used to evaluate how change in the final outcome is mediated by change in intervening variables (Cole & Maxwell, 2003
, MacKinnon et al.,2007
, Maxwell & Cole, 2007
). However, there are complex issues that impact how effective a specific study will be for testing causal hypotheses. The amount of follow-up time and the lag between predictors and outcomes in particular are important considerations that apply to this specific study and present limitations with respect to conclusions about causal mediation.
The overall duration of this study was only 36 months, and it seems likely that changes in Aβ impact cognition, brain structure and function, and even changes in tau, with a lag of more than three years. For example, there are recent estimates that Aβ production in AD may begin decades before clinical changes (Negash et al., 2011
). It is also conceivable that cognitive changes could lag brain function and structure changes by more than three years. FDG PET studies have shown reduced cortical metabolism in genetically at risk, cognitively normal individuals who are substantially younger than the typical age of onset of the earliest clinical symptoms of AD (Reiman et al., 2005
). Consequently, the relatively short duration of follow-up in this study could limit ability to detect the full effects of earlier variables in the hypothesized sequence on later variables. Despite this important limitation, CSF Aβ and CSF tau individually had robust relations with brain imaging and cognitive measures. This indicates that a cross-sectional snapshot of the Aβ and tau profile at baseline is informative about the baseline values of the other variables and about change in those variables, and shows that this study was able to detect biological and cognitive effects of Aβ and tau. However, Aβ and tau did not explain cognitive change beyond effects of brain function and structure. The observed pattern of results provides evidence to support mediation of Aβ and tau effects on cognition by brain function and structure. That is, Aβ and tau individually were related to cognition and were related to brain function and structure, brain function and structure were related to cognition, but Aβ and tau were not related to cognitive change independent of brain function and structure. While these results are consistent with hypothesized mediation effects, research with substantially longer follow-up that corresponds to the time frame for transition from normal levels of Aβ and tau to clinical AD would provide a more definitive test.
One proposed mechanism for the association between increased Aβ and neuronal death involves a greater neural susceptibility to oxidative stress (Butterfield et al., 2002
). These results are also consistent with CSF Aβ's hypothesized role as one of the earliest biomarkers for AD. However, while it is certainly plausible that Aβ precipitates the clinicopathological symptom cascade of AD, it is clear that elevated levels of Aβ are not sufficient to produce clinical AD since many nondemented, cognitively-intact older adults show CSF, neuroimaging, and autopsy evidence of elevated levels of Aβ (Negash et al., 2011
). Therefore, models of Aβ ‘s role in AD neuronal degeneration and cognitive deterioration must also address the lack of cognitive deterioration in nondemented cases. Results of this study suggest that the most likely explanation is that Aβ in isolation does not cause cognitive impairment, but does initiate a cascade of brain changes that leads to subsequent cognitive impairment; Aβ in the absence of downstream changes is not sufficient to cause cognitive decline.
The observations from this study that CSF Aβ and CSF tau made independent contributions to brain changes and that tau had effects on baseline cognition that could not be explained by brain structure and function are not entirely consistent with the Jack (2010) model. The level of analysis that was possible in this study could impact these findings. Our measures of Aβ and tau were aggregate measures, summed across the whole brain, that likely represent complex effects of multiple production and clearance mechanisms. Similarly, our measures of ventricle volume and FDG PET were aggregates involving the whole brain. Much of the earlier research supporting the amyloid hypothesis and the Jack model is based on molecular and cellular changes, and it is possible that such changes, occurring as hypothesized, might not show the same temporal relations when applied to the whole brain. This clearly is an issue that warrants further research. While the CSF measures of Aβ and tau are inherently aggregate measures, a high level of specificity of neuroimaging variables by brain regions and functional systems is possible within ADNI, and these measures might be useful for more refined studies of how AD progresses. For example, the highly influential Braak and Braak model for staging of AD neuropathology (Braak, Braak, & Bohl, 1993
) could be tested using longitudinal, in-vivo data to address progression of changes in region-specific medial temporal and neocortical brain measures, how they relate to CSF Aβ and tau, and how they correspond to predicted domain-specific changes in episodic memory and executive function.
The effect of CSF tau on baseline cognition independent of brain structure and function also merits comment. Human and animal models have implicated tau's potentially greater role than Aβ in cognitive impairment (e.g., Oddo et al., 2006; Dowling et al., 2010); however, few studies have directly compared the contribution of imaging biomarkers and tau upon cognition over time. Baseline values in a study using a sample that is heterogeneous with respect to disease progression (normal, MCI, clinical AD) represent a cross-sectional snapshot of the accumulated effects of any biological changes that have occurred prior to the initial assessment. It is possible that the tau effects on baseline cognitive measures represent a cumulative effect of tau over a longer time frame than was covered by the longitudinal part of this study, and consequently, this may provide a more sensitive indicator of the effect of tau on longitudinal change. If this is the case, one might expect independent tau effects on longitudinal change in a study with longer follow-up. The inherent problems of using cross-sectional data to make inferences about longitudinal relationships (Maxwell & Cole, 2007) temper any conclusions that can be made about this finding, and this clearly is an open question for further research.
The well-documented heterogeneity in longitudinal trajectories as people age (Mungas et al., 2010, Wilson et al., 2002) presents a challenge in terms of finding variables that explain and predict who will and will not lose cognitive ability. This study included a diverse sample that, by design, incorporated heterogeneous trajectories; that is many MCI and most AD cases would be expected to decline at relatively rapid rates, while most normals and some MCI would likely show lesser to no cognitive decline. The basic approach in this study was to use continuous measures of CSF and neuroimaging biomarkers as measures of AD severity that would be useful for explaining this heterogeneity. Underlying this approach is the notion that AD severity is an important determinant not only of who does and doesn't decline cognitively, but also of rate of decline. Our study used linear models to characterize and explain trajectories, and this presents a limitation to the extent that non-linear trends are present.
A number of other studies have taken a somewhat different approach to explaining variability in longitudinal trajectories. These studies generally have compared non-demented individuals who develop MCI or dementia during follow-up with individuals who do not develop dementia, and have used change point analyses to model non-linear trends (Howieson et al., 2008; Jacqmin-Gadda, Commenges, & Dartigues, 2006; Johnson et al., 2009; Yu & Ghosh, 2010). Change in clinical diagnoses over time has been used in these studies to identify more homogenous subgroups, time to diagnosis has guided modeling of differential linear and non-linear trends in these groups, and ultimately, these studies explain heterogeneity of cognitive trajectories. Variables like CSF Aβ and tau might prove to be similarly useful for identifying more homogenous subgroups within the general heterogeneity of cognitive trajectories. For example, studies of differential trajectories within homogenous groups defined by CSF Aβ and tau could contribute to understanding the overall biological sequence across the spectrum of AD related changes. ADNI is ideally suited to this type of study, and this is a logical next step to follow this study.
There are important limitations of the present study. Representativeness of the ADNI sample is an important concern. The sample is a highly selected cohort of well-educated participants and it is not a community-based sample. Therefore, caution must be exercised in generalizations of implications drawn from the ADNI sample to more diverse community populations. A second limitation, previously identified, is that the duration of longitudinal follow-up in this study was relatively brief in comparison with current estimates of the temporal progression of AD related changes, which may limit our ability to characterize longitudinal relationships. A third, related limitation is the use of linear models to characterize and explain longitudinal change. Non-linear models might be more appropriate for describing the development of AD related change over longer periods, but the short follow-up time in this study limits the use of non-linear models. A fourth limitation is the restricted age range of the sample. There is some evidence that biomarkers may be sensitive at even earlier age ranges, and this could not be investigated with the current data. Finally, the lack of comprehensiveness of biomarkers investigated in the present study is a limitation. Amyloid imaging might be especially relevant, but has been collected on only a limited number of ADNI participants to date. Neuroimaging of tau is potentially important but currently is not available.
Despite these limitations, this study makes unique contributions. The use of a large multidimensional and longitudinal dataset afforded the opportunity to systematically test hypotheses about the temporal relationships among multiple biomarkers. To our knowledge, this is the first systematic, longitudinal assessment of the temporal sequencing of biomarker classes in AD. Future studies are needed to clarify the broader implications of this study for understanding and measuring AD progression in clinical and community samples.