Estimating preclinical amyloid positivity in the community: a case study transporting estimates from ADNI to ARIC

Emma K. Stapp, Elizabeth Rose Mayeda, Elizabeth A. Stuart, Rebecca F. Gottesman, Dean F. Wong, Thomas H. Mosley, Michael Griswold, M. Maria Glymour, Scott C. Zimmerman, Melinda C. PowerAlzheimer 2024 · DOI

TL;DR

This paper explores a statistical method called "transport" to estimate the prevalence of preclinical amyloid positivity in the general US population, using data from a specialized research cohort (ADNI) and applying it to a more representative cohort (ARIC-PET). The study found that while this method shows promise, it currently struggles to accurately capture prevalence across diverse demographic groups, highlighting the challenges of generalizing findings from convenience samples.

What Problem Does This Paper Address?

Estimating the prevalence of preclinical amyloid positivity – a key biomarker for Alzheimer's disease that appears before cognitive symptoms manifest – in the general US population is crucial for understanding disease burden and planning public health interventions. However, directly measuring this in large, representative samples is incredibly difficult and expensive, often requiring PET scans. This leaves a significant gap in our knowledge. The authors propose using a statistical technique called "transport" from causal inference. This method aims to "transport" findings from a deeply-phenotyped, but likely not perfectly representative, convenience sample (like ADNI) to a more representative target population (like ARIC-PET). The core problem is to assess whether this "transport" method can accurately estimate amyloid positivity prevalence in the target population, serving as a proof of concept for its utility.

Methods

The researchers employed a statistical "transport" method to estimate amyloid positivity prevalence. The core idea is to adjust for differences in how participants were selected into the ADNI study versus the ARIC-PET study.

  1. Participant Selection Modeling: They used Gradient Boosted Trees (GBM), a machine learning algorithm, to estimate the probability of an individual being selected into the ADNI study versus the ARIC-PET study. This model considered harmonized sociodemographic and clinical covariates common to both datasets.
  2. Inverse Odds of Sampling Weights: These estimated selection probabilities were then used to create "inverse odds of sampling weights." These weights are applied to the ADNI data to make it more representative of the ARIC-PET population.
  3. Prevalence Estimation: By applying these weights to the ADNI participants, the researchers "transported" the estimated prevalence of amyloid positivity to the ARIC-PET population.
  4. Validation: The transported prevalence estimates were then compared to the observed prevalence of amyloid positivity directly measured in the ARIC-PET study. This comparison was done overall and stratified by age, sex, race, education, and APOE e4 status.

The novelty lies in applying this causal inference "transport" framework to the specific problem of estimating preclinical amyloid positivity prevalence across different study populations.

Datasets

  • Alzheimer's Disease Neuroimaging Initiative (ADNI): A deeply-phenotyped convenience sample used as the source for transporting estimates.
  • Atherosclerosis Risk in Communities Study PET Amyloid Imaging Study (ARIC-PET): A more representative target sample used for validation.

The paper does not introduce new benchmarks but rather uses existing, well-established datasets for its validation.

Key Insights and Ideas

  • The Challenge of Generalizability: This study powerfully illustrates the inherent difficulties in generalizing findings from specialized research cohorts (like ADNI) to broader, more diverse populations. Even with sophisticated statistical methods, bridging this gap is not straightforward.
  • Selection Bias is a Major Hurdle: The "transport" method's success hinges on accurately modeling the selection process into the source study. The paper suggests that the available data might not be rich enough to fully capture the complex selection biases that lead to differences between convenience and representative samples.
  • Demographic Nuances Matter: The failure of the transport method to accurately estimate prevalence across all subgroups highlights that selection biases are not uniform. Certain demographic groups (e.g., white, non-Hispanic individuals) were better represented or selected differently, leading to more accurate transport for them.
  • Potential for Transport Methods: Despite the limitations, the study doesn't dismiss transport methods entirely. It suggests they could be more useful when applied to samples with known sampling strategies or when the source and target populations are more similar.

Key Results

  • Underestimation of Prevalence: In most demographic groups, the "transported" prevalence of amyloid positivity from ADNI to ARIC-PET substantially underestimated the observed prevalence in ARIC-PET.
  • Overestimation for Specific Groups: The exception was for white, non-Hispanic individuals and those with at least one APOE e4 allele, where the transported prevalence was slightly overestimated.
  • Raw ADNI vs. Transported ADNI: The prevalence estimates derived from transporting ADNI were generally further from the observed ARIC-PET prevalence than the raw (unadjusted) prevalence of amyloid positivity in ADNI itself, with the exception of white, non-Hispanic individuals and those with mild cognitive impairment.

Limitations and Future Work

  • Insufficient Data for Selection Modeling: The authors acknowledge that the available data may be insufficient to fully capture the selection biases into deeply-phenotyped convenience samples. This is a critical limitation for the transport method's accuracy.
  • Under-representation in Research: The transport method may be less effective for groups who are under-represented in research, as their selection processes might be harder to model with limited data.
  • Future Directions: The paper suggests that transport estimators will likely have greater utility when applied to samples derived from known sampling strategies (rather than convenience samples) and when there is increased diversity within dementia research cohorts. Future work could focus on identifying richer covariate data or alternative transport methods that can better account for complex selection mechanisms.

Summary updated

Abstract

Abstract Background Estimates of the prevalence of preclinical amyloid positivity in the US general population are of great interest to the field, but difficult to measure and thus unavailable in representative studies. A statistical approach from causal inference, ‘transport’, may allow for improved generalizability of findings from a sample of persons from one population to another. We aimed to explore the feasibility and validity of extending results from a deeply‐phenotyped convenience sample, the Alzheimer’s Disease Neuroimaging Initiative (ADNI), to a representative target sample, the Atherosclerosis Risk in Communities Study PET Amyloid Imaging Study (ARIC‐PET) — with “proof of concept” defined by the performance of the transport estimator in recovering the observed prevalence of amyloid positivity in ARIC‐PET. Method Eligible ARIC‐PET and ADNI participants were either white or Black and had normal cognition or mild cognitive impairment. Amyloid positivity was defined using study‐specific cutoffs for standardized uptake value ratio (SUVR). Probability of selection into ADNI (vs. into ARIC‐PET), given harmonized sociodemographic and clinical covariates, was estimated using gradient boosted trees (GBM). The resulting study participation probability scores were used to transport the prevalence of amyloid positivity from ADNI to ARIC using inverse odds of sampling weights. Estimates of amyloid positivity prevalence derived from ADNI and from transporting from ADNI to ARIC‐PET were compared to the observed prevalence in ARIC‐PET, overall and by age, sex, race, education, and APOE e4 status. Result Transported prevalences substantially underestimated observed prevalences of amyloid positivity in most groups, with the exception of white, non‐Hispanic individuals and persons with at least one APOE e4 allele, where prevalence was slightly overestimated. Prevalence of preclinical amyloid positivity based on transporting ADNI was further from the prevalence in ARIC‐PET than the raw prevalence of amyloid positivity in ADNI except for white, non‐Hispanic individuals and those with mild cognitive impairment. Conclusion Transport estimators rely on adequately modeling the selection process. Available data may be insufficient to capture selection biases into deeply‐phenotyped convenience samples, particularly for groups who are under‐represented in research. Transport estimators will have greatest utility when applied to samples derived from known sampling strategies, and with increased diversity in dementia research.
BibTeX
@inproceedings{stapp2024estimating,
  author = {Emma K. Stapp and Elizabeth Rose Mayeda and Elizabeth A. Stuart and Rebecca F. Gottesman and Dean F. Wong and Thomas H. Mosley and Michael Griswold and M. Maria Glymour and Scott C. Zimmerman and Melinda C. Power},
  title = {{Estimating preclinical amyloid positivity in the community: a case study transporting estimates from ADNI to ARIC}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.089765},
}
Figure