Estimating preclinical amyloid positivity: A case study transporting ADNI to ARIC

Abstract

INTRODUCTION

Alzheimer's disease and related dementias researchers have benefited from deeply phenotyped clinical samples; however, there is a critical need for estimates that generalize to diverse, representative populations. Use of a statistical approach from causal inference, “transport,” may allow generalization of findings from clinical samples to other target populations. Here we explore the feasibility and validity of extending results from a clinical sample, the Alzheimer's Disease Neuroimaging Initiative (ADNI), to a community‐based target sample, the Atherosclerosis Risk in Communities Study Positron Emission Tomography Amyloid Imaging Study (ARIC‐PET) using transport estimation and a standard approach, direct standardization, which itself can be considered a simple transport.

METHODS

Eligible ARIC‐PET (n = 343) and ADNI (n = 821) participants were White or Black, with normal cognition or mild cognitive impairment (MCI; 26.5% ARIC‐PET, 56.4% ADNI). Estimates of amyloid positivity prevalence were derived from transporting from ADNI to ARIC‐PET or standardizing ADNI to ARIC–ET using only sociodemographic characteristics and apolipoprotein E (APOE) ε4 status. Resulting estimates were compared to observed prevalences in ARIC‐PET, overall and by age, sex, race, education, APOE ε4 status, and cognitive status.

RESULTS

Approximately half of transported prevalences were closer to observed ARIC‐PET prevalences than the crude ADNI prevalences, including in all five subgroups in which crude prevalences differed substantially across cohorts. However, for many subgroups, transported prevalences were substantially further from observed ARIC‐PET prevalences than crude ADNI prevalences. Standardization produced more variable estimates, which were not systematically closer to observed ARIC‐PET prevalence than the crude ADNI prevalence estimates. Restriction to more homogenous samples did not improve performance of either method.

DISCUSSION

Although transport performed better than direct standardization in this example, available data appear insufficient to generalize findings from convenience samples to less selected samples with high confidence using either method. Recruitment of diverse, representative samples will likely be needed to derive population‐level statistics given limitations of legacy samples.

Highlights

• Prior prevalence estimates of amyloid positivity are variable.

• Transport may improve generalization from samples to target populations.

• We standardized and transported estimates to a target with known prevalence.

• Neither worked reliably; transport performed better than standardization.

• Clinical convenience samples are inadequate to derive population‐level statistics.

1. BACKGROUND

Sample participants in deeply phenotyped clinical studies recruited via convenience sampling typically differ across many characteristics from the target population to whom such research is meant to be applied. This divergence can lead to prevalence or effect estimates that are not readily generalizable to the population of interest. For example, our prior work ¹ has shown that associations between dementia risk factors, cognitive test outcomes, and imaging outcomes differ significantly and meaningfully between one such deeply phenotyped convenience sample, the Alzheimer's Disease Neuroimaging Initiative (ADNI), ² , ³ and a community‐based sample, the Atherosclerosis Risk in Communities Study Positron Emission Tomography Amyloid Imaging Study (ARIC‐PET). ⁴ The full potential of deeply phenotyped clinical convenience samples will not be realized until inferences can validly be extended to the at‐risk population.

Use of “transport” estimators—statistical approaches that seek to extend inferences from a given sample to broader or different target populations—may allow for improved generalizability of findings. ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ For example, investigators recently used a weighting tool to generalize race/ethnicity‐specific prevalence estimates of cognitive impairment from a cohort of health plan members to the California population of older adults without dementia. ⁷ Broader use of these approaches could greatly expand our understanding of dementia due to the availability of highly selected, deeply phenotyped clinical convenience samples and the challenges of collecting biomarker and imaging data from large, population‐based samples. However, while theoretically sound, it remains unclear whether this approach can be validly applied in real‐world data, given the strong assumptions that are required for transportability. ⁵ , ⁶ , ⁸ , ¹⁰ , ¹⁶

Here, our objective was to explore the feasibility and validity of using transport estimators to generalize estimates of prevalence of amyloid positivity, a neuroimaging biomarker of Alzheimer's disease (AD) pathology, in persons without dementia from a clinical convenience sample to a sample of community‐dwelling older adults drawn from a community‐based cohort study—the ultimate goal being to evaluate feasibility and validity of using of these tools to generate future population‐level estimates. Specifically, we applied a doubly robust estimator to transport the prevalence of amyloid positivity in older adults with normal cognition or mild cognitive impairment (MCI) in ADNI, a selected convenience sample designed to reflect historical randomized trial samples, to ARIC‐PET, a community‐based sample of older adults with normal cognition or MCI in which the outcome of interest was observed and could thus be used as a benchmark of performance. The extent to which the estimated prevalence of amyloid positivity (i.e., transported from ADNI) approximates the observed prevalence in ARIC‐PET provides a unique opportunity to check whether the underlying assumptions were met and provides insight into the feasibility of using transport methods to extend inference from clinical samples to less selected target populations, including the general population, where it is not feasible to check these assumptions. Because amyloid positivity has been observed to vary across sociodemographic, genetic, and cognitive strata, ⁴ , ¹⁷ , ¹⁸ and because the utility of transport estimators will depend not only on recovery of overall but also subgroup‐specific estimates, we transport and compare both overall and stratum‐specific prevalence. Finally, we compared transport estimates to estimates from a widely used approach for generalizing findings, direct standardization, which itself is a simple form of a transport estimator, implemented here based solely on sociodemographic characteristics and apolipoprotein E (APOE) ε4 status.

2. METHODS

2.1. Source data

Derivation of our samples and harmonization efforts enabling this and prior work on generalizability have been described in detail elsewhere. ¹ We provide details relevant to the current work here.

ADNI (see Supplemental Methods) is a convenience sample of > 2000 participants aged 55 to 90 years enrolled with normal cognition, MCI, or early AD at ≈ 60 sites with specialized imaging technology and diagnostic expertise in the United States and Canada since 2004. ¹⁹ Participants were generally in good health at the time of enrollment, and attended a screening visit, a baseline visit, and subsequent follow‐up visits every 6 to 12 months. Data collected during the screening and baseline visits included socio‐demographic characteristics, genetics, bloodwork, functional status, neuropsychological tests, and brain imaging, including amyloid PET scans. Data used in the preparation of this article were obtained from the ADNI database (adni.loni.usc.edu). For up‐to‐date information, see www.adni-info.org.

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using traditional databases, meeting abstracts, and presentations of statistical and applied sources. Previous estimates of amyloid positivity were heterogeneous and attempts to produce population‐representative estimates have used direct standardization. We did not find examples of using other transport algorithms for estimation of such population‐representative prevalences or efforts to establish feasibility or validity of such approaches.

2. Interpretation: Transport performed better than direct standardization, but neither approach performed consistently across subgroups. Results caution against using direct standardization to estimate population preclinical amyloid positivity and suggest that generalization of such estimates from clinical convenience samples is neither valid nor feasible.

3. Future directions: Clinical convenience samples may not be adequate sources for deriving population‐level statistics. Transport may be feasible across diverse population‐based samples where sampling strategies are known and attention is paid to drivers of non‐response; additional work will be required to establish feasibility and validity.

Between 1987 and 1989, ARIC enrolled 15,792 adults ages 45 to 64 years from four US communities. Each of the four sites (Jackson, Mississippi [MS]; Washington County, Maryland [MD]; Forsyth County, North Carolina [NC]; and selected suburbs of Minnesota [MN]) used a variation of random sampling to identify potential participants (see supporting information). ²⁰ Thus, at enrollment, ARIC participants were considered to be representative of their communities. At Visit 5 (2011–2013), study clinicians provided final diagnoses of cognitive status (normal cognition, MCI, dementia) after algorithmic assessment based on cognitive and functional tests. ²¹ The ARIC‐PET enrolled a subset (n = 346) of the ARIC‐Neurocognitive Study that took place at ARIC Visit 5, ultimately including participants from the NC, MS, and MD sites with completed magnetic resonance imaging (MRI) who were cognitively normal or had MCI (see details in Supporting Information). ⁴

2.2. Eligibility criteria

For this cross‐sectional analysis, we used ARIC Visit 5/ARIC‐PET data and ADNI baseline and screening data from any ADNI phase from 2004 through 2019. The ARIC sample was limited to White and Black participants at Visit 5 with normal cognition or MCI who participated in ARIC‐PET (n = 343). Similar inclusion criteria were applied to the combined ADNI screening/baseline visit data—White or Black race, normal cognition or MCI, and available PET amyloid imaging data (n = 821).

2.3. Outcome

Both ADNI and ARIC‐PET used florbetapirAV‐45 and similar imaging procedures (e.g., FreeSurfer software, see Gianattasio et al., ¹ Gottesman et al., ⁴ and Landau and Jagust ²² for cortical regions of interest), with standardized protocols and imaging analysis across sites within studies. ⁴ , ²³ Our outcome of interest was amyloid beta (Aβ) positivity, defined as achieving PET imaging standardized uptake value ratios (SUVR) greater than study‐specific cutoffs (1.11 in ADNI, 1.20 in ARIC), following convention and prior studies. ¹ , ⁴ , ²²

2.4. Covariates

All measures were harmonized across ADNI and ARIC as described previously. ¹ Socio‐demographic characteristics included age, sex (men vs. women), self‐reported race (Black vs. White), education level (greater than high school diploma [HS]/General Educational Development [GED] test vs. up to and including HS/GED), and marital status (married vs. not married). Cognitive performance tests administered in a similar manner across both studies included the Mini‐Mental State Examination (MMSE), animal naming (60 seconds), word fluency (letter F; 60 seconds), and the Boston Naming Test (BNT; 30 items). ²⁴ , ²⁵ , ²⁶ , ²⁷ , ²⁸ For use in analyses, scores on the BNT < 20 were rescored as 20 (i.e., Winsorized). Measures of functional status included a summary of four items from the Functional Activities Questionnaire (FAQ) that were worded identically or nearly so between ADNI and ARIC. ²⁹ We considered participant cognitive status (MCI vs. normal cognition) at ARIC Visit 5 or the ADNI baseline visit, ascertained according to study protocols (see supporting information). APOE ε4 status was dichotomized as the presence of at least one APOE ε4 allele versus only ε2/ε3 alleles. Vascular risk factors included history of hypertension and study‐measured total serum cholesterol, triglycerides, and systolic and diastolic blood pressure.

2.5. Overview of statistical analyses

Here, we explore the feasibility and validity of using a transport estimator—doubly robust augmented inverse‐odds weighting—to generalize from a highly selected, clinical sample to a community‐based sample. We also compare its performance to direct standardization, a classic approach that is, itself, a simple form of transport and that has been used previously to estimate the prevalence of preclinical amyloid positivity. ³⁰ , ³¹ Specifically, we used these methods to estimate amyloid positivity prevalence in a target population—a community‐based sample (i.e., ARIC‐PET)—using data from a convenience sample (i.e., ADNI), and compared the results to the known prevalence of amyloid positivity in the target population (i.e., ARIC‐PET). In addition, because our prior work comparing associations across ARIC and ADNI suggested greater comparability as sample homogeneity increased, ¹ we conducted sensitivity analyses repeating this process after restricting to subsets representing majority classes within ARIC‐PET and ADNI: (1) White individuals and (2) White individuals with normal cognition.

2.6. Analytical dataset preparation

Separately for ARIC and ADNI, we used multiple imputation by chained equations to impute missing covariate data: < 1% for marital status, systolic and diastolic blood pressure, and word fluency; 2% for APOE ε4; higher levels for FAQ (7%), cholesterol or triglycerides (19%), and BNT (11%). After a burn‐in of 25 iterations, we extracted a single imputed dataset for use in analyses, having conducted basic diagnostics to ensure there were no problems with implementation (e.g., no incompleteness, values within expected ranges); we then stacked the ADNI and ARIC datasets, adding an indicator for dataset for use in subsequent analyses.

2.7. Transport approach: doubly robust augmented inverse‐odds weighting

Transport approaches require modeling either the outcome and/or the probability of participation/selection; these two models can be combined to provide an estimator that is robust against misspecification of either model such that if either model is correct, the resulting estimate is consistent even if one of the two is wrong. ⁶ , ¹¹ , ¹² Thus, we used a doubly robust augmented inverse‐odds weight estimator (DR‐AIOW) ³² with maximum likelihood estimation, for the outcome model, implemented in Stata using the package teffects. ³³ , ³⁴ , ³⁵ DR‐AIOW is an efficient‐influence function estimator that includes an augmentation term in the outcome model (here, amyloid positivity prevalence) to correct the estimator if the “treatment” (here, participation/selection in ADNI [vs. ARIC‐PET]) model is mis‐specified. The DR‐AIOW estimates the participation/selection model and computes inverse weights, estimates separate regression models of the outcome for each participation/selection level to obtain participation/selection level‐specific predicted outcomes, and computes weighted means of the participation/selection level‐specific predicted outcomes, with standard errors that account for the three‐step process. These steps provide consistent estimates because “treatment” (participation/selection) is assumed to be independent of the outcome after conditioning on the covariates, analogous to d‐separation of selection from the outcome. ⁵ , ¹⁶ Logistic models were specified for prediction of both participation/selection and the outcome model; all socio‐demographic and clinical covariates described above in section 2.4 were included in both the participation/selection and outcome model commands. In sensitivity analyses restricted to White or White with normal cognition participants, the outcome and participation/selection models naturally excluded race, or race and cognitive status, respectively.

2.8. Classic approach: standardization based on demographic characteristics

A classic approach to generalizing results to a target population is the use of direct standardization. Standardization can be considered a non‐doubly robust version of transport relying only on specifying an outcome model. However, unlike transport methods, in which guidance suggests using all predictors of selection or outcome status, as typically implemented, standardization efforts typically incorporate only a few demographic characteristics. Here, we standardized ADNI to ARIC‐PET (i.e., ARIC‐PET provides the standard/population distribution) creating strata based on dichotomized versions of core sociodemographic characteristics (age, >/≤ 73 years; sex, male/female; race, Black/ White; education, HS or GED or less/greater than HS or GED), and APOE ε4 status (present/absent) using individual level weights based on the prevalence of stratum membership. We combined some strata across APOE ε4 status to avoid empty cells (see details and examples in supporting information). In sensitivity analyses restricted to White participants with normal cognition, we also combined older/male/White/HS or GED or less/with or without APOE ε4 into a single stratum.

2.9. Covariate balance

For both transport and standardization approaches, we assessed covariate balance pre‐ and post‐weighting via standardized mean difference (SMD) ⁷ , ¹⁴ for all covariates in our primary, stacked ADNI and ARIC‐PET dataset (supporting information). In the literature, an |SMD| < 0.25 is often taken to indicate adequate covariate balance. ³⁶

2.10. Prevalence estimation

We estimated amyloid positivity in ADNI generalized to ARIC‐PET using DR‐IAOW post‐estimation and standardization, overall and stratified by age, sex, race, education, APOE ε4 status, and cognitive status. We then compared these to both the crude ADNI prevalence and the observed prevalence in the ARIC‐PET study. In sensitivity analyses, we repeated this process after restricting to White individuals and White individuals with normal cognition.

3. RESULTS

Compared to the ARIC‐PET sample, our ADNI sample was less racially diverse and ADNI participants were more likely to be married and have greater educational attainment (Table 1). By design, ADNI was composed of similar numbers of persons with normal cognition and MCI; yet, despite having greater proportions of persons with MCI, at least one APOE ε4 allele, and functional activity limitations compared to ARIC‐PET, ADNI participants had slightly higher mean cognitive performance scores and lower prevalence of amyloid positivity.

TABLE 1.

Sample characteristics across ADNI and ARIC participants of Black or White self‐reported race, without dementia, with amyloid PET neuroimaging.

	ARIC‐PET	ADNI
Variable, mean (SD) or N (%)	N=343	N=821
Socio‐demographic characteristics
Age in years	75.9 (5.4)	72.1 (6.9)
Male	149 (43.4%)	410 (49.9%)
Black	146 (42.6%)	37 (4.5%)
Education: greater than HS/GED	169 (49.3%)	722 (87.9%)
Married a	193 (56.3%)	611 (74.4%)
Cognitive status
Normal cognition	252 (73.5%)	358 (43.6%)
MCI	91 (26.5%)	463 (56.4%)
Functional and clinical characteristics
Functional Activities Questionnaire a ≥1 (out of 4)	21 (6.1%)	151 (18.4%)
Serum cholesterol a (mg/dL)	181.2 (39.3)	192.1 (37.7)
Triglycerides* ≥150 (mg/dL)	80 (23.3%)	320 (40.0%)
Hypertension history (ARIC up to visit 5)	257 (74.9%)	377 (45.9%)
Systolic blood pressure a (mm Hg)	128.7 (16.0)	132.8 (16.9)
Diastolic blood pressure a (mm Hg)	65.4 (10.6)	73.8 (9.5)
At least one APOE ε4 allele a	103 (30.0%)	331 (40.3%)
Cognitive test outcomes
MMSE score	27.0 (2.4)	28.5 (1.6)
Boston naming score a	24.7 (3.6)	27.6 (2.4)
Animal naming score	15.7 (4.7)	19.5 (5.4)
Word fluency score a	11.2 (4.2)	14.0 (4.7)
PET imaging measure
Amyloid positive (SUVR ADNI > 1.11, ARIC > 1.2)	175 (51.0%)	381 (46.4%)

Abbreviations: ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; ARIC‐PET, Atherosclerosis Risk in Communities Study Positron Emission Tomography Amyloid Imaging Study; HS/GED, high school diploma/General Educational Development test; MCI, mild cognitive impairment; MMSE, Mini‐Mental State Examination; PET, positron emission tomography; SD, standard deviation; SUVR, standardized uptake value ratios.

^a Missing data multiply imputed.

Prior to implementing our transport estimator or standardization, SMDs suggested substantial differences across ADNI and ARIC‐PET (Figure 1). After direct standardization, covariate balance improved marginally, especially for the demographic variables used to standardize; however, measures of cognitive status and vascular health remained imbalanced. The transport estimator DR‐AIOW achieved a better overall covariate balance, with SMDs within |0.25| for all covariates except hypertension history, with many covariates achieving SMDs closer to |0.1|.

FIGURE 1.

Covariate balance as measured by standardized mean differences between ADNI and ARIC‐PET shown for ADNI unweighted and after applying transport estimator weights and direct standardization. Dotted lines indicated a commonly used threshold of <|0.25| for adequate covariate balance. ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; ARIC‐PET, Atherosclerosis Risk in Communities Study Positron Emission Tomography Amyloid Imaging Study; DR‐AIOW, doubly robust augmented inverse odds weights; FAQ, Functional Activities Questionnaire; HS/GED, high school diploma/General Educational Development test; MCI, mild cognitive impairment; MMSE, Mini‐Mental State Examination; PET, positron emission tomography.

In our primary analyses, transported prevalences from ADNI were closer to observed ARIC prevalences than crude ADNI prevalences overall and for half of considered subgroups, which included the five subgroups in which crude ADNI and ARIC‐PET prevalences differed most substantially (i.e., women; Black; no APOE ε4; normal cognition; MCI; Table 2, Figure 2). However, transported prevalences were substantially different from observed ARIC‐PET prevalences for other subgroups, and the confidence intervals of transported prevalences did not contain observed ARIC‐PET prevalences in 3 of the 12 subgroups considered (i.e., men; White; at least one APOE ε4). Standardization produced estimates that often differed substantially from both the crude ADNI prevalence estimates and the observed ARIC‐PET prevalence estimates and were often accompanied by wide confidence intervals. Moreover, standardization estimates were usually further from ARIC‐PET than the DR‐AIOW transports.

TABLE 2.

Estimated prevalence of amyloid positivity as observed in ARIC‐PET and ADNI and with ADNI generalized to ARIC‐PET using a transport estimator and direct standardization, overall and by sociodemographic and clinical strata.

Prevalence estimate (95% CI)	Observed prevalence, ARIC‐PET (n = 343)	Observed prevalence, ADNI (n = 821)	Transported prevalence, ADNI to ARIC‐PET using DR‐AIOW	Standardized prevalence, ADNI to ARIC‐PET
Overall	51 (46, 56)%	46 (43, 50)%	54 (51, 56)%	48 (38, 59)%
Stratum‐specific
Age in years
Under 74	41 (33, 50)%	41 (36, 45)%	43 (39, 47)%	25 (17, 36)%
74 and over	57 (50, 64)%	55 (49, 60)%	61 (57, 64)%	63 (49, 74)%
Sex
Female	55 (48, 62)%	48 (43, 53)%	55 (51, 59)%	41 (28, 55)%
Male	46 (38, 54)%	45 (40, 50)%	52 (48, 56)%	58 (43, 71)%
Race
White	42 (35, 49)%	47 (43, 50)%	46 (43, 50)%	46 (40, 53)%
Black	64 (56, 72)%	43 (27, 60)%	64 (60, 68)%	50 (27, 73)%
Education
Education: HS/GED or less	55 (47, 62)%	56 (46, 66)%	58 (55, 62)%	44 (28, 62)%
Education: greater than HS/GED	47 (40, 55)%	45 (42, 49)%	49 (45, 53)%	51 (38, 63)%
APOE ε4 status
No copies of APOE ε4 allele	44 (37, 50)%	30 (26, 34)%	42 (39, 45)%	29 (21, 38)%
At least one APOE ε4 allele	68 (59, 77)%	71 (66, 76)%	81 (79, 84)%	79 (64, 90)%
Cognitive Status
Normal cognition	45 (39, 51)%	34 (29, 39)%	48 (45, 51)%	40 (26, 55)%
Mild cognitive impairment	68 (58, 78)%	56 (52, 61)%	69 (65, 74)%	54 (39, 69)%

Note: Percentages of amyloid positivity are within a given stratum within a dataset, not of the dataset overall.

Abbreviations: ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; ARIC‐PET, Atherosclerosis Risk in Communities Study Positron Emission Tomography Amyloid Imaging Study; CI, confidence interval; DR‐AIOW, doubly robust augmented inverse odds weights; HS/GED, high school diploma/General Educational Development test; PET, positron emission tomography.

FIGURE 2.

Visualization of estimated prevalence of amyloid positivity as observed in ARIC‐PET and ADNI and with ADNI generalized to ARIC‐PET using a transport estimator and direct standardization, overall and by socio‐demographic and clinical strata. ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; ARIC‐PET, Atherosclerosis Risk in Communities Study Positron Emission Tomography Amyloid Imaging Study; CI, confidence interval; DR‐AIOW, doubly robust augmented inverse odds weights; HS/GED, high school diploma/General Educational Development test.

In sensitivity analyses in which samples were restricted to White participants, DR‐AIOW produced amyloid positivity estimates that were slightly closer to ARIC‐PET than crude ADNI prevalences in about half of estimates (Table 3); for the remainder of the estimates, including the overall, confidence intervals for the transported amyloid prevalence did not include the observed ARIC‐PET amyloid prevalence. Direct standardization produced estimates that were further from the observed ARIC prevalence than the crude ADNI prevalence in half of subgroups, while being closer or similar to them for the other half; standardized estimates were generally further from the observed prevalence than transported estimates, although their wide confidence intervals often included the observed estimate. Further restriction to White participants with normal cognition did not improve performance of DR‐AIOW or direct standardization (Table 4).

TABLE 3.

Estimated prevalence of amyloid positivity as observed in ARIC‐PET and ADNI and with ADNI generalized to ARIC‐PET using a transport estimator and direct standardization, overall and by sociodemographic and clinical strata, in White participants.

	Observed prevalence, ARIC‐PET (n = 197)	Observed prevalence, ADNI (n = 784)	Transported prevalence, ADNI to ARIC‐PET using DR‐AIOW	Standardized prevalence, ADNI to ARIC‐PET
Overall	42 (35, 49)%	47 (43, 50)%	46 (43, 50)%	46 (40, 53)%
Stratum‐specific
Age in years
Under 74	39 (27, 50)%	41 (36, 46)%	36 (30, 41)%	32 (25, 41)%
74 and over	43 (35, 52)%	54 (49, 60)%	53 (49, 57)%	55 (46, 64)%
Sex
Female	44 (34, 53)%	48 (43, 53)%	47 (42, 52)%	44 (35, 54)%
Male	39 (29, 49)%	45 (40, 50)%	46 (40, 51)%	49 (40, 58)%
Education
Education: HS/GED or less	44 (35, 54)%	57 (47, 68)%	50 (46, 55)%	48 (37, 60)%
Education: greater than HS/GED	38 (28, 49)%	45 (41, 49)%	42 (36, 47)%	44 (40, 48)%
APOE ε4 status
No copies of APOE ε4 allele	35 (28, 43)%	30 (26, 34)%	36 (33, 39)%	36 (28, 45)%)
At least one APOE ε4 allele	60 (46, 74)%	72 (67, 77)%	77 (73, 81)%	76 (67, 83)%
Cognitive Status
Normal cognition	36 (28, 44)%	34 (29, 39)%	40 (36, 44)%	32 (23, 42)%
Mild cognitive impairment	57 (43, 70)%	56 (52, 61)%	64 (58, 70)%	56 (47, 65)%

Note: Percentages of amyloid positivity are within a given stratum within a dataset, not of the dataset overall.

Abbreviations: ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; ARIC–PET, Atherosclerosis Risk in Communities Study Positron Emission Tomography Amyloid Imaging Study; DR‐AIOW, doubly robust augmented inverse odds weights; HS/GED, high school diploma/General Educational Development test; PET, positron emission tomography.

TABLE 4.

Estimated prevalence of amyloid positivity as observed in ARIC‐PET and ADNI and with ADNI generalized to ARIC‐PET using a transport estimator and direct standardization, overall and by sociodemographic and clinical strata, in White participants with normal cognition.

	Observed prevalence, ARIC‐PET (n = 144)	Observed prevalence, ADNI (n = 336)	Transported prevalence, ADNI to ARIC‐PET using DR‐AIOW	Standardized Prevalence, ADNI to ARIC‐PET
Overall	36 (28, 44)%	34 (29, 39)%	39 (35, 43)%	29 (20, 41)%
Stratum‐specific
Age in years
Under 74	30 (18, 42)%	28 (21, 34)%	25 (20, 30)%	21 (13, 32)%
74 and over	40 (30, 51)%	42 (34, 50)%	48 (44, 53)%	34 (19, 53)%
Sex
Female	39 (28, 50)%	39 (32, 46)%	46 (41, 51)%	40 (27, 54)%
Male	32 (20, 44)%	26 (19, 33)%	31 (25, 36)%	15 (8, 27)%
Education
Education: HS/GED or less	39 (27, 50)%	31 (14, 48)%	38 (33, 43)%	23 (9, 48)%
Education: greater than HS/GED	33 (22, 45)%	34 (29, 39)%	41 (35, 47)%	35 (30, 42)%
APOE ε4 status
No copies of APOE ε4 allele	34 (25, 42)%	24 (18, 29)%	31 (28, 35)%	24 (14, 37)%
At least one APOE ε4 allele	45 (27, 64)%	57 (47, 67)%	69 (62, 75)%	55 (39, 70)%

Note: Percentages of amyloid positivity are within a given stratum within a dataset, not of the dataset overall.

Abbreviations: ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; ARIC‐PET, Atherosclerosis Risk in Communities Study Positron Emission Tomography Amyloid Imaging Study; DR‐AIOW, doubly robust augmented inverse odds weights; HS/GED, high school diploma/General Educational Development test; PET, positron emission tomography.

4. DISCUSSION

The ability to repurpose deeply phenotyped samples to make valid inferences about broader target populations using modern statistical approaches would realize the full potential of such samples to advance science and medicine. However, transport methods require that we understand and model either the outcome or the selection process. We are often doing the work because we do not fully understand the outcome, so we are left assuming we can model selection. We assume that after accounting for measured characteristics, the remaining influences on selection are simply chance, or at least factors that would not modify the quantity we want to know. However, these assumptions may not hold, and if they do not, then transport (or standardization, a simple form of transport) will not produce the answers we seek.

With this goal in mind, we sought to examine the validity of extending prevalence estimates from a clinical convenience sample, ADNI, to a community‐based target sample, ARIC‐PET—with “proof of concept” defined by the performance of our transport estimator (or standardization) in recovering the observed prevalence of the outcome, amyloid positivity, in ARIC‐PET/the target sample. Unfortunately, the transported prevalences of amyloid positivity in ADNI were not systematically closer to those observed in ARIC‐PET than the unweighted (observed) ADNI prevalence across all or even the vast majority of subgroups considered. This suggests that the underlying assumptions, most notably the assumption that we have captured data on important factors influencing selection into convenience samples, were not met in this example. This raises doubts on whether transport estimates can be feasibly and validly used to extend inference from clinical convenience samples to desired target populations. This is disappointing. Estimating the prevalence of preclinical amyloid positivity is an important step in understanding the burden of AD and planning health‐care delivery, ³⁷ but there is substantial heterogeneity of estimates across samples, making population‐level inference difficult. ³⁸ Importantly, the Alzheimer's Association has stated that they will not include prevalence estimates of preclinical AD in their annual Facts and Figures report until there is a validated calculation of the number of individuals in this stage. ³⁷

We remain hopeful that transport will be a valuable tool moving forward—other models using different variables and datasets may be able to meet the assumptions to transport estimates of this outcome. Transport estimators have been used successfully in human immunodeficiency virus research, such as for transporting results of randomized controlled treatment and prevention trials, although findings may differ across study subsets and the general population. ⁹ , ³⁹ , ⁴⁰ , ⁴¹ , ⁴² Transport estimators have also been used to infer or extend estimated effects from a school‐based Positive Behavioral Interventions and Supports program to the general population, ⁴³ from one site in the Moving to Opportunity experiment to another, ¹³ , ⁴⁴ from individuals in the Coronary Artery Surgery Study randomized trial to all trial‐eligible individuals, ⁴⁵ and from the ACCORD randomized controlled trial to the US population of adults with diabetes. ⁴⁶ However, a recent study demonstrated that direct transport of the effects of standard of care for metastatic colorectal cancer among participants in the HORIZON III trial to a target population of Medicare beneficiaries was not valid. ¹⁵ Confidence in meeting the assumption that one can model either selection or the outcome correctly may be easiest to justify in the context of randomization and/or replicable/known sampling strategies (e.g., probabilistic, to directly develop sampling weights).

Importantly, our findings reinforce the importance of developing new samples with a multiplicity of diversity, selected for deep phenotyping with known sampling strategies from a representative sampling frame. As these samples are developed, we also encourage thinking deeply about and measuring the “why” and “why not” of participation so the selection process can be modeled. Generalization methods require presence of representative people to “stand in” for the people about whom we wish to derive inferences. Past deeply phenotyped samples were predominantly White with higher socioeconomic status than the general population. ⁴⁷ Our analyses illustrate that it is not possible to continue to rely on such samples to address the diversity crisis in AD and related dementias research, under the assumption that the under‐represented members of these samples would be sufficient to “stand in” for others with similar characteristics or identities. Our analyses suggest that the underlying assumption that we understand selection into legacy convenience samples is incomplete, even for those of the majority class. As such, we cannot expect to understand the selection process by which the handful of non‐majority class individuals entered the sample or assume that this small number of individuals can adequately represent sub‐populations that have not historically been well represented in research. Simply put, there are no shortcuts to increasing representation of currently underrepresented populations in research if we want generalizable findings; clinical samples and convenience samples, while valuable for other purposes, will not produce generalizable statistics or findings.

Our study has strengths and limitations. We chose a question of great importance to the field as our example—the prevalence of amyloid positivity in the population of those without dementia—and used two datasets with large numbers of amyloid PET scans, ADNI and ARIC‐PET. We acknowledge that ARIC‐PET itself is not a population‐representative sample, as a subset of the original study population that survived until Visit 5 and enrolled into a neuroimaging study. As such, crude prevalences from ARIC‐PET may not be directly generalizable to the communities from which they were originally recruited. However, as it is one of the few amyloid PET samples based in a community‐based cohort, it is an appropriate sample given our focus of providing “proof‐of‐concept” validation. Our sample sizes were relatively small leading to challenges inherent with finite samples, including issues with identifiability due to positivity violations (which required collapsing strata in standardization), limitation to the use of logistic regression in implementation of our transport estimator (machine learning approaches led to overfitting given the small sample size), and wide and overlapping confidence intervals for most estimates leading to comparisons focused on point estimates. However, these will be common issues faced by most researchers hoping to generalize from clinical convenience samples, which are often small. We report results from a single transport method; however, we chose a doubly robust method that was appropriate for our data. We also implemented standardization using only a few characteristics; other implementations are possible and may have different performance. We used established, study‐specific SUVR cut‐offs for defining amyloid positivity. Alternative approaches for defining amyloid positivity (e.g., based on Centiloids) may be preferable for future work in other samples.

In conclusion, our results suggest that efforts to generalize estimates of dementia‐relevant statistics from clinical convenience samples using transport estimators or standardization may not yield accurate statistics. Transport estimators may have greatest utility when applied to representative, diverse samples derived with known sampling strategies and attention to not only whether but also why someone does or does not participate in a study. The development of such samples is a major challenge—and commensurately significant opportunity—to drive forward dementia research and, ultimately, prevention and care.

CONFLICT OF INTEREST STATEMENT

Emma K. Stapp reports funding from the Brain and Behavior Research Foundation, the State of Maryland's Prince George's County Department of Family Services, and the Office of Research Excellence at the George Washington University Milken Institute School of Public Health. Elizabeth Rose Mayeda reports funding from NIH/NIA, the California Department of Public Health, and the Toffler Foundation. Elizabeth A. Stuart reports grant funding from the National Institutes of Health, PCORI, and Eli Lilly; consulting fees from Eli Lilly; and expert witness testimony on behalf of the plaintiffs in the talc MDL legislation. Rebecca F. Gottesman reports no disclosures. Scott C. Zimmerman reports owning stock in Eli Lily & Co., Merck & Co., Johnson & Johnson, Gilead Sciences LLC, Abbott Laboratories, AbbVie, Inc., and CRISPR Therapeutics. Dean Wong reports no disclosures. Thomas H. Mosley reports funding from NIH. Michael E. Griswold reports funding from NIH. M. Maria Glymour reports funding from NIH/NIA and the Robert Wood Johnson Foundation. Melinda C. Power reports research grants from the United States (U.S.) National Institutes of Health, U.S. Department of Defense, and Prince George's County Department of Health. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All ARIC and ADNI participants provided informed consent, and procedures were approved by each study site's institutional review boards (IRB). This analysis was determined to be not human subjects research by the George Washington University IRB.

Supporting information

Supporting Information

Supporting Information

Stapp EK, Mayeda ER, Stuart EA, et al. Estimating preclinical amyloid positivity: A case study transporting ADNI to ARIC. Alzheimer's Dement. 2025;11:e70158. 10.1002/trc2.70158