Abstract
Background and Objectives
Proxy respondents are an important tool in survey research, especially among people with cognitive impairment. However, proxy respondents may be unable to accurately answer subjective survey instruments for cognitively impaired persons. This study investigates the mediating effect of proxy status on the relationship between cognitive impairment and subjectively rated health.
Research Design and Methods
Respondents from the 2018 wave of the Health and Retirement Study (N = 17,146) were included, and the Baron and Kenny method assessed the potential mediating role of having a proxy on subjectively rated health. Subjectively rated health and proxy status were dichotomously coded, and relationships between cognitive impairment, proxy status, and subjectively rated health were assessed using logistic regression.
Results
Findings indicate that cognitive impairment is inversely associated with subjectively rated health, and it is directly associated with having a proxy. They also indicate that having a proxy is inversely related to subjectively rated health. When including proxy status and cognitive impairment in the same model, cognitive impairment no longer predicted subjectively rated health. This indicates that proxy status perfectly mediated the relationship between cognitive impairment and subjectively rated health.
Discussion and Implications
Measuring the experiences of people with cognitive impairment is challenging in survey research due to the limiting features of cognitive impairment. Although having a proxy respondent helps address attrition-related measurement challenges in survey research, findings indicate that proxies report worse subjectively rated health for people with cognitive impairment compared to individuals with comparable levels of cognitive impairment without a proxy. Future research may benefit from exploring optimal proxy-respondent characteristics.
Keywords: Baron and Kenny method, Cross-sectional study, Dementia, HRS
Single-item self-rated health remains one of the most frequently used survey instruments within social and clinical research. Self-rated health represents respondents’ holistic assessment of health (Idler & Cartwright, 2018; Jylhӓ, 2009). Therefore, multiple biopsychosocial characteristics covary with self-rated health scores. Covariates commonly used to predict self-rated health include disease and disability status and severity, demographic characteristics, and psychosocial characteristics such as temperament and feelings of belonging and community (Ge et al., 2018; Idler & Cartwright, 2018; Kim, 2021; Stephan et al., 2020, 2021).
Physical, mental, and emotional health challenges diminish self-rated health scores compared to the absence of these challenges (Ferraro, 1980; Jylhӓ, 2009). Furthermore, different populations emphasize distinct characteristics when rating health. For example, self-rated health is usually inversely related to age (Idler & Cartwright, 2018; Vogelsang, 2018; Vogelsang et al., 2019). However, younger people with physical health limitations show poorer self-rated health than older people without limitations (Idler & Cartwright, 2018). Extensive research in the United States shows generally better self-rated health among hegemonic populations including those characterized as white, young, wealthy, better educated, and with more and stronger social ties (Idler & Cartwright, 2018; Stephan et al., 2020, 2021).
Cognitive flow theory provides an explanation for different self-ratings of health among individuals with identical health conditions (Knӓuper & Turner, 2003). Cognitive flow theory suggests that an individual’s health judgments are “mental representations of personal health consisting of semantic and episodic knowledge, as well as information about change” (Jylhӓ, 2009, p. 308). Semantic knowledge is knowledge attained through personal experience, and episodic knowledge is ones’ ability to re-experience past events. This suggests that self-rated health may vary among individuals with identical diagnoses but different social contexts because episodic knowledge and expectations regarding one’s experiences differ (Jylhӓ, 2009; Vogelsang, 2018).
One analysis of social participation among people with disabilities shows that people with congenital disabilities rated better health than people with adult-onset disability, but both groups reported lower social participation than people without disabilities (Verbrugge & Yang, 2002). This evidence supports the importance of historical perspectives and social comparisons in self-rated health measures conceptualized by cognitive flow theory. Cognitive load theory, similarly, supports that cognitive impairment diminishes the respondent’s ability to respond to survey questions.
Cognitive load theory explains the cognitive process required to assess one’s own self-rated health measures. Cognitive load theory posits a person’s cognitive capacity and working memory is measurable and finite whereas people with greater cognitive loads experience greater challenges when using working memory (Sweller, 2011). People with cognitive impairment and, by extension, greater cognitive loads may fail to report their health successfully because working memory is required to critically evaluate prior health expectations (Chao et al., 2020; Jylhӓ, 2009; Paas et al., 2016; Sweller, 2011). Therefore, people with cognitive impairment may be required to utilize a proxy to complete tasks requiring higher cognitive loads.
Although cognitive flow theory and cognitive load theory propose how respondents reply accurately to survey questions, these theories are limited in explaining how proxy respondent differences may influence the accuracy of surveys with their participation. Proxies may be used to answer surveys for individuals unable to answer reliably for themselves, yet there remain several challenges when using proxy responses in survey research. First, survey instruments may require rephrasing to hold a similar meaning between proxies’ answers and respondents’ answers (Pickard & Knight, 2005). Second, self-rated health instruments may not appropriately measure comparable concepts when asking subjective questions of proxies instead of survey respondents (Kelfve et al., 2013). Third, proxies generally rate greater impaired psychological and emotional well-being relative to respondents without proxies (Kelfve et al., 2013; Neumann et al., 2015).
Longitudinal studies are vulnerable to attrition bias due to acquired functional and cognitive limitations (e.g., Wu et al., 2013). Therefore, proxies may be used by people incapable of completing subsequent survey waves (Wu et al., 2013). In the Health and Retirement Study (HRS), having consistent contact with the respondent and understanding their health status, family circumstances, and financial situation are essential characteristics for a proxy (HRS, 2021). Usually, spouses, partners, or respondents’ children act as proxies, but other proxies may be identified by the respondent (HRS, 2021). Proxies are not necessarily consistent across waves of the HRS, contributing to additional discrepancies and inconsistent reporting for subjective measures.
Still, using proxies has shown great success in improving the quality of survey research by keeping individuals in the sample who otherwise could not participate due to their conditions (Kelfve et al., 2013; Weir et al., 2011). This representation is especially critical in longitudinal studies where researchers are most interested in persons with declining health over time; older, more disabled, and sicker people leave non-random, critical data gaps when lost to attrition (Vetter & Mascha, 2017). Nevertheless, speculation remains about proxies report accurately for respondents (Neumann et al., 2015; Pickard & Knight, 2005) and whether responses differ systematically depending on proxy- versus self-response.
When answering for a respondent, proxies rely on their observations of the respondent and their own judgment when making estimates that inform their responses (Todorozv & Kirchner, 2000). Cognitive flow theory posits that respondents’ self-reports on surveys are based on a range of factors, including understanding their own circumstances, lived experiences, and reasons for answering subjective questions (Jylhӓ, 2009). The nature of subjective survey questions requires respondents to recall personal contexts to accurately respond. However, proxies responding on behalf of respondents would, theoretically, diminish the accuracy of reporting subjective domains since they are not privy to the information and experiences of the represented respondents (Lopez et al., 2023; Pickard & Knight, 2005; Todorozv & Kirchner, 2000).
Different perspectives between proxies and respondents limit concordant answers to survey questions (Pickard & Knight, 2005; Todorozv & Kirchner, 2000). The reasons for this lack of concordance are due, in part, to the questions respondents and proxies are asked to answer. For example, “Would you rate your health as excellent, very good, good, fair, or poor?” would be reworded for proxies as, “Would you rate [respondent’s name]’s health as excellent, very good, good, fair, or poor?” or “Would [respondent’s name] rate their health as excellent, very good, good, fair, or poor?”. Each question evaluates the respondent’s health using the same scale, but the first wording pulls directly from the respondent’s own experiences, expectations, and observations; the second wording rates the respondent’s health from a proxy’s perspective of the respondent’s circumstance; and the final wording rates a proxy’s perspective of how the respondent would rate their health if asked directly (Pickard & Knight, 2005). One study evaluated the differences between self-, proxy-, and physician-rated health for a group of survey respondents and found extreme variation between raters (Smith & Goldman, 2011). It was hypothesized that these differences arose due to the different value systems and perspectives of the raters (Smith & Goldman, 2011).
Best practices for identifying proxy-respondent measurement differences include using intraclass correlation coefficients in conjunction with differences in means testing between methods (van der Linden et al., 2009). The standard reliability criteria suggest intraclass correlation coefficients greater than 0.75 indicate good reliability with higher scores indicating greater concordance between measurement strategies (van der Linden et al., 2009). However, collecting responses from both proxies and respondents may be impossible due to the cognitive or functional limitations of respondents. Additionally, larger surveys may not feasibly ask for proxies and respondents to answer identical survey questions (Rahi, et al., 2019). Therefore, data may not be available to evaluate measurement differences between proxies and respondents. As such, alternative strategies need to be developed to assess the extent to which proxies and self-responses result in different answers to survey questions across respondents.
Challenges with using proxy respondents for subjectively rated health questionnaires may limit the correct representation of subjective-rated health in people with cognitive impairment in clinical and policy-related decisions. Still, proxies may offer a greater benefit than the risk to survey research in assessing subjective-rated health where such assessments may not otherwise be possible. Extant evidence suggests that proxies and respondents generally agree on objective assessments of physical health, functioning, and cognitive status (Neumann et al., 2015). Furthermore, proxies offer avenues for survey researchers to address attrition bias by representing respondents otherwise lost due to cognitive challenges (Li et al., 2015).
The aim of this study is to explore the mediating role that having a proxy plays in subjective-rated health measurements for individuals with cognitive impairment. Proxy respondents are frequently removed in survey research, and addressing how proxy responses are measured and reported in survey research may improve the representation of respondents with the worst cognitive and physical health.
Method
Data
The HRS is a nationally representative, longitudinal panel study of Americans over the age of 50 (Sonnega & Weir, 2014). The HRS collects data biennially using face-to-face interviews, leave-behind questionnaires, and telephone interviews across the contiguous United States. Longitudinal cohorts at baseline included community-dwelling individuals aged 51–61 years residing in the United States with new cohorts added every 6 years (3 waves). Respondents receiving institutional care are retained in subsequent waves after baseline (HRS, 2021).
This study used data from the 2018 wave of the HRS. The 2018 core sample consists of 17,146 respondents from seven different cohorts with an overall response rate of 68.7% for individuals and 71.4% for households (Health & Retirement Study, 2021). Only 0.75% of the analytic sample contained missing values. Missing data were managed using twenty iterations of multiple imputations. This study included all responses from the HRS where values for self- or proxy-rated health, cognitive impairment, or model covariates could be ascertained. No other exclusion criteria were applied, in order to include as many respondents as possible. The final sample included all 17,146 respondents in the core sample after all missing values were imputed.
Measures
Subjectively rated health measures holistic and individual perspectives of a respondents’ health. Self-rated health was collected using the Self-Rated Health-5 instrumentation wording, “Would you say your health is excellent, very good, good, fair, or poor?” (Jylhӓ, 2009). Proxy respondents were asked the modified proxy-patient perspective wording, “Would [respondent’s name] say their health is excellent, very good, good, fair, or poor?.” Proxy- and self-rated health for each respondent were measured on a 5-point Likert scale and relabeled as subjectively rated health. Subjectively rated health was reverse coded with higher values representing better health (1-Poor, 5-Excellent). A dichotomous subjectively rated health variable was recoded (1-Excellent, Very good, Good; 0-Fair, Poor) and used for multivariate assessments. The proxy status variable identifies the rater for the respondent’s subjective health.
Cognitive impairment was measured using the Langa–Weir method (Crimmins et al., 2011; Langa et al., 2017). Respondents with a proxy do not complete the cognitive battery of questions in the Health and Retirement Study. Therefore, the Langa–Weir method allows for reliable comparison of cognitive status between respondents with, and without, proxies (Crimmins et al., 2011). Cut-points were established using equipercentile equating to harmonize cognitive impairment scales between the Langa–Weir criteria and cognitive statuses estimated in the Aging, Demographics, and Memory Study (ADAMS), which used the Mini-Mental State Examination, Boston naming test, digit span, Symbol Digit Modality Test, animal fluency, word list three trial learning, construction praxis copying, Trial Making Test, Wechsler Memory Scale, Fuld Object Memory Test, Shipley vocabulary test, and the WRAT 3 blue reading test (Langa et al., 2005).
The instruments used for respondents without proxies included immediate and delayed recall tests of memory, series 7s subtraction test of working memory, and backward counting to test attention and processing speed (Crimmins et al., 2011). Immediate recall involved asking the respondent to recite as many words as they could remember from a list of 10 words immediately after showing them the word list. Delayed recall asked respondents, after a time, to recall as many words as possible from the same list. Scores for both immediate and delayed recall range from 0 to 10 based on a count of correctly recalled words. The series 7 assessment involves asking individuals to start from 100 and count backward by 7 for 5 iterations. Respondents who incorrectly subtract 7 would fail that iteration, but individuals who subtract 7 successfully from their last given response would succeed that iteration. Backward counting involved respondents counting backward from 20 to 10. Respondents who succeeded on the first try were given 2 points toward cognitive performance, respondents who needed to try again scored 1 point toward cognitive performance, and respondents who could not complete the backward count in two tries scored 0. A scale, ranging from 0 to 27, was generated summing responses to the above instruments. Langa et al., (2017) cutoff points were used to categorize respondents without proxies’ cognitive impairment scores into cognitively impaired (0–11) and not cognitively impaired (12–27) groups.
Cognitive impairment among respondents with proxies was measured using a composite score from multiple instruments including proxy assessment of cognitive status (0 [Excellent]- 4 [Poor]), 5 instrumental activities of daily living (ADL) (managing money, taking medication, preparing hot meals, making phone calls, and shopping for groceries; 0 [Excellent] to 5 [Poor]), and the respondent’s ability to complete the survey independently (0-No help needed, 1-Some help needed, and 2-incapable of completing independently). A summed scale, ranging from 0 to 11, was generated, with cutoff points for cognitively impaired (3–11) and not cognitively impaired (0–2) applied (Crimmins et al., 2011). The cognitive measures employed in this study to assess differences between proxy and non-proxy respondents have been validated and used in prior studies using the HRS (Crimmins et al, 2011; Langa et al., 2017).
A binary variable was used in the analysis to indicate whether a proxy was used or not (1—Yes, 0—No). Proxies are offered to respondents unable or unwilling to complete the HRS survey or interview independently (Wu et al., 2013). Proxies should be knowledgeable of the respondent’s health, family circumstances, and financial situation (HRS, 2021). Proxies are often selected by the respondent directly, or they are listed as their contact person for the research study (HRS, 2021).
The covariates controlled for in this analysis were activities ADL), age, gender, race, education, comorbidities, marital status, cohabitation, assets, and having ever been diagnosed with Alzheimer’s disease. Age was a continuous variable counting years since birth. Race and ethnicity were recoded into four groups including non-Hispanic White, non-Hispanic Black, non-Hispanic Other, and Hispanic with the non-Hispanic White group serving as the reference group. A dichotomous gender variable was used (0—male, 1—female). Marital status included a categorical variable with three groupings (0—never married, 1—married/partnered, 2—separated/divorced/widowed), with never married serving as the reference group. Similarly, educational attainment included categorical groupings (0—less than high school, 1—GED, 2—high school graduate, 2—some college, and 3—college degree holder or more) with less than high school serving as the reference category for analyses. ADLs were generated from a composite score of binary responses (1—not independent, 0—independent) to whether the respondent could independently bath or shower, dress, get in and out of a chair, walk, toilet, and eat (scale range 0–6; Edemekong et al., 2022). Comorbidities were coded from a count of binary responses (1—Yes, 0—No) to whether the respondent was ever diagnosed with high blood pressure, diabetes, cancer (excluding skin cancer), stroke, lung disease, heart disease, depression, arthritis, weak immune system, high cholesterol, trouble with pain, and sleep disorders. Cohabitation was represented as a binary (0—No, 1—Yes) response to whether the respondent lives with anyone else. Alzheimer’s disease was similarly assessed using binary (1—Yes, 0—No) responses to ever being diagnosed with Alzheimer’s disease. Assets were assessed via self-report, and results were winsorized at the 1st and 99th percentile (range: −$1,000,000 to $4,234,306) to address extreme values. All continuous variables were mean centered and assessed for skewness to meet the normality assumption.
Analysis
Descriptive statistics were generated based on imputed values. Additionally, bivariate analyses between proxy and non-proxy groups were completed and compared using χ2 and two-way t tests. Spearman ranked correlation coefficients were generated between subjectively rated health and all model variables using Spearman ranked correlation coefficients. Significance was reported at the 5% level.
The Baron and Kenny (1986) method was used to determine the mediation of proxy status on the relationship between cognitive impairment and subjectively rated health. The relationships that must be established include (a) cognitive impairment on subjectively rated health (Model 1), (b) cognitive impairment on proxy status (Model 2), and (c) proxy status on subjectively rated health (Model 3). Last, a full model that tests the association between cognitive impairment and proxy status on subjectively rated health (Model 4). For mediation to be present, the effect of cognitive impairment needs to be diminished or changed from a significant predictor to an insignificant predictor of subjectively rated health when including proxy status in the model. Ordered logistic regression, as the appropriate statistical method for assessing subjectively rated health as an outcome (Min, 2013), was attempted to generate models where subjectively rated health was the outcome, but the models failed the post hoc Brant test, indicating a violation of the parallel regression assumption. To address this, subjectively rated health was recoded to a binary variable, and logistic regression models were used in place of ordered logistic models where subjectively rated health was the outcome. Similarly, logistic regressions assessed the associations where proxy status was the outcome. Control variables were included in all models.
Results
Descriptive statistics for the total and proxy and non-proxy subsamples are shown in Table 1. The majority of survey respondents were not cognitively impaired, non-Hispanic White race, female, had greater than a high school education, were married, and lived with another person. As expected, a higher proportion of respondents with proxies had cognitive impairment (52.94%) compared to respondents without proxies (1.68%), and respondents with a proxy had significantly worse rated health (45.10% in good health) compared to respondents without a proxy (71.25% in good health). Respondents with proxies were significantly older (M = 76.30, SE = 0.49) compared to respondents without proxies (M = 66.57, SE = 0.09), and respondents without proxies were significantly less likely to be male (40.59%) compared to respondents with proxies (51.43%). Consistent with prior literature, respondents without proxies were more educated with a higher proportion attaining college degrees (25.40%) compared to respondents with proxies (14.93%).
Table 1.
Descriptive Statistics by People with and Without Proxies in the 2018 HRS Core Data After Multiple Imputation
| Full sample (n = 17,146) |
Respondents with proxies (n = 663) |
Respondents without proxies (n = 16,483) |
||||||
|---|---|---|---|---|---|---|---|---|
| Variables | M or % | (SE) | M or % | (SE) | M or % | (SE) | χ2 or t | |
| Cognitive impairment (%) | ||||||||
| No cognitive impairment | 96.34 | — | 47.06 | — | 98.32 | — | 47,000 | *** |
| Cognitive impairment | 3.66 | — | 52.94 | — | 1.68 | — | 47,000 | *** |
| Race (%) | ||||||||
| White non-Hispanic | 56.91 | — | 58.07 | — | 56.86 | — | 0.37 | |
| Black non-Hispanic | 21.63 | — | 19.61 | — | 21.71 | — | 1.67 | |
| Other non-Hispanic | 4.84 | — | 5.13 | — | 4.83 | — | 0.14 | |
| Hispanic | 16.62 | — | 17.19 | — | 16.60 | — | 0.17 | |
| Male (%) | 41.01 | — | 51.43 | — | 40.59 | — | 31.00 | *** |
| Education | ||||||||
| Less than high school diploma | 16.66 | — | 33.48 | — | 15.98 | — | 140.60 | *** |
| GED | 5.30 | — | 4.98 | — | 5.31 | — | 0.15 | |
| High school graduate | 26.36 | — | 28.21 | — | 26.29 | — | 1.22 | |
| Some college | 26.69 | — | 18.40 | — | 27.02 | — | 24.23 | *** |
| College degree | 24.99 | — | 14.93 | — | 25.40 | — | 37.26 | *** |
| Marital status (%) | ||||||||
| Single, never married | 6.49 | — | 2.71 | — | 6.64 | — | 16.12 | *** |
| Married | 61.10 | — | 58.52 | — | 61.21 | — | 1.94 | |
| Separated/widowed | 32.40 | — | 38.76 | — | 32.15 | — | 12.70 | *** |
| Cohabitating (%) | 76.11 | — | 74.66 | — | 76.16 | — | −0.02 | |
| Alzheimer’s disease (%) | 1.53 | — | 23.53 | — | 0.65 | — | 2200 | *** |
| ADLsa | 0.46 | (0.01) | 2.06 | (0.09) | 0.39 | (0.01) | −37.40 | *** |
| Agec | 66.95 | (0.09) | 76.30 | (0.49) | 66.57 | (0.09) | −21.83 | *** |
| Comorbiditiesb | 1.98 | (0.02) | 2.35 | (0.08) | 1.97 | (0.02) | −4.92 | *** |
| Self-rated health (good health; %) | 70.24 | 45.10 | 71.25 | 208.51 | *** | |||
Notes: ADL = activities of daily living; GED = general education development; HRS = Heath and Retirement Study; M = mean; SE = standard error.
N = 17,146.
aRange (0–6).
bRange (0–12).
cRange (27–104) years.
*p > .05. **p > .01. ***p > .001.
Spearman bivariate correlations (Table 2) indicate that cognitive impairment and proxy status are both significant, weak, and negative correlates of subjectively rated health. Other significantly negative correlates of subjectively rated health include race, marital status, diagnosis of Alzheimer’s disease, ADLs, age, and comorbidities, with the strongest negative correlate with subjectively rated health being ADLs (ρ = −0.45, p < .001). Positive correlates of subjectively rated health include gender, education, cohabitation, and assets, with the strongest positive correlate to subjectively rated health being assets (ρ = 0.30, p < .001).
Table 2.
Bivariate Spearman Correlation Coefficients
| Variables | Self-rated health | Sig. |
|---|---|---|
| Cognitive impairment | −0.14 | *** |
| Proxy status | −0.11 | *** |
| Race | −0.16 | *** |
| Gender | 0.03 | *** |
| Education | 0.27 | *** |
| Marital status | −0.07 | *** |
| Cohabitating | 0.05 | *** |
| Alzheimer’s disease | −0.11 | *** |
| ADLs | −0.40 | *** |
| Age | −0.03 | *** |
| Comorbidities | −0.40 | *** |
| Assets | 0.30 | *** |
Notes: ADL = activities of daily living; Sig. = significance.
N = 17,146.
*p > .05. **p > .01. ***p >.001.
Table 3 shows all models in accordance with the Baron and Kenny Method to determine mediation. Logistic regression models were used where proxy status and dichotomous subjectively rated health variables were the outcomes of interest.
Table 3.
Model Results for Association Between Cognitive Impairment, Proxy Status, and Subjectively Rated Health Using 2018 HRS CORE Data
| Model 1 (N = 17,146) | Model 2 (N = 17,146) | Model 3 (N = 17,146) | Model 4 (N = 17,146) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Variables | B | SE | Sig. | B | SE | Sig. | B | SE | Sig. | B | SE | Sig. |
| Cognitive impairment | ||||||||||||
| No cognitive impairment | — | — | — | — | — | — | — | — | — | |||
| Cognitively Impairment | −0.34 | 0.12 | ** | 3.28 | 0.13 | *** | −0.20 | 0.13 | 0.116 | |||
| Proxy | ||||||||||||
| Without proxy | — | — | — | — | — | — | ||||||
| With proxy | −0.42 | 0.11 | *** | −0.35 | 0.12 | ** | ||||||
| Covariates | ||||||||||||
| Race | ||||||||||||
| White non-Hispanic | — | — | — | — | — | — | — | — | — | — | — | — |
| Black non-Hispanic | −0.28 | 0.05 | *** | −0.32 | 0.14 | * | −0.29 | 0.05 | *** | −0.29 | 0.05 | *** |
| Other non-Hispanic | −0.34 | 0.10 | *** | 0.38 | 0.22 | 0.09 | −0.33 | 0.10 | *** | −0.33 | 0.10 | *** |
| Hispanic | −0.76 | 0.06 | *** | −0.44 | 0.16 | ** | −0.76 | 0.06 | *** | −0.76 | 0.06 | *** |
| Agea | 0.01 | 0.002 | *** | 0.03 | 0.004 | *** | 0.01 | 0.002 | *** | 0.01 | 0.002 | *** |
| Gender | ||||||||||||
| Male | −0.10 | 0.04 | * | 0.77 | 0.11 | *** | −0.09 | 0.04 | * | −0.09 | 0.04 | * |
| Female | — | — | — | — | — | — | — | — | — | — | — | — |
| Education | ||||||||||||
| No degree | — | — | — | — | — | — | — | — | — | — | — | — |
| GED | 0.43 | 0.09 | *** | −0.57 | 0.24 | * | 0.42 | 0.09 | *** | 0.42 | 0.09 | *** |
| High school grad. | 0.72 | 0.06 | *** | −0.43 | 0.14 | ** | 0.72 | 0.06 | *** | 0.72 | 0.06 | *** |
| Some college | 0.96 | 0.06 | *** | −0.70 | 0.15 | *** | 0.95 | 0.06 | *** | 0.95 | 0.06 | *** |
| College degree | 1.33 | 0.07 | *** | −1.01 | 0.17 | *** | 1.33 | 0.07 | *** | 1.32 | 0.07 | *** |
| ADLsa | −0.47 | 0.02 | *** | 0.33 | 0.03 | *** | −0.47 | 0.02 | *** | −0.47 | 0.02 | *** |
| Marital status | ||||||||||||
| Single, never married | — | — | — | — | — | — | — | — | — | — | — | — |
| Married | 0.24 | 0.09 | ** | 1.32 | 0.32 | *** | 0.25 | 0.09 | ** | 0.25 | 0.09 | ** |
| Separated or widowed | 0.05 | 0.08 | 0.59 | 0.80 | 0.31 | ** | 0.05 | 0.08 | 0.54 | 0.05 | 0.08 | 0.56 |
| Assetsa,b | 0.00 | 0.00 | *** | 0.00 | 0.00 | * | 0.00 | 0.00 | *** | 0.00 | 0.00 | *** |
| Comorbiditiesa | −0.42 | 0.01 | *** | −0.13 | 0.03 | *** | −0.43 | 0.01 | *** | −0.43 | 0.01 | *** |
| Alzheimer’s disease | −0.62 | 0.19 | *** | 1.58 | 0.20 | *** | −0.63 | 0.19 | *** | −0.56 | 0.19 | ** |
| Cohabitating | −0.06 | 0.06 | 0.36 | 0.16 | 0.18 | 0.36 | −0.06 | 0.06 | 0.37 | −0.06 | 0.06 | 0.28 |
| Model F-values | 193.97 | *** | 89.30 | *** | 194.07 | *** | 183.34 | *** | ||||
| Constant | 1.29 | 0.10 | *** | −4.81 | 0.34 | *** | 1.29 | 0.10 | *** | 1.30 | 0.10 | *** |
Notes: ADL = activities of daily living; GED = general education development; HRS = Health and Retirement Study; SE = standard error; Sig. = significance.
N = 17,146.
Model 1 tests cognitive impairment and covariates on subjectively rated health.
Model 2 tests cognitive impairment and covariates on proxy status.
Model 3 tests proxy status and covariates on subjectively rated health.
Model 4 tests proxy status and cognitive impairment and covariates on subjectively rated health.
Logistic regression was used for all models: log odds units.
aContinuous model variables were mean centered.
bAssets’ coefficient as rounded to 0 across all models, and it remained positively associated across all models.
*p < .05. **p < .01. ***p < .001.
Model 1 tested the association between cognitive impairment and covariates on subjectively rated health. On average, those with a cognitive impairment rate of 0.34 lower logged odds on subjectively rated health measures compared to respondents without cognitive impairment (p < .05). All racial categories indicate negative associations compared to White non-Hispanic participants, with Hispanic respondents indicating, on average, the worst subjectively rated health compared with White non-Hispanic respondents (B = −0.76, p < .001). Those with more education generally rated their health better than those who did not graduate high school. Individuals with more ADL challenges and comorbidities indicated worse subjectively rated health (B = −0.47, p < .001; B = −0.42, p < .001).
Model 2 used logistic regression to test the association between cognitive impairment and all model covariates on proxy status. On average, those with cognitive impairment indicated greater utilization of a proxy compared to those without cognitive impairment (B = 3.28, p < .001). Graduating high school indicated diminished odds of utilizing a proxy, with the most diminished odds for college degree holders (B = −1.01, p < .001). Being married and being diagnosed with Alzheimer’s disease were strong predictors of utilizing a proxy (B = 1.32, p < .001; B = 1.58, p < .001).
Model 3 tested the association of proxy status and all model covariates on dichotomized subjectively rated health. On average, respondents with a proxy indicated worse health (B = −0.42, p < .001) compared to those without a proxy. All model control variables remained similar in strength, direction, and significance to those reported for Model 1.
Model 4 tested cognitive impairment, proxy status, and all model covariates on subjectively rated health. Cognitive impairment was not a significant predictor of subjectively rated health when controlling for proxy status (B = −0.20, p = .12). However, on average, those with a proxy continued to report worse health compared to those without a proxy (B = −0.35, p < .01). All other associations remained similar in significance, strength, and directionality to those indicated in Models 1 and 3.
The results indicate proxy status fully mediates the relationship between cognitive impairment and subjectively rated health (see Figure 1). Models 1 and 2 established that cognitive impairment predicts subjectively rated health and proxy status; Model 3 established that proxy status was a significant predictor of subjectively rated health. When including cognitive impairment and proxy status in the same model predicting subjectively rated health (Model 4), proxy status remained a significant predictor of subjectively rated health, but cognitive impairment was no longer a significant predictor.
Figure 1.
Barron and Kenny established mediation by proxy status on the relationship between cognitive impairment and subjectively rated health. (a) Relationship of cognitive impairment on proxy status. (b) Relationship of proxy status on subjectively rated health. (c) Relationship of cognitive impairment on subjectively rated health. (cʹ) Relationship of cognitive impairment on subjectively rated health mediated by proxy status. * p < .05. ** p < .01. *** p < .001.
Discussion
Having a proxy respondent helps address attrition related to measurement challenges in survey research involving people with cognitive impairment, but little research has explored the accuracy of proxies when reporting for people with cognitive impairment in surveys. Researchers are unable to directly assess the accuracy of proxy respondents in large surveys because survey questions are only asked of the proxy or the respondent, limiting comparability and concordance between respondents. This study thus aimed to understand the relationship between cognitive impairment and subjectively rated health, and how having a proxy may affect this relationship using the HRS. The study found that cognitive impairment was inversely associated with subjectively rated health; proxy status was inversely associated with subjectively rated health; and cognitive impairment was directly associated with having a proxy respondent. When testing both cognitive impairment and proxy status on subjectively rated health, cognitive impairment was no longer a significant predictor of subjectively rated health. This met the criteria for perfect mediation. These findings demonstrate that the subjectively rated health of people with cognitive impairment is related to having a proxy, suggesting that having a proxy may bias subjective health responses.
Consistent with prior literature (Devore et al., 2017; Petersen et al., 2009), this research found that people with cognitive impairment generally rate poorer subjective health than people without cognitive impairment. Moving beyond prior literature, these results indicate proxies reported worse subjectively rated health for respondents than respondents themselves. People with cognitive impairment indicated greater odds of utilizing a proxy in the HRS. This is likely due to the survey’s design. Large surveys are designed to help address attrition bias by representing respondents in future waves of data collection who otherwise would not be represented in the survey.
Cognitive load theory posits that cognitive load precedes the task to be completed (Sweller, 2011). Insufficient cognitive load leads to insufficient completion of demanding tasks. In this study, cognitive impairment was chosen as the predictor variable to remain consistent with cognitive load theory, and reporting subjectively rated health, as the demanding task, was used as the outcome variable. Importantly, the cognitive ability of respondents is one of the criteria used for determining the need for proxy assistance. Therefore, this assessment used cognitive impairment to predict proxy status and, in turn, subjectively rated health as well. Subjectively rated health was also used as the outcome variable because proxies may more readily assess holistic health rather than specific health domains (Jylhӓ, 2009). Given the cross-sectional nature of the present study and the potentially reciprocal nature of the relationships involved, future assessments should consider subjectively rated health as the predictor and cognitive impairment as the outcome. Future research should also consider expanding upon this research by investigating proxy mediation of specific health domains and cognitive impairment while remaining sensitive to the subjectivity of health-related survey measures.
Assessing self-rated health is challenging for people with cognitive impairment. Forgetfulness and memory loss prevent respondents with cognitive impairment from accurately judging their past and present health expectations (Jylhӓ, 2009). The proxy may similarly be unaware of the respondent’s past or present health expectations while also overvaluing observed health behaviors (Pickard & Knight, 2005; Smith & Goldman, 2011; Todorozv & Kirchner, 2000). Therefore, one would expect an overestimation of poorer proxy-rated health measures for people with cognitive impairment compared to self-rated health for people with cognitive impairment. This was apparent in the study results. People with cognitive impairment without a proxy reported similar self-rated health to people without cognitive impairment without a proxy, whereas proxies rated worse health for respondents with cognitive impairment than self-rated health reported by people without cognitive impairment. Individuals who require a proxy to support their responses in survey research are reportedly less healthy than those without a proxy. However, we would expect individuals with cognitive impairment without a proxy to also report worse self-rated health than people without cognitive impairment (Walker et al, 2004). This was not the case. Therefore, it is possible that the severity of cognitive impairment for people requiring a proxy explains their reported worse subjectively rated health. Further research should explore how the severity of cognitive impairment may affect proxy-rated health compared to self-rated health among individuals with comparable severity of cognitive impairment.
It is challenging to measure the experiences and expectations of people with cognitive impairment in survey research. Either the features of cognitive impairment prevent respondents from representing themselves accurately, or functional and cognitive limitations prevent their participation at all in survey research (Walker et al, 2004). Proxies may help address attrition bias due to cognitive impairment (Wu et al., 2013), but proxies may also inaccurately represent people with cognitive impairment on subjective measures of health (Pickard & Knight, 2005).
Large surveys may be limited in applying best practices to measure the accuracy of reporting by either respondent with cognitive impairment or their proxies. Best practices refer to using intraclass correlation coefficients with self- and proxy-rated health for the respondent and determine if the difference in reported health is due to respondents or proxies (van der Linden et al., 2009). The HRS does not collect self- and proxy-rated health responses, relying on proxies to completely represent the respondent. However, the implications of this study suggest that misrepresentation of subjectively rated health scores for people with cognitive impairment by proxies exists within the HRS. Efforts should thus be taken to improve how people with cognitive impairment and their proxies accurately participate in survey research. Proxy status explained away the effect of cognitive impairment on subjectively rated health, which suggests that proxy status may have an undue influence on subjectively rated health ratings among cognitively impaired participants with proxy respondents. Therefore, similar findings may be expected for other proxy reported, subjective survey instruments.
Further research should explore the extent to which proxy bias and attrition bias introduce error into longitudinal studies, with efforts made to address these errors and to ensure a more accurate representation of people with cognitive impairment. Longitudinal studies provide key insights into time-varying processes such as healthy aging, onset and fluctuation of cognitive function, and onset and management of disability. By addressing measure errors presented in these data, stakeholders can make better-informed decisions for health-related practices and policies for individuals with cognitive impairment.
Limitations
These analyses and conclusions were limited due to the data available in the HRS. Respondents with and without proxies significantly differed from one another. Unsurprisingly, respondents with a proxy-rated greater cognitive impairment and worse health than those without a proxy, and respondents with a proxy reported more comorbid conditions, greater age, and were less educated than respondents without a proxy. These analyses would benefit from the inclusion of other neurological or cognitive impairment conditions (such as Parkinson’s disease, history of traumatic brain injury, or delirium, to name a few), but which were excluded due to data limitations. Future analyses should employ datasets that permit comparison of self-rated health with proxy-rated health at similar levels of cognitive impairment. Furthermore, the HRS does not collect data on the level of familiarity proxies have with their respondent, which are listed criteria for being a proxy. Also, the HRS lacks specific demographic data on proxy respondents that may prove important for future research.
Proxies fill a key role by reporting for respondents who otherwise would not respond to the survey, but the HRS does not report whether the respondent supported the proxy when answering specific survey items on their behalf or the level of such support (Sonnega & Weir, 2014). Therefore, these analyses assumed proxies answered completely independent of the respondent’s support. This assumption likely provides strength to the analyses, for proxies with support in answering subject health questions should report more accurately than proxies without respondent support. It would be useful to know the extent to which proxies answer survey items independent of respondent’s support.
These analyses were further limited in exploring the features of proxies and their contribution to confounding subjectively reported health of people with cognitive impairment. In the HRS, proxies are expected to know the respondent’s health, finances, and personal information to a reasonable extent to qualify as a proxy (HRS, 2021), but it is unclear how these criteria are proven or enforced. Most proxies for people with cognitive impairment were spousal proxies, and, when stratifying by proxy–respondent relationship, too few observations were available to statistically assess the relationship of non-spousal proxies to spousal proxies. Additionally, this exploration of proxy characteristics remained outside the scope of this project’s aims. Therefore, future research should explore how proxy characteristics are associated with subjectively rated respondents’ health among people with cognitive impairment, specifically including variables describing the proxy-respondent relationship, proxy demographic characteristics, and the proxy’s own self-rated health. Multilevel modeling may be necessary for further assessments, nesting prox–-respondent dyads to compare between- and within differences, where dyadic data is available.
Last, these cross-sectional analyses are limited in drawing firm conclusions due to temporal unknowns. Physical and cognitive health declines may occur in conjunction with one another. Future research should build off of the present study by examining subjectively rated health, cognitive functioning, and proxy status over longer periods of time, in addition to considering potential reciprocal relationships between subjectively rated health and cognitive health and the characteristics of proxy respondents themselves. Although this study cannot make conclusive claims about whether proxies reliably report the subjective health of respondents with cognitive impairment in survey data, this study contributes to the literature by demonstrating that subjectively rated health in persons with cognitive impairment is fully related to having a proxy, suggesting having a proxy respondent might bias results.
Conclusion
Researchers should aim to accurately represent people with cognitive impairment in survey research. Proxies remain a useful tool for addressing attrition bias, but they may over represent poorer health for people without cognitive impairment compared to people with cognitive impairment. Therefore, future research should develop methods to best survey people with cognitive impairment in addition to comparing errors in survey research due to proxy bias compared to attrition bias among people with cognitive impairment. Survey designers should similarly be sensitive to these errors within their surveys of people living with cognitive impairment and build survey instrumentation to use intraclass correlation coefficient methods. This would mean asking proxies- and respondent-rated health and other subjective measures to better discern the accuracy of reported data.
Acknowledgments
We thank Alison Rataj for her expertise and assistance throughout the writing and analyses of our study.
Contributor Information
Andrew G Alberth, Department of Gerontology, University of Massachusetts Boston, Boston, Massachusetts, USA.
Edward Alan Miller, Department of Gerontology, University of Massachusetts Boston, Boston, Massachusetts, USA; Department of Health Services, Policy, and Practice, Brown University, Providence, Rhode Island, USA.
Nina M Silverstein, Department of Gerontology, University of Massachusetts Boston, Boston, Massachusetts, USA; Department of Gerontology, Gerontology Institute, University of Massachusetts Boston, Boston, Massachusetts, USA.
Jeffrey Stokes, Department of Gerontology, University of Massachusetts Boston, Boston, Massachusetts, USA.
Yan-Jhu Su, Department of Gerontology, University of Massachusetts Boston, Boston, Massachusetts, USA.
Funding
None.
Conflict of Interest
None.
Data Availability
The HRS is sponsored by the National Institute of Aging (grant NIA U01AG009740) and is conducted by the University of Michigan. This study was not pre-registered.
References
- Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51, 1173–1182. 10.1037//0022-3514.51.6.1173 [DOI] [PubMed] [Google Scholar]
- Chao, Y. P., Wu, C. W., Lin, L. J., Lai, C. H., Wu, H. Y., Hsu, A. L., & Chen, C. N. (2020). Cognitive load of exercise influences cognition and neuroplasticity of healthy elderly: An exploratory investigation. Journal of Medical and Biological Engineering, 40(3), 391–399. 10.1007/s40846-020-00522-x [DOI] [Google Scholar]
- Crimmins, E., Kim, J. K., Langa, K. M., & Weir, D. R. (2011). Assessment of cognition using surveys and neuropsychological assessment: The health and retirement study and the aging, demographics, and memory study. Journals of Gerontology B: Psychological Sciences and Social Sciences, 66B(SI), il62–i171. 10.1093/geronb/gbr048 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Devore, E., Fong, T., Marcantonio, E. R., Schmitt, E. M., Travison, T. G., Jones, R. N., & Inouye, S. K. (2017). Prediction of long-term cognitive decline following delirium in older adults. Innovation in Aging, 1(Suppl_1), 1326–1326. 10.1093/gerona/glx030 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Edemekong, P. F., Bomgaars, D. L., Sukumaran, S., & Schoo, C. (2022). Activities of daily living. StatPearls. StatPearls Publishing. [PubMed] [Google Scholar]
- Ferraro, K. F. (1980). Self-ratings of health among the old and the old-old. Journal of Health and Social Behavior, 21, 377–383. 10.2307/2136414 [DOI] [PubMed] [Google Scholar]
- Ge, L., Ong, R., Yap, C. W., & Heng, B. H. (2018). Effects of chronic diseases on health-related quality of life and self-rated health among three adult age groups. Nursing & Health Sciences, 21, 12585. 10.1111/nhs.12585 [DOI] [PubMed] [Google Scholar]
- Health & Retirement Study. (2021). Produced and distributed by the University of Michigan with funding from the National Institute on Aging (grant number U01AG009740). [Google Scholar]
- HRS. (2021). Proxy selection in the health and retirement study. Accessed February 16th, 2022.
- Idler, E., & Cartwright, K. (2018). What do we rate when we rate our health? Decomposing age-related contributions to self-rated health. Journal of Health and Social Behavior, 59(1), 74–93. 10.1177/0022146517750137 [DOI] [PubMed] [Google Scholar]
- Jylhä, M. (2009). What is self-rated health and why does it predict mortality? Towards a unified conceptual model. Social Science & Medicine, 69(3), 307–316. 10.1016/j.socscimed.2009.05.013 [DOI] [PubMed] [Google Scholar]
- Kelfve, S., Thorslund, M., & Lennartsson, C. (2013). Sampling and non-response bias on health-outcomes in surveys of the oldest old. European Journal of Ageing, 10(3), 237–245. 10.1007/s10433-013-0275-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Kim, J. -H. (2021. ). The neighborhood effect of cognitive function on self-rated health: A population-based observational study. Archives of Gerontology and Geriatrics, 93, 104285. 10.1016/j.archger.2020.104285 [DOI] [PubMed] [Google Scholar]
- Knäuper, B., & Turner, P. A. (2003). Measuring health: Improving the validity of health assessments. Quality of Life Research: An International Journal of Quality of Life Aspects of Treatment, Care and Rehabilitation, 12(1), 81–89. 10.1023/a:1023589907955 [DOI] [PubMed] [Google Scholar]
- Langa, K. M., Larson, E. B., Crimmins, E. M., Faul, J. D., Levine, D. A., Kabeto, M. U., & Weir, D. R. (2017). A comparison of the prevalence of dementia in the United States in 2000 and 2012. JAMA Internal Medicine, 177(1), 51–58. 10.1001/jamainternmed.2016.6807 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Langa, K. M., Plassman, B. L., Wallace, R. B., Herzog, A. R., Heeringa, S. G., Ofstedal, M. B., Burke, J. R., Fisher, G. G., Fultz, N. H., Hurd, M. D., Potter, G. G., Rodgers, W. L., Steffens, D. C., Weir, D. R., & Willis, R. J. (2005). The aging, demographics, and memory study: Study design and methods. Neuroepidemiology, 25(4), 181–191. 10.1159/000087448 [DOI] [PubMed] [Google Scholar]
- Lopez, A., Tinella, L., Caffò, A., & Bosco, A. (2023). Measuring the reliability of proxy respondents in behavioural assessments: An open question. Aging Clinical and Experimental Research, 35, 2173–2190. 10.1007/s40520-023-02501-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- Li, M., Harris, I., & Lu, Z. K. (2015). Differences in proxy-reported and patient-reported outcomes: assessing health and functional status among medicare beneficiaries. BMC Medical Research Methodology, 15(1), 1–10. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Min, H. (2013). Ordered logit regression modeling of the self-rated health in Hawai’i, with comparisons to the OLS model. Journal of Modern Applied Statistical Methods, 12(2), 371–380. 10.22237/jmasm/1383279720 [DOI] [Google Scholar]
- Neumann, P. J., Araki, S. S., & Gutterman, E. M. (2015). The use of proxy respondents in studies of older adults: Lessons, challenges, and opportunities. Journal of the American Geriatrics Society, 48(12), 1646–1654. 10.1111/j.1532-5415.2000.tb03877.x [DOI] [PubMed] [Google Scholar]
- Paas, F., Tuovinen, J. E., Tabbers, H., & van Gerven, P. W. (2016). Cognitive load measurement as a means to advance cognitive load theory. In Educational psychologist (pp. 63–71). Routledge. [Google Scholar]
- Petersen, R. C., Roberts, R. O., Knopman, D. S., Boeve, B. F., Geda, Y. E., Ivnik, R. J., Smith, G. E., & Jack, C. R. (2009). Mild cognitive impairment: Ten years later. Archives of Neurology, 66(12), 1447–1455. 10.1001/archneurol.2009.266 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Pickard, S., & Knight, S. (2005). Proxy evaluation of health-related quality of life: A conceptual framework for understanding multiple proxy perspectives. Medical Care, 43(5), 493–499. 10.1097/01.mlr.0000160419.27642.a8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Rahi, S., Alnaser, F. M., & Abd Ghani, M. (2019). Designing survey research: Recommendation for questionnaire development, calculating sample size and selecting research paradigms. Economic and Social Development: Book of Proceedings, 37, 1157–1169. [Google Scholar]
- Smith, K. V., & Goldman, N. (2011). Measuring health status: Self-, interviewer, and physician reports of overall health. Journal of Aging and Health, 23(2), 242–266. 10.1177/0898264310383421 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Sonnega, A., & Weir, D. R. (2014). The Health and Retirement Study: A public data resource for research on aging. Open Health Data, 2(1), 1–4. 10.5334/ohd.am [DOI] [Google Scholar]
- Stephan, Y., Sutin, A. R., Luchetti, M., Aschwanden, D., & Terracciano, A. (2021). Self-rated health and incident dementia over two decades: Replication across two cohorts. Journal of Psychiatric Research, 143, 462–466. 10.1016/j.jpsychires.2021.06.036 [DOI] [PubMed] [Google Scholar]
- Stephan, Y., Sutin, A. R., Luchetti, M., Hognon, L., Canada, B., & Terracciano, A. (2020). Personality and self-rated health across eight cohort studies. Social Science & Medicine, 263, 113245. 10.1016/j.socscimed.2020.113245 [DOI] [PubMed] [Google Scholar]
- Sweller, J. (2011). Cognitive load theory. In Psychology of learning and motivation (Vol. 55, pp. 37–76). Academic Press. [Google Scholar]
- Todorozv, A., & Kirchner, C. (2000). Bias in proxies’ reports of disability: Data from the National Health Interview Survey on disability. American Journal of Public Health, 90(8), 1248–1253. 10.2105/ajph.90.8.1248 [DOI] [PMC free article] [PubMed] [Google Scholar]
- van der Linden, F. A., Kragt, J. J., Hobart, J. C., Klein, M., Thompson, A. J., van der Ploeg, H. M., Polman, C. H., & Uitdehaag, B. M. (2009). The size of the treatment effect: Do patients and proxies agree? BMC Neurology, 9, 12. 10.1186/1471-2377-9-12 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Verbrugge, L. M., & Yang, L. S. (2002). Aging with disability and disability with aging. Journal of Disability Policy Studies, 12(4), 253–267. [Google Scholar]
- Vetter, T. R., & Mascha, E. J. (2017). Bias, confounding, and interaction: Lions and tigers, and bears, oh my! Anesthesia & Analgesia, 125(3), 1042–1048. 10.1213/ANE.0000000000002332 [DOI] [PubMed] [Google Scholar]
- Vogelsang, E. (2018). Feeling better at this age? Investigating three explanations for self-rated health Improvements among the oldest-old. Gerontologist, 58(5), 825–834. 10.1093/geront/gnx149 [DOI] [PubMed] [Google Scholar]
- Vogelsang, E., Raymo, J., Liang, J., Kobayashi, E., & Fukaya, T. (2019). Population aging and health trajectories at older ages. Journals of Gerontology, 74(7), 1245–1255. 10.1093/geronb/gbx071 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Walker, J., Maxwell, C., Hogan, D., & Ebley, E. (2004). Does self-rated health predict survival of older persons with cognitive impairment? Journal of the American Geriatrics Society, 52(11), 1895–1900. 10.1111/j.1532-5415.2004.52515.x [DOI] [PubMed] [Google Scholar]
- Weir, D., Faul, J., & Langa, K. (2011). Proxy interviews and bias in the distribution of cognitive abilities due to non-response in longitudinal studies: A comparison of HRS and ELSA. Longitudinal and Life Course Studies: International Journal, 2(2), 170–184. 10.14301/llcs.v2i2.116 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Wu, Q., Tchetgen, E. J. T., Osypuk, T. L., White, K., Mujahid, M., & Glymour, M. M. (2013). Combining direct and proxy assessments to reduce attrition bias in a longitudinal study. Alzheimer Disease and Associated Disorders, 27(3), 207. 10.1097/WAD.0b013e31826cfe90 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The HRS is sponsored by the National Institute of Aging (grant NIA U01AG009740) and is conducted by the University of Michigan. This study was not pre-registered.