Surveys of the recipients of health care are increasingly important means of assessing the care provided by health plans, hospitals, physicians, and other entities (Lied et al. 2003; Tai-Seale 2004; Darby, Hays, and Kletke 2005). At the same time, an aging population means that more patients are unable to answer surveys themselves. The use of proxy respondents provides a practical solution to survey nonresponse and missing data attributable to limitations in the ability of patients, beneficiaries, or nursing home residents to respond themselves. By asking representatives to respond on behalf of the patient, survey researchers need not omit the experiences of the least healthy and most vulnerable. For these reasons, most health care surveys allow the use of proxy respondents. Yet, questions inevitably arise about whether proxies give responses comparable to what might have been self-reported. One might ask whether there is systematic bias or substantial decrement in reliability from the use of proxy responses that erodes much of the apparent gains from reducing the selective omission of those needing assistance in responding. Prior research on proxy response has largely focused on differences between self- and proxy responses reporting on the experiences of the same individual. Much of this work has taken place with small convenience samples of patients, often with severe health problems (e.g., end-of-life, schizophrenia, or lung cancer) (Kutner et al. 2006; Hoe et al. 2007; Wennman-Larsen et al. 2007). In such a design, any differences between the self- and proxy responses are because the proxy is reporting on the same experiences in a different way than self-respondents. While typically in poor health, targets in these studies are by definition able and willing to provide self-responses, and thus differ fundamentally from individuals who require proxy assistance. Those who are unable to respond for themselves are not represented in these prior studies and may differ in other important ways. This prior research finds that proxies can both over- and underestimate morbidity and disability (Magaziner et al. 1988; Sneeuw et al. 1997; Shaw, McColl, and Bond 2000; Todorov and Kirchner 2000; Duncan et al. 2002; Tang and McCorkle 2002; Pickard et al. 2004) and other aspects of health-related quality of life (Hays et al. 1995; Andresen, Vahle, and Lollar 2001; Becchi et al. 2004; Higashi et al. 2005; Kutner et al. 2006; Hoe et al. 2007). Proxy reporting differences may be attributable to different cognitive and perceptual strategies to answering questions (Todorov and Kirchner 2000; Stineman et al. 2004; Lynn Snow et al. 2005). Proxy respondents rely on observable factors, such as counts or the presence or absence of a symptom (Lynn Snow et al. 2005), suggesting smaller discrepancies between self- and proxy-reports for objective or observable measures than for subjective measures (Whiteman and Green 1997; Todorov and Kirchner 2000; Sneeuw, Sprangers, and Aaronson 2002; Pickard et al. 2004; Stineman et al. 2004). The relationship of the proxy to the intended respondent may also influence the accuracy of proxy responses. Some research suggests that spouses and proxies who live with the intended respondent provide responses that are closer to those of the intended respondent than do other family members (Shaw et al. 2000; The Medical Research Council Cognitive Function and Ageing Study 2000). This finding may be a function of both the opportunity for direct observation and of a similarity in perspective attributable to similarity in age, education, and other factors (Qian and Preston 1993) that are known to influence evaluations of health care (Elliott et al. 2001; Zaslavsky et al. 2001). In nursing homes and other institutional settings, spouses and family members are often less readily available than nonrelative adults, such as health care workers. Several studies examined the similarity of these proxy responses to those from spouses and other relatives (Rubenstein et al. 1984; Becchi et al. 2004; Kane et al. 2005; Kutner et al. 2006). The answer may depend on the subjectivity of the measure. For example, compared with relatives, nurse proxy reports on number of instrumental activities of daily living are closer to patient self-reports (Rubenstein et al. 1984), but nonrelative and relative proxies reports on nursing home resident quality of life were about equally close to resident self-reports. When beneficiaries are unable to provide responses independently to the CAHPS® Medicare Fee-for-Service (MFFS) and Medicare Managed Care (MMC, now Medicare Advantage or MA) surveys, proxies are permitted to participate in ways that range from assistance (by reading the questions, writing down the answers the beneficiary gives, translating the questions into the beneficiary's language, or helping in some other similar manner) to serving as a proxy respondent (i.e., answering about the beneficiary's experiences in place of the beneficiary). The current practice in the CAHPS MFFS is to use case-mix adjustment (CMA) with “assisted” and “proxy respondent” cases distinguished from “unassisted” cases, controlling for age, self-rated health, and education (Zaslavsky, Zaborski, and Cleary 2000; Elliott et al. 2001; Zaslavsky et al. 2001). This approach has found consistently less positive evaluations when proxy respondents were used (Zaslavsky, Zaborski, and Cleary 2000; Elliott et al. 2001; Zaslavsky et al. 2001), but the reasons for these differences have not been investigated. When subgroups, such as unassisted and proxy respondent cases, are very dissimilar, CMA and other regression methods by themselves may produce biased estimates of the effect of proxy use if the regression model is misspecified. This may occur because proxy status is not typically randomly assigned and standard regression methods give equal weight to all cases, including those cases with almost no chance of membership in the group to which they are being compared. To address this limitation, we employed propensity score weighting (PSW) to focus the comparison on the subgroup of beneficiaries who most resembled the beneficiaries who employed assistance or proxy respondents (Hirano and Imbens 2001). The propensity score is the probability that an individual belongs to a naturally occurring treatment group based on the individual's characteristics (Rosenbaum and Rubin 1983). This approach approximates inference under experimental assignment of treatment group under the assumption that there are no omitted variables in the propensity model relevant to selection into treatment. In practice, the technique may greatly reduce bias due to selection of the less healthy into the proxy treatment group and result in a more accurate measure of the effects of proxy use even when the assumption is not fully met, because it makes the regression model less sensitive to misspecification (Robins, Hernan, and Brumback 2000). Similarly, the use of covariates in regression in combination with PSW increases the robustness to misspecification of the propensity model (Robins et al. 2000). The PSW technique has been demonstrated to substantially reduce unadjusted differences in between self- and proxy reported health among Medicare beneficiaries (Ellis, Bannister, and Cox 2003). We hypothesize that PSW will also reduce the estimated size of proxy effects on evaluation of care received. Our study investigates proxy effects so that we can better measure the care received by vulnerable recipients of health care. We explore the extent to which observed differences in ratings and reports of care are likely to be a result of actual patient experience and unobserved selection. We also assess whether rating differences vary as a function of characteristics of the survey items and by the nature of full proxy respondents. This study extends the existing literature by (1) estimating proxy effects among those who cannot or will not self-respond in dyad studies and (2) examining the extent to which mere assistance may influence survey responses. Additionally, we compare proxy estimates obtained from the standard regression-based approach with estimates obtained using PSW, bringing a new approach to the study of proxy effects on responses.