The AHEAD data set
To evaluate whether dual use based on inpatient services increases the risk of mortality, data are taken from the Survey on Assets and Health Dynamics among the Oldest Old (AHEAD). Because over-sampling was used to increase the number of African Americans, Hispanics, or Floridians in the AHEAD, the data are weighted to adjust for the unequal probabilities of selection due either to the multi-stage cluster sampling design and/or the over-sampling. The AHEAD data set was selected for three reasons. First, it is a nationally representative probability sample that includes 2,911 men and 4,536 women who were 70 years old or older at their baseline interviews in 1993–94. Because only 57 women (1%) were veterans, these analyses are restricted to men. Among the men, 1,574 (54%) are veterans. This provides a large, nationally representative sample, evenly distributed between veteran and non-veteran men. Second, the AHEAD baseline survey data have been linked to Medicare claims from January 1989 through December 1996. Third, the AHEAD survey and Medicare claims data have also been linked to the National Death Index (NDI) through December 2002. This provides a nine-year window for examining the association of an indirect dual use measure based on inpatient services with mortality, during which 52% (1,524) of these men died.
The dual use measure
Unfortunately, the AHEAD is not linked to VHA claims. Thus, an indirect measure of dual use based on inpatient services was constructed. This was done by building on the extant literature addressing differences between self-reports and administrative records of health services use. It has been well established that concordance between the two is not perfect, and that the discordance is not easily predicted [29–37]. The magnitude of the discordance is primarily influenced inversely by the salience of the event, and directly by the length of the recall period [33–35]. Hospital episodes are considered to be the most salient events, and for them a 12-month window provides a reasonable balance between recall abilities and the incidence of hospitalization [29, 37]. Discordance has also been shown to be positively associated with the number of events during the reporting period, and to a lesser extent with demographic, social, and health factors, although these associations have not been consistently observed [37]. Elsewhere, we have shown that in the AHEAD, the concordance of self-reports and Medicare claims is high for both any (vs. none; κ = .763) and the precise number of (κ = .663) hospital episodes over a 12-month window [38].
Building on this literature, the indirect measure of dual use based on inpatient services was constructed as follows. At baseline, each AHEAD man was asked whether he was hospitalized overnight during the previous 12 months, and if he had been, how many times this occurred. Using each AHEAD man's baseline interview date, corresponding data were harvested from his Medicare claims. The self-reports were then compared to the claims results. If the AHEAD man reported at least one more hospital episode than was found in his Medicare claims he was considered an over-reporter.
Although straightforward, this approach has four limitations that, although addressable, warrant further mention. First, this approach ignores any dual use that occurs solely in the outpatient setting. As indicated above, however, the veridicality of such measures is substantially less than that obtained by focusing just on the inpatient setting [29–37]. Indeed, in the AHEAD, we have shown that the concordance of self-reports and Medicare claims is low for both any (vs. none; κ = .248) and the precise number of (κ = .347) physician visits over a 12-month window [38]. Second, this approach ignores the extent of the over-reporting. In these data, however, 79% of those who self-reported more hospital episodes than were found in their Medicare claims over-reported by only one hospital episode. Third, this approach may confound dual use with the proclivity for hospitalization. That is, by definition, all dual users have reported that they were hospitalized. This confound, however, can be addressed by including in the analysis a binary marker for whether the AHEAD man reported being hospitalized. Fourth, this approach does not actually identify dual use, because dual use can only occur among veterans.
Isolation of the dual use effect, however, can be readily achieved by constructing a set of four dummy variables that reflects the cross-classification of over-reporting with veteran status. The analyses reported here include three of these markers – veterans who over-reported (and thus are considered to be dual users of VHA and Medicare), veterans who accurately reported, and non-veterans who over-reported (but could not be dual users of the VHA and Medicare). The omitted or reference group is that of non-veterans who accurately reported their number of hospital episodes.
In this approach, the effect of the dummy variable for veterans who over-reported their number of hospital episodes accurately reflects the mortality risk of dual use of VHA and Medicare based on inpatient services. The hypothesis is that this effect will be statistically significantly greater than unity, reflecting the increased mortality risk associated with dual use of VHA and Medicare. It is further hypothesized that the effect of the dummy variable for non-veterans who over-reported will not be statistically significantly different than unity, because non-veterans have no access to VHA.
Mortality
Vital status was obtained by linking the AHEAD to the NDI. The NDI files indicate whether each AHEAD man died, and if so, provide the month and year of death. Also provided, but not used in these analyses, are indicators of the probability of the match using standard criteria [39] developed by the National Center for Health Statistics (NCHS), and detailed ICD9-CM codes for the cause of death.
Selection bias
Of the 2,911 AHEAD men, the survey data was provided by a proxy-respondent (usually the spouse) for 391 (13%). Because the literature [29–38] on reporting discrepancies assumes self-respondents, analyses were restricted to the 2,520 AHEAD men who were self-respondents. Linkage to the Medicare claims was not available for an additional 954 AHEAD men (38%). Of these, 182 refused to consent to having their Medicare claims accessed, and for the remainder (772 men) sufficiently accurate information to facilitate the linkage process was not available. All 954 AHEAD men whose self-reported survey data could not be linked to their Medicare claims were excluded from the analyses. Thus, of the 2,911 AHEAD men, the analyses reported here were restricted to the 1,566 (54%; weighted N = 1,522) who were self-respondents and whose survey data was linked to their Medicare claims.
The exclusion of so many AHEAD men from these analyses raised the potential for selection bias. This potential for selection bias was addressed using propensity score methods. Developed by Rubin, [40] popularized by Rosenbaum and Rubin, [41] and illustrated by D'Agostino, [42] propensity scores are traditionally obtained by using multiple logistic regression to model a binary outcome reflecting group assignment in observational (vs. randomized controlled trial) studies. Here, propensity scores were used to model selection bias. To obtain the selection bias propensity scores, a multivariable logistic regression was conducted using all appropriate baseline covariates, including veterans' status, to predict exclusion of self-respondents from the analytic sample. The fit of the propensity score model was reasonably robust (C-statistic [43] = .638), and there was no evidence of heteroscedastic error (Hosmer-Lemeshow statistic [44] p value = .934). Thus, adding the obtained predicted probabilities of exclusion to the final analytic model among the restricted sample should be an appropriate adjustment for potential selection bias (at least in terms of the data available for analysis). This propensity score regression approach, however, assumed additive linearity in the relationships of interest. Therefore, as an added safeguard, the results were replicated using the more popular stratification approach. In this approach, the final model (excluding the propensity score) was re-estimated separately within strata based on the propensity score. If the results were robust across these strata, greater confidence could be had in the selection bias adjustment process.
Exploring the etiological mechanism
Although the main focus of this article was whether dual use of the VHA and Medicare systems, indirectly indexed by veterans' over-reporting their number of hospital episodes, was associated with mortality, a secondary interest involved the etiological mechanism through which this association occurred. It was assumed that (a) dual use decreased the likelihood of receiving continuously coordinated health care, (b) the lack of continuously coordinated health care increased the risk of subsequent hospitalizations for ACSCs, and (c) these three factors (dual use, the lack of continuously coordinated care, and hospitalization for ACSCs) increased the risk of mortality. Left unspecified was whether the effect of dual use was direct, indirect (through its intermediary [e.g., falling domino] effects on the lack of continuously coordinated care and the increased risk of subsequent hospitalization for ACSCs), or a combination of the two. To begin exploring these issues, a final adjustment in the modeling process was made for subsequent hospitalizations for ACSCs. This was done by creating a binary marker indicating whether the subject had one or more hospitalizations for ACSCs after their baseline interview but before January 1, 1997. This marker was coded one for subjects with one or more such hospitalizations based on AHRQ's computerized criteria, [21, 22] and zero otherwise.
Although a measure of continuity of care based on the Medicare Part B (outpatient) claims has been developed for use in these data, [45] it could not be incorporated into these analyses. The reason was that this measure would have yielded biased estimates of continuity of care among dual users because it would have ignored their VHA outpatient visits, as those data were not available. Not being able to adjust for continuity of care severely constrained the ability to further explore the etiologic mechanism through which dual use was associated with mortality.
Covariates
In addition to adjusting for potential selection bias, fifteen covariates were included to ensure that the estimated association of dual use with mortality was fully independent from other background factors. These covariates included age, race, education, income, assets, activities of daily living (ADLs), instrumental ADLs (IADLs), self-rated health, five chronic diseases, cognitive ability, and depressive symptoms. Age was measured in years. Race was measured by a set of three dummy variables for Hispanics, African Americans, and other non-Caucasians (with Caucasians as the reference group). Education was measured by a dummy variable contrasting high school graduates (and above) with those having less education. Income was measured by a binary marker for having less than $15,000 in annual income. Household wealth was measured as the sum of all reported assets net of debt, and was coded by a binary marker for having $19,000 or less in total wealth. ADLs were measured by a count of the number of five items (e.g., bathing) that the subject reported having any difficulty performing. Similarly, IADLs were measured by a count of the number of five items (e.g., meal preparation) that the subject reported having any difficulty performing. Self-rated health was measured by a set of four dummy variables for excellent, very good, fair, or poor responses to the standard question asking subjects to rate their health (with a good response as the reference group). Five binary variables (1 = yes, 0 = no) were used to indicate whether the subject reported having ever been told by a physician that he had cancer, diabetes, heart disease, lung disease, or a stroke. Cognitive ability was measured using the 7-item version of the Telephone Index of Cognitive Status, which ranged from 0 (worst) to 15 (best) (TICS-7) [46]. Depressive symptoms were measured as the number of symptoms endorsed using an 8-item version of the CES-D [47].
Analytic approach
Because the month and year of death are known, proportional hazards models were the appropriate statistical approach for estimating the effect of dual use on mortality [48]. A series of proportional hazards models were estimated that initially assessed the crude effect of the set of three dummy variables reflecting the cross-classification of veteran status with over-reporting, and then serially decomposed that effect. The decomposition approach involved four stages that serially introduced (a) the binary marker for reporting any hospital episodes in the year prior to baseline, (b) the binary marker for whether post-baseline hospitalizations for ACSCs occurred, (c) the fifteen covariates, and finally, (d) the propensity score adjustment for potential selection bias. Sensitivity analyses were then conducted in which the final model, excluding the propensity score adjustment for selection bias, was re-estimated within each propensity score strata.
Institutional review
Because the research reported here involved the linkage of public use data files containing the AHEAD survey data with restricted data from the NDI files and Medicare claims, three layers of institutional review and approval were obtained. The first involved review and approval of the research and restricted data protection plans associated with the main NIH grant (R01 AG022913) by the AHEAD's Data Confidentiality Committee (DCC). These were approved by the AHEAD DCC on February 20, 2003 (#2003–006). The second layer of review and approval involved the University of Iowa Institutional Review Board (UI-IRB). The UI-IRB approved the original protocol on March 24, 2003, and has subsequently approved the protocol at all annual reviews (including appropriate modifications to incorporate the second NIH grant – R03 AG027741 – which specifically focused on dual use). The third layer of review and approval involved the Centers for Medicare and Medicaid Services (CMS). CMS approved the Data Use Agreement (DUA 14807) to access the Medicare claims for this research on March 3, 2005.