Fredric D. Wolinsky The University of Iowa June 29 J 29, 2010 Academy Health Annual Research Meeting Collaborators Suzanne Bentler Maksym Obrizan Paula Weigel Mike Jones Jason Hockenberry Brian Kaskie Robert Ohsfeldt Gary Rosenthal Robert Wallace Support Over the years, my work on this issue has been generously supported by several NIH Grants: K04-AG-00328 R01-AG-06618 R01 AG 06618 R37 AG-09692 R01 AG AG-022913 022913 R03 AG-027741 R21 AG AG-030333 030333 R21 AG-031307 Prior Hospitalizations and the S bseq ent Use of Health Ser Subsequent Services ices For at least four decades decades, it has been shown that having been hospitalized in period “A” (pre-baseline) predicts the use of health services in period “B” (post-baseline) At least five explanations of why have been offered: The propensity (or capacity) to institutionalize Learned (or patterned) behavior Reflects poorly or unmeasured health needs Frailty (circling the drain, terminal drop) Lack of support, access, continuity of care, etc. Re-Hospitalization as an Offshoot of Prior Hospitalizations Re-hospitalization is a serious and widespread problem In I 1984 A Anderson d and d St Steinberg i b reported t d th thatt 22 22.5% 5% off all hospitalized Medicare beneficiaries were re-admitted within 60 days of their index hospital episode Last year, Jencks et al. showed that 60-day re- hospitalization rates among Medicare beneficiaries had risen to 28.2%, reflecting reduced lengths of stay in index hospitalizations and the rise of ambulatory surgery hospitalizations, Number (and Percent) of Medicare Re-hospitalizations and Deaths (adapted from Jencks et al., 2009 NEJM). Interval after Discharge Patients at Risk at Beginning of Period Cumulative Rehospitalizations by End of Period Cumulative Deaths without Re-hospitalization by End of Period 0-30 days 2,961,460 (100.0) 579,903 (19.6) 103,741 (3.5) 31-60 days 2,277,816 (76.9) 834,369 (28.2) 134,697 (4.5) 61-90 days 1,992,394 (67.3) 1,006,762 (34.0) 151,901 (5.1) 91-180 days 1 802 797 (60 1,802,797 (60.9) 9) 1 325 645 (44 1,325,645 (44.8) 8) 177 234 (6.0) 177,234 (6 0) 181-365 days 1,458,581 (49.3) 1,661,396 (56.1) 200,852 (6.8) >365 days 1,099,212 (37.1) Predictors of 30-Day Re-Hospitalization (adapted from Jencks et al., 2009 NEJM). Variable Hospital's ratio of observed to expected hospitalizations National re-hospitalization rate for DRG Number of re-hospitalizations since October 1, 2003 0 1 2 ≥3 Length of stay sta >2 times that expected for DRG 0.5-2 times that expected for DRG <0.5 times that expected for DRG Race Black Other Disability End-stage renal disease R Receipt i t off Supplemental S l t lS Security it IIncome Male sex Age <55 yr 55-64 55 64 yr 65-69 yr 70-74 yr 75-79 yr 80-84 yr 85-89 yr >89 yr Hazard Ratio (95% Confidence Interval) 1.097 (1.096-1.098) 1.268 (1.267-1.270) 1.00 1.378 (1.374-1.383) 1.752 (1.746-1.759) 2.504 (2.495-2.513) 1.266 (1.261-1.272) 1.00 0.875 (0.872-0.877) 1.057 (1.053-1.061) 1.00 1.130 (1.119-1.141) 1.417 (1.409-1.425) 1 117 (1 1.117 (1.113-1.122) 113 1 122) 1.056 (1.053-1.059) 1.00 0.983 (0.978-0.988) (0.978 0.988) 0.999 (0.989-1.009) 1.023 (1.012-1.035) 1.071 (1.059-1.084) 1.101 (1.089-1.113) 1.123 (1.111-1.136) 1.118 (1.105-1.131) An Alternative Approach: Prior Hospitalizations as Health Shocks Conceptually consistent with the human capital and health capital perspectives of Becker and Grossman Grossman, such that health shocks can be defined as stochastic changes g in health stocks It is generally assumed that health shocks are unanticipated drops in the value of one’s health stock Prior hospitalizations are good indicators of health shocks because they reflect “sea changes” Health shocks capture the dynamic nature of health stocks, especially when treated as time dependent covariates Overview of Our Approach AHEAD is a large, nationally representative sample whose participants were > 70 years old at their baseline interviews in 1993-1994 Medicare claims for calendar years 1993 1993-2007 2007 were linked for post-baseline health shock surveillance Baseline interviews p provided an array y of covariates Propensity score methods adjusted for selection bias Data Data weighted to adjust for multi multi-stage stage cluster and overover sampling of African Americans, Hispanics, and Floridians Baseline response rate of 80 80.4% 4% (N=7 (N=7,447) 447) Analyses restricted to 5,511 AHEAD subjects who were: self-respondents self respondents at baseline linkable to their Medicare claims not enrolled in managed Medicare at baseline Participants were censored at the time of either of two competing p g risks—death or enrollment in managed g care Selection Bias To adjust for potential selection bias due to the three exclusions, we used propensity score methods to reexclusions weight the data Multiple logistic regression modeled inclusion in the analytic sample using baseline interview data The fit of this model was good g C statistic = .72 Hosmer-Lemeshow statistic = .15 Selection Bias… Within each propensity score (predicted probability) decile we determined the average participation rate (P) The inverse (1/P) was used to re-weight the data This gives greater influence to participants in the analytic sample most like those not included The propensity score weights were adjusted so that the final weighted N was equal to the actual number of participants in the analytic sample (i.e., 5,511) Measurement and Modeling Surveillance started at the baseline interview date Hip fractures fractures, strokes strokes, and heart attacks were identified using appropriate ICD9-CM diagnostic codes These outcomes had to occur at least one day after the baseline interview Multivariable proportional hazards regression with competing risks were used Model building g and evaluation used standard p procedures Covariates were selected separately for each outcome Baseline Covariates Demographics Socioeconomics Geographic Factors Health H lth B Behaviors h i Disease History Functional F ti l Status St t Cognitive Status Health Shocks Time-dependent covariates were switched “on” when the person was admitted to a hospital for a primary ICD9-CM diagnosis other than the model outcome It stayed “on” on for n days after discharge, and was then switched “off” It could be switched back “on” at the onset of another pre-censoring hospital admission for something other than the outcome in the model Sensitivity analyses determined which of several values of n was most predictive of the target outcome Health Shocks… Health shocks were experienced by about 80% Hip fractures were experienced by 8 8.9% 9% (N=491) Strokes were experienced by: 9.9% 9 9% (N=545) using high sensitivity algorithm 6.8% (N=374) using high specificity algorithm Heart attacks were experienced by 8 8.8% 8% (N=483) There were 40-45K person-years of surveillance (mean of 7 7-8 8 years per-person) per person) Descriptive Results mean age was 77 mean income was 38% were men $25,417 $25 417 25% had arthritis 9% had angina 13% had cancer 12% had diabetes 46% had hypertension 10% were African American 4% were Hispanic 41% were widowed 25% had only been to grade school The Bottom Line on Health Shocks Adjusted Hazard Ratios (AHRs) for Time-Dependent Prior Hospitalization Marker by Selected Catchment Periods for Hip p Fracture,, Stroke and Heart Attack Hip Fracture (n=5,511) Stroke Sensitivity (n=5,511) 7 days 2.03 14 days Catchment Periods Stroke Specificity (n=5,511) 5.36 First-Ever Stroke Sensitivity (n 5 039) (n=5,039) 5.80 2.90 First-Ever Stroke Specificity (n 5 039) (n=5,039) 3.37 2.51 4.65 4.98 3.10 30 days 1.91 3.53 3.60 60 days 1.93 3.04 90 days 1.97 6 months Heart Attack (n=5,511) First-Ever Heart Attack (n=5,137) 4.66 5.38 3.41 3.90 4.20 2.82 2.94 2.86 2.91 3.23 2.73 2.95 2.81 2.76 2.98 3.13 2.78 2.98 2.58 2.39 1.84 2.75 2.89 2.49 2.66 1.99 1.93 1 year 1.79 2.37 2.50 2.09 2.19 1.82 1.74 1.5 years 1.68 2.16 2.18 1.89 1.87 1.79 1.74 2 years 1.56 2.01 2.05 1.78 1.79 1.65 1.60 Sources: J Gerontol: Med Sci, 2009;64A:249-255 (hip fracture) BMC Geriatrics, 2009;9(17):1-11 (stroke) J Gerontol: Med Sci, 2010;65A769-777 ((heart attack)) Discussion and Implications Health shocks significantly and substantially improved the fit of all of three risk models The health shocks did not diminish the effects of the known risk factors, but represent new (previously unmeasured) risks Health shocks are NOT the index visits for re-hospitalizations Peak health shock effects occur in the first two weeks, suggesting ti clear l value l iin post-discharge t di h monitoring it i and d intervention during transition periods Future work should examine whether some health shocks are more important for specific health outcomes than others and as always, always may the Lord’s Blessings be upon you, as they have been upon me