We examined the unadjusted relationship between ADI percentile and 30-day rehospitalization, overall and by primary disease. Based upon the empiric ADI data, the most disadvantaged neighborhoods made up the top 15% of the distribution. To better assess for within-group differences, we divided this most disadvantaged 15% into three equally sized 5% groupings representing the third-most, the second-most and the most disadvantaged 5% of neighborhoods. The remainder of neighborhoods (85%) were grouped into a comparator category. We examined frequencies of patient and index hospital characteristics for each grouping.
We used logistic regression to assess the relationship between ADI grouping and 30-day rehospitalization. Next, to assess the full spectrum of ADI impact, we divided the distribution into 20 equally-sized neighborhood groupings of increasing ADI (5% each), and used logistic regression to assess the relationship between ADI grouping and rehospitalization. To investigate the within-hospital ADI effects (43 (link)), we employed conditional (44 (link)) and random effects logistic regression (45 (link), 46 (link)). To assess for differences in disease grouping and rural-urban effects, the relationship was assessed using logistic regression models stratified by disease grouping and RUCA code. Patient numbers in stratified analyses were smaller, so we analyzed the most disadvantaged 15% of neighborhoods as a single group.
Control variables were drawn from theoretical models of rehospitalization (47 (link)) and included patient HCC score tertile, comorbidities, length of stay, discharge to skilled nursing facility, age, gender, race, Medicaid status, disability status and RUCA code of primary residence; and index hospital medical school affiliation, for-profit status and discharge volume tertile. We calculated adjusted risk ratios, predicted probabilities, and 95% confidence intervals from these models on the basis of marginal standardization, as per methods by Kleinman and Norton (48 (link)) and by Localio (49 (link)). All models were estimated twice—once accounting for hospital-level and patient-level clustering, and again using robust estimates of the variance. Since no differences were noted, we present the more conservative robust estimates. All analyses were performed using SAS 9.3 (SAS Institute. SAS Statistical Software. 9.3 ed. Cary, NC: SAS Institute; 2011) and STATA 12 (StataCorp. Stata Statistical Software. 12.0 ed. College Station, TX: StataCorp LP; 2011).
We used logistic regression to assess the relationship between ADI grouping and 30-day rehospitalization. Next, to assess the full spectrum of ADI impact, we divided the distribution into 20 equally-sized neighborhood groupings of increasing ADI (5% each), and used logistic regression to assess the relationship between ADI grouping and rehospitalization. To investigate the within-hospital ADI effects (43 (link)), we employed conditional (44 (link)) and random effects logistic regression (45 (link), 46 (link)). To assess for differences in disease grouping and rural-urban effects, the relationship was assessed using logistic regression models stratified by disease grouping and RUCA code. Patient numbers in stratified analyses were smaller, so we analyzed the most disadvantaged 15% of neighborhoods as a single group.
Control variables were drawn from theoretical models of rehospitalization (47 (link)) and included patient HCC score tertile, comorbidities, length of stay, discharge to skilled nursing facility, age, gender, race, Medicaid status, disability status and RUCA code of primary residence; and index hospital medical school affiliation, for-profit status and discharge volume tertile. We calculated adjusted risk ratios, predicted probabilities, and 95% confidence intervals from these models on the basis of marginal standardization, as per methods by Kleinman and Norton (48 (link)) and by Localio (49 (link)). All models were estimated twice—once accounting for hospital-level and patient-level clustering, and again using robust estimates of the variance. Since no differences were noted, we present the more conservative robust estimates. All analyses were performed using SAS 9.3 (SAS Institute. SAS Statistical Software. 9.3 ed. Cary, NC: SAS Institute; 2011) and STATA 12 (StataCorp. Stata Statistical Software. 12.0 ed. College Station, TX: StataCorp LP; 2011).
