Patients were recruited as soon as possible after admission to the study units (including those transferred from other hospital units). Patients were enrolled at their first admission during the study period. Excluded were: patients admitted from a long-term care setting, those who were expected to be discharged to a long-term care setting, and those patients (or caregivers) who were unable to speak and read in English or French (See flowchart in Fig. 1).
Patient consent was sought for an in-hospital structured interview and a telephone follow-up after discharge to ask about discharge experiences (reported elsewhere). If clinical staff judged that the patient was not capable of informed consent (for physical or mental reasons) these patients were invited to provide assent for study staff to contact their caregiver (unpaid family member or friend), who was invited to participate as a proxy respondent for the patient.
After consent, we conducted a short (5–10 minutes) structured interview to collect the following information on the patient: level of education; language spoken at home; country of birth (Canada vs other); receipt of local authority (CLSC) home care services; and the ISAR-R.  (Appendix A). ISAR-R was used as a covariate in the current study; results were not disclosed to hospital staff. If a patient responded positively to the ISAR-R question on hospitalization during the previous 6 months, we asked whether they had been admitted to a different hospital. The answers enabled us to estimate the likely under-reporting of readmissions.
Patient data were linked to the hospital discharge databases of their admission hospital using their medical record number. The databases were used to extract hospital admissions during the 90 days after discharge, as well as discharge diagnoses during the 12 months before admission, to compute the Charlson Comorbidity Index , a widely-used measure of multimorbidity. (As noted above, insufficient resources were available to use provincial administrative data.)
Sample derivation and statistical analysis.
A total of 711 patients were identified by research staff as eligible and available to approach, and were invited to participate in the study. Of these, 485 (68.2%) consented and completed the in-hospital interview. Of these, 398 (82.1%) were discharged back home; 12 had one or more missing ISAR-R revised items, leaving 386 with complete ISAR-R revised data and formed the main analysis sample (see flowchart in Fig. 1).
To assess the representativeness of the main analysis sample (group A), we compared demographic and other characteristics with those of the following 3 groups: B - those with one or more missing ISAR-R responses (n = 12); C - those excluded because they had not been discharged home (n = 87); D - other patients aged 65 + discharged home from the same units during the study period (n = 2,878). The latter group was defined from hospital administrative data on all patients aged 65 + discharged home, excluding those in the structured interview sample; for those with more than one admission during the study period, we randomly selected one admission. It was not possible to identify patients in group D who would have been ineligible due to language or cognitive impairment with no proxy informant.
To compare the study sample with the excluded groups, we computed Chi-square tests for categorical variables and t-tests for continuous variables; the Kruskal-Wallis test was performed for skewed distributions . The Bonferroni correction was applied at alpha 0.05 to account for multiple testing (level of significance after Bonferroni correction = 0.016) .
The performance of the ISAR-R on 30-day and 90-day readmission outcomes was assessed with sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), computed for each outcome and ISAR-R cut-point. Area Under the ROC (receiver operating characteristic) Curve (AUC) and 95% confidence intervals were computed for each outcome to estimate the overall performance of the ISAR-R score [15, 16]. For each outcome, the AUC (95% confidence interval) of the ISAR-R score was computed by age group, informant (patient vs caregiver), language and previous hospital admission at different hospital; AUC values were compared between the subgroups .
The associations between the 6 ISAR-R items and each outcome (30-day and 90-day readmission) were analyzed with logistic regression. For each outcome, two multivariable models were fitted; the first model included all ISAR-R items; the second model also included the covariates (hospital of index admission, service, age, previous hospitalization at different hospital, language spoken at home, informant). Odds ratios and 95% confidence intervals were computed for all ISAR-R items; the C-statistic (concordance statistic) was computed from the logistic regression model . The C-statistic is equivalent to the AUC; values closer to 1 indicate better performance of the model at correctly classifying outcomes.
The performance of two shorter versions of the ISAR-R, excluding items with lower predictive value for the outcomes, was analyzed following the same approach used for the full ISAR-R. All the analyses were conducted with Stata (version 15.1, Stata Corp, College Station, TX).