We used administrative hospital abstract data collected by the Department of Health of the Canadian province of Manitoba, housed in de-identified form at the Manitoba Centre for Health Policy (MCHP). Manitoba has a population of 1.2 million; its two urban areas contain 61% of the population. The government-funded health care system covers all provincial residents. The administrative data contains comprehensive health-related information, and have been linked to other data including census-based socioeconomic information, and vital status. MCHP data has been used and validated extensively to study a wide range of medical outcomes [21, 22].
All Manitoba residents ≥18 years of age, discharged alive after admission to provincial acute care hospitals from April 1, 1990 to February 28, 2009 were identified. Because a patient can undergo inter-hospital transfer within a single episode of hospital care, we identified two abstracts as representing a transfer and therefore part of a single hospital episode if: (i) hospital entry of the later abstract was within one calendar day of the previous hospital separation, and (ii) an acute care hospital was the “discharge to” location of the earlier abstract, and/or the “admitted from” location of the later abstract. The exception was that two hospital abstracts were considered as separate hospital episodes if the earlier one indicated that the patient left AMA. Same day surgeries were not considered admissions in this dataset.
In Manitoba, hospital abstracts are collected in each hospital by centrally trained data abstractors using uniform definitions, data collection methods, and data entry software. A specific discharge code for AMA is used. Episodes of care were designated as AMA or non-AMA depending on the presence of this code in the last hospital abstract of each hospital episode.
Age, sex, hospital, fiscal year of admission, and postal code of residence were obtained from the first hospital abstract for each episode of hospital care. Fiscal years run from April 1st to the following March 31st and will be referred to by the calendar year of the start of the period. Length of stay was calculated from the admission and discharge dates and times of the hospital episode. Whether the individual had a prior AMA discharge during the five years before the current admission was also identified.
Socioeconomic status (SES) was based on average household income within geographic dissemination areas, based on the 2001 Canadian census; in Manitoba these area-level census tracts contain an average of 550 persons. These were separately divided into quintiles for rural and urban residents, with 1 being the lowest and 5 the highest income quintile. People living in areas where the Canadian census does not calculate an average household income formed an eleventh SES category called “not calculated” (NC). Most people in the NC category are residents in nursing homes, other chronic care facilities, or penitentiaries.
Diagnosis was derived from the “Most Responsible Hospital Diagnosis”  i.e. the diagnosis responsible for the majority of the hospital stay, obtained from the final hospital abstract of each hospital episode. Diagnoses were initially grouped into the 18 main ICD-9-CM chapter headings . Headings with low counts were collapsed, and specific diagnostic entities of interest were extracted from major headings; for a total of 23 diagnoses or diagnostic categories. The only discharges that were excluded were the small number of hospital episodes that lacked discharge diagnoses.
Whether the hospital episode included a major surgical procedure was identified. From 2004 onwards, this information was based on Canadian reporting standards for hospital abstracts, the Case Mix Group (CMG) system. Before 2004, it was based on the similar Diagnosis Related Groups system [25, 26].
Hospitals were categorized into: the three urban tertiary hospitals in Manitoba, the four urban community hospitals grouped together, and the rural hospitals grouped together. To assess for changes over time, years were groups as: 1990–1993, 1994–1998, 1999–2003, and 2004–2008.
For co-morbidities, the 31 conditions described by Elixhauser et al.  were identified from all diagnoses in the hospital discharge abstracts, using the coding described by Quan et al. . For this purpose we included the index hospitalization and all hospital diagnoses for all hospitalizations within one year backwards in time [29, 30]. Although the 31 conditions separately codified diabetes with and without chronic complications, in our data these two were not accurately distinguished prior to 2006, so the two subcategories were collapsed together.
We performed external validation of the AMA designation in the administrative data, using 291 hospital abstracts where true AMA status was identified by reading the final nurse and physician progress notes in the hospital charts. These charts were chosen in an approximate 1:2 ratio of AMA:nonAMA as indicated by independently acquired data used by our Department of Medicine. All 198 patients who did not leave AMA were correctly identified as such in the abstracts (specificity 100%, 95% C.I., 98.2-100%). However, only 81 of 93 patients who left AMA were correctly coded in the abstracts as having done so (sensitivity 87%, 95% C.I., 78.5-93.2%). In such a cohort, the indication of AMA status in the hospital abstracts would have a positive predictive value of 100%, and a negative predictive value of 99.86%.
To identify independent factors independently associated with patients leaving the hospital AMA, we constructed multivariable logistic regression models. Independent variables were the hospital diagnosis, co-morbidities, sex, age, hospital type, time period, SES, and whether a major surgical procedure was performed. Anticipating that having left AMA before would have such a strong association with subsequently going AMA that it might confound analysis of other factors, two multivariable regression models are presented -- one including and the other excluding an independent variable representing whether patients had any prior AMA episodes.
Though the unit of measure for this analysis was individual episodes of hospital care, these are not all independent since many individuals had multiple episodes. To account for this clustering of data, we used General Estimating Equations (GEE) , with an exchangeable correlation structure and robust (empirical) standard errors. We assessed for multicollinearity among the independent variables using the variance inflation factor, with values under 4 considered acceptable . We report parameter estimates from these models as odds ratios (OR) with 95% confidence intervals (95% CI). We compared regression models using the QIC parameter for GEE models, where lower values indicate a better fit .
An important issue in dealing with clustered data is that independent variables may have different between-person and within-person associations with the outcome. We allowed for this in the regressions by considering two separate versions of independent variables [34, 35]. For example, the coefficient of the within-person age variable indicates how the probability of leaving AMA varied with age for a given person; the coefficient of the between-person age variable indicates the difference in probability of leaving AMA between different people of different ages, each of whom had a single hospitalization. For most of the independent variables only the between-individual version of the variable was included. However, both versions were included for age, and whether the hospitalization included a major surgical procedure.
Univariate comparisons were done using Chi-square tests and t-tests, as appropriate. All analyses were done using SAS version 9.1 (SAS Institute, Cary, NC). A p-value of 0.05 was considered significant.
This proposal was approved by the Research Ethics Board of the University of Manitoba and the Health Information Privacy Committee of the Manitoba Government.