Introduction

An estimated 37,000 individuals are incarcerated in Canada on any given day and these individuals tend to be disproportionately burdened by poor health, including high rates of infectious disease, substance use, and mental illness[1–6]. Most data on the health of persons experiencing incarceration is based on cross-sectional surveys collected while in custody; there is little longitudinal research on the health of incarcerated persons outside of correctional facilities either prior to or post incarceration[7–9]. The one exception is mortality: several studies report a high risk of death immediately following custodial release[10, 11].

There are increasing calls to enrich our understanding of the health of persons who experience incarceration[12–15]. Such individuals typically draw from the most vulnerable and marginalized groups, groups that are less likely to access health care. Incarceration may improve health, providing an opportunity to diagnose and manage illness in traditionally under-serviced populations. On the other hand, it has been argued that incarceration may engender illness through environments (e.g. close person to person contact) and behaviours (e.g. risky sexual activities) that may lead to health implications. In order to truly understand if and how incarceration impacts health, longitudinal data are needed. Currently, there are no existing Canadian datasets that capture the health of federally incarcerated persons pre and post custody.

Record linkage could be a cost-effective strategy to generate data on the health of persons experiencing incarceration. Linkage studies capitalize on data collected for other purposes and the indicators in the resultant repositories enable research that would not have been possible from any data source alone[16]. Linkage studies can increase capacity to conduct research in vulnerable populations and such an approach has been used to examine the mortality of persons incarcerated post-release[17].

We performed a record linkage study, joining together Canadian federal correctional data and administrative health data. The resultant database is powerful and unique: one of only a few population-based, longitudinal datasets to capture a wide range of health and correctional indicators from incarcerated individuals in the world[17]. The present study describes the record linkage and examines the characteristics of individuals whose correctional records were more likely to be matched to administrative health data as well as the characteristics of those linked deterministically (i.e. exact matches).

Materials and Methods

Data linkage

All linkage was performed using the Automatch software[23]. We used surname, given name, sex, date of birth, and residential postal code to link OMS records to the RPDB. These data elements are assumed to be accurate in both databases and contain no missing fields. We linked records using both deterministic and probabilistic approaches. Deterministic linkage requires perfect agreement on specified data fields; these data fields should be sufficiently unique so that only one person can be ascribed to the record. Probabilistic linkage is used when linking data fields may not be unique. In the face of such complexity, statistical probabilities, called linkage weights, are generated reflecting the likelihood two records are a true match[23].

In the present study, linkage occurred in five different passes: records linked on the most recent pass were removed from the linkage pool. The first pass used deterministic linkage, requiring an exact match on surname, given name, date of birth and sex. This combination of data elements produced a correct linkage rate of over 98% in another Canadian study[24]. The second through fifth passes used probabilistic linkage, requiring agreement on fewer data fields or strings of data (e.g. first three letters of a surname). Match requirements for each pass are found in Table 1. At each probabilistic linkage pass, linkage weights were generated by the software. Weights were based on the m and u probabilities which are the conditional probability that a field agrees, given the pair is a true match and the conditional probability a field agrees, given the pair is a true mismatch, respectively. A total linkage weight is calculated from these m and u probabilities[23]. Linked pairs with weights falling above pre-determined thresholds were considered automatic matches whereas pairs with weights falling below thresholds were considered non-matches. Linked pairs with weights falling within the threshold range were considered possible matches and were manually reviewed by two data covenantors. Postal codes were considered during the manual review. Rules were created a priori to determine if manually reviewed pairs would be accepted or rejected as matches and are listed in S1 File.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Data fields used to link Correctional Service Canada data to the Registered Persons Database, by pass number

https://doi.org/10.1371/journal.pone.0161173.t001

Because of the matching algorithm and decision rules, it was possible that multiple unique correctional records could be matched to the same health record on the same pass. When this occurred, we considered the true match to be the record with the highest linkage weight, i.e. the record that was statistically more likely to be a true match.

Results

Out of 56,867 OMS records, 22,844 (40.2%) were linked to the RPDB and 34,023 (59.8%) were not. After assuming duplicate matches with the lowest linkage weights were false positives (n = 29), the linkage rate was 40.1% (22,815 of 56,867).

Looking exclusively at individuals admitted to an Ontario-based facility, the linkage rate increased to 98% (14,953 of 15,248): only 295 individuals were not linked (2%). Outside of Ontario, 18.9% of records were linked (7,862 of 41,619). The overwhelming majority of matched records were linked deterministically (i.e. on the first pass): 78.6% of all matched records (17,941 of 22,815) and 85.6% of all Ontario-based records were linked deterministically (13,057 of 15,248). Within Ontario, 8.1% (n = 1235), 3.6% (n = 549), 0.6% (n = 100), and <0.1% (n = 12) of all records were linked on the second through fifth passes, respectively.

The mean number of OMS records per unique person (i.e. aliases) was 2.56 (SD = 2.00) and 3.23 (SD = 2.83) for men and women incarcerated in Ontario. Aliases were less common among individuals incarcerated outside of Ontario: the average number of names was 2.20 (SD = 2.26) for men and 1.98 (SD = 2.15) for women. The maximum number of surnames was 53.

Table 2 shows the distribution of matched records, by linkage rate, in Ontario. We found that women, minority populations, and individuals with more severe alcohol use were less likely to be linked than their respective comparator (column 6). We found that individuals with a greater number of aliases or a higher risk of reoffending were more likely to be linked.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Characteristics of federally incarcerated persons in the province of Ontario (1998–2011), by linkage outcome (N = 15,248).

https://doi.org/10.1371/journal.pone.0161173.t002

Table 2 (column 7) also presents the odds ratios for being matched deterministically compared with probabilistically. These odds did not vary significantly by age, marital status, or substance use histories. Those identifying as either black or ‘other’ races were less likely to be linked deterministically compared with those identifying as white. Having more aliases and having a higher criminal history risk rating was also associated with lower odds of deterministic linkage.

Discussion

We linked records from almost 15,000 Ontario-based, federally incarcerated individuals to a province-wide health registry. This translated to 98% linkage rate and most records (86%) were linked deterministically. Our observed deterministic match rate is comparable to rates reported in other Canadian studies using the same data fields[24].

While the high linkage rate for Ontario-based persons is reassuring, we found that women, minority populations, and substance users were less likely to be linked to health records. Variations in linkage by participant characteristics are not surprising and have been reported elsewhere[31]. Reasons for lower linkage and the consequent underrepresentation of certain groups in the final database are not clear. A possible explanation for the lower linkage of women may relate to surname changes arising from marriage or divorce. We included a surname record for all documented aliases, but if different databases did not have the same surname on file, records could not have been linked. A possible explanation for the lower linkage rate of Indigenous Canadians may relate to jurisdictional differences in health care delivery. In Canada, the federal government is responsible for the delivery of some health care to Status Indians, including primary and emergent care on reserve. Although Status Indians are eligible for provincial health care, they may or may not be registered, i.e. in the RPDB. We acknowledge that systematic exclusions of sub-populations may bias estimates generated from these data[32–35]. However, our relatively large sample size means that we are more likely to conclude minor differences between groups are statistically significant, raising the question of how different truly are the 2% of men and 3% of women who were not linked? This will require attention as we move to using these data in future research.

We also found that certain groups were more likely to be linked, including those with a higher static risk rating. Although initially surprising, indicators of past criminal history contribute to the overall static risk rating and being involved with the criminal justice system could itself facilitate contact with health care. Individuals are routinely seen by a physician within a period of days to weeks of admission to a provincial correctional facility. Institutional staff will request health card numbers for anyone who is in provincial custody for long enough to be seen by a physician who does not already have a health card number. Accordingly, individuals with a higher risk rating may be more likely to be in the RPDB because of their criminal history.

We also observed a positive relationship between the number of aliases and data linkage. Aliases are ubiquitous in correctional populations: women incarcerated in British Columbia reported an average of 2.7 surnames in a 2005 study[36]. Similarly high numbers of aliases have been reported in incarcerated men and are in line with what we report herein[37]. Intuitively, it makes sense that those with more names have a higher likelihood of being linked to health data. The quandary lies in how to manage these individuals in linkage studies; specifically, how does one maximize rejection of false matches (i.e. specificity) and acceptance of true matches (i.e. sensitivity)? Some researchers have excluded individuals with aliases in excess of a specified threshold[36], although such an approach may decrease the linkage sensitivity and, as a result, misestimate the burden of health outcomes in incarcerated persons[22]. Future linkage studies in correctional populations should continue to explore the optimal approach to including those with several names.

We also compared records linked deterministically to those linked probabilistically. On most indicators, there were no systematic variations by linkage method. Because there is less certainty if probabilistically linked data are true matches, a conservative approach may be to remove these individuals from the linked dataset. Future studies using these data should conduct sensitivity analyses to fully understand the implications of including or excluding probabilistically linked records.

Our correctional data repository will enable the generation of much-needed knowledge of health pre and post incarceration. Because the OMS records were linked to a registry of all individuals eligible for health care in Ontario, this repository can be used to compare the health of incarcerated persons to the general population. We can also use these data to examine variations in health within specific sub-populations of interest (e.g. Aboriginal). In regards to the latter, the OMS contains a variety of standardized instruments that would not ordinarily be captured with administrative data (e.g., confounders, mediators).[16] These powerful data have the potential to contribute substantially to the field of health research in incarcerated persons.

The limitations of this study should be considered. The OMS contains records for all Canadians admitted to federal custody between 1998 and 2011, yet we only had access to Ontario health data. We have no way of knowing the true denominator for federal inmates eligible for health care in Ontario from these data. Because region of admission is often determined by the region the criminal offence was committed and where sentencing occurred, we restricted the majority of analyses to those admitted to an Ontario facility. The assumption being if region of crime and region of residence are interchangeable, those admitted to an Ontario facility should be in the RPDB. It is unclear if the 295 unlinked Ontario-based records were true missed linkages or if they belonged to people who resided outside of Ontario and could not have been linked. On a related note, 20% of inmates admitted to a federal facility outside of Ontario were linked to the RPDB, suggesting some had residence in Ontario over their lifecourse, and perhaps prior to incarceration.

Conclusion

Through data linkage, we created a powerful platform to enable future research on the health of Canadian individuals who were incarcerated, pre- and post-incarceration. Our repository of almost 15,000 individuals federally incarcerated in Ontario includes demographic measures and validated instruments on substance use. Linkage to the RPDB, the registry of persons eligible for health care in Ontario, will enable the exploration of the health utilization and status of individuals using a variety of different indicators derived from physician billings, hospital separations, and emergent care databases. Although linkage rates in Ontario were high (98%), some populations were proportionately under-represented, like women and minorities.

Supporting Information

Acknowledgments

The Centre for Urban Health Solutions is part of the Li Ka Shing Knowledge Institute of ST. Michael’s Hospital. This study was supported by the Institute for Clinical Evaluative Sciences (ICES), which is funded by an annual grant from the Ontario Ministry of Health and Long-Term Care (MOHLTC) and also supported by the Dalla Lana School of Public Health at the University of Toronto. The opinions, results and conclusions reported in this paper are those of the authors and are independent from the funding sources. No endorsement by ICES or the Ontario MOHLTC is intended or should be inferred. Parts of this material are based on data and information compiled and provided by CIHI. However, the analyses, conclusions, opinions and statements expressed herein are those of the author, and not necessarily those of CIHI.