ED Utilization Among Recently Released Prisoners
We merged data from several sources for the present study. First, the Rhode Island Department of Corrections (RIDOC) provided data for 6,046 sentenced adults released from state correctional facilities between January 1, 2007 and December 31, 2008 ("Dataset A"). These data included demographic data, admission and release dates and ZIP code of residence for each individual. The Rhode Island Department of Corrections is unique in that it operates a unified correctional system. All sentenced individuals are housed in 1 of 7 facilities located on a single campus that is located approximately 6 miles from the state's urban center and its academic medical center. RIDOC housed approximately 3900 individuals in 2008, and 77% of released individuals returned to the counties served by study hospitals.
RIDOC data was linked to the electronic health record of a major hospital system in Rhode Island ("Dataset B"). The system's three hospitals include the state's urban, tertiary care hospital ("Hospital B") and together are responsible for approximately 50% of ED visits in the state. We identified all ED visits occurring within 1 year of each ex-prisoner's first release during the study period. Data included intake, service and discharge records. Data were linked using first name, last name and date of birth. A research analyst with extensive experience working with electronic health record data performed data linkage and extraction electronically. These data were de-identified once this linkage was made.
To obtain data on visits by the Rhode Island general population, the Rhode Island Department of Health (RIDOH) provided data on all ED visits in the hospital system from January 1, 2007 to December 31, 2009 ("Dataset C"). Data included patient age, gender, race, ethnicity, residence, diagnosis (ICD-9), year of visit, treatment facility and ZIP code of residence. No unique identifiers were included in these data and therefore visits could not be linked to individuals across facilities or over time. We obtained data on population size and unemployment rates from the 2000 United States Census ("Dataset D"). We linked census data with ex-prisoner and general population visit data using ZIP codes. We excluded visits by individuals outside of Rhode Island and nearby Bristol County, MA as they were deemed unlikely to access the hospital system of interest.
Finally, we combined visit-level data from datasets A, B, C and D to create the final sample, which included 333,369 ED visits.
We created three dependent variables at the level of the ED visit, indicating whether each visit had a primary diagnosis of one of three types of diagnosis. For the first dependent variable, we measured whether a visit had a primary diagnosis of a mental health disorder. To create this variable, we used the New York University Emergency Department (NYU ED) Algorithm, which uses International Classification of Diseases, Clinical Modification, Ninth Revision (ICD-9-CM) codes to classify ED visits into categories. We created a dichotomous variable in which visits categorized as mental health-related by this algorithm were coded affirmatively. For the second dependent variable, we measured whether a visit had a primary diagnosis of a substance use disorder. Two of the NYU ED algorithm categories were used to create this variable: alcohol and other substance use-related visits. We coded visits affirmatively if the algorithm indicated that a visit was related to alcohol or other substance use. For the third dependent variable, we measured whether a visit had a primary diagnosis of an ambulatory care sensitive condition. These conditions include several common physical health-related conditions such as asthma, hypertension, and diabetes. We coded visits with an ICD-9-CM code indicating a primary diagnosis of an ambulatory care sensitive condition affirmatively.
The study's independent variable was ex-prisoner status, defined as an index release from the state's correctional facility within the year prior to the ED visit. In these analyses, we do not differentiate between those visits occurring while an individual was living in the community and visits occurring while re-incarcerated during the year following the index release.
At the individual-level, we included variables for age (measured as a continuous variable), gender, race/ethnicity (black, Hispanic, white, other race), and the hospital facility in which the visit occurred. We excluded visits by individuals under 18 and over 70 years of age from the general population sample to ensure appropriate comparison with the ex-prisoner sample, which did not include children and included few older adults. Indicator variables for year controlled for changes in ED visitation patterns over time. At the ZIP code-level, we measured unemployment rate (measured as tenths of a percentage point) as a surrogate measure of both economic disadvantage and rate of uninsurance. Finally, we measured community population at the ZIP code-level. As these population data were highly positively skewed, a natural logarithmic transformation was performed to decrease the influence of extreme values.
We first performed descriptive statistics within the ex-prisoner cohort (N = 1434). We determined the timing of first ED visit after release, both overall and for the three diagnosis types of interest. We examined the relationship between first release from prison and first ED visit and used the chi-square test to assess associations between the timing of first ED visits and several relevant individual-level characteristics. We next compared visits by the ex-prisoner and general populations across several patient- and community-level characteristics. We used the chi-square test for differences in categorical variables and analysis of variance (ANOVA) for differences in continuous variables. We used Satterthwaite corrected t-tests to account for highly unbalanced variances across the two groups due to the large difference in the number of ED visits among the ex-prisoner and general populations.
We created random effects logistic regression models to examine the association between ex-prisoner status and the proportion of ED visits within ex-prisoner and general population groups for three outcome conditions. We assumed a logistic distribution with a logit-link function. To account for potential correlation among individuals living in the same community, we assumed an exchangeable covariance structure among patients from the same ZIP code. We created three separate models to investigate the relationship between ex-prisoner status and each of the three outcomes of interest: mental health-related visits, substance use-related visits and ambulatory care sensitive condition-related visits. We adjusted for patient gender, race/ethnicity, age, visit year, visit facility at the individual-level as well as unemployment rate and total population at the level of the ZIP code. We explored interactions between the independent variable, ex-prisoner status, and patient age, gender and race/ethnicity. We found no significant interactions and so did not include these terms in the final models.
We report results as odds ratios with 95% confidence intervals. We performed all statistical analyses using SAS version 9.3 and STATA MP version 11. The study was approved by the Miriam Hospital Institutional Review Board and by the Rhode Island Department of Corrections Medical Research Advisory Group.
Methods
Study Protocol/Data Sources
We merged data from several sources for the present study. First, the Rhode Island Department of Corrections (RIDOC) provided data for 6,046 sentenced adults released from state correctional facilities between January 1, 2007 and December 31, 2008 ("Dataset A"). These data included demographic data, admission and release dates and ZIP code of residence for each individual. The Rhode Island Department of Corrections is unique in that it operates a unified correctional system. All sentenced individuals are housed in 1 of 7 facilities located on a single campus that is located approximately 6 miles from the state's urban center and its academic medical center. RIDOC housed approximately 3900 individuals in 2008, and 77% of released individuals returned to the counties served by study hospitals.
RIDOC data was linked to the electronic health record of a major hospital system in Rhode Island ("Dataset B"). The system's three hospitals include the state's urban, tertiary care hospital ("Hospital B") and together are responsible for approximately 50% of ED visits in the state. We identified all ED visits occurring within 1 year of each ex-prisoner's first release during the study period. Data included intake, service and discharge records. Data were linked using first name, last name and date of birth. A research analyst with extensive experience working with electronic health record data performed data linkage and extraction electronically. These data were de-identified once this linkage was made.
To obtain data on visits by the Rhode Island general population, the Rhode Island Department of Health (RIDOH) provided data on all ED visits in the hospital system from January 1, 2007 to December 31, 2009 ("Dataset C"). Data included patient age, gender, race, ethnicity, residence, diagnosis (ICD-9), year of visit, treatment facility and ZIP code of residence. No unique identifiers were included in these data and therefore visits could not be linked to individuals across facilities or over time. We obtained data on population size and unemployment rates from the 2000 United States Census ("Dataset D"). We linked census data with ex-prisoner and general population visit data using ZIP codes. We excluded visits by individuals outside of Rhode Island and nearby Bristol County, MA as they were deemed unlikely to access the hospital system of interest.
Finally, we combined visit-level data from datasets A, B, C and D to create the final sample, which included 333,369 ED visits.
Study Measures
We created three dependent variables at the level of the ED visit, indicating whether each visit had a primary diagnosis of one of three types of diagnosis. For the first dependent variable, we measured whether a visit had a primary diagnosis of a mental health disorder. To create this variable, we used the New York University Emergency Department (NYU ED) Algorithm, which uses International Classification of Diseases, Clinical Modification, Ninth Revision (ICD-9-CM) codes to classify ED visits into categories. We created a dichotomous variable in which visits categorized as mental health-related by this algorithm were coded affirmatively. For the second dependent variable, we measured whether a visit had a primary diagnosis of a substance use disorder. Two of the NYU ED algorithm categories were used to create this variable: alcohol and other substance use-related visits. We coded visits affirmatively if the algorithm indicated that a visit was related to alcohol or other substance use. For the third dependent variable, we measured whether a visit had a primary diagnosis of an ambulatory care sensitive condition. These conditions include several common physical health-related conditions such as asthma, hypertension, and diabetes. We coded visits with an ICD-9-CM code indicating a primary diagnosis of an ambulatory care sensitive condition affirmatively.
The study's independent variable was ex-prisoner status, defined as an index release from the state's correctional facility within the year prior to the ED visit. In these analyses, we do not differentiate between those visits occurring while an individual was living in the community and visits occurring while re-incarcerated during the year following the index release.
Study Covariates
At the individual-level, we included variables for age (measured as a continuous variable), gender, race/ethnicity (black, Hispanic, white, other race), and the hospital facility in which the visit occurred. We excluded visits by individuals under 18 and over 70 years of age from the general population sample to ensure appropriate comparison with the ex-prisoner sample, which did not include children and included few older adults. Indicator variables for year controlled for changes in ED visitation patterns over time. At the ZIP code-level, we measured unemployment rate (measured as tenths of a percentage point) as a surrogate measure of both economic disadvantage and rate of uninsurance. Finally, we measured community population at the ZIP code-level. As these population data were highly positively skewed, a natural logarithmic transformation was performed to decrease the influence of extreme values.
Data Analysis
We first performed descriptive statistics within the ex-prisoner cohort (N = 1434). We determined the timing of first ED visit after release, both overall and for the three diagnosis types of interest. We examined the relationship between first release from prison and first ED visit and used the chi-square test to assess associations between the timing of first ED visits and several relevant individual-level characteristics. We next compared visits by the ex-prisoner and general populations across several patient- and community-level characteristics. We used the chi-square test for differences in categorical variables and analysis of variance (ANOVA) for differences in continuous variables. We used Satterthwaite corrected t-tests to account for highly unbalanced variances across the two groups due to the large difference in the number of ED visits among the ex-prisoner and general populations.
We created random effects logistic regression models to examine the association between ex-prisoner status and the proportion of ED visits within ex-prisoner and general population groups for three outcome conditions. We assumed a logistic distribution with a logit-link function. To account for potential correlation among individuals living in the same community, we assumed an exchangeable covariance structure among patients from the same ZIP code. We created three separate models to investigate the relationship between ex-prisoner status and each of the three outcomes of interest: mental health-related visits, substance use-related visits and ambulatory care sensitive condition-related visits. We adjusted for patient gender, race/ethnicity, age, visit year, visit facility at the individual-level as well as unemployment rate and total population at the level of the ZIP code. We explored interactions between the independent variable, ex-prisoner status, and patient age, gender and race/ethnicity. We found no significant interactions and so did not include these terms in the final models.
We report results as odds ratios with 95% confidence intervals. We performed all statistical analyses using SAS version 9.3 and STATA MP version 11. The study was approved by the Miriam Hospital Institutional Review Board and by the Rhode Island Department of Corrections Medical Research Advisory Group.
Source...