Development of a Comorbidity Score Based on Institutionalization: Comparative Performance with a Comorbidity Score Modeled on Mortality

Background: Comorbidity scores based on the prediction of 1-year mortality risk are used in pharmacoepidemiologic and pharmacoeconomic research to control for unmeasured confounding. However, admission to a long-term care institution may be a more salient and discriminant outcome on which to base comorbidity scores. Objective: To develop a prescription claims-based comorbidity score to predict institutionalization in the community-dwelling elderly population, and to compare its performance against existing mortality comorbidity scores. Methods: A geriatric institutionalization comorbidity index was derived from data of a retrospective cohort of 61,172 community-dwelling older adults identified through the Quebec claims databases (2000-2009). Predictors of institutionalization were identified through clinical expertise and by nested case-control analysis. The Framingham Heart Study method was used to develop the score. The performance of the score was assessed through the c-statistic in a separate validation cohort of 26,216 persons and compared with the performance of a mortality score, the Geriatric Comorbidity Index. The robustness was assessed in a cohort of elderly individuals with dementia. Results: Drugs associated with an increased risk of institutionalization were: antipsychotics, antidepressants, hypoglycaemic agents, statins, benzodiazepines and antihypertensives. The mean score assigned to cases was significantly different from that of controls. The c-statistic for the Institutionalization Comorbidity Index was 0.79 (95% CI: 0.77-0.83) compared to 0.75 (95% CI: 0.73-0.78) for the mortality index. The score was robust when in the dementia cohort Comorbidity Index was 0.81 95% CI: 0.78-0.84). Conclusion: A score that predicts institutionalization in the community-dwelling elderly population offers improvement over existing comorbidity scores. It may therefore be used in research conducted in this population, especially for drug effectiveness and health economic studies which often involve institutionalization as the outcome of interest. *Corresponding author: Yola Moride, Faculty of Pharmacy, Université de Montréal, CP 6128, Succursale Centre-Ville, Montreal Qc H3C 3J7, Canada, E-mail: yola.moride@umontreal.ca Received October 24, 2012; Accepted November 23, 2012; Published November 25, 2012 Citation: Béland SG, Tannenbaum C, Ducruet T, Préville M, Moride Y (2013) Development of a Comorbidity Score Based on Institutionalization: Comparative Performance with a Comorbidity Score Modeled on Mortality. J Gerontol Geriat Res 2:116. doi:10.4172/2167-7182.1000116 Copyright: © 2013 Béland SG, et al.. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.


Introduction
Accounting for a large portion of health care expenditures, institutionalization is considered to be an important patient outcome in real world studies conducted in the elderly population [1]. Furthermore, it is well known that institutionalization is associated with several adverse outcomes such as poor quality of life, loss of independence and even mortality [2]. Institutionalization and admission to long-term care represents an important outcome in clinical studies involving the elderly population, as well as a costly endpoint in health economic studies. As the numbers of older adults increase, the demand for long term care services will increase too.
Comorbidity scores have previously been developed in order to address unmeasured confounders in claims databases. Health care claims databases are a critical source of data from which to conduct health services and epidemiologic research. These databases are easy to obtain, relatively inexpensive and contain a large bank of populationbased health information. Database linkage leads to the accumulation of longitudinal data on prescriptions and medical services as well as health outcomes in large populations. However, the advantages of these databases for their use in research are offset by the absence of clinical information which is often not recorded in claims data.
Among the most widely comorbidity score used are the von Korff Chronic Disease Score (CDS), based on prescription claims [3], and the Charlson Index, based on medical services claims [4]. Both involve scores that predict the risk of mortality in the following year [5][6][7][8][9]. However, as shown by Mikaeloff et al. mortality may not be the most relevant outcome in certain sub-populations, such as that of the infant population, due to its very low frequency [10]. In contrast, in the elderly population a more salient outcome to mortality is disability-free survival or its corollary, admission to long-term care. Comorbidity indices published in the literature typically include asymptomatic medical conditions, such as hypertension, that predict mortality but exclude many diagnoses related to quality of life, such as urinary incontinence and depression. Schneeweiss  depends on several factors, such as the choice of study outcome and the target population [11]. In the elderly population, a score based on mortality is most appropriate to address research questions related to survival as opposed to outcomes, such as disability-free survival, institutionalization, hospitalization, morbidity, quality of life and health care costs [12]. As such, institutionalization may be a more relevant outcome than death when constructing comorbidity scores for the elderly, for cost reasons as well as those related to patient preference. Older adults report fear of losing their autonomy, their dignity and to be abandoned by their family. Because of this, institutionalization is not really an option but more a must.
Persons aged 65 years and older are expected to account for 30% of the population by the year 2050 [13]; hence costs associated with the aging population will continue to grow.
It is important to develop tools that are specific and relevant to elderly population since many pharmacoepidemiologic and pharmacoeconomic studies will be performed in the next few years. In addition, the elderly have been identified by the FDA as a population vulnerable to drug-induced adverse effects [14]. Yet, they are rarely included in pre-approval randomized controlled trials. Consequently, at the present time, evidence on the benefits and harms of drugs in this population is mainly generated through real world studies In a previous study, we have developed the Geriatric Comorbidity Score (GCS), which is based on the risk of mortality in the following year. This prescription claims-based score was developed using the Framingham Study methodology [15]. Potential predictors of mortality were initially selected through a comprehensive literature review and clinical expert opinions. A multivariable logistic regression analysis was then conducted to retain the final predictors to be included in the score. Our validation assessment has shown that the GCS yields higher performances than the von Korff CDS in the sub-population of community-dwelling elderly. We hypothesize that such performance may be further improved with the use of institutionalization as the outcome of interest instead of mortality. The objectives of our study were to develop a prescription claims-based comorbidity score to predict the risk of institutionalization in a community-dwelling elderly population, and to compare the performance of such score with that of the Geriatric Comorbidity Score, which is based on mortality. As secondary objective, we wanted to assess the factors associated with institutionalization.

Design Overview
The development of the score was based on a nested casecontrol analysis. This design was selected for its ability to control for confounding variables in a time window that is proximal to the occurrence of the outcome [16]. A cohort study design was not retained for the development stage since time dependent analyses of exposure variables may result in complex scores, which may not be suitable for use in most clinical health services or epidemiologic models. However, since our score will be used in many different designs, we tested the performance of our new institutionalization score using a cohort study design.

Data sources
The public drug program of the province of Quebec (Canada) covers the great majority (>97%) of residents of the province age 65 and over. The program is administered by the Régie de l'assurance maladie du Québec (RAMQ), and the resulting prescription claims database includes the following information on each dispensing: drug name, drug class (using the American Hospital Formulary Classification), dosage, number of units, prescribed duration (from which can be derived prescribed daily dosage), date of dispensing, prescriber's specialty. The program covers medications included in the drug formulary and dispensed in an outpatient setting; over-thecounter medication or those dispensed in-hospital are not recorded in the RAMQ prescription database.
Health coverage is universal in Quebec, i.e. all residents are covered regardless of age or income. The resulting database includes information on medical services that are billed on a fee-for-service basis whether they are rendered in outpatient clinics, emergency rooms, or hospitals. Among the variables included in the medical services database are: date of service, physician's specialty, diagnosis (according to the ICD-9 classification), type of service (medical act) and location of service. In addition, through the beneficiaries' databases, data on age and gender are available. Databases may be linked through the health insurance number which is unique for each patient and remains unchanged over time.

Definition of the cohort
The study population consisted of a cohort of 87,392 communitydwelling elderly ages 66 and over who were randomly sampled from members of the public drug program from 1 st January 2001 to 31 st December 2009. Age 66 was chosen as the lower limit since a one-year prescription history was required for all patients. Patients who were users of cholinesterase inhibitors were excluded because, in Quebec, the reimbursement of these drugs is conditional on their effectiveness, i.e. criteria for reimbursement include assessment of cognitive status and decline. Inclusion of these drugs in the development of the score would have likely resulted in a protopathic bias. Also, it is important to note that dementia is one of the most important predictor of institutionalization. By including demented patients, the majority of the variance would have been taken by this variable.
To optimize statistical efficiency, 70% of the cohort was used for the construction of the score and 30% for its validation.

Follow-up
Cohort members were followed until the first of the following events: i) institutionalization; ii) end of their coverage in the public drug program due to discontinuation of residency; iii) mortality; iv) end of the study period (31 st December 2009).

Identification of cases and controls
Patients institutionalized during the follow-up period were retained as cases. Institutionalization was identified through the location of services billed by physicians. The date of occurrence was the first date for which a service was billed at an institution, and was retained as the index date. For each case, on the date of institutionalization, up to 4 controls were selected through risk set sampling. Controls were matched on the year of entry in the main cohort, age (65-74, 75-79, 80-84, 85 and over), and gender. The same index date as the case was assigned to the matched controls.

Independent variables
Potential predictors of institutionalization were selected based on their clinical relevance and evidence from the literature concerning their association with institutionalization [17][18][19][20][21]. Some of those factors cannot, however, be considered in a claims-based tool since several patient characteristics, such as marital status, social status, income, are not available in administrative databases. The following conditions were retained as potential predictors: diabetes, urinary incontinence, depression, anxiety, sleep disturbances, Parkinson's disease, hypercholesterolemia, pulmonary disease, chronic infection, hypertension and cardiovascular and cerebrovascular disease. For each of these conditions, drug classes used for treatment were identified and were considered as potential predictors of institutionalization. Drug dispensings were used as opposed to diagnoses in the medical services database since, in the Quebec claims databases, they have previously been shown to be more reliable to ascertain the presence of a medical condition than medical claims [22]. Also, in this context (comorbidity score), developing a score based on drug users would be of particular interest, as not all databases contain medical diagnostic information and as this is of inconstant validity, especially in elderly.

Time window
Exposure to the selected drugs was assessed through the presence of at least one dispensing in the prescription database during the year prior to index date. A sensitivity analysis was conducted using a 3 month time window prior to the event which yielded similar results (data not shown). Since most comorbidity in this population consists of chronic diseases, drug dispensing received in a 3 month time window was highly correlated with drug dispensing in a one-year time window.

Statistical analysis
Development of the score: Bivariate analyses were conducted to identify potential predictors of institutionalization in the construction cohort. Drugs that were significantly associated with institutionalization were considered as potential predictors and were retained for multivariable analyses at a significance level of 0.20. Multivariable conditional logistic regression analysis was then conducted, with independent variables identified through backward stepwise selection. To ensure the stability of the model, forward and bidirectional selection were also used, and the retained variables compared. However, in order to keep all important clinical variables, in the development of a score, it is important to use a level of significance which is higher than evaluation study. The final predictors were those variables which were significantly associated with an increase or decrease in the probability of institutionalization with the level of statistical significance set at 0.20. A weighted score was assigned to each predictor, using the Framingham Heart Study Method [15]. Weights correspond to a multiplication of the Beta coefficient obtained from the final multivariable logistic regression model with a constant in order to obtain an integer number. The constant consists of an arbitrary number that yields integers instead of decimals. In that case, the retained constant was 0.25 a risk factor is associated with a positive score while a protective factor is associated with a negative score. An overall risk score was then derived through summation of individual scores for each individual included in the construction cohort.
Validation of the score: nested case-control design: A nested casecontrol design was also used to assess the performance of the score in the validation cohort. The predictive accuracy of the institutionalization score was assessed through measures of calibration and discrimination of the multivariable logistic model and the point scoring system. Calibration and predictive accuracy were determined by comparing the predicted and observed number of patients institutionalized in the validation cohort. Discrimination was assessed using the area under the Receiver Operating Characteristic (ROC) curve in the validation cohort. The c statistic was used as a measure of performance. The c statistic is the fraction of patients with the outcome among pairs of patients where one has the outcome and one not; the patient with the highest prediction being classified as the one with the outcome. Hence, when a model provides no information, the value of the c statistic is 0.5. Hosmer and Lemeshow suggest that a c statistic over 70 provides acceptable predictive validity [23]. Confidence interval on the c statistic was obtained using a bootstrap method.
Validation of the score: cohort design: The new comorbidity score based on the prediction of institutionalization was also validated using a cohort study design. Subjects were considered exposed if they were using drugs of interest (the ones which are included in the calculation of the score) at the time of entry in the cohort. The retained time window is the same than in the nested-case control approach was 1-year. The score was then applied in the cohort and the c-statistic was calculated.
Sensitivity analysis: assessment of the robustness of the score: Since dementia is one of the most important predictors of institutionalization and it is not taken into account in the development of the score, it was decided to test the robustness of the score in a demented elderly population as sensitivity analysis. To be included in the cohort with dementia, elderly must be new users of cholinesterase inhibitors between 1999 and 2009. A cohort of 37,138 patients was constructed. The institutionalization score was then applied in that specific cohort and the c-statistic was used to assess the performance.
All data were analyzed using the SAS statistical packages (SAS 9.2 for Windows, SAS Institute Inc., Cary, NC, USA).

Ethical Considerations
No patient or physician identifiers were provided to the researchers; only encrypted identifiers were used throughout the study. The study was approved by the Ethics Committee of the University of Montreal Hospital Center.

Results
The development and validation cohorts were similar in terms of demographic characteristics ( Table 1). The majority of patients was between the ages of 65 and 69 (55.85% and 56.51%, respectively) and was females (57% for both cohorts).
As shown in table 2, factors which significantly increased the risk of institutionalization were the use of antipsychotics, antidepressants, and hypoglycaemic agents. Protective factors included: use of statins, antihypertensives, and benzodiazepines. Finally, despite a plausible clinical association between some drugs and the risk of institutionalization, the effect of such variables were not statistically significant in the multivariable model (use of osteoporotic drugs, respiratory drugs, diuretics, glaucoma drugs, proton-pump inhibitors, cardiovascular drugs, antiparkinsonians, urinary incontinence drugs, anti-thrombotic agents). Table 3 presents the results of individual scores for each of predictors of institutionalization. Being the most important factors impacting on institutionalization, use of antipsychotics obtained the highest score with 5 points followed by antidiabetics with 2 points.

Validation of the score
Results of the validation component of the study are presented in table 4. The score range between -19 and 8. The mean score for cases is statistically higher than that for controls, which indicates that the score discriminates between the two groups. When applied to the validation cohort, the c statistics is 0.79 (CI: 95% 0.77-0.83). Figure 1 show that the area under the ROC curve for the institutionalization score is 0.736 and the null hypothesis was rejected.

Comparison with the mortality score
The c statistics obtained using the GCS based on mortality, when applied in this validation cohort are presented in table 4. The c statistics is 0.75 (95% CI: 0.73-0.77), which statistically different than the 0.79 (95% CI: 0.78-0.83) obtained with the institutionalization score.

Sensitivity analysis
The c statistic obtained by the institutionalization score in the demented elderly cohort is 0.81 (95% CI: 0.78-0.84) and is not statistically different than the score obtained in the general elderly population (Figure 2). When tested in a non-demented elderly cohort but using a cohort study design, the c-statistic of the model including age and gender is 0.81 (95% CI: 0.79-0.83). The odds ratio of the score is 1.22 which represents an increasing of institutionalization of 22% by point in the score scale (Figures 3 and 4).

Discussion
Through this study, we have developed a comorbidity score based on institutionalization in the community-dwelling elderly population. It was shown that very few prescribed medications are significantly associated with institutionalization. Although institutionalization is a relevant outcome in this population, it remains fairly rare (only 2180 events or 2.5% of the study population were identified). Despite the relatively large sample size, some drugs were not associated with a statistically significant effect, even though they were clinically relevant. This is probably due to the relatively low frequency of use of these drugs in this population. This is the case for drugs used to treat urinary incontinence, for example. In the literature, it is well know that urinary incontinence is one of the major factors which lead to institutionalization [21]. However, our study was unable to confirm    Table 4: Mean score and mean risk for institutionalization and mortality and c-statistics. this in a large population of community-dwelling elderly, since few elderly receive prescribed treatment for this condition (3% of cases and 2.1% of controls).
Comorbidity scores to control for prescription channelling are often used in pharmacoepidemiology [24]. Iezzoni stated that the predictive performance of comorbidity scores depends on several factors such as the clinical conditions included in a score and their relative weights, the distribution of comorbidity in the source population, the study outcome, and the accuracy of the administrative data. Our study focused on the impact of the choice of the outcome [25].
The institutionalization score that we have developed yields adequate performances, comparable to those of other scores that predict mortality that are found in the literature. Comparability of performances between the institutionalization and mortality scores may be explained by the fact that, in the literature, it was shown that the most important factors that predict both institutionalization and mortality are age and gender [11,24]. In the elderly population alone,  this fact may be emphasizing by the interaction between age and gender.
Our results showed really good performances of the institutionalization score in the demented elderly population. This fact may be explained by the higher prevalence of both institutionalization and exposures. Also, since the demented population is older than general population, the high performances may be explained by the variances explained by age.
One of the strengths of our study is that the same method was used to develop scores of mortality and institutionalization. Our study is the first to demonstrate that a score based on institutionalization may be a reliable predictor of overall health status in the elderly population. The study population was very large (10.9% of the Quebec elderly) and very representative of the community-dwelling elderly population of Quebec. Also, it is the first study to derive a score using a nested case-control approach; all others used a cohort design. This design allowed us to ensure that drug exposure preceded the occurrence of institutionalization. Another major strength of the study was the fact that the initial predictive model was based on clinical relevance using a geriatrician's expertise, and was supported by a comprehensive literature review. Most other scores are data-driven only or used only clinician's opinions. A limitation of our study is the fact that we used a proxy to ascertain institutionalization. However, misclassification is probably minimal given that the observed rate in our study (2.5%) is similar to that of the community-dwelling Quebec elderly population (3%) [26]. Also, our score is based on prescription claims instead of diagnostic or procedure codes. However, knowing that this is a chronic disease score, one can believe that the majority of patients are treated. Also, this allowed controlling for the contribution of the treatment. It is the reason why some variables are protectors.
In conclusion, a comorbidity score to predict institutionalization that performs as well as a comorbidity score that predicts mortality was developed. Depending on the outcome in a original research, this tool may therefore be helpful to support pharmacoepidemiologic research conducted in this population. As example, if the outcome of interest is a clinical outcome like hospitalization, a score based on institutionalization may be more representative of the state of illness of patients.