Indicators of Preventable Drug-related Morbidity in Older Adults

OBJECTIVES: To determine the incidence of preventable drug-related morbidity (PDRM) in older adults in a provider-sponsored network and identify risk factors for PDRM. METHODS: The study was based on a retrospective review of an integrated health care database, using 52 newly developed clinical indicators of PDRM. The incidence of PDRM was determined by identifying individuals in the database who matched an outcome and pattern of care associated with an indicator. Risk factors were determined through a forward inclusion logistic regression model. The subjects in this study were 3,365 older adults enrolled in a hospital based health care system in Florida in 1997. The principal outcome measure was identification of individuals who matched a PDRM indicator and risk factors for PDRM. RESULTS: Ninety-seven enrollees who matched one or more of 52 PDRM indicators were found in 3,365 older adults, for an overall incidence rate of 28.8 per 1000. The top 5 indicators of PDRM were responsible for 46.8% of all PDRMs found. Regression analysis identified 5 risk factors: 4 or more recorded diagnoses, 4 or more prescribers, 6 or more prescription medications, antihypertensive drug use, and male gender. CONCLUSIONS: This study demonstrated that clinical indicators can be used in a managed care organization to identify seniors who have experienced a PDRM. The risk model should better prepare managed care organizations to proactively identify patients at risk for PDRM and to optimize medication use in older adults.


O R I G I N A L R E S E A R C H
nappropriate medication use and adverse drug-related outcomes in older adults are becoming increasingly serious problems for many managed care organizations (MCOs). Zhan and colleagues recently observed that 21.3% of seniors in the ambulatory setting in the United States are prescribed a potentially inappropriate medication. 1 Drug-related problems (like inappropriate medication use) in seniors are related to poorer health outcomes such as lower health-related quality of life. 2 Many of these adverse outcomes are preventable. For example, Bero and Lipton at the University of California followed 706 older adults discharged from a California hospital. 3 Within 6 months of discharge, 247 (35%) reentered the hospital. About one fifth (45) of the readmissions were drug-related. The majority (76%) of the problems identified were potentially preventable according to medical record audit. These preventable drug-related morbidites (PDRMs) can be caused by factors other than inappropriate prescribing. Underuse of technologies (including medications) is a significant contributor to adverse outcomes. 4 Others argue that lack of ongoing monitoring is a principal cause of PDRM. 5,6 Thus, there is a need for a comprehensive study that quantifies the degree of the problem of PDRM in older adults in the managed care setting. Moreover, such a study should examine these multiple causes of PDRM.
As with any epidemic disease, prevention is the most efficient and humane strategy. Reducing PDRM would significantly improve the safety and quality of medical care provided by MCOs, while at the same time reducing average per-patient costs. Risk factors may be used to help allocate scarce health care resources and as potential indicators to identify patients who need interventions. In Tamblyn' s review of the literature of geriatric pharmacotherapy, categories of risk factors for an adverse outcome were identified: the specific drug prescribed, qualities of the physician, qualities of the patient, and health care system and practice. 7 While risk factors are known for many adverse drug-related outcomes in seniors, such as adverse drug reactions, unfortunately, risk factors that are specific for PDRM in seniors in the managed care setting have not yet been identified. If such risk factors were known, they could be used by MCOs to optimally design interventions that would target high-risk patients.
Identifying the incidence of and risk factors for PDRM may also help MCOs to reduce health care resource utilization.
A recent article in this journal observed that seniors who were prescribed potentially inappropriate medications, when compared to those who were not prescribed such medications, had significantly higher utilization of health care resources (emergency room, inpatient, and outpatient visits) and costs (facility, provider, and overall). 8 Previous research has concluded that those older adults who received sedative-hypnotics with doses exceeding guidelines had increased hospital costs and longer lengths of stay as compared to those who did not receive these drugs or whose dosages did not exceed the guidelines. 9 Bates and colleagues determined that patients with a preventable adverse drug event had an average increase of 4.6 days in length of stay and $5,857 in total cost. 10 Therefore, it appears that there is a significant cost associated with adverse outcomes of drug therapy, and the patients who experience these outcomes consume more health care resources. Of interest to managed care administrators and health care professionals is the relationship of PDRM to health care resource utilization in the managed care setting. At least theoretically, these adverse outcomes, being preventable, can be greatly reduced.
This study used clinical indicators of PDRM, linking both suboptimal patterns of care and outcomes, in a health care database. This database contained administrative and clinical data for older adults who were enrolled in a hospital-based health care system. The objectives of this study were to (1) determine the incidence of PDRM in older adults in this provider-sponsored network and (2) identify risk factors for PDRM.

■■ Methods
The study was based on a retrospective review of the database of a hospital-based integrated health care system. Ethics approval for the study was obtained from the University of Florida Health Science Center Institutional Review Board.

Study Population
Our study population was drawn out of a larger pool of enrollees in a hospital-based health care plan in Florida with a Medicare contract. In order to be eligible for the health care plan, the enrollees had to live in one of 3 specific counties in Florida and had to be enrolled in Medicare Part B and continue to pay the Medicare Part B premium. Individuals who elected the Medicare hospice benefit and those with end-stage renal disease were not eligible for enrollment in the health plan. For this study, our inclusion criteria were: (1) individuals who were enrolled in the plan anytime during 1997 and (2) those enrollees who completed the Personal Wellness Profile (PWP) Senior Assessment (approximately 50% of the total plan members). Completion of this instrument was included in our inclusion criteria as the PWP Senior Assessment contained a majority of the variables that were included as possible risk factors for PDRM. The PWP Senior Assessment is an instrument that was given to all plan enrollees to complete upon enrollment. It is an instrument used to identify seniors at high risk for health-related problems. It has been previously used by other health plans and its predictive validity has been verified. [11][12][13]

Study Database
The data used in this study consisted of (1) all claims made in the outpatient and inpatient settings for this population that were already collected as a natural part of the administration of this health plan, (2) data on all prescriptions filled for the plan enrollees in the ambulatory setting provided by the pharmacy benefit manager (PBM), and (3) the PWP Senior Assessment instrument results. All claims processed and surveys completed between January 1 and December 31, 1997, were included in the study. A central database containing these 3 data sets was constructed by a medical artificial intelligence company. This database was completed on April 15, 1998, with approximately 95% of claims from 1997 processed at this time. A unique patient identifier was used to link these 3 data sets. All patient names were masked to protect patient confidentiality.

Incidence of PDRM
In the first phase of this study, 52 clinical indicators of PDRM in older adults were developed. This has been described previously in this Journal. 14 The database was searched for patients who had an ICD-9 diagnosis code corresponding to an outcome included in one of the 52 PDRM indicators. Then, for each patient with the outcome, the computerized record was reviewed to determine whether the patient had received care corresponding to the pattern of care associated with the outcome in the indicator. If both the outcome and pattern of care matched the specific indicator, then it was counted as an occurrence of PDRM.

Risk Factors for PDRM
The peer-reviewed medical literature on drug-related morbidity in older adults since 1967 was reviewed for possible risk factors for PDRM. Peer-reviewed medical and pharmacy articles and referenced texts were included in the literature review. Once a possible risk factor was identified, it was determined whether measurement of this risk factor was possible in our study database. Out of this process, 18 possible risk factors for PDRM were selected for inclusion in a forward inclusion logistic regression model to determine which of the 18 hypothesized risk factors were associated significantly with PDRM. The entry level was set at P=0.05. These 18 hypothesized risk factors and the data sources for their measurement are contained in Table 1.
Additional logistic regression models were then run, with the risk factors from the first model entered a priori to adjust for their effects on PDRM. Other demographic variables contained in the study database were allowed to enter the model to see if they added significantly to the prediction, based on statistical significance in a bivariate analysis between patients who did, and who did not, have PDRM. These additional demographic variables and the data sources for their measurement are contained in Table 2. Due to the large number of variables that were considered, they were dichotomized where possible. Because there was a theoretical basis for including the risk factors from the first model, it was felt that the final model for PDRM must include all of these risk factors, even if it explained less of the variance of PDRM. Thus, this process incorporated both statistical and theoretical criteria for deciding which terms to include in the model and this helped to focus attention on those variables that fit into the conceptual framework and that had the greatest independent effect on PDRM. For the risk factors in the final regression model, the parameter estimate, standard error, and chi-square probability were all calculated. The odds ratios and 95% confidence interval were also determined for each risk factor to provide an estimate of the relative risk of having PDRM given the presence of the risk factor. SAS (SAS Institute Inc., 1993) and JMP IN (Sall and Lehman, 1996) were used to create the regression models.

■■ Results Demographics
Enrollment into the health plan began in January 1997, with approximately 7,000 enrollees by December 1997. Approx-imately 50% of these enrollees completed the PWP Senior Assessment instrument, and 3,365 patients met our inclusion criteria for the study. Table 3 contains the demographic characteristics for the study population.

Incidence of PDRM
When the 52 clinical indicators of PDRM were applied to the study database, 1,005 patients were found who had outcomes consistent with one of the indicators. When each of the 1,005 patient records was searched for the related pattern of care, we found 158 events meeting an indicator. This represented 97 patients, as several patients met more than one clinical indicator. The overall incidence of PDRM was 28.8 per 1,000.

Risk Factors for PDRM
Through the use of forward inclusion logistic regression with PDRM as the dependent variable and the 18 hypothesized risk factors as independent variables, a 5-variable risk model was produced (Table 4). This model indicates that individuals with 4 or more recorded diagnoses were 2.93 times more likely to have PDRM than those with 3 or fewer diseases (P=0.0001; 95% CI, 1.81-4.76). Those with antihypertensive drug use are at a much greater risk (2.02 times) for having PDRM (P=0.0118; 95% CI, 1.16-3.52) as are seniors taking 6 or more prescription medications (1.92 times) (P=0.0266; 95% CI, 1.08-3.42). One other variable, 4 or more prescribers, also placed individuals at a greater risk (1.31 times) for developing a PDRM (P=0.0001; 95% CI, 1.16-1.47). While female gender was included as a hypothesized risk factor for PDRM, the odds ratio for female gender was less than one (0.52) (P=0.0056; 95% CI, 0.823-0.322), meaning that females were at a far lower risk of developing PDRM than males.
The amount of variance explained by the prediction model is quite good. While R 2 is not recommended for use in logistic regression, an analogue called RL 2 has been proposed. 15 RL is a measure of the proportional reduction in chi-square, and it varies between 0 and 1 (where 1=the model predicts the dependent variable perfectly). For this prediction model, the RL 2 is 0.562. The correlation matrix for the final 5 variables in the model suggests that it is free of multicollinearity. Overall, the correlations (pair-wise relationships) between the variables are quite low. Only 2 correlations were greater than ± 0.20: 4 or more prescribers and antihypertensive drug use (0.2995) and 4 or more recorded diagnoses and 6 or more prescription medications (0.2965). A backward elimination procedure was also performed, which confirmed the forward inclusion model results.
It was thought that there might be some additional demographic variables that are risk factors for PDRM. If a variable was significantly associated with PDRM in the bivariate analysis (P<0.05) then it was included in the regression model along with the 5 variables that were identified as risk factors in the original model. A forward inclusion procedure was again used with the entry level set at P=0.05. Further regression models did not greatly increase the amount of explained variance of PDRM so it was decided that the original regression model with the 5 risk factors was the optimal model for predicting PDRM.

■■ Discussion
A few specific clinical indicators were responsible for a large percentage of total PDRMs found. Overall, the top 7 indicators of PDRM were responsible for almost half of all PDRMs found (46.8%) ( Table 5). The most frequently occurring indicator (outcome=secondary myocardial infarction; pattern of care= history/diagnosis of myocardial infarction, no use of ASA and/or a beta-blocker) was found 24 times (15.2% of all PDRMs). The second most frequently occurring indicator (outcome= ERvisit/hospitalization use to hyperglycemia; pattern of care= use of an oral hypoglycemic agent, hemoglobin A1c level not done at least every 6 months) was found 18 times (11.4% of all PDRMs). As shown in Figure 1, 23 clinical indicators did not occur even once in the study population, while 14 occurred from 1 to 3 times. Figure 2 displays the number of PDRMs experienced by each patient in the study. It appears that just as a small proportion of patients and diseases are responsible for a large proportion of health care costs, a small proportion of indi-Demographic Characteristics of the Study Population (N=3,365) cators are responsible for most PDRMs. A majority (62.9%) of the patients who had PDRM matched only one indicator of PDRM. Almost 19% (18.6 percent) of patients with PDRM matched 3 or more indicators and one patient matched 5 indicators. This finding should not be too surprising. There is some evidence that patients who experience an adverse drug event are at higher risk for a second adverse drug event 16,17 therefore, the same could be true for PDRM. Some patients may suffer from general medical mismanagement. For example, if the patient has multiple prescribers who are not communicating their therapeutic plans to one another, the patient may be taking a dangerous combination of medications, placing that patient at greater risk for a PDRM.
The overall incidence of PDRM, 28.8 per 1,000 (2.9%), is really a lower-bound estimate of the incidence of PDRM since the 52 indicators used in this study do not represent all possible PDRMs that occur in older adults. Further investigations may focus on refining and/or developing additional PDRM indicators in an attempt to identify more patients with PDRM. This would help to increase the sensitivity of using these indicators together as a group to detect PDRM in a population of older adults.
The application of the prediction model for PDRM is beyond the scope of this study, but some general principles may still be stated. Boult and colleagues argue that the first step of any geriatric evaluation and management program is the identification of high risk. 18 Patients with multiple risk factors for PDRM could be identified by MCOs or individual physicians and then proactively managed to help prevent PDRM and allocate resources in the most efficient manner. For example, we know that physician-pharmacist-patient communication is a necessity of the medication use system, and, therefore, if a patient has 4 or more prescribers, attempts should be made to improve communication by reducing the number of prescribers or better coordinate therapy. The final PDRM model demonstrates that a wide variety of factors influence PDRM, not just drugs themselves or certain diseases. This supports the idea that the medication use system is influenced by numerous factors. In fact, only one of the 5 risk factors is a drug class (antihypertensive drug use). Other researchers in the future may investigate whether the risk factors identified are truly causal or predictive.

■■ Limitations
Some potential limitations pertain to the identification of risk factors for PDRM. Only risk factors associated with PDRM in older adults were considered. Risk factors may differ for other populations, and it may differ for PDRMs in older persons that were not investigated in this study. While the list of possible risk factors considered in this study was more thorough than any other study in the peer-reviewed medical literature, there could be additional risk factors that may contribute to the regression models. Some potential risk factors for PDRM, such as an abnormal potassium level, drug interactions, specific dosages of drugs, and a patient belief that the drugs were responsible for hospitalization, have been previously shown to be risk factors for adverse drug events but could not be tested in this study due to limitations of the study database. For example, although the study database listed all laboratory tests performed on any given patient, it did not contain the actual value of the test. Since administrative health claims data were used in this study, the limitations associated with the use of this type of data (not all clinical data present, misclassification bias in ICD-9-CM coding, etc.) apply to this study.
Other limitations pertain to the population used in this study. These clinical indicators of PDRM were applied to a population enrolled in one health care plan and who completed the PWP Senior Assessment tool. Since enrollment in the plan was optional, the study population may not be representative of the geriatric population in general. As well, those older adults who elected to complete the PWP Senior Assessment instrument may be different from those who did not complete the instrument. However, a previous study of Medicare beneficiaries concluded that older persons who participate in screening services (such as completing a risk assessment questionnaire) did not differ significantly in their health behaviors from older persons not participating in the preventive services. 19 A study that included 217 noninstitutionalized older persons in Sweden also concluded that those older persons who do, and do not, participate in health promotion activities do not differ in health status. 20 Older persons enrolled in a managed care Medicare-risk health plan may differ demographically and may have different health care resource utilization and outcomes than those older persons not in these plans. One study that compared older persons with joint or chest pain in traditional Medicare and Medicare-risk health plans found little significant difference, although the patients in the managed care environment had reduced utilization of services and poorer improvement of symptoms in one of 4 outcomes considered. 21 In contrast, another study that compared 10 HMOs with Medicare-risk contracts to 10 traditional fee-for-service plans found that enrollment in an HMO was not significantly associated with functional status or medical visits. 22 There is much controversy surrounding the matter of how much of the measurable effects of managed health care are due to potential enrollment bias. 23,24 Future research should replicate this present study, and these indicators will need to be extended to include other, nonsenior, patient populations, validated for positive and negative predictive values via chart review and updated as clinical practice and standards of care progress.

■■ Conclusion
Managed care pharmacists and administrators may be able to use the results of this study to reallocate resources in the most effective and efficient way possible. The identification of the significant risk factors for PDRM should support the development of a rational basis for planning and implementing interventions to reduce drug-related morbidity and mortality in older adults. Furthermore, the patterns of care included in the PDRM indicators may be useful as a clinical tool. They could be used to identify patients who should receive prompt, preventive, clinical follow-up. Future studies may be able to use this knowledge to perform pharmaceutical care interventions that will have the potential to deliver the greatest good and test the results of these "targeting" pharmaceutical care interventions.

FIGURE 1
The Number of PDRMs Experienced by Each Patient (N=3,365)