Identifying Patterns of Potentially Preventable Emergency. Department Utilization by American Children. Kathleen M. Alber.

Transcription

1 Preventable ED Utilization by Children Identifying Patterns of Potentially Preventable Emergency Department Utilization by American Children Kathleen M. Alber A Thesis Submitted in Partial Fulfillment of the Requirements for the Degree of Master of Science in Data Mining Department of Mathematical Sciences Central Connecticut State University New Britain, Connecticut January 2007 Thesis Advisor Dr. Daniel T. Larose Department of Mathematical Sciences

2 Preventable ED Utilization by Children 2 Identifying Patterns of Potentially Preventable Emergency Department Utilization by American Children Kathleen M. Alber An Abstract of a Thesis Submitted in Partial Fulfillment of the Requirements for the Degree of Master of Science in Data Mining Department of Mathematical Sciences Central Connecticut State University New Britain, Connecticut January 2007 Thesis Advisor Dr. Daniel T. Larose Department of Mathematical Sciences Key Words: Emergency Department Utilization, Children, National Survey of Children s Health, Medical Home

3 Preventable ED Utilization by Children 3 ABSTRACT Emergency department care and primary care are ideally distinct parts of the health care delivery system. In theory, each answers a specific and different health care need. However, in practice this distinction blurs. Many visits to hospital emergency departments are potentially preventable by timely and appropriate primary care. This paper employs a descriptive data mining approach to the analysis of data collected by the National Survey of Children s Health 2003 with the purpose of identifying global and local patterns of potentially preventable emergency department utilization by American children. Understanding the factors influencing the decision to seek emergency department care is an essential step in ensuring adequate and appropriate health care for all children. Using SPSS Clementine data mining software, classification models are employed to discover nationally significant and geographically specific patterns of potentially preventable emergency department utilization by children. This paper identifies numerous patterns. Globally, requirement for care, age, and insurance type were found to be the most significant predictors of the target behavior. Local patterns characterize several subsets of the population who are significantly associated with potentially preventable emergency department utilization. These characterizations often logically suggest theories which explain a group s association with such behavior. In many cases, further directed research would be required to confirm and clarify these assumptions. These are noted throughout this paper. Access to quality primary care logically influences the choice of emergency department care to treat a potentially preventable condition. The American Academy

4 Preventable ED Utilization by Children 4 of Pediatrics has developed the concept of a medical home and detailed the specific requirements for care defining its existence. This study describes a method of utilizing survey responses to measure each child s association with a medical home. Association rule mining is used to characterize groups of children according to association with a medical home and investigate potentially preventable emergency department utilization in the context of access to primary care. Differences in characteristic patterns of health care utilization for these groups are evident, suggesting that efforts to encourage the appropriate use of primary care in place the emergency care would most effectively be tailored according to level of access to primary care.

5 Preventable ED Utilization by Children 5 TABLE OF CONTENTS ABSTRACT... 3 INTRODUCTION... 6 RELATED RESEARCH... 9 METHODS Data Preprocessing Exploratory Data Analysis Target Variable PPA_Visits Predictive Attributes Non-Predictive Attributes Data Preparation for Modeling Clustering Classification Modeling National Model Insurance Type Model State Models Association Rule Mining RESULTS DISCUSSION REFERENCES Appendix A: Modeling Attributes Appendix B: Percent of Children Associated with PPA Visits by State Appendix C: Sources and Derivation of State Characteristic Attributes

6 Preventable ED Utilization by Children 6 INTRODUCTION Each year, approximately 30 million visits to American hospital emergency departments are made by children aged 18 years or younger (Institute of Medicine [IOM], 2006; McCraig & Burt, 2005). Many of these visits are avoidable, addressing conditions which would more appropriately be treated in a primary care setting. Such visits contribute to overcrowding of emergency departments and are indicative of the use of emergency department services as a proxy for primary care. It is generally recognized that health care delivery is optimized on both an individual and system-wide level when children receive appropriate and consistent primary care directed by a primary care professional and utilization of hospital emergency departments is limited to unpreventable medical emergencies. The positive effect of comprehensive pediatric primary care on children s health has been documented (Alpert, Robertson, Kosa, Heagarty, & Haggerty, 1976). Optimally, the primary care professional has a relationship with the patient and therefore is familiar with the child s medical history and family issues and provides not only preventative and illness care but also follow-up attention and focused advice as necessary. The parent who is comfortable in his/her relationship with the child s primary care professional is more likely to seek advice and care for problems before they escalate to a more serious and urgent status. Emergency department care is focused on addressing an immediate concern rather than providing ongoing care. It is a more expensive and often less medically appropriate alternative to primary care for children. A recent Institute of Medicine report (IOM, 2006) noted that many hospital emergency departments are not well equipped to handle pediatric patients. Many

7 Preventable ED Utilization by Children 7 don t have the essential supplies on hand for these smallest patients and the majority lack medical personnel with specialized pediatric training. Several previous studies have sought to discern factors related to the choice of emergency rather than primary care for non-urgent conditions (Cunningham, Clancy, Cohen, & Wilets, 1995; Doobinin, Heidt-Davis, Gross, & Isaacman, 2003; Phelps et al., 2000; Walls, Rhodes, & Kennedy, 2002). Use of emergency department resources for the treatment of non-urgent cases is one indicator of a breakdown of pediatric primary care delivery. Another, with more potentially serious medical consequences, is exemplified by the ill child who is not provided with appropriate, timely primary care treatment for a non-urgent condition thus resulting in a deterioration of health status which requires emergency care. To illustrate, Dombkowski, Stanley, and Clark (2004) found that asthmatic children who were regularly followed and appropriately medicated were less likely to require emergency intervention. Another study (Johnson & Rimsza, 2004) identified several emergent conditions, including epileptic convulsions, severe ear, nose, and throat infections, and bacterial pneumonia, with which children who received primary care were four times less likely to be associated. Putting aside the beneficial effect of the primary care professional as a health/safety educator, accidents, injuries and poisonings are validly considered not preventable by primary care professional intervention. For the purpose of this study, all others are considered at least potentially preventable. This includes most nonemergent conditions and emergent conditions which may not have escalated to an urgent nature had proper earlier intervention been sought in a primary care setting.

8 Preventable ED Utilization by Children 8 This study identifies factors associated with children likely to be brought to an emergency department for potentially preventable conditions. It employs a data mining approach to the analysis of data collected for the National Survey of Children s Health 2003 [NSCH] which includes data pertaining to 102,353 children less than 18 years of age. Residents of each of the fifty states and Washington, DC are represented. Survey responses were provided by a parent or caregiver and describe the referenced child s demographic and economic conditions as well as profiling health status and access and utilization of health care resources. It is noted that factors which affect children s emergency department utilization patterns may differ from those which contribute to adult behavior in this regard. Usually, the decision about whether or where to seek care for a child is made not by the child himself, but by a parent or other caregiver and the factors influencing that decision are often more complicated than those required for an adult determining his or her own personal medical requirements. Therefore, when examining potential contributing factors, caregiver and family characteristics which might influence the decision-making process are considered as well. As the name implies, the NSCH data is national in scope allowing a more comprehensive study of utilization patterns than most previous research which analyzed data collected on a local or statewide level. This thesis initially focuses on national patterns of potentially preventable utilization of emergency department resources by children and then investigates geographic variation through the comparison of factors associated with these patterns on the state level.

9 Preventable ED Utilization by Children 9 RELATED RESEARCH There have been numerous previous studies which investigated aspects of emergency department utilization by children. For the most part, the data analyzed by these samples represent a limited population, often a single hospital and/or exclusively publicly insured patients. Several studies have sought to identify factors associated with non-urgent use of hospital emergency departments by children. Phelps et al. (2000) investigated the relationship between non-urgent visits and caretaker characteristics including stated reasons for bringing the child to the emergency department. This study suggested the importance of types of family unit and insurance as well as the modeling effect of where the caregiver was taken as a child for ill care. Newcomb (2005) focused on variables measuring Medicaid children s access to primary care as well as caretaker characteristics. The author concluded that non-urgent visits to the emergency department by a localized group of publicly insured children were at least partly due to lack of ready access to primary care. Doobinin et al. (2003) surveyed the parents of children brought to a single urban pediatric emergency department with non-urgent illnesses to determine their reasons for choosing the emergency department for care. Considered were issues of convenience as well as parental discrimination of urgency and knowledge of insurance requirements. Another research approach focuses on the identification of factors associated with children who visit an emergency department for any reason and/or the effects of the implementation of certain policies on emergency department utilization patterns. These studies classify according to whether or not a child uses the

10 Preventable ED Utilization by Children 10 emergency department, not the nature or urgency of the complaint. Much of the research found had a narrowly defined purpose. For example, Pollack et al. (2004) sought to discover sources of variation in emergency department utilization by Medicaid-insured Michigan children with chronic or complex health conditions while Fredrickson, Molgaard, Dismuke, Schukman, & Walling (2004) studied Medicaidinsured asthmatic children in rural Kansas with similar goals. Kotagal et al. (2002) explored the relationship between primary care and emergency department utilization in early infancy by studying Medicaid-insured newborns in Ohio. Likewise, the potential correlation between continuity/consistency of primary care and emergency department utilization has been the focus of studies (Christakis, Wright, Koepsell, Emerson, & Connell, 1999; Gill, Mainous, & Nsereko, 2000; Ryan, Riley, Kang, & Starfield, 2001). While others took a dichotomous approach to classification of emergency department visits, Weinick, Billings, and Burstin (2002) introduced a four class model which differentiated not only between urgent and non-urgent visits, but further between emergent conditions which could have been treated or prevented in a primary care setting and those which required emergency department care and were not preventable. Much research is limited to or focused on children who are insured under Medicaid for a variety of reasons including the established tendency of Medicaid insured children to visit emergency departments with greater frequency than children of other insurance classifications (Dombkowski et al., 2004; McCraig & Burt, 2005; Phelps et al., 2000). This greater utilization may be due in part to some Medicaid

11 Preventable ED Utilization by Children 11 delivery systems which do not financially penalize patients who make use of the more expensive emergency department alternative for non-urgent illnesses (Phelps et al., 2000). Tailoring studies to Medicaid enrollees also has the advantage of practical applicability of results. When research demonstrates problem areas, public policies can be implemented to encourage and facilitate more appropriate utilization of medical resources. Managed care Medicaid programs are an example of such policies and determination of their success in this regard has been the focus of several studies (Alessandrini, Shaw, Bilker, Perry, Baker, & Schwarz, 2001; Dombkowski et al., 2004; Piehl, Clemens, & Jones, 2000). Research using Medicaid data is perhaps also prevalent because state Medicaid data bases provide a rich, convenient, relatively available source of data to be analyzed. As was noted above, most of the research obtained for this review had a narrow geographic focus. One exception was a study by Luo, Liu, Frush, & Hey (2003) who investigated whether type of insurance coverage affects the likelihood that a child visits the emergency department. This study utilized the 1997 Medical Expenditure Panel Survey, a national survey. The authors studied data pertaining to 10,193 children from across the country and included potential predictor variables describing basic demographic information and type of insurance. They achieved some different results than many of the more limited studies indicating that a more diverse dataset may provide different insight. Ryan et al. (2001) also pointed out the importance of diversifying study populations. Noting that data for most previous studies originated in urban hospital/clinic settings, they focused their study on rural adolescents.

12 Preventable ED Utilization by Children 12 Certain correlations with various forms of emergency department utilization were noted by multiple studies. The first, mentioned above, is between insurance type and emergency department utilization by children. The heavier utilization of emergency department services by Medicaid insured children was noted by most studies. Another is the child s age. Infants and adolescents were observed to be more likely visitors to hospital emergency departments than children of the ages in between these two groups (Johnson & Rimsza, 2004; Pollack et al., 2004). A care-giver s marital status has also been found to be predictive of the tendency to bring a child to the emergency department for care (Kotogel et al., 2002; Phelps et al., 2000). Additionally, the mother s level of education has been associated with the use of emergency department services (Alessandrini et al., 2001; Ryan et al.,2001). These correlations were considered in designing this study and interpreting results.

13 Preventable ED Utilization by Children 13 METHODS This study utilizes SPSS Clementine 10.0 data mining software to analyze the NSCH 2003 data. The primary data mining task is descriptive in nature with the goal of identifying patterns and trends of preventable emergency department utilization by children. Bivariate analysis was employed to explore the relationship between potentially pertinent survey response attributes and the target behavior. Additionally, Clementine decision tree and association rule modeling tools were utilized to allow the identification and clarification of multivariate patterns. Data Preprocessing The NCHS data set includes 301 variables describing diverse aspects of children s health. Many of these are unrelated to hospital emergency department utilization for potentially preventable conditions and therefore were disregarded by this study. Retained for consideration were variables describing (a) demographic characteristics including age, gender, family structure, primary language, race, and caregiver education; (b) health status; (c) insurance type; (d) family income; (e) health care access and utilization; (f) family risk behaviors; and (g) geographical home. Assessing the quality and completeness of the data is a crucial step in the successful modeling process. Missing attribute values in the NSCH data are prevalent for several reasons. The design of the survey is such that many questions are asked only when responses to other questions indicate the appropriateness. For example, if a respondent indicates that the child has not received any medical

14 Preventable ED Utilization by Children 14 attention in the referenced year, further questions regarding frequency or type of medical care received are omitted. Consequently, records associated with such children are missing values for these attributes. Two sections of the survey are agespecific. Each respondent answers only the questions targeted to the referenced child s age group resulting in missing values for attributes associated with questions of the section applying to the other age group. Additionally some missing values are the result of the respondent s inability or refusal to answer a question or errors in capturing or storing the response. Missing values were handled as follows: 1. Where there was a logical means of discerning the appropriate replacement for the missing data, that replacement was made. In the example above, a response indicating that a child received no medical attention in the referenced year clearly implies zero values for attributes reflecting the number of visits for preventative primary care, sick care, and emergency department care. 2. Where the correct replacement was less obvious but a reasonable approximation could be determined, that replacement was made. For 57% of the attributes, the majority class was assigned. Others required more involved deductive processes which are detailed in the exploratory data analysis section which follows. 3. Where an appropriately logical replacement value could not be discerned or approximated, the record was removed from consideration. This resulted in the removal of a very small subset (1.3%) of records which were missing values for attributes such as gender or education level.

15 Preventable ED Utilization by Children 15 Each variable in the data set is identified by a question code. For ease of modeling, attributes were renamed so as to be understandable. Additional attributes were derived using the information conveyed in NSCH responses. Appendix A lists the 37 attributes utilized in modeling and references the corresponding data set question code(s). The NCHS survey implemented a design of top-coded variables which effectively suppress outliers. Exploratory Data Analysis A significant challenge in the application of data mining to the NSCH data set is the derivation of an effectual attribute set. NSCH attributes are answers to specific survey questions. Some can be individually considered as potential predictors, but most attributes are derived using a selection of responses. Target Variable PPA_Visits The target behavior is the use of a hospital emergency department for potentially preventable conditions. The NSCH data includes three variables which were used to derive a single attribute, PPA_visits, to classify each record according to whether the referenced child made at least one such visit. The first refers to the total number of emergency department visits. If that number is one, a second attribute indicates whether the visit was in response to an accident injury or poisoning. If more than one, a third variable holds the number of visits in response to accident injury or poisoning.

16 Preventable ED Utilization by Children 16 PPA_visits is set as false if survey responses indicate that the referenced child made no visits to an emergency department or that all such visits were due to accidents, injuries, or poisonings. The target variable is true if the referenced child made one or more visits to an emergency department which were not due to accident, injury, or poisoning. Using these criteria, 11.62%, or 11,890 of the NSCH respondents are identified as having visited the emergency department for potentially preventable reasons. Predictive Attributes Age Analysis of the NSCH data indicates that patterns of potentially preventable hospital emergency department utilization vary by age, with younger children generally making more such visits than those who are older. Survey respondents specified the age of the referenced child in years. For the purposes of this study, ages were binned into groups: (a) infants of ages 0 or 1 years, (b) preschoolers between the ages of 2 and 5 years, (c) young school age children between the ages of 6 and 9 years, (d) middle school aged children between the ages of 10 and 12 years, and (e) adolescents between the ages of 13 and 17 years. Among the children referenced by the survey, infants were most likely to be linked with a potentially preventable emergency department visit. Close to one-quarter of infants were associated with the target behavior in the studied year. As age increases, the tendency to make such a visit decreases. Only 7.37% of the adolescents referenced in the survey indicate a trip to the emergency department for a problem not related to an accident, injury or poisoning.

17 Preventable ED Utilization by Children 17 Table 1 Percent of Children Associated with a Potentially Preventable/Avoidable Emergency Department Visit by Age Group Age Group Percent Associated with PPA Visit Caregiver Structure The decision to seek emergency care for a child is usually made by that child s caregiver(s). It has been suggested that the number and relationship of caregivers residing with the child influences that decision. The NSCH provides two variables which can be used to ascertain the caregiver structure of the child s family unit. The RELATION field refers to the relationship of the person providing the information for the survey. By survey design, this respondent is the parent or guardian who lives in the household and knows the most about the health and health care of the child. The TOTADULT3 field specifies the number of adults living in the household. Using these two variables, a new attribute is derived which indicates whether the child resides with (a) a single mother, (b) a single father, (c) a single other caregiver, (d) two adults, or (e) three or more adults. Phelps et al. (2000) noted that single caregivers were more likely bring a child for a non-urgent emergency department visit and speculated that this may be due to lack of input from another adult in the evaluation of the necessity of such care. Children of single mothers represented in the NSCH data are indeed more likely to be associated with a potentially preventable visit to a hospital emergency department than those living in households with two or more adults. This is also true of children

18 Preventable ED Utilization by Children 18 with a single caregiver who is neither mother nor father. However this data indicates that single fathers are the least likely caregivers to seek preventable emergency care for their children. This suggests a gender difference in a caregiver s inclination to seek emergency care for a child. Caregiver Education The NSCH data includes indication of the highest education level attained by any member of household. For the studied children, a higher level of household member education is associated with a lesser likelihood of a visit to an emergency department for potentially preventable reasons. The difference is most notable for those who have continued their education past high school graduation. Links exist between education level and income and insurance type, suggesting the possibility of education level functioning as a covariant of these factors. However data investigation showed that education level is indeed an independent factor in the decision to seek potentially preventable emergency department care for a child, particularly in middle income and/or privately insured families. Primary Language/Interpreter Of those interviewed in this nationwide survey, 7.56% identified some language other than English as their primary home language. These households are more likely to be associated with a child s visit to an emergency department for potentially preventable conditions. An additional survey question was posed of respondents who indicated a non-english primary language to ascertain whether an interpreter would be required for effective communication between a patient/caregiver and medical personnel. By segmenting the group of records

19 Preventable ED Utilization by Children 19 indicating a non-english primary language according to the need for an interpreter, a clear association between this requirement and potentially preventable emergency department utilization was established. In fact, children from families with primary language other than English who do not require an interpreter to effectively communicate with medical personnel are approximately equally likely to visit the hospital emergency department with potentially preventable problems as children who come from homes where English is the primary language. In contrast, those who require an interpreter are almost twice as likely to make such as visit as their peers with greater facility for the English language. Race/Ethnicity The NSCH identifies four racial designations: white, black, multiple race or other. A new classification, unknown, is created to apply to children of respondents who do not know or refuse to provide a racial designation. A separate survey answer indicates whether the child is of Hispanic or Latino origin. Data exploration led to the discovery that most of the children classified as of unknown race are Hispanic. This suggests that the absence of Hispanic designation as one of the survey choices for race left respondents uncertain about the appropriate response to describe the race of the referenced child. Consequently, a new variable was derived combining the two survey responses in order to allow Hispanic as a possible primary race/ethnicity. This new variable includes in the multiple race category those children who were identified as black or white race and also of Hispanic origin. All other records indicating Hispanic origin are categorized as Hispanic.

20 Preventable ED Utilization by Children 20 Distribution of data by race with regard to the target variable indicates that black children are most likely to be brought to a hospital emergency department with potentially preventable conditions. Hispanic and multiple race children are slightly less likely than black children to make such visits, but more likely than white children. Insurance Type Multiple survey responses provide information as to the existence and type of health care insurance coverage for the referenced child. A single attribute was derived combining this information to identify each referenced child as covered by Medicaid, private, or no insurance. Examining this variable with respect to associated potentially preventable emergency department visits indicates a significantly greater tendency for such visits by Medicaid insured patients. Table 2 Rate of Potentially Preventable Emergency Department Utilization by Insurance Type Insurance Type % Associated with a PPA Visit Medicaid 19.65% Private 9.23% Uninsured 9.48% This strong association between Medicaid insurance and potentially preventable emergency department utilization is consistent across all three major racial/ethnic categories. Concerning other insurance classifications, it is noted that uninsured Hispanic and white children are slightly less likely to be brought to an emergency department with preventable conditions than those who are privately insured. Black children without health insurance on the other hand are even more

21 Preventable ED Utilization by Children 21 likely to make a potentially preventable visit to an emergency department than their peers with private insurance. Table 3 Percent of Children Associated with a Potentially Preventable/Avoidable Emergency Department Visit by Race/Ethnicity and Insurance Type RACE/ETHNICITY NONE MEDICAID PRIVATE Black Hispanic White Family Income The NSCH used Department of Health and Human Services guidelines to derive the household poverty status for each respondent based upon household income and number of people residing in the household. Each record was assigned to one of eight poverty status categories. A new variable was derived for this study collapsing this information into three income categories: (a) lower less than 150% of poverty level, (b) middle - 150% to 300% of poverty level, or (c) upper - over 300% of poverty level. Approximately nine percent of the records have no poverty status designation because the respondent did not know or refused to provide family income information. Prior to data collection, survey designers assumed that these responses would be most associated with actual family incomes in excess of 300% of the poverty level and subsequent income related questions were asked based upon that assumption. However, as evidenced by Figure 1, the distribution of these nonassigned ($null$) records with regard to insurance status suggests that they are more associated with the middle income group. The percentage associated with Medicaid

22 Preventable ED Utilization by Children 22 insurance is particularly telling. Medicaid is allocated primarily on the basis of income. While income requirements vary by state, it is logically expected that there would be few Medicaid recipients in the upper income category. Additionally, the lack of medical insurance is atypical of upper income families and yet a significant minority of the unassigned records indicates no coverage. Figure 1. Association of insurance type with income group. While it can be assumed that the respondents who did not know or refused to specify household income are not exclusively representative of a single income group, it appears that they are most representative of the middle income group and therefore, these records were assigned as such for the purposes of this study. There is a strong association between family income and potentially preventable emergency department visits. The likelihood of such visits increases as income level decreases. However, given the aforementioned relationship between family income and Medicaid coverage, this study considered the possibility of this association as a covariant effect of insurance type. Records were segmented by insurance type. Within each insurance type, preventable emergency department utilization was most strongly associated with the lower income group and that association weakened as income increased. In fact, this pattern was most pronounced among the privately insured which comprise close to 70% of the surveyed population.

23 Preventable ED Utilization by Children 23 Thus it was concluded that both income level and insurance type separately influence the likelihood of potentially preventable emergency department visits. General Health The NSCH data includes an attribute which reflects the survey respondent s description of the referenced child s general health as (a) excellent, (b) very good, (c) good, (d) fair, or (e) poor. Fortunately, a large majority of children, 87.11%, enjoy excellent or very good health. However, general health status is clearly linked with the propensity for potentially preventable emergency department visits. Children associated with less well conditions are more strongly linked with these visits. Figure 2. Percent of children associated with a potentially preventable/avoidable emergency department visit by general health specification. Primary Care Variables As this study seeks to identify factors associated with hospital emergency department visits which might have been prevented by appropriate and timely

24 Preventable ED Utilization by Children 24 utilization of primary care resources, variables which measure necessity, access, and quality of primary care are of great interest. Medical home. It is logically assumed that access to quality primary care would be a deterrent to potentially preventable emergency department utilization. The concept of a medical home is advanced by the American Academy of Pediatrics [AAP] (Children s Health Topics: Medical Home, n.d.). Historically, the existence of a medical home was defined solely by association with a primary care physician. The NSCH data suggests that such an association does not have a significant effect on the likelihood of a potentially preventable visit to a hospital emergency department. However, the AAP has expanded the definition of the concept of a medical home to be primary care which is accessible, continuous, comprehensive, family-centered, coordinated, compassionate, and culturally effective. Under this definition, the existence of a medical home requires not just the presence of a doctor, but the availability of high quality primary care. By utilizing information obtained in response to NSCH survey questions, a consideration of medical home which is more in keeping with this expanded definition can be attained and considered as potentially predictive of the target behavior. Bethell, Read, and Brockwood (2004) discussed using population-based national data sets to ascertain the degree of association with a medical home for children with special health care needs. They described the necessity and challenge of the development of a quantitative measure of medical homeness and discussed how survey responses could be used to make at least an approximate measurement of medical home for children referenced in the surveys.

25 Preventable ED Utilization by Children 25 Their work led to the inclusion of several questions in the NSCH survey pertaining to the discernment of a medical home. Building upon this work, this study utilized responses to these and other survey questions to develop a scoring method to provide an indication of each referenced child s association with a medical home. The AAP definition identifies seven dimensions to be considered when determining the existence of a medical home for a child. The NSCH survey instrument was analyzed and questions identified which provide some measure of six of these seven dimensions. It was determined that no measure of family-centeredness was available from this data. This process is detailed in Table 4. Obviously, these are not comprehensive or ideal measures of each dimension. For example, a primary care practice which employs bilingual doctors and nurses would be arguably more culturally competent than one which calls in an interpreter as needed. However, this exercise in determining the existence of a medical home is limited by the information available in the survey data which provides no indication of the linguistic abilities of medical practice personnel. The goal is to achieve the best approximation of a child s association with a medical home that is possible given the constraints of the questions asked.

26 Preventable ED Utilization by Children 26 Table 4 NSCH Questions Addressing the Seven Dimensions of Medical Home Dimension Question Description accessible S5Q06A How often can help be obtained from personal MD on phone? S5Q07A How often is immediate care available from personal MD? family-centered none continuous S5Q01 Does child have a personal doctor or nurse? comprehensive PREV_CARE During the previous year, did the child make (derived from S5Q08A, S5Q08B, S4Q03R) the appropriate number of preventative care visits to their personal doctor according the schedule recommended by the American Academy of Pediatrics for children of their age? S4Q07 During the previous year, did the child receive all the medical care necessary? S5Q02 How often does personal MD spend enough time with child? S5Q04 How often does personal MD explain things in an understandable way? S6Q28 Did personal MD ask about learning, development or behavioral concerns? S6Q29 IF LDB concerns, did personal MD give information to address those concerns? coordinated S5Q09B If specialist was needed, did personal MD help get specialist care? S5Q10B If special services, equipment or other health care was needed, did personal MD help get it? compassionate S5Q09C If specialist was needed, how often did personal MD talk to patient/caregiver about what would happen at specialist visit? S5Q10C If special services, equipment or health care was needed, how often did personal MD talk to culturally competent S5Q13A patient/caregiver about it? If an interpreter was required, how often was one available to assist with personal MD visits? A medical home score variable was derived to represent each child s relationship with a medical home on a scale of The existence of one or more medical professionals identified by the caregiver as the referenced child s personal

27 Preventable ED Utilization by Children 27 doctor or nurse is considered a necessary, though not sufficient, measure of medical home. Therefore records indicating the lack of association with a personal doctor or nurse were assigned a medical home score of zero. Not all of the fourteen measures of medical home described above are collected from each survey respondent. However, there are five core measures which are collected concerning every child who has identified a personal doctor or nurse and has sought medical care of any kind in the survey reference period. The other nine variables contain information collected only if the child required certain health care in the referenced period. For example, Question S5Q06A is asked only if the caregiver indicates that phone assistance was required and sought and Question S5Q13A is asked if the primary household language is other than English and the caregiver indicated that an interpreter was required for adequate communication with the medical professional. Therefore, in order to assure accurate and comparable medical home scores both for healthy children who used few non-preventative medical resources in the referenced period as well as those for whom more resources were required, the medical home score was designed to include two components. The first is a base score which incorporates the more universal measures of continuous, comprehensive medical care. Each of these is assigned a value and the values are summed to attain a base measure of medical home. For children over five years of age who needed and received only preventative care during the referenced year and required no interpreter, the medical home score is determined solely on the basis of these five variables. Those who have received their recommended preventative care from a health care professional who always devotes

28 Preventable ED Utilization by Children 28 ample time and provides understandable guidance receive a perfect medical home score of 100. Table 5 Measures of Care Contributing to Base Medical Home Score Measure Description Value Personal_MD Child has a personal doctor or nurse. 40 Prev_Care Child received AAP recommended # of 15 preventative visits, Got_Care Child received all necessary medical care during 15 MD_Time MD_Explain the past year. How often did doctor spend enough time with the child? How often did doctor explain things in an understandable way? always 15 usually 10 sometimes 5 never - 0 always 15 usually 10 sometimes 5 never - 0 The second component adjusts the medical home score to reflect inadequate provision of primary care services addressing individual health care needs. For example, comprehensive medical care for children under five requires that the personal doctor provide guidance about learning, developmental, and behavioral concerns. If he/she neglects to provide this, the medical home score is decremented. Likewise, if the child (or caregiver in support of the child) required assistance by phone, immediate care for illness or injury, referral to and help with a specialist or special medical equipment or care and any of these needs were not met by their personal medical practice or clinic, the value of the medical home score is decreased. Each of these measures of failure to provide a medical home is assigned a value and the total medical home score is decremented by that value if response to the pertinent survey

29 Preventable ED Utilization by Children 29 question indicates the appropriateness of this action. If the response indicates that the need was adequately met or if the question did not apply to the referenced child, the medical home score is not adjusted. The medical home score then is defined as the total of the base measure of comprehensive care less adjustments. Table 6 Adjustments to Base Measure of Medical Home Measure Description Decrement Value MD_Phone_Help Got_Immed_MD How often did doctor provide help by phone when needed? How often did doctor provide immediate care when needed for illness or injury? LDB_MD_Conc Doctor did not ask about learning, development or behavioral concerns for child under 5. LDB_MD_Info Doctor did not provide information to address stated concerns about learning, development or behavior for child under 5. Spec_PMD_Access Doctor did not help patient get specialist care when needed. Spec_PMD_Explain How often did doctor talk to patient/caregiver about what would take place during specialist visit (if specialist required)? SCare_PMD_Access Doctor did not help patient get special equipment or care when needed. SCare_PMD_Explain How often did doctor explain about needed special equipment or care? Got_Interp How often was an interpreter provided if one was needed to adequately communicate with doctor? never 15 sometimes 10 usually 5 always- 0 never 15 sometimes 10 usually 5 always never -5 sometimes 5 didn t go - 5 otherwise never 5 sometimes 5 didn t get 5 otherwise - 0 never -5 sometimes 5 otherwise - 0

30 Preventable ED Utilization by Children 30 The medical home score provides a measure of the strength of the association of the referenced child with a medical home. Fortunately, for those who have a personal doctor or nurse and therefore a non-zero medical home score, higher scores are more prevalent than lower. Figure 3. Distribution of survey respondents according to medical home score. It was hypothesized that children with stronger associations with quality medical homes would be less likely to utilize the emergency department for potentially preventable conditions. In general, this hypothesis was not supported by the data. However, for the very small segment of the survey population comprised of children associated with a personal doctor or nurse and medical home scores less than 50, there is a link between lower medical home score and greater likelihood that the child has made at least one potentially preventable visit to the emergency department. This is illustrated by Figure 4. Only approximately 1% of the children referenced by the survey belong to this group. These children identify a personal health care provider

31 Preventable ED Utilization by Children 31 but apparently receive very few primary care services from him/her. In fact, over 98% of these children do not receive the recommended preventative care. They also receive considerably less sick care outside of a hospital. Figure 4. Percent of children associated with a potentially preventable/avoidable emergency department visit by medical home score. While the medical home score provides a measure of the child s access to quality primary care, for the vast majority of survey respondents it is not a particularly strong indicator of potentially preventable emergency department utilization. A more predictive relationship exists between the base medical home score, the first component of the medical home score calculation, and the target behavior. Recall that this base score measures the referenced child s consistent access to a personal health care professional who offers recommended preventative care and provides adequate time and explanations as necessary. It is also dependent on whether the child receives all necessary medical care, but does not reflect the adjustments indicating the degree to which special health care needs are met. By definition, a base medical home score is either zero, if the child has no personal doctor or nurse, or

32 Preventable ED Utilization by Children if there is an associated personal medical professional. Figure 5, the normalized histogram of base medical home scores, shows the proportionate association with potentially preventable emergency department visits. Figure 5. Percent of children associated with a potentially preventable/avoidable emergency department visit by base medical home score As the base medical home is more predictive of the target behavior, it is adopted as the measure of access to quality primary care for the purpose of this study. Children associated with base medical home scores of 75 or more are considered as having a medical home. Note the difference in behavior with regard to potentially preventable emergency visits for those children who claim to have a personal doctor but based upon their base medical home score do not have an established medical home. This 8.65% of the population are far less likely to visit the emergency department with such problems than those with a medical home. They are also far less likely to make such a visit than children who have no personal doctor. A medical home variable was derived which indicates a child s level of access to adequate basic primary care. Each record is categorized according to association

33 Preventable ED Utilization by Children 33 with (a) a personal medical professional who provides a medical home, (b) a personal medical professional but no medical home, or (c) no personal medical professional. Attributes measuring the specialized care features which were reflected as deductions in the more comprehensive medical home calculation are considered as separate potential predictors, with the exception of those relating to learning, developmental and behavioral concerns which were found to be unrelated to the target behavior. Specialized care features. Figures 6 and 7 demonstrate how improved access to medical advice by phone or immediate primary care for illness or injury decreases the likelihood of a potentially preventable emergency department visit. Figure 6. Association of access to medical phone advice with potentially preventable/avoidable emergency department use. Figure 7. Association of access to immediate care for illness or injury with potentially preventable/avoidable emergency department use.

34 Preventable ED Utilization by Children 34 On the other hand, patients with access to a primary care provider who consistently provides explanations as to the necessity for and process of obtaining required specialized equipment are more likely to make a potentially preventable visit to a hospital emergency department. It is possible that greater attention is devoted to those patients with the most significant health care needs and these would be the patients most likely to require emergency care. Figure 8. Association of frequency of primary care provider s explanation of specialist care with potentially preventable/avoidable emergency department use. Likewise, children with a primary care provider who coordinates access to a specialist or specialized care or equipment are more, not less, apt to make a potentially preventable emergency department visit. Figure 9. Association of primary care provider s assistance in obtaining specialist with potentially preventable/avoidable emergency department. While the availability and nature of primary care influences the likelihood of a potentially preventable emergency department visit for children who require

35 Preventable ED Utilization by Children 35 specialized care, figures 6 through 9 clearly indicate that the children least associated with such visits are those who require or seek no special care. Sick care/sick days. Sick care refers to the number of times the referenced child saw a medical professional for non-preventative reasons outside of a hospital in the twelve months prior to the survey. The data indicates a clear relationship between number of sick visits and a potentially preventable emergency department visit. The more often a child seeks medical care outside of the hospital, the more likely he/she will seek potentially preventable care in a hospital emergency department. This relationship applies throughout the range of number of sick care visits. This suggests that this attribute primarily indicates level of wellness rather than discriminate between primary or emergency care when a child s illness is primary care treatable. Very few of the surveyed children used the emergency department exclusively in place of primary sick care. Only 1.8% of the referenced children are associated with a potentially preventable emergency department visit but not a sick care visit to a primary health care provider. Sicker children who require more medical care are more likely to seek or be sent by a medical professional to the emergency department for some of that care. In fact, nearly half the referenced children who made more than 12 sick care visits also made at least one potentially preventable emergency department visit. The number of school days missed due to illness provides another measure of wellness for all school-age children, including those who are not typically brought to a medical professional for treatment of illnesses. As the number of sick days

36 Preventable ED Utilization by Children 36 increases, so does the likelihood of a potentially preventable emergency department visit. Health Conditions NSCH survey questions investigated the existence of particular health conditions which require treatment. Ideally this treatment would be provided in a primary care setting. However, when primary care is not sought or provided in a timely and effective manner, children dealing with these conditions are more likely to require emergency department services identified by this study as potentially preventable. Asthma and allergies (respiratory, food, or skin) are two of these conditions. Each of these is associated with more than ten percent of the study of population and is therefore considered individually as a possible risk factor for potentially preventable emergency department utilization. Several other conditions occur in smaller segments of the pediatric population. Each one is associated with a greater likelihood of potentially preventable visits to emergency departments. However, they are optimally not considered individually because their numbers are too small to be statistically significant. Instead, they are grouped into a single flag, Health_Conditions which is true if the child exhibits one or more of these conditions. Asthma. Of the children studied by the NSCH, 11.92% were reported to have been informed by a health care professional at some point that they had asthma. Caregiver responses indicated that 8.5% of the children still had the condition at the time of the survey. This latter group can be segmented according to whether the referenced child experienced an episode of asthma (asthma attack) within the 12 months prior to the

37 Preventable ED Utilization by Children 37 survey. Such an episode serves as an indicator of unsuccessful management of the child s asthma. For the purposes of this study, children identified by their caregiver as no longer having asthma were grouped with those who never had asthma in the no asthma category. As is quantified in Table 7, records corresponding to children who have recently suffered an asthma attack are considerably more likely to be also associated with a potentially preventable emergency department visit. This is hardly surprising. Asthma attacks which require emergency department care are classified as such as they theoretically could be prevented by appropriate, regular treatment and medication. Table 7 Percent of Children Associated with a Potentially Preventable/Avoidable Emergency Department Visit by Asthma Classification Asthma Classification Percent Associated with a PPA Visit Asthma with recent attack Asthma with no recent attack No asthma The link between regular, appropriate medication and decreased necessity for emergency department intervention for children with asthma was reported by Dombkowski et al (2004). The NSCH data is limited in the information it provides concerning adherence to an asthma medication regimen. Survey responses provide an indication of the elapsed time since the child with asthma last received medication for this condition. There is no distinction as to whether the medication is designed to prevent asthma attacks or respond to them or the frequency or regularity with which it is administered. For the referenced children, more recent medication is associated with an increased likelihood of potentially preventable emergency department

38 Preventable ED Utilization by Children 38 utilization. It is noted that children with more severe asthma are more likely to receive frequent medication. Additionally, those with poorly managed asthma are likely to frequently require medication in response to episodes. Consequently, children with more severe or unmanaged forms of the disease are more represented among the most recently medicated. Therefore, rather than measuring the effect of appropriate medication in preventing unnecessary emergency department attention for asthmatic conditions, it appears that this attribute functions as an indicator of asthma severity and management of the condition. The data also includes a separate variable which reflects the respondent s subjective assessment of the severity of the referenced child s asthma. There is the clear and logically expected association between greater perceived severity and increased likelihood of a potentially preventable emergency department visit. In summary, while this data does not allow the opportunity to study the effect of a regular sustained program of medication as it relates to the necessity of emergency care, it does provide indicators of asthma severity and management. More severe and/or less managed asthma are strongly associated with potentially preventable emergency care. Allergies. Children with respiratory, food, or skin allergies are more likely to visit a hospital emergency department with a potentially preventable problem than their nonallergic peers. This is particularly true for those with food allergies. In this survey, 23.14% of the 3945 children with food allergies made potentially preventable visits to

39 Preventable ED Utilization by Children 39 an emergency department. This is almost twice the percentage of the general survey population associated with a visit of this nature. Other health conditions. Other health conditions which afflict fewer children are also associated with significantly greater potentially preventable emergency department utilization. Table 8 indicates the prevalence of these conditions and the percentage of those afflicted who are associated with a preventable emergency department visit. Table 8 Percent of Children Associated with a Potentially Preventable/Avoidable Emergency Department Visit by Health Condition Condition Percent of Survey Population Afflicted Percent Associated with a PPA Visit Diabetes Depression Bone, Joint, or Muscle Problem Developmental Delay or Physical Impairment Severe Headaches or Migraines >= 3 Ear Infections, Past Year None Of the surveyed children, 15.37% have one or more of these conditions. They are identified in the modeling process by the derived flag attribute Health_Conditions. Condition severity. Children with mild forms of the aforementioned conditions are less likely to require potentially preventable emergency care than those for whom the conditions are more severely manifested. The Condition_Severity variable reflects the severity of the most severe condition for children afflicted with allergies or any of the conditions combined in the Health_Conditions flag.

40 Preventable ED Utilization by Children 40 Figure 10. Association of condition severity with potentially preventable/avoidable emergency department use. Exercise The survey questioned the number of days per week that a child six years of age or older vigorously exercised. For children who exercise, the number of days does not have a marked effect on the likelihood of a potentially preventable emergency department visit. However, children who consistently refrain from exercise are significantly more likely to make such a visit. Consequently, a flag attribute was derived to indicate if the referenced child of at least six years of age engages in regular exercise. Weight The NSCH data includes a derived variable which identifies the referenced child as (a) underweight, (b) normal weight, (c) at risk for becoming overweight, or (d) overweight. The overweight category includes children at or above the 95 th percentile of BMI-for-age. Approximately 18% of the referenced children are classified as overweight. Children under six years of age who are identified as overweight are not significantly more likely to be associated with a potentially preventable visit to an emergency department. However overweight school-age children are 45% more likely to make such a visit than their peers of other weight classifications. Overweight school-age children, particularly those ten years of age

41 Preventable ED Utilization by Children 41 and older, are also more likely to refrain from regular exercise. However these children are no more or less likely to use emergency department services for potentially preventable reasons than their overweight peers who exercise. The referenced child s weight is considered in modeling in the form of a derived flag variable which indicates whether or not the overweight classification was applied by the NSCH formula. Geographical home State. The NSCH survey was designed to provide approximately equal representation of children from each of the 50 states plus the District of Columbia. Additionally, survey design ensured that each of these 51 subsets would be independent data sets allowing for statistically accurate subsetting by state (NSCH Survey Design, 2005). The percentage of records associated with a potentially preventable emergency department visit varies significantly by state. Appendix B contains a table ranking the states from least association to greatest. States associated with the most extreme patterns of potentially preventable emergency department utilization were profiled in order to gain insight into shared characteristics which may influence these patterns. The first consideration was disproportionate populations of groups which were previously established to be more represented among those who engage in the target behavior. Poverty status, education and race have been shown to be predictive of potentially preventable emergency department utilization. State deviation in prevalence of the target behavior related to differences in population concentrations with regard to these factors merely

42 Preventable ED Utilization by Children 42 confirms the consistent importance of these factors in influencing this behavior. Additionally, differing access to primary and emergency care within each state may contribute to variations in emergency department utilization. This study included a comparison of the number of children per practicing pediatrician and the number of hospitals per square mile in each state. These attributes were derived using information obtained from multiple sources as described in Appendix C. Table 9 Comparison of Characteristics for States Associated with the Lowest Rates of Potentially Preventable Emergency Department Utilization State %Pop Below Poverty Level %HS/College Grad % Black Child per Pediatrician # of Hospitals per 1000 square miles VT / UT / CT / NE / CO / Table 10 Comparison of Characteristics for States Associated with the Greatest Rates of Potentially Preventable Emergency Department Utilization State %Pop Below Poverty Level %HS/College Grad % Black Child per Pediatrician # of Hospitals per 1000 square miles MS / WV / LA / DC / KY / Tables 9 and 10 confirm the relationship between poverty, education and the target behavior on a state level. Consistent with national patterns, the states

43 Preventable ED Utilization by Children 43 associated with the highest rates of potentially preventable emergency department utilization have proportionately greater populations identified as living below the poverty level or lacking in educational attainment. Conversely, emergency departments in states with considerably fewer poor and less educated residents see fewer potentially preventable cases. In general, states with the highest rates of potentially preventable emergency department utilization are home to significant black populations and blacks are much less represented in those with lower rates. Connecticut and West Virginia are exceptions to this rule indicating that race is independently less influential in determining potentially preventable emergency department utilization in these states. DC is somewhat atypical of states showing greatest association with potentially preventable emergency department utilization. While its black population and percentage living below the poverty level clearly associate it with similarly grouped states, it boasts more college graduates than any state in the country. The nature of this district is that two very different groups coexist within its boundaries: the very poor as well as those associated either directly or indirectly with the national government who tend to be well educated, compensated, and insured. However further investigation indicates that within DC each group engages in the target behavior in a manner consistent with similarly characterized groups outside the district. The national average for children per clinically active pediatrician is There is no consistent deviation from this average among either of the groups of states which represent the extremes of the target behavior suggesting that this

44 Preventable ED Utilization by Children 44 measure of accessibility to a pediatrician has no apparent individual effect upon the level of utilization of hospital emergency departments for potentially preventable reasons. On the other hand, there is evidence that the concentration of hospitals within a state is significant. In general, the states most associated with the target behavior have more hospitals per 1000 square mile area than those least associated with it. Again Connecticut, which is second only to the District of Columbia in hospitals per thousand square miles and yet boasts among the lowest of rates for potentially preventable emergency department visits, is an exception. Note however that Connecticut also has a large number of pediatricians with respect to the number of children within the state, indicating that health care in general is very available in this state. It is logical that if a hospital is conveniently located, the choice to seek care there instead of at an alternative facility is more attractive. However if pediatricians are also readily available, the convenience factor may be less of an influence. Given the suggestion that these two factors may have a combined effect on a state s rate of potentially preventable emergency department visits, new variables were derived to include in modeling these measures of availability of practicing pediatricians and hospital emergency care. Values were calculated for each state and then the new variables assigned to each record according to the associated state of residence. Character of residence. The NSCH data includes an attribute which indicates whether the referenced child resides within a metropolitan statistical area (MSA). This designation is applied only to records of children who reside in states with sufficient populations in both

45 Preventable ED Utilization by Children 45 MSA and non-msa areas. There is a slightly greater tendency for children in nonmetropolitan areas to make a potentially preventable visit to a hospital emergency department. Non-Predictive Attributes Several attributes were considered but discovered to be lacking any significant association with the target behavior. These include gender, age position within the household, caregiver country of origin, and geographic region. Additionally, the presence of smoker in the household was considered. In general, this data suggests weak positive correlations between children who reside with a smoker and potentially preventable visits to a hospital emergency department and between such children and the existence or aggravated severity of respiratory allergies or asthma. Unfortunately, a survey design error resulted in the omission of this question when surveying caregivers of children less than six years of age for the first six months (approximately one third) of the period in which the survey was administered (NSCH Design, 2006). Consequently the survey data includes a group of approximately 13% of the response records, primarily associated with infants and toddlers, for which the existence of a smoker in the household is unknown. Rather than remove these records when modeling, this decision was made to disregard this attribute. Data Preparation for Modeling In preparation for modeling, the NSCH data set was randomly split into training and test data sets, each of which were determined to be characteristic of the entire set. The training data was balanced in order to assure sufficient representation

46 Preventable ED Utilization by Children 46 of the relatively rare target class which indicates a potentially preventable emergency department visit. The balanced training data set is composed of more than 16,000 records with relatively equal representation of target attribute values. Clustering K-means clustering was employed to form two clusters which are characterized by Table 11. Comparing the clusters, the first is less associated with the target behavior. In general, the children represented by this cluster are healthier with fewer allergies, asthma, and other health conditions. Fewer needed immediate care by a primary care provider or specialist. Almost all of the children who have no primary care doctor are assigned this cluster. Hispanic children are more than twice as likely to belong to this cluster. This group also includes the majority of those who require an interpreter to communicate with a medical professional. Records assigned to cluster 2, on the other hand, are significantly more likely to be associated with a potentially preventable emergency department visit. This cluster is characterized by sicker children who are more likely to suffer from allergies, asthma, or other health conditions, particularly those with moderate to severe forms of these conditions. Children referenced by the records in this cluster were more likely to need immediate or specialized medical care or medical assistance by phone. Almost all are associated with a medical home and are insured, either under Medicaid or private plans. Given the disparate concentrations of records associated with a potentially preventable or avoidable emergency department visit in the two clusters, cluster assignment can be used as a predictive indicator of target class.

47 Preventable ED Utilization by Children 47 Table 11 Cluster Characterization using Full Data Set CLUSTER 1 CLUSTER 2 % associated with PPA visit Indicators of Wellness mean # sick care visits mean number of sick days mean general health % with allergies % with multiple allergies % with asthma % with health conditions % with severe health condition % who get no regular exercise % who are overweight % who required specialist % who required specialized equipment or care % who required immediate care Access to Health Care % with medical home % Medicaid insured % uninsured % who sought medical advice by phone % who need interpreter Demographics Median age group Median Income Group middle middle % with post high school education % Hispanic % Black Classification Modeling Exploratory data analysis identified factors which are individually related to potentially preventable emergency department utilization and detailed the nature of

48 Preventable ED Utilization by Children 48 those relationships. Data mining classification modeling techniques allow the investigation of more sophisticated multivariate patterns. Since the goal of this study is the discovery and description of trends, the choice of modeling technique was made with consideration of transparency of results. Decision tree algorithms produce interpretable rules which detail the interaction of factors influencing the target behavior. In addition to providing a global description of the data set with regard to the target behavior, decision tree rules can be used to identify and characterize interesting subpopulations which are disproportionately associated with this behavior. Algorithms which construct decision trees include Classification and Regression Tree (CART) (Breiman, Friedman, Olshen, & Stone, 1984) as well as C5.0 and its predecessors ID3 and C4.5 (Quinlan, 1993). An alternative classification approach would be the application of a covering algorithm such as PRISM or RIPPER which produce lists of rules which could be used to identify multivariate patterns of potentially preventable emergency department utilization. National Model A CART model was built using the full, balanced training set to gain insight into nationally significant patterns of potentially preventable emergency department utilization. CART employs a binary, recursive partitioning of the NSCH data set. This application of CART analyzes how best to utilize the 38 (including cluster) predictor variables to split the data into smaller groups of records according to shared class. For this data, there are two target classes. A record is either associated with potentially preventable emergency department utilization or it is not. At each node,

49 Preventable ED Utilization by Children 49 CART performs an exhaustive search of the variable set to determine the split which minimizes the Gini impurity index. This index, which was proposed by Breiman et al. (1984), measures the extent to which the segments of data assigned to each child node deviate from the ideal of homogeneous target values and is calculated as follows: #classes Gini L = 1 - (L j /T L ) 2 (1) j=1 #classes Gini R = 1 - (R j /T R ) 2 (2) j=1 where: Impurity Index = (T L * Gini L + T R * Gini R )/n (3) n = the number of records at parent node T L = the number of records in the left child node T R = the number of records in the right child node L j = the number of records of class j in the left child node R j = the number of records of class j in the right child node CART builds a tree by adding branches which decrease the impurity of the parent node until no further reduction in impurity is possible or until stopping criteria have been satisfied. In designing a classification tree, depth is a crucial consideration. A tree which is grown beyond the optimal length may overfit the data, achieving admirable classification accuracy on the training data but generalizing poorly to new data. Conversely, insufficient depth limits the trees predictive capability and often disregards rules which demonstrate interesting patterns of behavior. To address this concern, stopping criteria were set to limit the depth of this tree to five levels under the root node.

50 Preventable ED Utilization by Children 50 Figure 11. National decision tree. By design, the root node attribute is the variable which most cleanly splits the data according to class. At each subsequent level of the tree, the attribute chosen is that which most cleanly splits the remaining data according to class. Consequently, the top splits of this decision tree illustrate some of the most important patterns of potentially preventable emergency department utilization by children.

51 Preventable ED Utilization by Children 51 Table 12 Top Level Split Attributes National Model Attribute First Split cluster Second Split age group Third Split insurance type sick care sick days Split cluster 1 cluster 2 < 6 years old school-age Medicaid none or private 0 sick care visits > 0 sick care visits <= 5 sick days off school > 5 sick days off school This model stresses the importance of the child s level of wellness in determining potentially preventable emergency department utilization. The most significant partition separates the generally healthy children assigned to cluster 1 from the less-well children, including most with asthma, allergies or health conditions which require immediate or special care, of cluster 2. The number of sick care visits outside of a hospital setting or days home sick from school are also considered in determining the referenced child s association with illness and hence potentially preventable emergency department utilization. That ill children are more likely visitors to an emergency department for any reason is not a particularly startling or actionable finding. However, this model also addresses factors which influence the decision to seek care in an emergency department setting for potentially preventable problems when the child is ill. Of considerable importance is the age of the child. CART distinguishes the behavior of school-age children in this regard from that of those younger. Infants and preschoolers are strongly associated with potentially preventable emergency department utilization. In contrast, school-age children are considerably less so.

52 Preventable ED Utilization by Children 52 Indicators of wellness are particularly important in isolating the group of school-age children who do visit the emergency department with potentially preventable concerns. The type or existence of medical insurance coverage also plays a role in a caregiver s decision to seek potentially preventable care in an emergency department. CART segments the population of generally healthy, cluster 1 children under the age of six according to this factor. Those with Medicaid are shown to be significantly more associated with such care than those who have no insurance or are privately insured. In fact, while CART utilizes this attribute to partition a specific subset, exploratory data analysis concluded that this relationship holds for the full data set. Additionally the level of education attained by a caregiver influences the decision as to where to seek care for a child. Children in cluster 2 generally require more medical care than other children. CART found that a caregiver with college education is more likely to refrain from utilizing emergency department services to obtain that care for their school-age child, particularly when the child is not significantly ill as is evidenced by no more than five days absent from school. Finally, this model detailed an effect of lack of association with a personal doctor. Generally healthy school-age children of cluster 1 without private health care insurance who have been ill enough to receive care outside the hospital are likely to also make a potentially preventable visit to the emergency department if they have no personal doctor. The support for a decision tree rule is the proportion of records which satisfy the split conditions defining the path from root to the given terminal node. The support for this rule is a scant 1.775% of the training data. Therefore,

53 Preventable ED Utilization by Children 53 further data investigation was undertaken to assure its generality. Confidence refers to the proportion of records satisfying the split conditions which are correctly classified by the rule. The noted confidence of 63.2% applies to the balanced training data. In order to perform a valid comparison, the confidence of this rule as measured on the unbalanced training records was contrasted with that measured on the test data set. Table 13 provides evidence that this rule is in fact generally applicable. Table 13 Comparison of Rule Confidence on Training and Test Set Rule if cluster 1 and school-age and any sick care and not privately insured and no personal doctor Unbalanced Test Set Training Set Still further investigation examined whether the effect of the lack of association with a personal doctor was limited to the population for which CART designed this rule. That population was segmented based upon four attribute values: (a) cluster 1, (b) school-age, (c) sick care visits > 0, and (d) Medicaid or no insurance. However, it was discovered that the effect of the lack of a personal doctor on potentially preventable emergency department utilization patterns applies more broadly to all records which indicate any sick care visits outside of a hospital, regardless of insurance type, age or cluster. Sick children who have no personal doctor are more likely to seek potentially preventable care in an emergency department. Figures 12 and 13 illustrate this pattern. Children without a personal doctor who do not visit a doctor outside of a hospital for sick care are less likely than those with a medical home to seek potentially preventable emergency department care. Children of every basic medical home status are more likely to make a potentially preventable emergency department

54 Preventable ED Utilization by Children 54 visit if they have sought sick care outside of a hospital. For those with no doctor or with a personal doctor but no medical home, this change in potentially preventable emergency department utilization rate is greater than for those with a medical home. Figure 12. Association with potentially preventable emergency department use by basic medical home category for children who make no sick care visits. Figure 13. Association with potentially preventable emergency department use by basic medical home category for children associated with 1 or more sick care visits. This suggests a more general rule than that described by CART. Table 14 Comparison of Medical Home Decision Rules Rule Antecedent Consequent Support Confidence CART if cluster 1 and school-age and any sick care visits and not privately insured and no personal doctor PPA Generalized if no medical home (no doctor or a doctor but no medical home) and any sick care visits PPA

55 Preventable ED Utilization by Children 55 This more general rule has greater support and confidence than that constructed by CART. Note that CART s more specific rule applied only to cluster 1. This is largely because the clustering process grouped more than 95% of records associated with the lack of personal doctor in cluster 1. While the lack of a personal doctor is associated with increased likelihood for potentially preventable emergency department utilization for children of both clusters, those few who are assigned cluster 2 overwhelmingly likely to make such a visit. Figure 14. Association with potentially preventable emergency department use by basic medical home category for children of cluster 2. Model Evaluation Classification accuracy provides a quantitative measure of this model s ability to determine association with potentially preventable emergency department utilization. Of the training data records, 75.06% were correctly classified. When applied to the test data, this model achieved an overall accuracy rate of 74.85%. The consistency of these two rates indicates the generality of the model, testifying to its ability to perform equally well on new data. Only 11.44% of the test data records indicate association with a potentially preventable emergency department visit while 88.56% are pre-classified as not associated with this target behavior. Thus, though this model incorrectly classifies 37.32% of the records which are associated with a potentially preventable visit, these

56 Preventable ED Utilization by Children 56 comprise a very small portion of the records classified as non-visitors as is reflected in the false negative rate of On the other hand, the 23.58% of non-visitors who are misclassified are a significant portion of those visitors resulting in a false positive rate of Figure 15. Predicted vs. actual association with potentially preventable emergency department visit national model. Figure 16 graphically illustrates the performance of this model which predicts potentially preventable emergency department utilization for 8516 (28.05%) of the test records. This group includes 62.68% of the records which are pre-classified as associated with the target behavior along with 23.58% of the records which would correctly have been assigned the opposite classification. This corresponds to a lift of Figure 16. Gains chart for national model.

57 Preventable ED Utilization by Children 57 If the model less liberally applied the positive classification, the effect would be to move the arrow to the left along the curve, decreasing the number of records incorrectly identified as associated with the target behavior (false positives), but also the number of those correctly classified as such. In this case, the trade-off for identifying a reasonable percentage of potentially preventable emergency department visitors is accepting the misclassification of a significant number of records which are pre-classified as non-visitors but follow patterns identified by the model to be associated with such a visit. Achievable accuracy for this model is limited by the attributes ability to measure necessity for potentially preventable care. While available attributes measure association with illness of any nature, for the most part they are incapable of discerning the severity or term of illness. For example, a sick care visit in response to a cold appears to the model as identical to one for treatment of pneumonia. Logically, children associated with illness of a more serious or persistent nature would be more likely to seek care in an emergency department. Furthermore, it is not unusual for children, including those who typically seek non-emergent or preventable care in an emergency department, to experience years in which they require no non-preventative medical attention. As the goal of this study is descriptive in nature, evaluation of the success of this endeavor is not exclusively measured by the overall model accuracy. In addition to understanding the interaction of factors which broadly describes potentially preventable emergency department utilization, this project seeks to identify specific actionable patterns which describe trends of population subsets. These can be

58 Preventable ED Utilization by Children 58 examined separately and their classification effectiveness measured by rule confidence. Table 15 details the rules generated from the national model with the most significant confidence. Table 15 Decision Rules Generated from National Model Antecedent Consequent Support Confidence if cluster 1 and school-age No_PPA if cluster 1 and school-age and no sick care No_PPA if cluster 1 and < 6years of age and Medicaid PPA if cluster 1 and school-age and any sick care and PPA not privately insured and no personal doctor if cluster 2 PPA if cluster 2 and < 6 years of age PPA if cluster 2 and school-age and > 5 sick days PPA if cluster 2 and school-age and <= 5 sick days and family education <= high school grad PPA Insurance Type Model Health insurance status has been established as a significant contributing factor in the decision to seek potentially preventable emergency department care. In an effort to better understand how behavior patterns differ among children with various types of health insurance, the CART interactive tree builder function was used to build a classification model in which the training data is first split according to insurance type so that separate branches individually model the training data associated with each insurance type. The cluster attribute was intentionally not included in this model. Clustering groups similar records together. The k-means algorithm when applied to the NSCH data grouped those records with attribute values which indicated similar degrees of wellness and requirements for immediate or specialized care. These factors are highly

59 Preventable ED Utilization by Children 59 predictive of the target behavior. Consequently, clustering was employed as a useful tool to increase the predictive accuracy of the national model. However, when cluster is used, the specific factors which influence the choice of classification are somewhat less transparent. Therefore, in an attempt to maximize descriptive value of this second model, cluster was omitted from the input list. The interactive tree builder function implemented in SPSS Clementine 10.0 allows the data miner to apply knowledge of the data in determining split attributes and values (Clementine 10.0 Users Guide, 2005, Appendix C). The interactive tree can be grown a level at a time. At each split, the data miner is afforded the opportunity to accept the split determined by CART as most capable of decreasing the impurity index or instead choose a competing predictor and/or customize the split conditions. In this decision, he/she has access to information about available predictors and the change in impurity available with each choice. For the NSCH data, the CART algorithm recognized Sick_Care as the best split attribute. Figure 17. Clementine defined split.

60 Preventable ED Utilization by Children 60 Given the modeling decision to investigate patterns of potentially preventable emergency department utilization by insurance type, this choice was overridden in favor of the Insurance attribute using the Select Predictor dialog box. Figure 18. Customizing split attributes using Clementine select predictor dialog box. Note that this dialog box provides information as to the potential improvement associated with each competing predictor. Improvement is the degree to which the impurity could be reduced by a split using that predictor. As the Gini index has been specified for this model as the measure of impurity, improvement is specifically defined as the difference between the Gini impurity index of the parent and that of the resulting child nodes. In this case, the choice of Insurance as a split variable in place of Sick_Care results in less improvement for this initial split but facilitates the descriptive goals of this model. As the tree grows, there are nodes at which multiple predictors promise very similar improvements. In these cases, the decision to override CART s first choice can result in a more descriptive or understandable tree with negligible effect on classification accuracy. Also, customizing splits can have a positive effect on the descriptive ability of the model. In the interactive insurance model, an MD_Phone_Help split was designed by CART to group attribute values usually and

61 Preventable ED Utilization by Children 61 not sought in one branch and never, sometimes, and always in the other. This split does not describe an understandable pattern of behavior. The split was customized, instead grouping usually with other attribute values indicating that medical advice was sought by phone, resulting in no discernible difference in accuracy and a far more understandable pattern of behavior. Figure 19. Customizing split values using Clementine define split dialog box. Another advantage of the interactive method of building a decision tree is the ability to test the effect of specific attributes on different subsets of the population. An attribute split which contributes to the identification of a pattern can be tested on a node in the same branch closer to the root to determine if the pattern can be generalized. Likewise, an attribute can be chosen to split the data of various tree branches in turn to compare the relative influence and effect on the target behavior when applied to different groups. While these splits may result in less effective

62 Preventable ED Utilization by Children 62 classification and therefore not become part of the final tree, they afford the opportunity to gain knowledge of the data. Interactive modeling allows the data miner to grow or prune a tree by level or branch providing more specific control over the design and depth of each branch and optimizing the descriptive value of the resulting tree. The three branches of the NSCH insurance type tree designed interactively are described in detail in the following sections. Medicaid Insurance Branch It has been noted that children covered by Medicaid health care insurance are significantly more likely than their privately insured or uninsured counterparts to visit hospital emergency departments with potentially preventable problems. Nationally, 11.62% of all NSCH respondents indicated such a visit. By comparison, 19.65% of those who are insured by Medicaid engaged in this target behavior. This strong association with potentially preventable emergency department utilization is reflected in the decision rules built by CART to classify records of Medicaid insured children. Figure 20. Medicaid decision tree branch.

63 Preventable ED Utilization by Children 63 This model classifies the majority of training records, 78%, as potentially preventable emergency department visitors. The strongest indicators of such a visit are evidence of an illness which was treated outside of a hospital setting and age. Table 16 Split Attributes Medicaid Insurance Branch Attribute First split after insurance sick care split Second split after age group insurance split sick care Third split after insurance split Fourth split after insurance split Fifth split after insurance split sixth split after insurance split age group got_imm_md race health conditions Split 0 sick care visits > 0 sick care visits < 6 years old school-age <= 3 sick care visits > 3 sick care visits < 6 years old school-age required immediate medical care required no immediate medical care black or other white, Hispanic, multiple race, or unknown none at least one All Medicaid insured children less than six years old are identified as likely associated with this target practice. School-age children who make no primary sick care visits are among the few who are not classified as likely to make a potentially preventable emergency department visit. Those who have made more than three such visits are confidently classified as likely to make a potentially preventable emergency department visit. Finally, those who have made one to three sick care visits are classified according to their need for immediate medical care, race, and the existence of health conditions. The necessity for immediate medical care in addition to sick care results in a classification as potentially preventable emergency department visitor.

64 Preventable ED Utilization by Children 64 Black children have previously been shown to seek potentially preventable emergency department care at a greater rate than children of other races. This model classifies all black children who have sought any primary sick care as likely associated with this practice. Table 17 summarizes rules generated from this model with significant confidence. Note that support refers to that of the model as a whole, not the individual branch. Table 17 Decision Rules Generated from Medicaid Branch Antecedent Consequent Support Confidence if any sick care visits PPA if 1-3 sick care visits PPA if > 3 sick care visits PPA if 1-3 sick care visits and < 6 years old PPA if no sick care visits and school-age No_PPA if 1-3 sick care visits and school-age and PPA required any immediate medical care if 1-3 sick care visits and school-age and required any immediate medical care and race = black or other PPA No Medical Insurance Branch The NSCH data indicates that children who lack any form of medical insurance make potentially preventable emergency department visits at a rate which is consistent with that of privately insured children, but less than half that of Medicaid insured children. This is the least represented insurance type in the NSCH data. Only 8.69% of the referenced children are uninsured. Consequently, the depth of this branch is limited by the relative scarcity of records.

65 Preventable ED Utilization by Children 65 Figure 21. Uninsured decision tree branch. As for Medicaid insured children, careful examination of the model indicates that uninsured children who use the emergency department for potentially preventable care are most recognizable by their utilization of other health care resources. Those who seek no primary care for illness in person or medical advice by phone either lack the necessity for such care or are hesitant to seek care due to financial concerns imposed by the lack of insurance coverage. These children typically do not seek potentially preventable care in an emergency department and are so classified by this model. Family income determines the classification of those who do visit a health care professional to address an illness. Uninsured children from upper income families are considerably less likely to make a potentially preventable emergency department visit than those from lower or middle income families. The former are classified by the model as non-visitors and the latter as visitors. Those of lower or middle income families who seek phone advice in addition to making one or more sick care visits are particularly likely to use emergency department services to address

66 Preventable ED Utilization by Children 66 a potentially preventable concern. The use of multiple health care resources may indicate more frequent, persistent, or serious illness. Table 18 Split Attributes Uninsured Decision Tree Branch Attribute First split after insurance sick care split Second split after md_phone_help insurance split Third split after insurance split income md_phone_help Split 0 sick care visits > 0 sick care visits sought medical advice by phone did not seek medical advice by phone lower, middle upper sought medical advice by phone did not seek medical advice by phone For uninsured children, CART recognizes that indication of necessity for care and income are more predictive of the target behavior than age. However, further data exploration confirms that younger children without health insurance do make more potentially preventable emergency department visits than older children who are similarly uninsured. Figure 22. Effect of age on potentially preventable emergency department utilization for uninsured children.

67 Preventable ED Utilization by Children 67 Table 19 summarizes rules generated from this model which achieved significant confidence. These illustrate the importance of evidence of necessity for care in determining association with a potentially preventable emergency department visit. Once again, support is measured for the entire model. The limited concentration of records associated with the lack of insurance results in corresponding small support for these rules. Table 19 Decision Rules Generated from Uninsured Branch Antecedent Consequent Support Confidence if no sick care No_PPA if no sick care and no medical advice by phone No_PPA if at least 1 sick care visit and lower or middle income and medical advice sought by phone PPA Private Insurance Branch More than 63% of the balanced training records reference children who have private health care insurance coverage. These children are the most likely to have access to adequate health care resources outside of a hospital. Figure 23. Association with basic medical home by insurance type. Nevertheless, they visit emergency department at a rate similar to that of uninsured children.

68 Preventable ED Utilization by Children 68 Figure 24. Private insurance decision tree branch. For Medicaid or uninsured children, the best discriminator between those who seek potentially preventable emergency department care and those who do not is the need for sick care of any kind. For those who are insured privately, it is the need for immediate medical care. Those with private medical insurance who require no immediate care are unlikely to visit the emergency department with a potentially preventable concern. In fact the CART branch dedicated to the classification of this group identifies only infants and preschoolers with less than excellent health who have sought primary sick care as likely to make such a visit. All the attributes used to partition the privately insured training data measure necessity for care or age. This branch of the decision tree classifies records as potentially preventable visitors to an emergency department only with evidence of extensive, immediate or specialized care or less than excellent general health. As with

69 Preventable ED Utilization by Children 69 the Medicaid branch, the threshold of required care to gain this classification is lower for infants and preschoolers. Table 20 Split Attributes Private Insurance Decision Tree Branch Attribute First split after insurance got_imm_md split Second split after insurance split Third split after insurance split Fourth split after insurance split Fifth split after insurance split md_phone_help sick care sick care age group age group gen_health spec_pmd_explain Split required immediate medical care required no immediate medical care sought medical advice by phone did not seek medical advice by phone 0 sick care visits > 0 sick care visits < = 6 sick care visits > 6 sick care visits < 6 years old school age < 6 years old school age excellent < excellent required a specialist did not require or made no visit to a specialist Table 21 summarizes the generated decision rules with significant confidence. Table 21 Decision Rules Generated from Private Insurance Branch Antecedent Consequent Support Confidence if no immediate care needed and no sick care No_PPA visits if no immediate care needed and no sick care No_PPA visits and school-age if immediate care needed and phone advice PPA sought and < 7 sick care visits and age < 6 if immediate care needed and medical phone advice sought and > 6 sick care visits PPA

70 Preventable ED Utilization by Children 70 Comparison of Insurance Type Branches All three branches of the insurance type model focus primarily on partitioning the data according to likely necessity for care. Regardless of insurance type, children are not brought to an emergency department unless care is necessary. What differs is the typical level of care required for classification as a potentially preventable visitor. Potentially preventable immediate concerns are those which could be treated with prompt attention in primary care or emergent conditions which might have been prevented by earlier primary care. Privately insured children characteristically seek emergency department care when the concern is immediate in nature and are more likely to receive all their non-immediate sick care in a primary setting when compared with those who are covered by Medicaid or no insurance. While uninsured children and those insured by Medicaid are also quite likely to seek immediate care in an emergency department when such care is required, many are associated with a potentially preventable emergency department visit but no need for immediate care. These children seek care for clearly non-urgent conditions in an emergency department. CART specifically identified black Medicaid insured children as associated with this type of utilization. Further investigation linked black children with this behavior across all insurance types and age groups. However, the disparity between black children and those of other racial/ethnic classifications in this regard is greatest among uninsured children. Black children without medical insurance are almost as likely to seek potentially preventable care in an emergency room as those insured by Medicaid. Uninsured white and Hispanic children are considerably less

71 Preventable ED Utilization by Children 71 likely to seek such care, behaving much more like their privately insured peers in this regard. Table 22 Comparison of Percentage of Test Records Associated with a Potentially Preventable Emergency Department Visit by Population and Insurance Type Population % Associated with PPA Visit Medicaid Insurance Private Insurance No Insurance Black White Hispanic Overall, Medicaid insured children are most likely to seek potentially preventable emergency department care. In fact, the rate at which infants and preschoolers are brought for such care is so significant that CART requires no evidence of illness to classify them as likely visitors. This model relies upon attributes which indicate utilization of primary care resources in an attempt to gauge level of wellness and necessity for care which might be sought in an emergency department. As with the general national model, these measures are imperfect indicators. Many children are treated repeatedly in a primary care setting and never seek emergency department care. Others utilize the emergency department in place of primary care. In an effort to understand whether the type of medical insurance affects the tendency to choose the emergency department exclusively for care, the percentage of potentially preventable emergency department visitors who use the emergency department as a sole source of primary care was calculated by insurance type. Records referencing these children indicate a potentially

72 Preventable ED Utilization by Children 72 preventable visit to a hospital emergency room but no sick care visits to any health care professional outside of a hospital. Table 23 Potentially Preventable Emergency Department Visitors who Use the Emergency Department as a Sole Source of Care Test Data Insurance Type PPA, No Sick Care PPA % ED as Sole Source of Care Medicaid No Insurance Private Uninsured children are most likely to seek care exclusively in an emergency department. Despite the strong association between Medicaid health care insurance and potentially preventable emergency department utilization, children covered by Medicaid are considerably less likely than their uninsured peers to receive all their primary care in an emergency department. Model Evaluation The insurance type model correctly classifies 73.17% of the test data. Like the national model, it misclassifies a significant number of records which are not associated with a potentially preventable emergency department visit, resulting in a false positive rate of It also fails to correctly classify 38.96% of those who are appropriately associated with such a visit but, given the predominance of the negative class, these records comprise a small percentage of those assigned the negative classification. Hence the false negative rate is only 0.063%.

73 Preventable ED Utilization by Children 73 Figure 25. Predicted vs. actual association with potentially preventable emergency department visit insurance model. Note that the national model and the insurance type model performed similarly. Both relied predominantly upon indicators of necessity for care and age in determining class. Cluster, which was chosen as the most significant split attribute by the national model, was intentionally disregarded in the insurance modeling process in the interest of improved transparency. Nevertheless, the accuracy of this model approached that of the national model. The combined lift chart of Figure 26 compares the performance of the two models. Arrows indicate the lift attained when each model was applied to the test data. Figure 26. Comparative lift of national and insurance models as applied to test data.

74 Preventable ED Utilization by Children 74 State Models The national and insurance type models identify the most significant nationally applicable patterns of potentially preventable hospital emergency department utilization. Segmentation of the training data by state affords the opportunity to investigate geographic variations in these patterns. Separate CART models were built for each of the fifty states and the District of Columbia. Each was tested for general applicability using an independent, hold-out test data set. The state models consistently reinforced the importance of indicators of wellness or necessity for care and age in predicting the target behavior. As these patterns have already been established and discussed, the focus of this section is limited to rules noted in the state models which suggest patterns which were not identified by the models of national scope. These CART rules are used as the starting point of an investigation which seeks to determine if the suggested patterns apply nationally or are uniquely characteristic of certain geographic areas. Where they do suggest differences, additional extensive data exploration further defines and clarifies those differences. This technique identified two factors associated with geographic variation in potentially preventable emergency department utilization. Additionally, understanding of generally applicable patterns of such utilization is broadened with the identification and analysis of rules included in individual state models but not in models built using the complete data set.

75 Preventable ED Utilization by Children 75 Geographic Variation Distance from hospital. Thirty-five states have significant enough population in both MSA and non- MSA areas to have this designation of residency applied to data collected there. When considered as a group, children who reside in non-msa areas are more likely to be from less educated, lower or middle income families and are either uninsured or insured by Medicaid. They are less likely to have a medical home and more likely to be white. They are also more likely to visit a hospital emergency department with a potentially preventable concern. However, when the data is segmented by state, there are three states which, although they adhere to these demographic patterns do not follow the typical pattern of potentially preventable emergency department utilization. Children residing in non-msa areas of Iowa, Nebraska, and Texas are less likely to make a potentially preventable visit than their peers in MSA areas. An investigation into shared characteristics which might explain this departure from common utilization patterns revealed that pediatricians are in short supply in these states. In fact, when compared to other states, these three have among the greatest values for number of children per clinically active pediatrician. Aside from this commonality, Nebraska and Iowa share many similarities with each other and few with Texas. Nebraska and Iowa have significantly more children living in non-msa areas than most other states. In these states, approximately half the children surveyed lived in rural areas. Both MSA and non-msa areas in Nebraska and Iowa boast a greater than average percentage of families with higher levels of education and income, privately insured children and those who have a medical home. Children of

76 Preventable ED Utilization by Children 76 color make up a significantly smaller percentage of the population, particularly in rural areas. Texas on the other hand, has a larger than average percentage of lower income citizens with no insurance or doctor and little education. As is the case for other states with both MSA and non-msa designations, these characteristics are somewhat less prevalent in MSA areas. However in Texas the relative disparity between MSA and non-msa areas in terms of the prevalence of these characteristics is considerably less than the average for all states with such designation. Hispanic and multiple race children are represented in both MSA and non-msa areas in Texas at 2-3 times their average representation. Children of black or white race are proportionately less represented than average. In fact, in MSA areas of Texas, white children comprise only 41.53% of the surveyed population. Texas is unusual among states with fewer clinically active pediatricians per child in that the vast majority of children are clustered in MSA areas. In general, the density of pediatricians is inversely proportional to the percentage of children living in non-msa areas. In Texas, 86.84% of the surveyed children reside in MSA areas, a percentage significantly greater than the average of 72.7%. Doobinin et al. (2003) investigated the significance of convenience in determining whether to utilize a hospital emergency department. They considered the tendency of urban children to seek emergency department care in place of primary care due to proximity and hours of availability. It is suggested that in rural areas far removed from hospitals, the opposite tendency is driven by the inconvenience of traveling great distances to obtain hospital care. In these areas it may be more convenient to obtain primary care which, given the scarcity of available pediatricians,

77 Preventable ED Utilization by Children 77 may be provided by general practitioners who lack specialized training in pediatrics. These primary care physicians, realizing the travel burden, may be hesitant to suggest a trip to a hospital emergency department if at all possible to treat the condition locally. It is also logical that parents residing in such areas would be cognizant of the inherent time delay in obtaining emergency department care and would therefore be more likely to address issues before they became emergencies. As the NSCH does not provide a means of determining the distance from a hospital for each child, the previous assumptions can not be conclusively supported or refuted by this data. However, the database created for the Mapping Health Care for America s Children Project (AAP, 2003) does present information regarding geographic location of clinically active pediatricians within each state. Pediatricians tend to cluster in MSA areas, particularly those with hospitals. There are significant geographic areas in Nebraska, Iowa, and Texas which not only lack a clinically active pediatrician, but also are far removed from any clusters of pediatricians which might indicate the availability of hospital care. This suggests that distance from emergency care may be one characteristic shared by non-msa residents of these three states. Additional data would be required to thoroughly investigate the correlation between distance from a hospital and rate of potentially preventable emergency department utilization. Need for an interpreter. Children for whom an interpreter is required for effective communication between caregiver/patient and medical personnel are more likely to be associated with a potentially preventable emergency department visit. The strength of this association

78 Preventable ED Utilization by Children 78 varies widely by state as does the representation of children concerned. While the NSCH data indicates a national average of 1.76% of children who require an interpreter, many states, most notably Maine, West Virginia, Montana, North Dakota, and Vermont, have few or no such children among those referenced by this survey. On the other hand, 7.11% of the children in CA require an interpreter for effective communication with health care personnel. Arizona, Nevada, Oregon, Rhode Island and Texas also have strong representation of this group. The likelihood that a child requiring an interpreter will make a potentially preventable visit to an emergency department in states with the greatest concentration of such children is below the national average of 20.45%. In states with moderate populations of non-english speaking families requiring interpreters, including Connecticut, Delaware, and New Jersey, children from such families are far more likely to be brought to an emergency department with potentially preventable complaints. In these states, over 30% of the responding children who require interpreters are associated with a potentially preventable visit. Furthermore, while this group of children is more associated with Medicaid than any other insurance, the type of medical insurance has no apparent effect on the tendency for them to seek care in a hospital setting. Perhaps children who do not speak English are more likely to find bilingual medical personnel or interpreters in a hospital emergency department. States with the largest concentrations of non-english speaking families may offer more primary care settings staffed with personnel who can effectively communicate in patients primary languages. However children in states with more moderate non-english speaking populations may find it difficult to obtain such care. Further research utilizing data

79 Preventable ED Utilization by Children 79 with measures of availability and location of multilingual medical personnel would be required to confirm this supposition. General Trends Most of the rules generated for the individual state models identified patterns which do not indicate geographic differences in behavior, but rather highlight patterns of behavior which generally apply to children across the nation. In many cases, the CART rule defined a very specific subset of children associated with the target behavior, which additional data analysis proved could be validly generalized to a larger population. Asthma. Numerous state models include rules predicting potentially preventable emergency department visits according to asthma related factors. Many models predict a visit to an emergency department with strong confidence if a child has had a recent asthma attack. Such an attack is the best available indicator of ineffectively managed asthma. Models included asthma rules applying to various subsets of the population. Further data investigation verified that unmanaged asthma is a strong indicator of potentially preventable emergency department utilization independent of other factors. Income and Insurance. The uninsured branch of the insurance type model noted a difference in target behavior for children who required some sick care according to family income. Those from families with income classified as upper are less associated with a potentially preventable emergency department visit. Several state models included rules which

80 Preventable ED Utilization by Children 80 supported this pattern which applies not only to the uninsured but to those who are privately insured as well. Family income does not have a significant effect on potentially preventable emergency department utilization for Medicaid insured children. While most of these children do come from lower income families, the minority from middle or upper income families engage in the target behavior at a rate consistent with those of lower incomes. Figure 27. Combined effect of income and insurance type on potentially preventable emergency department utilization. Race and age. Several state models utilized race to segment data. Rules consistently separate black children from white children in classifying according to the target behavior.

81 Preventable ED Utilization by Children 81 However, classification of Hispanic children is dependent on age. Hispanic infants make potentially preventable emergency department visits at rate very similar to black infants. As age increase, the likelihood of a visit decreases for all races, but does so most dramatically for Hispanics. By adolescence, Hispanic children behave most similarly to white children. Figure 28. Combined effect of race and age on potentially preventable emergency department utilization. Medical home and sick care. The effect of access to a medical home on potentially preventable emergency department utilization differs according to necessity for care as measured by sick care

82 Preventable ED Utilization by Children 82 visits outside of a hospital. Healthy children with no medical home who require no sick care are less likely to engage in the target behavior than similarly healthy children with a medical home. However, when consideration is limited to those who require sick care, the absence of a medical home is associated with greater likelihood of a potentially preventable emergency department visit. Figure 29. Combined effect of medical home status and amount of sick care on potentially preventable emergency department utilization. Caregiver structure and age. Exploratory data analysis found that children of single mothers are more likely to make a potentially preventable emergency department visit than those who are cared for in two-parent homes while those raised by a single father are least likely to make such a visit. CART devised rules which suggest age plays a role in determining the influence of caregiver structure on the decision to seek potentially preventable emergency care for a child. Further data investigation showed that indeed

83 Preventable ED Utilization by Children 83 single fathers of infants behave similarly to single mothers in this regard. Preschoolers with single fathers seek such care at a rate less than that of children with single mothers but still greater than those cared for by two parents. The characteristic hesitance of single fathers to seek potentially preventable emergency department care for their children is evident only for school-age children. Figure 30. Combined effect of caregiver status and age on potentially preventable emergency department utilization

84 Preventable ED Utilization by Children 84 Caregiver education and medical advice by phone. The effect of consistently available medical advice by phone on potentially preventable hospital emergency department utilization was noted to be dependent on caregiver level of education. Children of caregivers with less than a high school education are no less likely to make a potentially preventable emergency department visit if consistent access to medical advice by phone is available. In more educated families, the availability of such advice is associated with decreased likelihood of such a visit. Figure 31. Combined effect of consistently available medical phone advice and education on potentially preventable emergency department utilization. Perhaps there is more effective communication between the doctor and caregiver or the doctor feels more confident in accepting the more educated caregiver s appraisal of the nature of the child s problem. Ineffective communication may result in the medical decision to direct that the child be brought to the emergency department or

85 Preventable ED Utilization by Children 85 the degradation of the child s condition resulting in the necessity of emergency care. Regular exercise and age. The school-age child who does not participate in regular exercise is more likely to make a potentially preventable emergency department visit than his/her peer who does exercise. While this pattern holds for all school-age children, for the NSCH data, it is most pronounced for those between the ages of 10 and 12. Figure 32. Combined effect of lack of exercise and age on potentially preventable emergency department utilization.