The current study used Medicaid administrative claims data from the Centers for Medicare & Medicaid Services (CMS) between 2000-2005 linked to participants of the 2001-2005 National Health Interview Survey (NHIS), utilizing Medicaid Analytic eXtract (MAX) files. These data are restricted use data and can only be accessed via the National Center for Health Statistics (NCHS) Research Data Center (www.cdc.gov/nchs/data-linkage/Medicaid-restricted.htm). The MAX files are administrative data consisting of calendar-year files that contain utilization and expenditure information for individuals enrolled in Medicaid/CHIP.⁶ More specifically, MAX files consist of administrative claims for Medicaid enrollees resulting from: 1) inpatient hospital services; 2) institutional long-term care records for services provided by four types of long-term care facilities: mental hospitals for the aged, inpatient psychiatric facilities for persons under age 21, intermediate care facilities for the mentally disabled, and nursing facilities; 3) filled prescription drugs, over-the-counter drugs, and other items dispensed by a free-standing pharmacy (non-hospital based); 4) physician and professional services, outpatient and clinic services, durable medical equipment, hospice, home health, and laboratory and x-ray results. Data for this analysis come from each of these four types of administrative claims.

The NHIS is a nationally representative survey of the civilian noninstitutionalized U.S. population. It is conducted continuously throughout the year by NCHS. The NHIS is an in-person interview conducted in the respondent’s home. In some instances, follow-up to complete the interview is via telephone. The survey consists of (a) a family core component, with questions asked about all members of the family; (b) a sample adult component, which collects additional information from one randomly-selected adult per family; and (c) a sample child component, which collects additional information about one randomly-selected child per family. The sample child component is completed by a knowledgeable adult respondent, usually the parent. Data for this analysis come from the sample child and family core components of the NHIS. For more information about the NHIS, visit https://www.cdc.gov/nchs/nhis.htm. Response rates for the 2001-2005 Sample Child component of the NHIS ranged from 78%-81%.

Details of the linkage methodologies and the linked data files have been published elsewhere.^6,23 In short, linkage eligibility was defined as children for whom an adult respondent, acting on behalf of their child, had not refused to provide their Social Security Number (SSN) or health insurance claim numbers, had provided sufficient personal identifier information about their child, and had an SSN verified by the Social Security Administration Enumeration Verification System. Linkage eligibility for children in the 2001-2005 NHIS ranged from 42%-62% and successful linkage rates of children ranged from 13%-20%.

There were 11,986 linkage eligible children aged 3-17 who had linked NHIS-MAX observations for at least one calendar year, representing 23% of the total NHIS sample from 2001-2005 (see Figure 1). The age range 3-17 years was selected due to the uncertainty of an accurate asthma diagnosis among very young children.²⁴ The analytic sample was restricted to children with Medicaid who had non-duplicative (i.e. records in only one state in a given year) continuous enrollment status for 12 months prior to the date of their NHIS interview in order to increase the likelihood that all services and prescriptions received would be documented, removing 7,184 children from the analytic sample. An additional 872 children were removed from the analytic sample because they had no claims with a valid diagnostic code in the 12 months prior to the date of their NHIS interview. The final eligible analytic sample consisted of 3,930 children (8% of the total NHIS sample from 2001-2005). The analytic sample contained children from 49 states and the District of Columbia.

Demographic and clinical characteristic differences between children in the analytic sample and those who were linkage eligible, but were not selected were compared utilizing bivariate logistic regressions that were weighted and accounted for the complex survey design of the NHIS (see Table 1).

Concordance measures among children in the analytic sample, which were also weighted and incorporated the complex survey design of the NHIS, included overall agreement between the survey-based asthma diagnosis and the claims-based asthma diagnosis. A prevalence-adjusted, bias-adjusted kappa (PABAK³³), was presented alongside Cohen’s kappa, as Cohen’s kappa may be skewed when the prevalence of the condition of interest is small within the population. Cohen’s kappa may also not fully account for biases that are introduced when differing methodologies are used for ascertainment, such as when comparing survey-based methods and claims-based methods.³⁴ Differences between subgroups of children on percent agreement were calculated using logistic regressions and differences in the proportion of survey-based and claims-based asthma diagnoses among subgroups of children were calculated using McNemar’s test via adjusted χ² tests.

In order to properly account for correlations that may exist between the factors of co-occurring conditions, service utilization, and demographics, structural equation modeling (SEM) was employed. SEM has some additional capabilities compared to other types of modeling (e.g. linear regressions), namely the ability to concurrently analyze both the association between multiple dependent factors while also exploring the loadings of observed items on these dependent factors, all while simultaneously accounting for the possibility of measurement error within items.³⁵

A three-factor model of the covariates of demographics (child’s sex, age, race/ethnicity, family income recorded as a percentage of the federal poverty level, educational attainment of the highest educated household member, geographic region of residence, MSA status of residence, family structure, respondent’s relationship to sample child), co-occurring conditions (hay fever, any kind of respiratory allergy, any kind of food or digestive allergy, eczema or any kind of skin allergy, frequent or repeated diarrhea or colitis, anemia, frequent or severe headaches (including migraines), three or more ear infections, and seizures), and service utilization (saw generalist, saw specialist, saw mental health professional, saw therapist, office visit, emergency department visit, and home visit in the past 12 months) were modeled directly onto agreement and allowed to correlate with each other. All covariates were modeled as binary, with a series of dummy variables for demographic variables. All SEM modeling was conducted within STATA 14.0,³⁶ utilizing a maximum likelihood estimator which could account for missingness on a given covariate, and were weighted and incorporated the complex survey design of the NHIS using subpopulation estimation for the analytic sample.

A sensitivity analysis utilized a weighted least squares estimator (asymptotic distribution free) for the aforementioned model to determine how missingness on various covariates may impact the SEM, and a secondary model with demographics modeled onto conditions and services, instead of directly onto agreement was also tested. For all models, a sub-analysis was conducted for any significant coefficients to further explore individual items. Finally, a secondary sensitivity analysis was conducted to assess the effect of excluding children with conditions known to use asthma medications for treatment (cystic fibrosis, bronchopulmonary dysplasia, and trachelomalacia).³⁷

Table 1 presents demographic characteristics, service utilization, and co-occurring conditions as measured through data from the NHIS of linkage eligible children stratified by inclusion (n=3,930) or exclusion (n=8,056) to the analytic sample and the total linkage eligible sample overall (n=11,986). Children included in the analytic sample were 49.6% boys and 60.8% were aged 3-10 years. About 46% of the children were non-Hispanic white, 29% were non-Hispanic black, 21% were Hispanic, and 4% were non-Hispanic other. Approximately two in five children lived in two-parent families (38.7%), and 84.6% of children lived in households with family incomes <200% of the federal poverty level. All children lived in the United States with the greatest portion living in the South, (40.5%), followed by the Midwest (24.0%), West (19.5%), and Northeast (16.1%). Most children had an office visit in the past 12 months (93.3%), or saw a provider in the past 12 months (94.3%). Approximately 3 in 10 children had some form of allergy (29.5%), while approximately 1 in 5 children had some other health condition (20.3%).

Children in the analytic sample were more likely to be non-Hispanic black, living in households below the 200% federal poverty level, without two parents, and in a large urban metropolitan statistical area when compared to children who were linkage-eligible, but not in the analytic sample. Children in the analytic sample were more likely to have a parent as the respondent for the NHIS interview than children not in the analytic sample. Additionally, children in the analytic sample were also more likely to have seen any provider, received care at any location, and been diagnosed with either an allergy or other health condition in the past 12 months.

For approximately 90% of children in the analytic sample (89.9%), the diagnostic status for asthma matched when comparing the survey-based definition to the claims-based definition (see Table 2). The prevalence-adjusted bias-adjusted kappa was 0.80 while Cohen’s kappa was 0.55. The prevalence of survey-based asthma diagnosis was 14.6% which was significantly higher than the claims-based asthma diagnosis of 10.6% (p<.001), a finding true of most comparisons across subpopulations. Agreement was higher for girls than boys (p<.05), while children who had received care in the past year had lower agreement than children who had not received care in the past year (p<.01), and children with allergies and other health conditions in the past 12 months had lower agreement than children without these conditions (p<.001).

A three-factor model (see Figure 2) was constructed to evaluate the relationship among a child’s demographics, service utilization, and co-occurring conditions and the concordance (modeled as agreement) of a child having a claims-based asthma diagnosis and a survey-based asthma diagnosis.

After controlling for demographics and services, co-occurring conditions had a significant negative effect on agreement (β = −0.25, p<.001). However, after controlling for demographics and co-occurring conditions, service utilization had no significant effect on agreement nor did demographics after controlling for co-occurring conditions and services. There was a significant positive correlation found between the factors of service utilization and co-occurring conditions (ρ = 0.59, p<.001), but there was not a significant correlation found between service utilization and demographics, nor co-occurring conditions and demographics.

A sensitivity analysis was performed to see what impact using a different estimator (weighted least square), which did not have the ability to account for missingness on a given covariate, would have on the main findings. This analysis revealed comparable results, but with less power. A secondary model of demographics modeled directly onto co-occurring conditions and services produced nearly identical coefficients – a likely product of the non-significant direct pathway between agreement and demographics [see Online Supplemental Figure].

Finally, the secondary sensitivity analysis conducted assessed the effect of excluding children with conditions known to use asthma medications for treatment (cystic fibrosis, bronchopulmonary dysplasia, and trachelomalacia (n=6). The main results did not change given this degree of misclassification.

The significant correlation between co-occurring conditions and agreement provided justification for a secondary analysis to further explore the impact of selected conditions on the concordance between asthma definitions (see Figure 3). An additional structural equation model with two factors was developed which included: allergies (hay fever or seasonal allergies, respiratory allergies, food or digestive allergies, and eczema or skin allergies), and other health conditions: (diarrhea or colitis, anemia, frequent or severe headaches (including migraines), frequent ear infections, and seizures). After controlling for other health conditions, allergies had a significant negative effect on agreement (β = −0.23, p<.001), but other health conditions did not have an effect on agreement after controlling for allergies. There was a significant positive correlation found between allergies and other health conditions (ρ = 0.60, p<.001).

Although the analytic sample encompasses a nationwide cohort of children within Medicaid with continuous enrollment over a twelve months period, this study inclusion restriction impacted the generalizability of the sample given typical breaks in coverage seen among children receiving Medicaid.^43–44 It should be noted that children who receive medical services through Medicaid who have been diagnosed with asthma are also not representative of the greater population of children with asthma. In fact, recent research has found significant differences in asthma management between children enrolled in a public or private health insurance plan, with children with public insurance being more likely to use the hospital emergency department and to discontinue preventive medications.⁴⁵ Nonetheless, children with Medicaid or some other public insurance account for almost half of children in the United States diagnosed with asthma.² Finally, it is important to note that reporting standards, cost-sharing practices, and managed care plans may differ from state to state, particularly over a five year period, resulting in discrepant proportions of cases represented in the final analytic sample. However, it is worth noting that 49 states and the District of Columbia contributed cases to the sample.

The definitions of both claims-based asthma and survey-based asthma in this study have limitations. While claims are generally an accurate reflection of diagnoses, it is possible that a diagnosis of interest may be missed either because the associated claim occurs beyond the scope of the study period (e.g. the child did not receive services for their asthma in a given year)²⁸ or the claim may never appear in the records because the child received services through a non-Medicaid provider.⁴⁶ Furthermore, MAX prescription claims are based on prescriptions being filled, and it is known that children with Medicaid are less likely to fill asthma prescriptions compared with children that do not have Medicaid.⁴⁷

Finally, for claims-based cases there was the potential for children to have asthma medication claims for treatment of non-asthma conditions. Results from a sensitivity analysis suggest that the main results are unchanged given this degree of misclassification. Survey-based asthma cases are also subject to misclassification due to the reliance on parent report that are subject to recall bias and social desirability.