Asthma Care Quality, Language, and Ethnicity in a Multi-State Network of Low-income Children

John Heintzman, MD, MPH; Jorge Kaufmann, ND, MS; Jennifer Lucas, PhD; Shakira Suglia, PhD; Arvin Garg, MD; Jon Puro, MS; Sophia Giebultowicz, MA; David Ezekiel-Herrera, MS; Andrew Bazemore, MD; Miguel Marino, PhD


J Am Board Fam Med. 2020;33(5):707-715. 

In This Article


Data Sources

We utilized data from the Accelerating Data Value Across a National Community Health Center Network (ADVANCE) Clinical Research Network.[23] The ADVANCE Clinical Research Network contains EHR data from the OCHIN (not an acronym) network and Health Choice Network (HCN) of CHCs in 21 states.

Population. Our study population included any child aged 3 to 17 years,[24] with a primary care office visit between January 1, 2012 and January 31, 2017 and with any diagnosis code (primary or otherwise) of asthma either during a clinic visit or recorded on their problem list (as these are the possible locations for a provider to designate an asthma diagnosis code). Our diagnoses were not compared with a diagnostic standard (such as pulmonary function testing): our goal was instead to measure the care quality of those children whom the clinic designated as having asthma at some point. For the children who met this 2012 to 2017 inclusion criteria, we then had access to all EHR data as far back as 2005.

Exclusion Criteria. Patients were excluded if they had a diagnosis of cystic fibrosis and if they did not have a valid body mass index (BMI) on the date of their asthma diagnosis as BMI is an important confounder of asthma course.[25,26]

Dependent Variables. Outcome variables were features of common asthma care measurable in discrete fields in the EHR. We considered 2 sets of outcomes: the accurate and timely diagnosis of asthma as a chronic condition, and the annual prescription rates of common asthma medications. Our first outcome was a binary indicator denoting whether asthma was documented on the problem list on the same day as the initial asthma diagnosis in the EHR. We utilized the recording of a chronic illness on a chart's problem list not as a standard of diagnosis, but as a fundamental, guideline driven quality care step in managing any chronic condition.[14,15,27–29] Specifically, it is standard in chronic disease to document chronic conditions in some kind of registry.[14,15] The "problem list" is the variable field that allows reporting and tracking this diagnosis across visits. Maintaining an accurate problem list is also a core measure of EHR Meaningful Use.[27] Next, for those patients without a same-day problem list documentation of asthma, we estimated time from any diagnosis to problem list recording. We also considered a binary outcome of whether asthma severity was ever recorded on the problem list. Asthma severity was only reflected in diagnosis coding (ICD-10) after October 1, 2015, so evaluating documentation of asthma severity was performed in a subset of children with a visit after this date. For the set of prescription outcomes, we analyzed annual rates of albuterol, inhaled corticosteroid, and oral corticosteroid prescriptions, all which are medications recommended in national guidelines, for daily control or to be prescribed or had on hand for exacerbations.[13] Similarly, as inhaled and oral steroids are more commonly indicated for children with persistent asthma,[13] we also used this visit restriction along with a requirement that patients have documented persistent asthma during this time. Analyses of albuterol were done on our entire sample with asthma.

Variables. Our primary independent variables were a combination of ethnicity and preferred language expressed as 3 groups: non-Hispanic white patients, Spanish-preferring Latino patients, and English-preferring Latino patients. While we use Latino/a because it is often preferred in our study population, the actual ethnicity information collected by clinics is Hispanic and non-Hispanic white.

Covariates. Patient characteristics were derived from the date of first asthma diagnosis in our record including age, sex, insurance status at first diagnosis, family income as measured by percent of federal poverty level, and BMI. In addition, we estimated the rate of ambulatory visits (for any reason) per year up to and including the first asthma diagnosis and included it as a covariate.

Statistical Analysis. We described patient characteristics in our total sample and between our 3 ethnicity/language groups. For the binary outcome of whether asthma was documented on the problem list the same day as diagnosis, we used generalized estimating equation (GEE) logistic regression model to estimate odds ratios comparing ethnicity/language groups adjusted for covariates. We also performed a similar GEE logistic model for the binary outcome of whether asthma severity was ever documented. We fitted GEE logistic models with a compound symmetry correlation structure and empirical sandwich variance estimator to account for clustering of patients by clinic.

For those patients who did not have asthma recorded on the problem list on the day of first diagnosis, we utilized state-stratified Cox regression with Breslow method for ties to assess ethnicity/language differences in time-to-problem-list-record from the date of first diagnosis, adjusted for the same covariates. A robust sandwich estimator was used to construct 95% CIs for the hazard ratios (HRs), accounting for the clustering of patients by clinic. The assumption of proportional hazards was assessed using Schoenfeld residuals and was deemed suitable.

For the set of analyses evaluating annual prescription rates of common asthma medications, we first provide unadjusted annual prescription rate estimates for each medication by ethnicity/language group. Next, for albuterol, corticosteroid inhalers, and oral steroids separately, we estimate using a Poisson-logit hurdle regression model[30] for both 1) the odds of receiving any prescription in the study period by ethnicity/language groups, and 2) among those with a prescription (ie, those who passed the hurdle), the rate ratios of prescriptions by ethnicity/language groups. Similar to zero-inflated Poisson models, these Poisson-logit hurdle regressions model 2 processes. The first process examines the decision to prescribe a medication or not (modeled through a logit model) and in the second process, for patients who were prescribed a medication, a Poisson model is used to evaluate the rates of prescriptions over the study period. For all models, we adjusted for the covariates listed above. In the albuterol analysis, we included an additional covariate for patients' maximum recorded level of asthma-severity (mild intermittent, mild persistent, moderate persistent, severe persistent, or not documented). In the corticosteroid inhaler and oral steroids models, we also included an additional covariate for maximum level of persistent asthma on record (mild, moderate, or severe persistent). All hurdle models utilized robust sandwich variance estimation to account for clustering of patients by clinic. All statistical tests were performed with a 2-sided type I error of 5%. Analyses were conducted in Stata version 15 (StataCorp, College Station, TX). This study was approved by the Institutional Review Board of Oregon Health and Science University.