Clinical and Socioeconomic Associations With Hospital Days and Emergency Department Visits Among Medically Complex Children: A Retrospective Cohort Study

Data for this study was extracted from 4 standardized Alberta health administrative data sets housed in the Provincial data repository: the Discharge Abstract Database (DAD),¹¹  which contains hospital discharge data; the National Ambulatory Care Reporting System (NACRS),¹²  which contains data on hospital ED visits; Alberta data repository Statistics,¹³  which lists all births and deaths in Alberta; and the Pharmaceutical Information Network,¹⁴  which captures all pharmaceuticals dispensed in a community setting. The administrative data sets housed in the provincial data repository provides a unique identifier that links individuals across these data sets.

CMCs were identified through DAD using the International Classification of Diseases, Tenth Edition (ICD-10)-based Complex Chronic Conditions system originally developed by Feudtner et al and adopted by Cohen et al^2,15  (and commonly elsewhere in the children with medical complexity literature¹⁶ ). Only CMCs who were residents (in the province), had an index discharge and diagnosis between 2010 and 2013, and were at most 18 years old at index admission were considered. For each eligible CMC, annual cumulative numbers of hospital days and ED visits from DAD and NACRS were obtained for the 5 years after index hospital discharge; these serve as our primary outcomes.

Age, sex, index hospitalization length, number of chronic medications (for the 12 months after index discharge), 9 indicators for chronic disease involving different body systems, the presence of technology assistance (TA) (eg, hemodialysis), and 7- and 30-day readmission after index discharge were included as demographic and clinical variables. Number of chronic medications was only available as a count provided by clinical practitioners (and not as individual medications) as was extracted from Pharmaceutical Information Network data for the 12 months after index discharge. All variables were measured at index admission unless otherwise specified.

Socioeconomic variables were derived from residential postal code at index discharge. The Alberta Health Services classifies regions in the province into 6 rurality categories ranging from metropolitan to remote rural and classifies residents into socioeconomic groups using the Pampalon Deprivation Index.^17,18  The Pampalon index incorporates 6 variables from Canadian census data at the dissemination area (DA) level into 2 factor scores that measure material and social deprivation. Specifically, these 6 components are average household income and the proportions of single-parent families; individuals who are divorced, widowed, or separated; individuals living alone; individuals who are unemployed; and individuals without a high school diploma. A DA is the smallest unit (400–700 people) of the Canadian census and is socioeconomically homogeneous. The use of mental health services by the mothers of CMCs in the 12 months before index admission were identified through ICD10 codes (available in Supplemental Table 5) through DAD and NACRS. CMCs and their mothers were linked through DAD via a maternal and newborn chart.

A multivariable mixed-effect hurdle model was used to estimate associations between clinical and socioeconomic factors and each outcome. This type of model has previously been used to examine health care utilization and expenditure¹⁹  and is useful when an outcome contains disproportionately many 0s (a key characteristic of our study data, eg, 58.2% and 43.1% of CMCs have no hospital days and ED visits, respectively, in the first year).

Each hurdle model consists of 2 submodels. A binary (logistic regression) submodel estimates the distribution of responses of 0. A conditional submodel estimates the distribution of strictly positive response values as a truncated negative binomial distribution (with the canonical log link) and includes random intercepts by patient to account for repeated (ie, annual) measurements.

Sex, age group (with categories following Cohen et al² ), clinical conditions (after the aforementioned ICD-10 scheme), TA, number of chronic medications (with ≥5 medications corresponding to polypharmacy and an established higher risk of adverse drug effects to CMCs²⁰ ), readmission indicators (conditional submodels only), residence rurality, and mental health service use were included as categorical predictors. See Table 1 for categorical levels and summaries. Log-transformed initial length of stay (LOS) and the 2 deprivation factor scores were included as continuous predictors. Previous studies have demonstrated differences in resource use between age²¹  and clinical⁶  groups and have suggested links between chronic medications and overmedicalization.²²  Other clinical factors as proxies for clinical severity and quality of care were also included.^2,23  As noted above, previous studies have found associations between resource use and socioeconomic factors,^8,24  which here were studied in conjunction with clinical factors.

The following potential second-order interaction effects were included after an assessment of clinical plausibility: between clinical indicators; clinical indicators and time; clinical groups and initial LOS; mental health service use and deprivation scores; number of chronic medications and time; and age and time. The latter 2 interaction effects were removed from the binary submodel for ED visits for computational reasons.

All analyses were performed with R (version 3.6.3)²⁵  via glmmTMB (version 1.1.3).²⁶  Tests of overall covariate significance were conducted as type-II Wald χ-square tests with a Benjamini–Hochberg correction for false discovery and comparisons between groups with Holm corrections via multcomp (version 1.4–17).²⁷  All formal tests had 2-sided alternative hypotheses and used a 0.05 significance threshold.

Research ethics approval for this study was granted by the University of Alberta Research Ethics Board (Pro00103550_REN2).

Over half of 12 621 CMCs identified for inclusion were newborns (<1 year old). Of the cohort, 82.3% had 1 complex chronic condition, 13.8% had 2, and 3.9% had 3 or more. Number of chronic medications was most commonly 0 (64.9%), 1 (15.8%), or 2 to 4 (15.8%). The majority of CMCs lived in metropolitan areas (58.7%) and a minority in rural (18.7%) or remote rural (3.8%) areas. See Table 1 for detailed summary statistics. In the first year after diagnosis, 1.5% of CMCs accounted for 28.0% of all CMC hospital days, whereas 1.7% of CMCs accounted for 15.7% of ED visits. In total, 809 (6.4%) CMCs died during index admission or the subsequent 5-year observation window; these CMCs were retained in the sample.

In total, 1516 CMCs (12.0%) who were missing socioeconomic status measures (1515) or had a sex of “other” (1) were excluded. The excluded CMCs were comparable to the remaining 11 105 on the basis of sex, age, clinical group, number of chronic medications, initial LOS, and mothers’ prior use of mental health services.

Age was a significant predictor in both conditional submodels (χ₄² ≥ 18.3, P value ≤ .004). Notably in the first year, newborns had significantly more conditional hospital days on average relative to the 1 to 4, 5 to 9, and 14 to 18 age groups (relative difference [RD] 1.375–1.913, SE 0.129–0.163, P value ≤ .04) and significantly more conditional ED visits relative to the 5 to 9, 10 to 13, and 14 to 18 groups (RD 1.318–1.603, SE 0.090–0.125, P value < .001). In the fifth year, no significant differences in cumulative number of hospital days were present between the age groups; patterns of differences in ED visits were comparable between the first and fifth years. For all comparisons, see Supplemental Table 3.

Figure 1 illustrates differences in the conditional number of hospital days between patients in exactly 1 clinical group with an initial LOS of 3 and 12 days (the approximate median and 80th percentile). CMCs in the neurology category appeared to have the steepest increase in conditional hospital days over the 5 years. Differences in hospital days between the groups were more apparent when initial LOS was large (see the following subsection). ED visit trajectories (Supplemental Fig 3) showed less variation across the clinical groups and with initial LOS.

Notably, for a 3-day initial LOS, the neurology group had significantly more comparable number of hospital days (RD 0.734–1.549; SE 0.086–0.232; P value ≥ .06 relative to gastrointestinal, malignancy, metabolic, and year 1 congenital or genetic; RD 1.363–2.468, SE 0.130–0.395; P value ≤ .002 otherwise) in years 1 and 5. The same was true for ED visits (RD 0.887–1.549, SE 0.072–0.435, P value ≥ .07 relative to gastrointestinal, hematology or immunodeficiency, malignancy, metabolic, year 1 renal, and year 1 congenital or genetic; RD 1.272–2.468, SE 0.073–0.395, P value ≤ .01 otherwise).

Initial LOS was a significant predictor in both conditional submodels (χ₁² ≥ 14.7, P value ≤ .001). Table 2 presents estimates of average changes in the conditional response associated with a 10-fold increase in initial LOS, which was statistically significant for all (RD 1.960–5.097, SE 0.161–0.610, P value < .001) but the gastrointestinal and hematology and immunodeficiency groups in the hospital days model and significant only for the gastrointestinal, hematology or immunodeficiency, and malignancy groups in the ED visit model (RD 0.663–1.364, SE 0.086–0.138, P value ≤ .05).

Initial LOS was also significant In the binary submodels (χ₁² ≥ 95.0, P value < .001), where these associations were nonsignificant or negative (odds ratio [OR] 0.520–0.897, SE 0.031–0.084, P value ≤ .007 for cardiology, congenital or genetic, malignancy, neurology, renal, and respiratory) in all groups except the congenital or genetic group in the hospital days model (OR 1.308, SE 0.084, P value < .001) (Supplemental Table 4).

Number of chronic medications was a significant predictor in both conditional submodels (χ₃² ≥ 193.1, P value < .001). Both conditional responses differed significantly between every pair of chronic medication groups. These differences were comparable between years 1 and 5. CMCs with more chronic medications had more conditional hospital days on average (between adjacent groups, RD 1.409–1.663, SE 0.089–0.170, P value < .001); the results for ED visits were similar (between adjacent groups, RD 1.149–1.280, SE 0.044–0.091, P value ≤ .009). Figure 2 illustrates this positive association.

Number of chronic medications was also a significant predictor in the binary submodels (χ₃² ≥ 838.2, P value < .001), which yielded similar interpretations regarding the absence of hospital days (between adjacent groups, OR 0.322–0.691, SE 0.032–0.069, P value < .001) or ED visits (OR 0.563–0.811, SE 0.018–0.053, P value < .001).

In the conditional submodels, residence rurality was significantly associated with cumulative ED visits (χ₃² = 583.5, P value < .001), but not with cumulative hospital days (χ₃² = 8.5, P value = .09). CMCs living in metropolitan areas had the fewest conditional number of ED visits, followed by those in urban (RD 1.087 vs metropolitan, SE 0.041, P value = .03), rural (RD 1.727 vs urban, SE 0.075, P value < .001), and remote rural (RD 1.114 vs rural, SE 0.122, P value < .001) areas.

Rurality was significant in both binary submodels (χ₃² ≥ 316.2, P value < .001), which suggested similar differences between ruralities in terms of the odds of having no hospital days (OR 1.342, SE 0.039, P value < .001 for metropolitan versus urban; OR 1.047, SE 0.036, P value = .17 for urban versus rural; OR 1.149, SE 0.056, P value = .009 for rural versus remote rural) and ED visits (OR 1.135, SE 0.037, P value < .001 for metropolitan versus urban; OR 1.758, SE 0.071, P value < .001 for urban versus rural; OR 1.207, SE 0.079, P value = .004 for rural versus remote rural).

Material (χ₁² ≥ 13.1, P value ≤ .003) and social (χ₁² ≥ 8.6, P value ≤ .001) deprivation were both significant predictors in the conditional submodels. For CMCs with hospital days, material deprivation had a significant association with hospital days only among those with mothers who did not use mental health services (RD 1.099 for a 1-SD increase, SE 0.036, P value = .02). On the other hand, both material (RD 1.057 for a 1-SD increase, SE 0.018, P value = .005) and social (RD 1.073 for a 1-SD increase, SE 0.018, P value < .001) deprivation had significant, positive associations with conditional ED visits among CMCs whose mothers did not use mental health services. Among CMCs whose mothers did use mental health services, only social deprivation and conditional ED visits had a significant association (RD 1.076 for a 1-SD increase, SE 0.030, P value = .019).

A key finding of this study is that clinical and socioeconomic factors are associated with resource use by CMCs in different ways and are more strongly associated with hospital utilization and ED visits, respectively.

Although high resource use is a characteristic of CMCs,¹  imbalance among our population suggests that certain subpopulations require more resources and support. This conclusion mirrors Gold et al,²¹  who identified a small group of CMCs with long index hospitalizations that accounted for the majority of CMC hospital days and spending across 44 US children’s hospitals. Targeting CMC subpopulations with high resource use may thus be more effective than targeting all CMCs in reducing hospital use.

Among CMCs with hospital days, those with a longer index hospitalization tended to have more hospital days relative to other clinical groups (consistent with Gold et al²¹ ). We hypothesize that the varying associations with initial LOS and other differences between clinical groups are attributable to clinical severity and the availability of support programs for patients and caregivers. That home nursing care reduces hospital use in CMCs²⁸  supports this hypothesis.

The generally positive association between initial LOS and hospital days is arguably intuitive: longer index hospitalization indicates greater need for resources and clinical complexity not resolved by initial care.^21,29  For the 1 exception to this trend in the congenital and genetic clinical group (in terms of the presence of hospital days), a larger initial LOS may indicate that longer index hospitalization tends to improve the long-term health of the child and that further hospitalization is not needed (eg, diaphragm and abdominal wall congenital defects that receive early surgical treatment can remain well managed). Our findings agree with the existing hypothesis that certain CMC subpopulations require more resources and may benefit from increased care coordination and other types of caregiver support programs.^29–31,32 

Although the association between number of chronic medications and hospital days is plausibly related to clinical acuity, other factors may drive this relationship. Pordes et al²² (pe478) found the CMC population to be at risk for overmedicalization (“the delivery of unnecessary and potentially harmful medical care”), which can propagate recurring hospitalizations.³³  Chronic medications may be a sign of complex needs but may also indicate overmedicalization. Further research is needed to examine specific medications prescribed for CMCs and identify whether overmedicalization is a potential cause of preventable hospitalizations.

Residence rurality at index admission is associated with number of ED visits: in our cohort, those living in remote rural areas had the highest number of ED visits, followed by those in rural, urban, and metropolitan areas. Possibly contributing to this pattern in the province is the fact that many primary care physicians also work in local EDs. It is feasible that families and caregivers use EDs if appointments are not readily available or if access to urgent care is otherwise limited. More generally, the limited availability³⁴  of quality home care, telephone access to knowledge providers, and next-day appointments³⁵  may contribute to higher ED utilization among rural CMCs.

Material and social deprivation have significant, positive associations with ED visits for CMCs whose mothers used mental health services before index admission. Since each measure rolls up multiple metrics, we cannot comment on the mechanism underlying this association. Nonetheless, social and material deprivation are important factors that must be accounted for when developing policy and planning health care programs. Our findings support the claim that improved social supports for families, such as funding for care coordination and caregiver education programs, may reduce ED visits and improve outcomes for CMCs.³²  However, we cannot determine the exact nature of the relationship between resource use, deprivation, and mental health service use since the type and duration of the supports accessed are not available in our data. Our findings, nonetheless, together with previous discussions in the literature on how mother mental health is associated with socioeconomic factors^36,37  and acute-care resource use during infancy,⁹  suggest the need for further investigation to better understand the role of maternal mental health in CMC resource use.

Overall, the results of this study are consistent with theoretical frameworks that emphasize the importance of both clinical and socioeconomic factors for understanding health care resource use. Theoretical frameworks, such as the Andersen Model and the Institute of Medicine model, highlight the importance of factors such as clinical characteristics, residency, and other socioeconomic factors as predictors of health care resource use.³⁸ 

All covariates were measured only at baseline, so an investigation with time-varying predictors would yield more-interpretable results. Studies with more-specific designs can further tease apart the associations and interactions observed in this work.

We suspect that missing socioeconomic measures stem from Statistics Canada withholding data for sparsely populated DAs; such DAs in our study are likely rural and may cause our analysis to underestimate (the significance of) comparisons with these ruralities and associations with socioeconomic status. Although DAs are homogeneous by design, their use introduces the potential for underestimated effects and associations.³⁹  Furthermore, our data does not capture CMCs receiving resources outside the scope of the provincial health care system (eg, patients leaving the province, supplemental caregiver support).

Mother use of mental health services data were missing primarily for CMCs born before 2005. Newborn linkage in DAD started in 2002 via maternal and newborn charts collected at delivery. We suspect that this data were missing because of noncollection (before 2002), time required to implement the new collection procedure (up to about 2005), and routine missingness common with administrative data. The specific ICD-10 codes (available in the supplementary materials) used to identify service use by mothers have not been validated in previous literature.

Other limitations common to retrospective research with administrative data stem from the socioeconomic measures used and provided by the Pampalon Index. These measures are at the DA level and can thus lose important, individual-level variability beyond that already lost by principal component analysis. Future studies may consider specific, patient-level factors, such as ethnicity or social supports, which are unmeasured in our data outside of social deprivation as a general proxy.

Our finding that clinical and socioeconomic factors respectively play more-important roles in hospitalizations and ED visits suggests that policies and programs should account for socioeconomic factors such as rurality and socioeconomic status. These findings can be generalized to other publicly funded health care systems (eg, in Canada and Europe) and possibly to Medicaid-funded hospitals in the United States. More research is needed to unravel the impacts of specific socioeconomic factors on families and caregivers so that systems can identify policies and interventions to improve CMC care.

We thank Taranpreet Kaur, lead maternal and child data analyst at Alberta Children’s Hospital (Calgary, Canada), who pulled and provided information on the data set used.