Individual and Area-level Factors Contributing to the Geographic Variation in Ambulatory Care Sensitive Conditions in Finland

Supplemental Digital Content is available in the text.

H ospitalizations for ambulatory care sensitive conditions (ACSC) is among the most commonly used proxy indicators to measure primary health care (PHC) performance. 1 ACSCs have been suggested as potentially avoidable by wellfunctioning PHC. 2 As equity in delivery of health care can be assessed by measuring geographic variation in different medical practices, 3 delivery of PHC can be assessed through geographic variation in ACSCs. Variation in ACSC is a common phenomenon, 4 but it is driven more by individual socioeconomic position (SEP) and health status than general practitioner (GP) workload. 5 Thus, the link between ACSCs and PHC performance remains controversial (Table 1).
ACSCs seem to associate with increased health needs of individuals. 6,7 From the patients' perspective ACSCs result from nonadherence to treatment; for example, due to combined lack of support and mental health issues. 8 This helps to understand why ACSCs are reduced by better patient support through specified family physicians, 9 payment models that reward comprehensive care 10 and continuity of care. 11,12 However, ACSC rates do not mirror the results of PHC clinical quality indicators 13 -and their connection with the number of GPs remains inconsistent. 9,14 Further, some individual and population characteristics contributing to variation in ACSCs might be partially addressable by PHC policies. Although the individual socioeconomic gradient in ACSC rates disfavors the poorest, these rates still differ depending on the incomes of residential areas. 15 And while individual comorbidities are a major predictor of ACSCs, 16,17 even area-level disease prevalence explains variation in ACSC rates. 18 Shorter travel time to PHC reduces ACSCs rates, 9,19 while shorter travel time to hospitals, 19 high rurality, 20 and high hospital bed supply 12,21,22 increase them.
Few studies have assessed over time the development of geographic distribution 23,24 or variation in ACSC rates, 25,26 but describe these only with area-level factors. 23,24 It is still unclear how a comprehensive array of arealevel factors over time contributes to geographic variation in ACSCs, when individual SEP and health status are adjusted for.

FINNISH CONTEXT
The mainly tax-funded Finnish health care system offers a good framework to assess variation in ACSCs, comprising universal access, hierarchical structure and a long tradition of collecting individual hospitalization data. Finland has a strong public PHC where GPs operate as gatekeepers to specialist care provided by public hospitals. 27 PHC is provided through ∼150 health centers, 28 which act also as community hospitals by providing inpatient care in GP led wards. Hospital care is provided through 20 hospital districts-each having mainly a single 24/7 emergency hospital. An individual receives PHC services from a single health center, which receives its specialist care through a single hospital district. However, occupational and private health care offer alternative routes to GP and specialist care outpatient consultations.

Outcome Variable
The Finnish Care Register for Health Care provided individual hospitalizations for the total Finnish population aged 20 years or more in 2011-2017. Of these we identified ACSCs using the UK definition 29 -with an addition of unspecified pneumonia (ICD-10 diagnosis code J18.9) as used previously in Finland. 23 Further, we divided ACSCs into subgroups of acute, chronic, and vaccine-preventable (Supplemental Digital Content 1, http://links.lww.com/MLR/C130). Competent authorities linked hospitalization data into individual sociodemographic data. We formed annual cohorts and applied municipality of residence to allocate individuals into health center areas according to arrangement of PHC; and into hospital districts according to arrangement of hospital care. To account for hospital With these cohorts we asked: (1) how geographic variance in ACSCs was distributed between PHC and hospital care; (2) which factors predicted ACSCs; (3) which area-level factors explained variance in ACSCs when analyzed separately; (4) what were the relative proportions of variance explained by area-level factors after adjustment for individual factors; and (5) how the heterogeneity in risk of ACSCs developed with each added factor?

Individual Factors
The annual FOLK databases provided data for sex, age, municipality of residence, and household income for each individual with ACSCs. For these individuals we selected 5 comorbidities [chronic heart failure, chronic obstructive pulmonary disease (COPD), diabetes, hypertension, and dementia] as suggested by Saver et al 16 from inpatient hospitalizations and specialist outpatient visits during the previous 5 years.

Area-level Factors
We chose area-level factors previously suggested to affect ACSCs, such as disease prevalence 18 and hospital supply. 12 These factors were allocated to health center areas and further categorized for the purposes of reporting-as summarized in Table 2. We categorized the proportion of the population aged 65 and above receiving pensioner's care allowance as disease burden because it captures both disease prevalence and functional limitations. It is granted when a chronic disease or disability limits daily living activities.

Statistical Methods
Age-standardized rates were calculated using the direct standardization. 36 We built on the analysis strategy presented by Falster et al 5 and allocated the annual cohorts to 3-level Poisson multilevel models: individuals were nested within 131 health center areas, which were in turn nested within 20 hospital districts. We analyzed separate models for total ACSCs and each of the 3 ACSC subgroups in 3 consecutive time periods: 2011-2012, 2013-2014, and 2015-2017. The total Finnish population of comparable age was applied as the population at risk and added into the models as an offset (range, 4020-519,853 in health centers and 22,410-1,288,747 in hospital districts).
By first analyzing an age-adjusted and sex-adjusted model (model 1) we estimated the random parameters (σ 2 ) representing the variance at health center and hospital district levels to which we compared all subsequent models. We added the factors stepwise into the models; and measured the effects of each addition with the proportional change in variance (PCV). 37 To analyze which factors predicted ACSCs we calculated incidence rate ratios (IRR). Finally, we calculated the areas' heterogeneity in risk of ACSCs with median rate ratios. 38 The statistical analyses were performed with R, release versions 3.5.1 39 -using the Laplace approximation method. 40   ACSCs indicates ambulatory care sensitive conditions; HC, health center; HD, hospital districts; MRR, median rate ratio; PCV, proportional change in variance, calculated as percentual decrease in variance between each model and model 1.

RESULTS
We observed 729,008 ACSCs in Finland in 2011-2017 (Table 3). The age-standardized ACSC rate in the adult population decreased from 2.67 (95% confidence interval, 2.66-2.68) per 100 person-years in 2011-2012 to 2.57 (2.56-2.58) in 2015-2017. In the age-adjusted and sex-adjusted models-model 1-of total ACSCs the hospital district level variance was approximately twice that of the health center level. These variances decreased slightly over time at both area levels, although the decrease was more pronounced at hospital district level.
In the final models (model 6) male sex, higher age, lower income, and higher number of comorbidities associated with higher IRRs in total ACSCs and all ACSC subgroups (Table 4). Area-level higher disease burden, hospital bed utilization rate, and median income were associated with higher IRRs in total ACSCs in all time periods-but a higher proportion of ACSCs in GP led wards only in 2013-2014 (Supplemental Digital Content 1, http://links.lww.com/MLR/ C130).
Each area-level factor explained variation in ACSCs when analyzed separately (Supplemental Digital Content 1, http://links.lww.com/MLR/C130). For total ACSCs our final model explained a little less than half (PCV = 39.8%-52.1%) the health center level variance and almost two thirds (PCV = 61.1%-74.2%) the hospital district level variance. The highest PCV occurred in 2011-2012 from when it decreased over time along with the variances in age-adjusted and sex-adjusted models. Although individual SEP and health status explained 18.7%-29.9% of health center level variance and 24.6%-35.7% of hospital district level variance, a combination of area-level disease burden and both arrangement and usage of hospital care accounted for an additional 14.1%-16.2% and 32.4%-33.0% of the respective variances. And while the distance to health services and number of GPs did not add much to the models, the areas' median income accounted for an additional 2.7%-6.5% and 4.2%-7.6% of these variances. Areas had some heterogeneity in risk of ACSCs which decreased with subsequent models (Table 5).

DISCUSSION
This study analyzed how individual and area-level factors over time contributed to geographic variance in ACSCs between 2 nested levels of health service providers in Finland. Among total ACSCs our final models explained less than half (PCV = 39.8%-52.1%) the variance between health center areas and almost two thirds (PCV = 61.1%-74.2%) between hospital districts. Even after adjusting for individual SEP and health status, area-level disease burden and both arrangement and usage of hospital care still explained 14.1%-16.2% and 32.4%-33.0% of these variances. The proportions explained were consistent over time. In age-adjusted and sexadjusted models hospital districts showed more variation than health center areas-a disparity which evened out after adjusting for the studied factors. This suggest that variation in age-standardized and sex-standardized ACSC rates could be driven more by factors related to hospital services rather than PHC.
Our findings support the previous studies stating that variation in ACSCs reflects health status, 5,16-18,41 SEP, 5,12,15,21 and factors related to hospital care. 12,19,21,[42][43][44] However, this study emphasized that these factors explained more of the variance occurring in hospital district level than in health center level.
The finding for area-level disease burden-a combination of disease prevalence and functional limitations-adds to earlier knowledge. 18 It explained additional variance independent of individual SEP and comorbidities. We suggest caution when interpreting ACSC rates adjusted with only disease prevalence-especially if this factor is used as a proxy for individual health status. We had a few possible hypotheses for this independent effect. At the hospital district level, it might reflect either different admission criteria between hospitals or systematically insufficient capacity of PHC to answer the high morbidity and disabilities. For health centers, it might reflect inadequate response of PHC in some areas; a possible link between ACSCs and PHC performance.
The finding that local GP led wards maintained variation in ACSCs was consistent with previous studies. 42,43 Countries applying ACSC rates as PHC performance indicators should consider if their arrangement of hospital care affects these rates. ACSCs also reflected areas' overall tendency for hospital utilization, suggesting unnecessary use of available hospital supply 34 rather than different GP referral practices. 45 The effects of distances to health services and the number of GPs were captured by other factors-such as the arrangement of hospital care. Our finding that higher arealevel income predicted higher ACSCs contradicts the previously reported low ACSC rates in wealthy areas. 15 These lower rates might not reflect better performing PHC, but rather other factors such as population's favorable health status.

Strengths and Limitations
The main strength of this study was that we were able to apply nested multilevel Poisson model to distinguish the variance in ACSCs between PHC and hospital care. Further, the individual hospitalization data used are of good quality 46and their comprehensive usage ensured the generalizability of our results in Finnish context. The relative proportions will differ when analyzed elsewhere, but the phenomena behind the analyzed factors are transferrable internationally.
Our study was limited by the lack of individual-level survey data to enrich the applied registers. Moreover, we did not analyze the effect of alternative routes to physician consultations in Finland: occupational and private health carebut believe that including individual and area-level incomes account for them. As the geographic diagnosis coverage in Finnish Register of PHC visits was partial, 47 we had to collect individual comorbidities from both specialist care outpatient visits and hospital discharges.

CONCLUSIONS
This study observed that administrative areas' disease burden and hospital care utilization patterns contributed to geographic variation in ACSCs-potential links to PHC performance. This followed adjusting ACSCs for individual SEP and health status, as well as for the country-specific arrangement of hospital care. Countries measuring PHC performance with ACSC rates should (1) consider that these rates might be driven more by hospital care than PHC; (2) interpret prevalence adjusted rates with caution; and (3) consider that the arrangement of their hospital care might affect not only ACSC rates, but also other factors thought to reflect only the provision of PHC services.