The Potentially Preventable Hospitalizations (PPH), also named Potentially Avoidable Hospitalizations, Ambulatory Care Sensitive Conditions or Prevention Quality Indicators (PQI), are hospital admissions –predominantly exacerbation of chronic conditions– that conceptually may be preventable with timely and appropriate outpatient care [1, 2]. The PPH are defined geographically (districts, municipalities, areas served by a primary care team or by a hospital, regions, etc.) based on the location of the patient’s residence and, succinctly, PPH could be defined as hospitalization rates for definite conditions from residents of delimited geographical areas, regardless of whether the hospital is located inside or outside the patient's area [2]. PPH rely on hospital discharge data, but are not intended as measures of in-hospital quality. Rather, they were built as indirect measures of accessibility problems to high quality outpatient care [3].

PPH, mainly in its PQI form, have been adopted (and adapted) by different national and international organizations [4–6] and currently are a common instrument for the evaluation of health care with expanded uses [7, 8]. In countries with health systems without universal coverage, the interpretation of PPH has been oriented towards the identification of gaps in insurance coverage, access to outpatient care and the volume of primary care supply [9–14]. In Europe, where insurance tends to be universal and primary care is extensively developed, interpretations have been directed towards the evaluation of the quality of ambulatory care, and often, specifically referred to the quality of the primary care level, disregarding issues related with the role of outpatient care provided by specialists, the hospital responsibility in the control of chronic patients (discharging patients more or less stable, deciding which patients are admitted based on their own admission policies, etc.) and, also, the critical importance of a proper coordination between the different levels of care [15].

In the Spanish National Health System context, with an extended network of healthcare centres structured in two levels (hospitals and specialized outpatient care, and primary care) and geographically ordered (Hospital Departments and Primary Care Basic Health Zones (BHZ)), it is assumed that PPH represent largely a problem of coordination in the necessary chronic diseases continuum of care, both between and within levels of care [15, 16]. This lack of coordination can affect similarly the full range of preventable hospitalizations, or may affect differently to each of the clinical conditions comprising PPH. In the former case, a common geographic pattern of the full range of PPH would reflect the quality and performance of healthcare providers, and could help to identify territories that handle homogeneously -better or worse - the most common chronic conditions causing PPH. In the latter, specific maps for each PPH would point out idiosyncratic organizational factors operating in the management of a particular PPH. It is expected both features to be present in our context, but in which extent has not been studied yet.

From a methodological perspective, given that the full range of preventable hospitalizations may share some of the aforementioned features, the conjoint analysis of them can be recommendable to gain in statistical power and interpretation. Additionally, the geographically structured nature of the data requires specific spatial methods to obtain unbiased estimates. Under the disease mapping setting, the Shared Component Model (SCM) [17] is a method that allows both the joint analysis of several diseases and the incorporation of spatial correlation, and has also shown to lead to improved inference over separate analysis of each outcome [18–21].

The aim of this paper is to explore the underlying pattern shared by six chronic PPH conditions in healthcare geographic BHZ of the Valencia Community (Spain) using Bayesian joint modeling, and assuming that the geographic areas serve as surrogate for a mix of the epidemiologic and medical practice risk factors that underlie any spatial variation in the pattern of hospital admissions.

Our study shows, primarily, the possibility of modeling the geographical distribution of the risk of admission for several Potentially Preventable Hospitalizations concomitantly, providing the latent pattern shared by all the PPH conditions analyzed and quantifying how well each PPH is represented on it. The distribution of this common latent pattern captures about the 36% of the total variability, and may be used to evaluate healthcare systems performance, as it allows detecting areas with homogeneously low or high risk of PPH admissions that may suggest hypotheses about possible causes (and possible interventions) to reduce the volume of PPH. In the case studied (Valencia region), for example, the spatial structure of the PPH common patterns shows a non-random pattern, with local effects and rural–urban trends, practically defining the borders of some Healthcare Areas, which suggests that PPH may have an important relationship with specialist care common to all BHZ of the same Area. Although this hypothesis cannot be tested with the current design, as it does not explicitly consider the possible role of Healthcare Areas, it is important to highlight it, because it weakens the common belief that PPH are almost exclusively a problem of primary care, incorporating specialist care (without excluding hospitalization, emergency departments and some intermediate units such as hospital at home) to the definition of the problem, and shifting the attention to the coordination between the two levels of care.

The analysis of the single PPH divergence distribution may allow the identification of areas where more specific organizational factors could influence the rates of a particular PPH. This would be probable if some of these conditions were managed under shared, and divergent, care between primary and specialty care. In any case, the marked differences for some BHZ in specific PPH rates regarding the predominant pattern of their corresponding Healthcare Area suggest that in these cases the attention should be focused on the primary care teams responsible for these deviating BHZ.

A different approach for analyzing different PPH conjointly was proposed by a working group from AHRQ in 2006 [34] and included in later versions of the AHRQ PQIs [35]. These composite PQIs were planned to improve the statistical precision of the individual PQIs (increasing the numerator, even at the cost of losing internal homogeneity) allowing for greater discrimination in performance among areas and improved ability to identify differences and potential determining factors in performance. The final proposal included three composite indicators (acute, chronic and all PQIs) by summing the hospitalizations across different conditions and dividing by the population (the number of hospitalizations is used as the “weight” for combining the different entities). In the latent variable method proposed in the present study, the shared pattern could also be considered as a composite measure, being the most prevalent conditions (COPD, CHF) more represented. Additionally, the SCM method allows the identification of causes of avoidable hospitalizations whose behaviour is significantly discrepant regarding this “composite indicator”, opening the possibility to new hypotheses, but at the expense of a more complex analysis than the composite AHRQ PQI measures.

Our study has some strengths and some limitations. Among the strengths it should be noted the organizational characteristics of the VHA, with hospital and primary care service areas administratively defined, facilitating the analysis of small areas and their interpretation in terms of healthcare providers and levels of care. Also, the use of a three-year period, allows having enough statistical power for low incidence PPH, improving the stability of the estimates. From a methodological point of view, besides providing separate maps for the shared and specific components of the risk surface, [19] and advantage of this approach is the ability of the model to pool data and borrow strength among multivariate health outcomes and across neighbouring geographical areas for more reliable small area risks prediction and inference [20]. Among the limitations, first, hospitalizations in private hospitals were not included, and even though we excluded populations with private insurance, we could not exclude patients with a double insurance. The significance and direction of this bias is difficult to estimate, but given the expected better socioeconomic situation of people with additional private insurance, and that Valencia private hospitals are relatively more specialized in elective surgery and deliveries than in chronic conditions, we think it may not be important. Nor were there included hospitalizations occurred outside the VHA because of inaccessibility of data, but we think this could only affect to a relatively small number of admissions on the causes under study. Second, hospital admission rates, as some authors have pointed out, [36, 37] even standardized by age and sex, do not fully account for the differences in disease prevalence between areas, or in the distribution of particularly vulnerable subpopulations, this last an aspect that has already been proved to be associated with PPH in many regions of Europe [38, 39] and USA [40]. Third, the quality of the information of the Minimum Basic Hospital Discharge Dataset has not been exhaustively analysed in the literature, and may not be optimal for research. Nevertheless, after having undergone a standardization procedure, it is considered one of the administrative information datasets with greatest scope and usefulness for research [41]. And last, the PPH definitions used were recently validated in Spain, [24, 25] an aspect which increases its internal validity, however this same consideration limits the contrast of our results with studies that used other lists (including other lists in Spain).

In summary, this paper shows that the shared component modelling offers useful information when analyzing PPH at community level. It provides a global picture of the geographical pattern shared by different PPH, improving our ability to evaluate healthcare systems performance, and at the same time it depicts the particularities of each PPH, an aspect that could be connected with particular organizational factors affecting specific territories.