Determinants of first-time utilization of long-term care services in the Netherlands: an observational record linkage study

The core of the dataset we used consisted of a large primary care database with information on chronic diseases as registered by general practitioners (GP), maintained by the Netherlands Institute for Health Services Research (NIVEL-PCD). More than 500 general practices participate in NIVEL-PCD, 10% of all Dutch general practices. Since all Dutch citizens are registered at a GP, our sample is largely representative for the Dutch population. We linked these data at a personal level with the national LTC-register, with data on LTC-use for the entire 20+ population, and with data on predisposing and enabling factors from several data sources available at Statistics Netherlands (CBS). Thus an artificial cohort was constructed (Table 1). Due to gaps in the LTC register before 2008 we could only use data from 2008 to the latest available year (2011). The remaining dataset consisted of 214,821 persons of age 20+ in 2008–2011. As an uninterrupted stretch of three years of contact data is deemed necessary to be able to establish the presence or absence of chronic disease, we analysed incident LTC-utilization in 2011 (see below) according to the determinants in 2008–2010. Health status (NIVEL-PCD as data source) was measured by the number of chronic diseases present at December 31st 2010. Whereas chronic diseases were defined using a list provided by O’Halloran [10]. This list had been successfully applied earlier to NIVEL-PCD-data [11]. See Additional file 1 for the full list. A patient was counted as prevalent for a chronic disease if at least one contact for this disease was registered during 2008–2010. The count of the number of diseases per patient varied between zero and nine.

Age, gender, ethnicity and household type were assessed on December 31st 2010, and home-ownership, household income and assets were determined for the year 2010 (Statistics Netherlands administrative databases as datasource). For ethnicity a classification adapted to the Dutch population was used. The main distinction is made on migration history. We have lumped different countries of origin with the same migration history together in one group, as is common in studies of the Dutch population. After the native Dutch the largest groups are formed by labour migrants from islamic countries (Morocco or Turkey) and by migrants from (former) Dutch Caribbean possessions (Surinam or Dutch Antilles). If a subject has at least one grandparent with a migrant background, he or she is included in this group.

Long-term care utilization was defined as the use (yes/no) of formal LTC services provisioned by Dutch healthcare organizations (National LTC-register as data source). An extensive overview of the Dutch system for LTC provision as of 2010 can be found in Mot [12]. A total of six different (combinations of) LTC services were distinguished. Table 2 provides a more detailed description, and links the Dutch LTC services to the health functions as used in the System of Health Accounts (SHA) of the OECD [13]. We counted utilization as incident if someone utilized LTC in 2011, but did not utilize any LTC service in the years 2008–2010. Patients with any previous utilization of LTC services in 2008–2010 were removed from the dataset, resulting in a study population of 197,100 cases.

Logistic regression was used to model the relation between the incident LTC service utilization in 2011 and the available predictors in 2008–2010. Several alternative operationalizations were available for some predictors. For instance, ‘household income’ could be operationalized as gross income, income after taxes or income standardized for household. In order to make an optimal choice between these alternatives, they were first tested in single-predictor models. The best-performing predictor or set of predictors were then chosen, as evaluated on the basis of the Akaike information criterion (AIC, see Additional file 2 for values), which penalizes models that are overspecified.

The final model was constructed by combining the best-performing predictors with a multivariate model. Table 3 lists a description of outcome variables and predictors used in the models.

The regression model can be specified as follows:

In which p _i represents our LTC-service outcome variables, x_1,i + … x_m the predictors and β₀ … β_m the model coefficients.

This model was applied successively to all six LTC-service outcome variables as the dependent variable. Although most users of LTC are 65+ of age, first-time use often starts earlier. We decided to run the model separately for 20+ and 65+ populations, so we could evaluate if this would result in markedly different models. The performance of the models was estimated using k-fold cross-validation, with k = 10 [14]. All subjects in the artificial cohort were randomly assigned to one of the 10 folds, and for each fold a seperate prediction was made based on the data of the other nine folds. From these predictions average specificity, sensitivity, accuracy and negative predictive value were estimateded.

As expected, age was a significant predictor in almost all models. In the 20+ population models, second-order age effects (age*age) are also significant; in most 65+ population models only first-order effects of age are significant. Only for LTC home-care nursing services age is not a significant predictor in the 65+ population. Females had a much higher probability of first-time use of any LTC-service than males, especially LTC home-care services. The odds of incident utilization of domestic LTC- services is twice as large for women (for 20+: OR = 2.10, 95% CI: 1.84–2.39, for 65+: OR = 2.04, 95% CI: 1.74–2.39).

Among the 20+ population being of non-Dutch origin was associated with a decreased likelihood of first-time utilization of LTC services. LTC-home or residential service was lowest for citizens born in Morocco or Turkey (OR = 0.79, 95% CI: 0.65–0.96, with the autochthonous population as reference). Among the elderly (65+) country of origin was not significant due to low numbers of people in these groups.

Living alone increased the proportion of first-time users of any LTC-home or residential service compared to living in a multi-person household, in both the 20+ (OR = 1.63, 95% CI: 1.52–1.75) and 65+ population (OR = 1.39, 95% CI: 1.26–1.54). The odds were higher for the use of domestic home-care services in both the 20+ (OR = 1.89, 95% CI: 1.67–2.15) and 65+ populations (OR = 1.66, 95% CI: 1.43–1.94).

Wealth was represented in our models by income and asset measures. A higher gross household income was associated with a decrease in first-time utilization of LTC services. For the first-time utilization of any type of LTC-service, an increase of 10,000 euro is associated with an OR = 0.91 (95% CI: 0.90–0.92) for the 20+ population and an OR = 0.94 (95% CI: 0.92–0.96) for the 65+ population. For LTC domestic home-care services the income effect is somewhat stronger for both the 20+ (OR = 0.87, 95% CI: 0.85–0.89) and 65+ population (OR = 0.88, 95% CI: 0.85–0.92). The best performing asset measure proved to be a classification of home-ownership into three groups: ‘homeowners’ (reference category), ‘social rent’ and ‘private rent’. Subjects classified under ‘social rent’ pay lower rents, because they have a low income compared to the market rates and get a tax break to remedy this. Subjects classified under ‘private rent’ pay rents at market prices.

Both types of rent were associated with a higher incident utilization of any type of LTC services. Social renting carries the highest odds (the 20+ population 2.45 (95% CI: 2.25–2.67); in the 65+ population1.46 (95% CI: 1.30–1.65)). The ‘private rent’ group also showed an elevated first-time use of any type of LTC-service compared to homeowners, but less than the social-rent category.

The higher utilization associated with renting a home applies to all types of LTC services, with the exception of personal and nursing care delivered at home, homeownership status not being a significant predictor for this type of care. This in contrast with domestic care services delivered at home, which is strongly associated with social renting (OR = 2.56 (95% CI: 2.21–2.97) for the 20+ population and OR = 2.06 (95% CI: 1.72–2.46) for the 65+ population).

The number of chronic diseases present was a significant predictor for incident utilization of LTC services in all models. For all types of LTC service combined, the odds ratio (OR) associated with an additional chronic disease was 1.45 (1.41–1.49) for the 20+ population and 1.31 (1.27–1.36) for the 65+ population. In the 20+ population this association seems weakest for residential LTC services, in the 65+ population for both residential LTC services and domestic home-care.

The full results of the cross-validation for all models are shown in Additional file 4. The sensitivity of the models ranges between 0.59–0.81 and the specificity between 0.62–0.84. All models perform very well in predicting a person not going to use LTC services, with a negative predictive value within the range 0.97–0.99 for all models. However, the precision in predicting who will use LTC services is low, with values of 0.01–0.13. The accuracy of the models was highest in the 20+ dataset, which correctly predict 0.75–0.84 of cases in the cross-validation exercise, but was lower (0.63–0.68) among the 65+ population.

An important application of a better understanding of which determinants drive LTC use is in improving predictions of future demand for LTC services. Modelling of future demand for LTC services is often based on demographic projections alone, keeping other factors such as disability prevalences equal [22, 23], or it is based on macro-economic simulation models [1, 22]. These models allow for crude projections of future demand. A more sophisticated model was developed by De Meijer [24], based on the relation between disability level, age and gender. This model has been applied to estimate the effect of ageing on expected LTC costs. The results of our study suggest that including other predictors in such models, such as the number of chronic diseases, homeownership and household size, could substantially improve such projections. For instance, including the number of chronic diseases as a predictor would make it possible to estimate the effects of different epidemiological scenario’s, such as alternative trends in disease prevalence, on future utilization of LTC services.

In this respect, it is of note that the number of prevalent chronic diseases seems to be interchangeable with ADL disability as a predictor of incident utilization of LTC services. Obviously, there is a strong correlation between ADL disability and the presence of chronic diseases, as confirmed by recent studies into the contribution of specific diseases to disability [25‐27]. That these alternative indicators of health status have similar predictive power, was clearly shown in the separate analysis we performed on the smaller dataset in comparing them. Our model of LTC utilization will thus have distinct advantages in comparison to the currently available predictive health care utilization models, especially in situations in which no information is available on ADL disability, but there is information on disease prevalence.

To our knowledge, no other model exists for predicting incident utilization of LTC services based on the prevalence of chronic disease. Another possible useful application of this model, besides projecting future demands, is screening a population for the current need of LTC services. It would also be interesting to employ and further develop the model in comparative studies, for instance by applying it to a country with a different long-term care system, relying less on formal services than the Netherlands. Or, alternatively, we might consider the long term effects of current trends in Dutch politics [28], in which access to LTC services is further restricted to people with the most severe disabilities and without assets or social support. It seems likely that this would lead to the need and enabling determinants becoming more strongly associated with LTC use, while the influence of predisposing predictors such as age and gender will decrease.

For this study we depended on administratively available quantitative data, which comes with some limitations. One of these is that we were not able to capture the utilization of LTC services when initiated by sudden health shocks, e.g. accidents, which do not show up in our selection of chronic conditions.

A further potential limitation concerns the representativeness of our dataset. Our study population consisted of citizens registered at GP practices participating in a network collecting data on primary care. Although large, this is not necessarily a representative sample of the Dutch population. The average age of 20+ study population was almost the same as (50.8) as the average reported by statistics Netherlands for the Dutch 20+ population (49.1) in 2010, but the average gross household income was higher than the Dutch average in 2010 (63,800 euro against 56,100).

Also the proportion of individuals using LTC services was comparable to the Dutch population. Thus 7.85% of our study population used LTC services in 2011, including those with previous utilization, which is close to the 8.46% reported by Statistics Netherlands for the whole Dutch 18+ population in the same year. We conclude that there are no indications of a selection bias that would restrict the generalizability of our findings.

A final and major strength of our study is that determinants and LTC utilization were measured independently of each other. Thus, we can be confident that our need determinant, the prevalence of chronic disease, was estimated in a manner free from an upcoding bias, because this prevalence was calculated from GP data and not known to the administrative bodies that regulate access to the LTC services.