Frailty transitions and health care use in Europe

Abstract

Objective

There is relative consensus that chronic conditions, disability, and time‐to‐death are key drivers of age‐related health care expenditures. In this paper, we analyze the specific impact of frailty transitions on a wide range of health care outcomes comprising hospital, ambulatory care, and dental care use.

Data Source

Five regular waves of the SHARE survey collected between 2004 and 2015.

Study Design

We estimate dynamic panel data models on the balanced panel (N = 6078; NT = 30 390 observations). Our models account for various sources of selection into frailty, that is, observed and unobserved time‐varying and time‐invariant characteristics.

Principal Findings

We confirm previous evidence showing that frailty transitions have a statistically significant and positive impact on hospital use. We find new evidence on ambulatory and dental care use. Becoming frail has greater impact on specialist compared to GP visit, and frail elderly are less likely to access dental care.

Conclusions

By preventing transitions toward frailty, policy planners could prevent hospital and ambulatory care uses. Further research is needed to investigate the relationship between frailty and dental care by controlling for reverse causation.

1. INTRODUCTION

The growth of population aging in developed countries has raised interest in its associated costs. Understanding the drivers of medical care utilization and spending is crucial for the efficient organization (and allocation) of health care services (and resources). However, the relationship between aging and health care expenditures has been the subject of much debate in the literature.1, 2, 3 Initially, Zweifel et al2 showed that age per se did not have a significant impact on health expenditures after controlling for proximity to death. Then, it has been suggested that comorbidities, chronic diseases, and mental health issues, whose prevalence increases with age, actually mediate the relationship between age and increased health care expenditures.4, 5, 6, 7, 8 In Canada, it has been shown that differences in pre‐existing chronic conditions contribute to health differences in health care expenditures.6 Furthermore, comorbidities, impairments, and proximity to death were found to be key mediators of age‐related expenditures in the UK.7 Other results from cross‐sectional data in Germany found that multimorbidity and mental health were consistently associated with total, inpatient, outpatient, and nursing costs.8

Recently, it has been advocated that frailty could be a key variable to consider when exploring the relationship between aging and increase in health care expenditures.9 Frailty has been a well‐established concept in the medical literature.10, 11, 12 It is considered as a distinct health dimension besides disability, chronic diseases, and functional limitations.13 Frailty is defined as a “clinically recognizable state of increased vulnerability resulting from aging‐associated decline in reserve and function across multiple physiologic systems.”12 Frailty is associated with loss of muscle mass and muscle quality referred to as sarcopenia, resulting from anatomic and biochemical changes in aging muscle.14 As such, it is highly correlated with aging. The estimated prevalence of frailty in Europe in 2004 was 4.1 percent among individuals aged 50‐64 years and 17.0 percent among 65 years and older subjects.15 The specific contribution of frailty to ambulatory care expenditures was calculated using French administrative data, peaking at about €1500 for frail individuals.9

From an economic point of view, the theoretical model underpinning the analysis of frailty in relation with health care utilization is the health deficit accumulation model.16, 17 This model posits that aging is associated with a loss of physical functions (eg, depreciation of health capital), and ability to cope with stressing factors. This should increase the demand for health care when people become frail, a prediction that is consistent with Grossman health care demand model.18 A growing body of the empirical literature has explored the specific impact of frailty on health care utilization.19, 20, 21 Using cross‐sectional data in Belgium, it has been shown that frail and prefrail individuals were more likely than robust elders to contact a GP, a specialist, or an emergency department.19 Using panel data in 10 European countries, two studies showed that frailty was significantly associated with hospital and ambulatory care use, after controlling for both socioeconomic and health status.20, 21 However, less is known about the correlation between frailty and dental care. Previous research underlined the importance of considering dental care access issues among the frail elderly population,22, 23, 24, 25 but the evidence on the impact of (pre‐) frailty on dental care use is scarce. The subject is important because by failing to consult for dental problems, prefrail individuals may increase their chances to lose weight and become frail because of lower nutritional intake.24, 26

Our aim in this paper is to provide new and more robust evidence on the impact of frailty transitions on health care use. Our contribution to the literature is threefold. First, we use more recent and exhaustive data from the five regular waves of the Survey of Health, Ageing, and Retirement in Europe (SHARE), thus providing more insights for the analysis of frailty transitions and more robust conclusions. Second, we analyze the impact of frailty on a wider range of health care use outcomes, namely probability of hospital admission, probability of ambulatory care visits (by distinguishing GP and specialist visits), and for the first time exploring its correlation with dental care use in a longitudinal setting. Third, from a methodological point of view, we estimate empirical models that are less prone to endogeneity as they account for various sources of selection into frailty: (1) unobserved time‐invariant characteristics through inclusion of individual fixed effects, (2) time‐varying observed variables, and (3) time‐varying unobserved shocks through the specification of a dynamic model with lagged dependent variable.27 Specifically, we formalize and discuss for the first time the conditions under which our estimates can be considered as causal.

2. DATA

2.1. Source and sample

We used data from the five regular waves of the Survey of Health, Ageing, and Retirement in Europe (SHARE) collected at different time points in 2004/05 (wave 1), 2006/07 (wave 2), 2011/12 (wave 41), 2013 (wave 5), and 2015 (wave 6). In these waves, SHARE provides information on socioeconomics, health, and health care use variables from representative samples of noninstitutionalized individuals over the age of 50 in eleven countries.28 The countries represent northern Europe (Denmark and Sweden), central Europe (Austria, France, Germany, Switzerland, Belgium, the Netherlands), and southern Europe (Spain, Italy, and Greece). We restricted the sample to the nine countries participating in all regular waves (thus excluding Greece and the Netherlands) and to individuals who participated in the first wave of the panel.

2.2. Dependent variables

Three types of health care utilization indicators were used as dependent variables in our models: (1) hospital admission, (2) ambulatory care use, and (3) dental care use. Hospital admission was included as a dichotomous variable measuring whether, during the last 12 months, the respondent had been in a hospital overnight (HOSP = 1/0). We used three indicators of ambulatory care use. The first indicator is having 1 + visit to a medical doctor (including emergency room or outpatient clinic visits) in the last 12 months (DOC = 1/0). In the first three regular waves of SHARE, it was also possible to distinguish between GP visit (GP = 1/0) and specialist visit (SPE = 1/0). Finally, we used an indicator of whether the respondent had at least one visit with a dentist in the last 12 months (DENT = 1/0). Note that this information was not collected in wave 4.

2.3. Independent variables

Our explanatory variable of interest was Fried's frailty scale, which defines frailty according to five dimensions10, 13: low energy, slowed walking speed, diminished appetite, low grip strength, and low physical activity. We followed prior work to compute the frailty scale15, 20, 21: One point was allocated for each fulfilled criterion, and a frailty score was computed by summing each criterion which contributed equally to the score (theoretical bounds: 0‐5). We set cutoff points in order to define three profiles: 0 for robust, 1‐2 for prefrail, and 3‐5 for frail.12, 20

We controlled for several socioeconomics and health status indicators identified as predictors of health care demand in previous work4, 29 and/or as potential correlates of frailty.20 Predisposing or enabling factors (according to Andersen's behavioral model)29 included age, gender, education, living with a spouse or having a partner, and a dichotomous measure of income adequacy based on a 4‐point scale indicating whether the household was able to make ends meet: with great difficulty/with difficulty (coded 1) vs fairly easily/easily (coded 0). This subjective indicator of economic deprivation has been found to be well correlated with material deprivation items in SHARE data.30 In addition, we used a dichotomous variable of self‐rated health being fair or poor (poor SRH), having two or more chronic diseases (chronic 2+), having two or more limitations in activities of daily living (ADL 2+), having two or more limitations in instrumental activities of daily living (IADL 2+), and having three or more depressive symptoms according to the Euro‐D measure.31

2.4. Sample attrition

Similarly to most panels focusing on elderly people, SHARE is characterized by an important attrition. In our sample, 52 percent of interviewees who participated in the first wave participated in all waves. In this situation, an important question arises, namely whether to use the balanced (ie, with no missing data) or the unbalanced dataset. Fixed effects (FE) estimates on the unbalanced panel are consistent only in the presence of exogenous attrition, that is, only if the determinants of attrition are uncorrelated with the time‐varying regressors.32 However, since individuals remaining in the panel are often healthier than attriters, sample selection is unlikely to be exogenous.

An overview of sample size, dropouts, and attrition rate by wave and according to respondents' characteristics is presented in Appendix S1 (Table S1). Individuals with poor health and, in particular, frail elders are more likely to be attriters. These results are confirmed by running Nijman and Verbeek' (1992) variable addition tests in order to test for exogenous attrition.33, 34 Rejection of one of the tests was interpreted as indicating nonexogenous selection, which was the case for nearly all models (Table S2). To be consistent, we estimated our FE models using the balanced sample, comprising 6078 individuals and 30 390 person‐wave observations. Results of models estimated on the unbalanced sample are provided in Tables S4 and S5 (Appendix S1).

3. ECONOMETRIC MODELS

3.1. General specification

Let $y_{\mathit{it}}^k$ denote a binary indicator of health care use for outcome k (k = HOSP, DENT, DOC, GP, SPE), individual i (i = 1,...,N) at time t (t = 1,...,T). Let ${\text{Frail}}_{\mathit{it}}=\left(\begin{array}{c}{\text{prefrailty}}_{\mathit{it}}\\ {\text{frailty}}_{\mathit{it}}\end{array}\right)$ denote a vector of dummy indicators of Frieds' frailty score being strictly higher than 0 (prefrail) or strictly higher than 2 (frail), X_it denote the full rank matrix of time‐varying control variables, c_i represents an unobserved individual specific term (capturing time‐invariant unobserved factors affecting health care use), and ε_it represents the idiosyncratic error term (capturing time‐varying unobserved shocks). The general model to be estimated is written as follows:

$$E(y_i{t}^k|{\text{Frail}}_{\mathit{it}},X_{\mathit{it}},c_i)=P(y_i{t}^k=1|{\text{Frail}}_{\mathit{it}},X_{\mathit{it}},c_i)={\delta }_{1:2}{\text{Frail}}_{\mathit{it}}+X_{\mathit{it}}\beta +c_i+{\epsilon }_{\mathit{it}}$$ (1)

3.2. Fixed effects (FE) specification

FE estimation of Equation (1) is obtained by OLS estimation on the within‐transformed data (thus identification relies completely on within‐group variations), which has two consequences. First, it allows interpreting the coefficients (δ ₁, δ ₂) in Equation (1) as the impact of frailty transitions on health care use. Second, it allows netting out the impact of unobserved time‐invariant factors (eg, genomic characteristics) that could be correlated with both frailty and health care use. Additional information regarding identification is provided in technical Appendix S2.

However, estimates from the FE model could be biased because of two types of endogeneity issues, namely omitted variable bias and reverse causality. First, in the case of hospital and ambulatory care use, time‐varying unobserved shocks that are included in ε_it (eg, accident, fall) may simultaneously affect frailty transitions and health care use. Second, in the case of dental care use, there could be reverse causality; that is, by failing to consult for dental problems, robust/prefrail individuals may increase their chances to lose weight and become frail because of lower nutritional intake. Indeed, it can be shown that, ceteris paribus, lack of dental care use is a significant predictor of frailty (see Table S6: Appendix S1). These two endogeneity issues cannot be completely removed in the absence of a well‐specified instrument. However, the source of bias can be partly reduced by using a correct dynamic specification of Equation (1).

3.3. Dynamic correlated random effects (CRE) model

A generalization of the static FE model to the dynamic case is presented in Equation (2):

$$E\left(y_{\mathit{it}}^k|y_{it-1},{\text{Frail}}_{\mathit{it}},X_{\mathit{it}},c_i\right)=\rho y_{it-1}^k+{\delta }_{1:2}{\text{Frail}}_{\mathit{it}}+X_{\mathit{it}}\beta +c_i+{\epsilon }_{\mathit{it}}$$ (2)

In the CRE specification, c_i is treated as a random effect and is replaced by its linear projection onto the means of the time‐varying regressors (see technical Appendix S2 for additional information). As shown in the seminal work of Mundlak,35 results of FE and CRE models were estimated on a balanced panel collapse (see Table S3 for empirical comparison using Hausman test). In the dynamic CRE model in Equation (2), $y_{it-1}^k$ is an autoregressive one‐period lag of the dependent variable k and $\rho$ denotes the associated coefficient, which reflects persistence or state dependence in health care utilization.36 In addition, ρ captures all time‐varying unobserved shocks that may simultaneously affect frailty and health care use, thus reducing the correlation between frailty and residual unobserved heterogeneity. Following previous research,21 we used information contained in the third retrospective wave of SHARE—namely an indicator variable of whether individuals had health problems in adult life—as the initial value of the dependent variable.37 The models were estimated by feasible generalized least square (GLS) with standard errors accounting for clustering at the individual level.

3.4. Dealing with the endogeneity of time‐varying controls

Our model accounts for all time‐invariant and some time‐varying characteristics (X_it) that could be associated with both frailty and health care use. Two types of time‐varying characteristics are included in X_it: (1) exogenous variables (ie, age and wave dummies) and (2) potentially endogenous variables (eg, living with partner, chronic diseases, ADL/IADL limitations) for which it is theoretically not possible to assert whether they are causes or consequences of frailty. On the one hand, including endogenous control variables that are consequences of frailty may block the causal pathway and lead to biased estimates of both the total and direct effect of frailty.38 In particular, if these endogenous variables are positively correlated with frailty, they will capture part of the correlation between frailty and health care use, such that controlling for these variables is likely bias downward the effect of frailty (assumption 1). On the other hand, if these variables are exogenous and exert a causal impact on frailty, then excluding them from the model would lead to an omitted variable bias, which likely will bias upward the effect of frailty (assumption 2). As it is not possible to determine with certainty which scenario is most plausible (most likely, there is a dynamic relationship between frailty and the included time‐varying endogenous variables), two models have been considered as follows: (1) excluding and (2) including these variables. Results of model (1) are interpreted as the total effect of frailty, and results of model (2) are interpreted as the partial effect of frailty.

3.5. Robustness analyses

We test the robustness of our main analyses by comparing the results of models (1) and (2) across different model specifications and estimation strategies: fixed effects OLS, conditional maximum likelihood (fixed effects logit), and random effects GLS, in the balanced and unbalanced samples. We also replicate FE analyses using SHARE weights designed to obtain a representative population of each country. These weights are calibrated to precisely reflect each country's age and gender proportions.39 All results are provided in Tables S4‐S7.

4. RESULTS

4.1. Descriptive statistics

Table 1 provides an overview of the distribution of the independent variables at different points in time in the balanced sample. Not surprisingly, the proportion of frail individuals is increasing in time as a result of aging of the population (+146 percent increase between wave 1 and wave 6). There is a similar increase over time in the levels of functional limitations (eg, +174 percent increase in prevalence of 2 + ADL limitations), chronic conditions (+134 percent increase), and, to a lesser extent, depressive symptoms (+15 percent increase). The socioeconomic indicators are stable over time.

Table 1.

Sample means of the independent variables by wave

	All waves	Wave 1	Wave 2	Wave 4	Wave 5	Wave 6	w6 vs w1 (%)
Time‐invariant characteristics
Female	0.563	0.563	0.563	0.563	0.563	0.563	0%
Secondary education	0.307	0.307	0.307	0.307	0.307	0.307	0%
Tertiary education	0.229	0.229	0.229	0.229	0.229	0.229	0%
Time‐varying characteristics
A. Sociodemographics
Age	67.5	61.9	64.2	68.5	70.5	72.4	+17%
Living with partner	0.618	0.623	0.643	0.627	0.609	0.586	−6%
Material deprivation	0.205	0.235	0.217	0.202	0.191	0.181	−23%
B. Need for care
Prefrail	0.390	0.379	0.382	0.405	0.389	0.397	+5%
Frail	0.098	0.059	0.067	0.099	0.122	0.145	+146%
Poor self‐rated health	0.303	0.220	0.284	0.322	0.333	0.354	+61%
Limit./ADL 2+	0.031	0.019	0.018	0.032	0.035	0.052	+174%
Limit./IADL 2+	0.049	0.022	0.027	0.045	0.064	0.084	+282%
Chronic conditions 2+	0.510	0.283	0.431	0.575	0.600	0.663	+134%
Depressive symptoms 3+	0.364	0.345	0.337	0.372	0.370	0.396	+15%

Reading grid: mean age has increased by 17% between wave 1 and wave 6.

Abbreviations: ADL, activities of daily living; IADL, instrumental activities of daily living.

Age‐ and sex‐standardized rates of health care use according to the frailty status are provided in Table 2. There is a strong positive correlation between frailty and hospital admission: 30.1 percent of frail individuals had at least one hospitalization in the past 12 months vs 9.7 percent among robust elders (of similar age and sex). Similarly, there is a strong positive correlation between frailty and ambulatory care: 97.5 percent of frail individuals had at least one doctor visit in the past 12 months vs 87.9 percent among robust elders. These differences seem to be more important concerning access to specialist vs primary care practitioner: There is a 20.9 percentage point difference in the probability of specialist visit among frail (vs robust) elders, as compared to a 14.0 percentage point difference in the probability of GP visit. On the contrary, there is a negative correlation between frailty and dental care use: 41.4 percent of frail elders had at least one visit to the dentist in the past 12 months vs 63.6 percent among robust elders.

Table 2.

Sex and age‐standardized rates of health care use according to frailty status

Health care use indicator	SHARE waves used	Obs.	Frailty status
			Robust		Prefrail		Frail
			Adj. rate	95% CI	Adj. rate	95% CI	Adj. rate	95% CI
Hospitalization	Waves 1, 2, 4, 5, 6	30 390	0.097	0.092‐0.102	0.166	0.159‐0.173	0.301	0.280‐0.324
Ambulatory care visit (all)	Waves 1, 2, 4, 5, 6	30 390	0.879	0.874‐0.884	0.926	0.918‐0.928	0.975	0.965‐0.980
GP only	Waves 1, 2, 4	27 480	0.797	0.799‐0.813	0.867	0.856‐0.871	0.937	0.925‐0.949
Specialist only	Waves 1, 2, 4	27 480	0.439	0.430‐0.447	0.512	0.503‐0.522	0.648	0.626‐0.671
Dentist visit	Waves 1, 2, 5, 6	26 884	0.636	0.627‐0.644	0.572	0.562‐0.581	0.414	0.389‐0.439

Abbreviations: Adj. rate, rate standardized on sex and age; 95% CI, 95 percent confidence interval.

4.2. Results of econometric models

The results of the static FE and dynamic CRE models of health care use estimated on the balanced sample are presented in Table 3 (hospital admission and ambulatory care), Table 4 (GP and specialist visit), and Table 5 (dentist visit). In all models except dental care, the sign and statistical significance of frailty are stable across the two econometric specifications. For parsimony, we only detail the estimates of the dynamic CRE models.

Table 3.

Panel data linear probability models of hospital and ambulatory care visit

Model specification	Hospital admission (1/0)				Ambulatory care visit (1/0)
	(1) Total effect of frailtya		(2) Partial effect of frailtyb		(1) Total effect of frailtya		(2) Partial effect of frailtyb
	Static FE	Dynamic CRE	Static FE	Dynamic CRE	Static FE	Dynamic CRE	Static FE	Dynamic CRE
Fried's frailty
Prefrail	0.050***	0.058***	0.038***	0.043***	0.013***	0.018***	0.008*	0.012**
Frail	0.134***	0.144***	0.090***	0.097***	0.016***	0.021***	0.005	0.009
Time‐varying controls
A. Exogenous controls
Age dummies	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Time fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Time × country	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
B. Potentially endogenous controls
Living with partner			0.004	0.007			−0.007	−0.008
Material deprivation			0.004	0.003			0.007	0.006
Poor SRH			0.084***	0.094***			0.029***	0.030***
Limit./ADL 2+			0.047**	0.042*			0.007	−0.003
Limit./IADL 2+			0.070***	0.072***			−0.011	−0.008
Chronic 2+			0.037***	0.041***			0.041***	0.047***
Depressive symptoms			0.008	0.007			0.006	0.003
C. Unobserved shocks
Lagged dep. variable		0.110***		0.097***		0.221***		0.197***
Initial condition		0.046***		0.031***		0.026***		0.016***
Time‐invariant controls
Female		−0.025***		−0.021***		0.007*		0.010**
Secondary education		0.006		0.009		0.003		0.005
Tertiary education		−0.008		0.001		0.009*		0.014***
Avg(Prefrail)c		0.075***		0.022**		0.048***		0.006
Avg(Frail)c		0.201***		0.066***		0.066***		−0.001
Avg(Partner)c				−0.003				0.015***
Avg(Deprivation)c				−0.018*				−0.017**
Avg(SHR)c				0.094***				0.026***
Avg(ADL2+)c				0.088**				−0.011
Avg(IADL2+)c				−0.018				−0.018
Avg(Chronic2+)c				0.047***				0.095***
Avg(Depress)c				0.016*				0.022***
Country fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Obs.	30 390	24 312	30 390	24 312	30 390	24 312	30 390	24 312

Abbreviations: CRE, correlated random effects; FE, fixed effects.

^a(1) total effect of frailty: model without endogenous health controls.

^b(2) partial effect of frailty: model including endogenous health controls.

^cMean of the time‐varying characteristic (Mundlak device) interpreted as between‐subject differences. Note that all the associated time‐varying regressors were mean‐centered in order to achieve this interpretation.

^d***<1%; **<5%; *<10%: Statistical significance.

Table 4.

Panel data linear probability models of GP and specialist visit

Model specification	GP visit (1/0)				Specialist visit (1/0)
	(1) Total effect of frailtya		(2) Partial effect of frailtyb		(1) Total effect of frailtya		(2) Partial effect of frailtyb
	Static FE	Dynamic CRE	Static FE	Dynamic CRE	Static FE	Dynamic CRE	Static FE	Dynamic CRE
Fried's frailty
Prefrail	0.012**	0.021***	0.008	0.012	0.027***	0.032***	0.017**	0.023**
Frail	0.021**	0.029**	0.010	0.004	0.073***	0.077***	0.046***	0.057***
Time‐varying controls
A. Exogenous controls
Age dummies	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Time fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Time × country	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
B. Potentially endogenous controls
Living with partner			−0.011	−0.008			−0.013	0.006
Material deprivation			0.009	0.017**			0.011	0.005
Poor SRH			0.037***	0.052***			0.071***	0.089***
Limit./ADL 2+			−0.011	−0.014			0.005	−0.037
Limit./IADL 2+			−0.002	−0.006			0.004	0.010
Chronic 2+			0.031***	0.018*			0.085***	0.064***
Depressive symptoms			0.003	0.002			0.006	−0.001
C. Unobserved shocks
Lagged dep. variable		0.225***		0.203***		0.206***		0.187***
Initial condition		0.027***		0.012**		0.080***		0.060***
Time‐invariant controls
Female		0.003		0.003		0.023***		0.031***
Secondary education		−0.003		−0.000		0.045***		0.047***
Tertiary education		−0.010		−0.004		0.097***		0.105***
Avg(Prefrail)c		0.062***		0.024**		0.105***		0.027*
Avg(Frail)c		0.073***		0.031**		0.158***		0.035
Avg(Partner)c				0.017**				0.030***
Avg(Deprivation)c				−0.007				−0.069***
Avg(SHR)c				0.021**				0.112***
Avg(ADL2+)c				−0.012				−0.051
Avg(IADL2+)c				−0.052***				−0.037
Avg(Chronic2+)c				0.119***				0.112***
Avg(Depress)c				0.029***				0.044***
Country fixed effects	Yes	Yes	Yes	Yes
Obs.	27 480	18 320	27 480	18 320	27 480	18 320	27 480	18 320

Abbreviations: CRE, correlated random effects; FE, fixed effects.

^a(1) total effect of frailty: model without endogenous health controls.

^b(2) partial effect of frailty: model including endogenous health controls.

^cMean of the time‐varying characteristic (Mundlak device) interpreted as between‐subject differences. Note that all the associated time‐varying regressors were mean‐centered in order to achieve this interpretation.

^d***<1%; **<5%; *<10%: Statistical significance.

Table 5.

Panel data linear probability models of dental care use

Model specification	Dental care (1/0)
	(1) Total effect of frailtya		(2) Partial effect of frailtyb
	Static FE	Dynamic CRE	Static FE	Dynamic CRE
Fried's frailty
Prefrail	−0.005	−0.010	−0.011	−0.014*
Frail	−0.012	−0.031**	−0.015	−0.031**
Time‐varying controls
A. Exogenous controls
Age dummies	Yes	Yes	Yes	Yes
Time fixed effects	Yes	Yes	Yes	Yes
Time × country	Yes	Yes	Yes	Yes
B. Potentially endogenous controls
Living with partner			0.006	0.008
Material deprivation			0.009	0.021*
Poor SRH			−0.016**	−0.018*
Limit./ADL 2+			0.015	0.002
Limit./IADL 2+			−0.044***	−0.039**
Chronic 2+			0.009	0.002
Depressive symptoms			0.027***	0.024***
C. Unobserved shocks
Lagged dep. variable		0.339***		0.336***
Initial condition		0.001		0.004
Time‐invariant controls
Female		0.027***		0.031***
Secondary education		0.077***		0.071***
Tertiary education		0.121***		0.111***
Avg(Prefrail)c		−0.025**		−0.019
Avg(Frail)c		−0.129***		−0.072***
Avg(Partner)c				0.024**
Avg(Deprivation)c				−0.073***
Avg(SHR)c				−0.033**
Avg(ADL2+)c				0.084**
Avg(IADL2+)c				−0.130***
Avg(Chronic2+)c				0.020**
Avg(Depress)c				0.022*
Country fixed effects	Yes	Yes
Obs.	26 884	20 163	26 884	20 163

Abbreviations: FE, fixed effects; CRE, correlated random effects.

^a(1) total effect of frailty: model without endogenous health controls.

^b(2) partial effect of frailty: model including endogenous health controls.

^cMean of the time‐varying characteristic (Mundlak device) interpreted as between‐subject differences. Note that all the associated time‐varying regressors were mean‐centered in order to achieve this interpretation.

^d***<1%; **<5%; *<10%: Statistical significance.

In the model of hospital use, becoming frail is associated with a total effect of 14.4 percentage point (pp., hereafter) increase in the probability of hospital use, and a 9.7 pp. increase once the effect of potentially endogenous controls has been netted out (partial effect displayed in model 2). As expected, in model 2, poor self‐rated health having 2 + ADL/IADL limitations and 2 + chronic diseases is further associated with increased hospital admission, but the marginal impact is lower compared to frailty. We find evidence of persistence in hospital admission. As expected, the initial conditions have a positive and statistically significant impact on contemporaneous hospitalization. Among the time‐invariant socioeconomic characteristics, only gender is significantly associated with hospital use: In model 2, being a female is associated with a 2.1 pp. decrease in the probability of hospital admission.

The magnitude and statistical significance of the impact of frailty transitions on ambulatory care use vary in the models (1) excluding or (2) including potentially endogenous covariates. Becoming prefrail is associated with a significant increase in ambulatory care use (total effect = 1.8 pp. increase and partial effect = 1.2 pp. increase). Becoming frail is associated with a significant effect on ambulatory care use only in model 1, with a total effect of 2.1 pp. increase (Table 3). We find evidence of persistence in ambulatory care use.

Table 4 reports the results of the analyses according to the type of medical practitioner (ie, GP or specialist). The impact of frailty transitions on care use is stronger for specialist compared to GP. Becoming prefrail or frail is significantly associated with an increased probability of GP visit, though with limited magnitude (2.1 and 2.9 pp. increase, respectively) and only in model 1 (not controlling for potentially endogenous covariates). However, becoming frail is associated with a total effect of 7.7 pp. increase—and a partial effect of 5.7 pp. increase—in the probability of specialist visit.

Among the time‐varying health controls variables, poor self‐rated health and 2 + chronic conditions are associated with increased ambulatory care use (either GP or specialist). Moreover, transition into material deprivation (within‐subject variability) is associated with higher likelihood of GP visit but no specialist visit. On the contrary, between‐subject variability in socioeconomic characteristics is significantly associated with specialist visit: Having a tertiary education is associated with a 10.5 pp. increase in the probability of specialist visit, and material deprivation is associated with a 6.9 pp. decrease in the probability of specialist visit.

The results of the models of dental care use are displayed in Table 5. Compared to previous results, the marginal impact of frailty transitions on dental care use (1) is negative, (2) differs in the static FE vs dynamic CRE model, and (3) is less impacted by the introduction of potentially endogenous variables. First, prefrailty transition is not significantly associated with dental care use in any of the models (at the 5 percent level). Second, becoming frail is significantly associated with lower dental care use only in the dynamic CRE specification. Third, the total and partial effect of frailty is of the same magnitude: Frailty transition is associated with a 3.1 pp. decrease in the probability of dental care use in model 1 and model 2. Among the potentially endogenous time‐varying characteristics included in model 2, having 2 + IADL limitations is negatively associated with dental care use, whereas depressive symptoms are positively correlated with dental care use. Moreover, there is significant persistence in dental care use: In model 2, past use increases the probability of contemporaneous use by 33.6 pp. Concerning the impact of the sociodemographic time‐invariant characteristics, being a female, having secondary or tertiary education, and living with a partner are all significantly associated with higher dental care use. On the contrary, material deprivation is negatively associated with dental care use, reflecting problems in access to dental care among economically less well‐off populations.

4.3. Results of robustness analyses

Our results are robust to competing econometric specifications: (1) estimations on the unbalanced panel, (2) use of logit link function in a conditional maximum likelihood estimation framework, and (3) weighted FE estimation using SHARE sampling weights to ensure representativeness of the surveyed population.39 Interestingly, the estimates are higher in absolute terms in alternative specifications (1) and (2) (see Tables S4 and S5 for model comparisons). For instance, becoming frail is associated with a 23.0 pp. increase (total effect) in the probability of hospitalization in the conditional logit model on the unbalanced panel (vs a 14.4 pp. increase in the dynamic CRE on the balanced panel). This shows the results of our main analyses on the balanced panel may be conservative.

5. DISCUSSION

This study provides an empirical and methodological contribution to the literature on the determinants of health care utilization among older people. To our knowledge, this study is the first to have used the five regular waves of the SHARE survey to analyze the effect of frailty transitions on a wide range health care indicators using econometric specifications that account for various sources of selection into frailty and thus are less prone to endogeneity bias. In additions, we compare the results across different specifications, (1) excluding or (2) including potentially endogenous (health) controls, the results of which are interpreted respectively as (1) the total effect or (2) the partial effect of frailty. However, care is needed in interpreting these results, as both estimates may be biased: upward for the “total effect” (because of omitted variable bias) and downward for the “partial effect” (because of endogenous controls capturing part of the relationship between frailty and the outcome). It is thus informative to present the results of both specifications, as they are likely to provide upper and lower bounds of the effect of frailty transitions on health care use. In our dynamic panel model specifications, transitions into frailty are associated with a marginal increase in the probability of hospitalization by 9.7 to 14.4 percentage points. Concerning ambulatory care use, the impact of frailty transition is economically unsignificant for GP visit, but is associated with a 5.7 to 7.7 percentage point increase in the probability of specialist visit. On the contrary, becoming frail is associated with a decrease in the probability of dental care visit by 1.5 to 3.1 percentage points. In all models, poor health, limitations in (instrumental) activities of daily living, or chronic diseases were significant predictors of health care use.

The results on the impact of frailty on hospitalization and ambulatory care use confirm previous findings. Using the first three waves of SHARE, it was shown that progression by one point on the frailty scale [0;5] was associated with an additional risk of hospitalization of about 2.1 percentage point.21 In our study, frailty was defined as scoring three or more on the frailty scale, that is, corresponding to an increase in three points (or more) on the frailty scale. Using static models estimated from the first three waves of SHARE, it was found that frailty was significantly associated with doctor and GP visits.20 Using a dynamic specification controlling for time‐invariant unobserved characteristics allowed us to provide more robust conclusions.

To our knowledge, our study is the first to assess the relationship between frailty and dental care use in a longitudinal setting. Previous work has underlined the importance of this subject, since frail elders suffer more from untreated dental issues than robust elders.22, 25 Nevertheless, our results should be interpreted with caution due to the problem of reverse causality; that is, people may become frail precisely because they do not consult a dentist and are thus more exposed to malnutrition.24 In our model, the negative estimate of the impact of frailty on dental care use may capture both the “pure” (causal) negative effect of frailty and also the negative “selection effect” into frailty (ie, the negative correlation between dentist care and frailty, see Table S6). Yet, in the absence of a plausible instrument, we cannot disentangle the two effects and our results should not be interpreted in a causal way. Still, controlling for past dental care in the dynamic specification is an important contribution to the literature, as it should theoretically allow capturing some of the sources of reverse causality. Despite this limitation, our results are consistent with the literature, showing for instance that functional limitations and poor oral health were associated with a lower probability of dental care among French institutionalized elders.40 In any case, even in the hypothetical scenario where all of the negative correlation between frailty and dental care was driven only by a selection effect, our results underline the need to increase access to dental care for frail elders and can be used to raise the awareness on that issue. The none significant effect of prefrailty may either be explained by the fact that reverse causation (prefrailty caused by absence of dental care) is lower or does not exist, or due to the fact that the “pure” negative effect of prefrailty on dentist visit is less important. This seems a reasonable assumption, as prefrails are less likely to suffer from mobility limitations associated with problems of access to dental care.

We found that socioeconomic variables (ie, living with a spouse, education, and subjective material deprivation) had little or nonsignificant impact on GP visit and hospital admission. These results are consistent with a previous analysis of health care use in the elderly German population.5 However, these variables had significant impact on dental care use. In particular, the result that material deprivation is associated with lower dental care use strengthens the need to promote better access to dental care in the elderly, for instance, by optimizing dental care insurance and/or increasing coverage.41 We found that only between‐subject differences in terms of material deprivation were associated with dental care use. This result could be partly explained by the low amount of within‐subject variability in socioeconomic status across waves. Another interpretation of the result is that habits or “predisposing” factors related to present or past economic conditions29 have higher impacts on dental care use than transitive economic shocks. In any case, that lower economic resources are associated with less access to specialists is consistent with theoretical and empirical evidence.4, 29

Our study faces several limitations. A well‐known shortcoming of the linear probability model is that the predicted probabilities are unbounded, thus making it unsuitable for making predictions. However, we were mainly interested in computing average marginal effects, and standard errors were adjusted for heteroskedasticity. We also checked the robustness of our results using nonlinear logit specifications, but the conclusion remained unchanged (Tables S4 and S5). Despite the richness of SHARE survey, the models were estimated only on a subset of the population being observed in all waves, which could create a selection bias. However, as our identification strategy relies on within‐individual variability, our estimates are consistent in the presence of selection on observables or unobservables insofar as the nonignorable nonresponse is due to time‐invariant characteristics.33 Another way to deal with attrition would be either to apply inverse probability weighting (IPW) on the unbalanced panel,34 or to use a competing‐risk model treating attrition as an endogenous variable. However, the conditional independence assumption underlying the IPW methodology could not be verified in our data (see technical Appendix S2). Moreover, absence of information regarding the date of hospitalization, ambulatory care, or dental care visit prevented joint modeling of health care use and attrition in a competing‐risk framework. Yet, several influential studies reported that attrition does not lead to serious bias (ie, affecting the magnitude of coefficients) even in the presence of large sample attrition.42, 43, 44 Another data limitation is lack of information about the reasons for hospitalization (eg, planned vs unplanned), which would be helpful distinguishing in empirical models. Finally, because of data limitations, we were not able to identify whether the use of hospital care and ambulatory care was adequate or not. Further studies should therefore explore whether frailty is correlated with unnecessary care use. Similarly, the decrease in dental care use associated with frailty may not necessarily indicate “inadequate” care/prevention, and this result should be therefore explored with caution.

6. CONCLUSION

Frailty is a multifactorial and dynamic concept. It is a nonabsorbing state, which means it is possible to design interventions to delay, if not reverse, frailty transitions. Our results suggest that delaying transitions toward frailty (eg, by promoting exercise and appropriate diet) may contribute to preventing hospital and ambulatory care uses. Further research is needed to investigate the relationship between frailty and dental care by controlling for reverse causation.

CONFLICT OF INTEREST

None to declare.

Supporting information

 

 

Sicsic J, Rapp T. Frailty transitions and health care use in Europe. Health Serv Res. 2019;54:1305–1315. 10.1111/1475-6773.13208