The relation between income and health
Examining a possible causal effect of income on the use of the general
practitioner in the Netherlands
Kayleigh Vonk
10157514
16/01/2014
University of Amsterdam
Faculty of Economics and Business
Specialization: Economics
Field: Econometrics
Abstract:
In the recent decade, the existence of a causality between income and health has been a debate. Most of the previous research concentrates on finding a causal effect of income on health, however with very different findings. This paper tries to add to the debate, by measuring health in the number of individual yearly contacts with the general practitioner (GP), rather than self-assessed health. Utilizing 9 waves of data of the DNB Household Survey, linear and nonlinear econometric models were estimated to find a possible causality running from income to health. I find a considerable and significant outcome of the income coefficient in bivariate, multivariate, linear and nonlinear models, implying a causal effect on the number of GP contacts. However, after controlling for general and health-related
characteristics at the same time, the effect of income is insignificant at the 10% level.
1. Introduction
Numerous research has been done on the existence of a causal relationship between income and health. Different studies have produced different findings on this issue. Most authors confirm the existence of a relation between income and health (see e.g. Mackenbach et al., 2004, Semyonov et al., 2013, and Larrimore, 2011.), although this does not directly imply a causality. Nevertheless, several authors verify a causality running from income to health (see e.g. Ettner, 1996, van Lenthe et al., 2013, and Fritzell et al., 2004), however differing in quantitative size of evidence. Furthermore, a few authors (see e.g. Smith, 1999 and Rahkonen et al., 2000) propose another causality, which runs from health to income. On the contrary, various authors do not find a causal relationship between income and health (see e.g. Frijters et al., 2005, Semyonov et al., 2013, Larrimore, 2011, van Doorslaer et al., 2004). The greater part of the previous research is done utilizing household income and self-assessed health (SAH) as the health proxy.
Furthermore, the previous study by van Doorslaer et al. (2004) uses the same health proxy as is used in this paper, namely the number of contacts with the general practitioner (GP). However, van Doorslaer et al. (2004) extend their study by taking an additional health proxy, opposite to this study, which is the number of contacts with medical specialists. In addition, individual income, rather than household income, is used, thus differing from this research. The authors conclude that a causal effect of income on the use of medical care is considerably more observed in the case of specialist services than in the case of the GP. In the latter case, a causality was almost unobservable.
and health, or more specific, on the existence of a causality running from income to health. I take a different perspective than the authors of most previous studies, by utilizing a different health proxy than SAH. Agreeing with the findings of Crossley and Kennedy (2002, p. 653), I believe that SAH is an unobjective measure of health. Individuals have different perceptions and might be unrealistically optimistic or pessimistic. Therefore, in this study, health is measured by the number of contacts with the GP per individual per year, similar to van Doorslaer et al. (2004). Income is measured by net equivalent household income, thus adjusted for family size. Furthermore, making use of 9 different data waves from the DNB Household Survey, I try to answer the following research question:
Does net equivalent household income has a causal effect on health measured in the number of individual yearly contacts with the general practitioner in the Netherlands?
In order to answer the research question, I perform eight different econometric models,
specified for three different dependent variables, namely the number of GP contacts, SAH and the logarithm of the number of GP contacts. Net equivalent household income is used as the foremost independent variable. Furthermore, I estimate two bivariate models, of which one takes income as a linear variable, and the other takes income as a nonlinear variable. Additionally, I include individual general and health-related characteristics to function as control variables. The method of Ordinary Least Squares (OLS) is used to carry out these econometric models.
The remainder of this paper is organized as follows. In section 2, I provide an extensive overview of the previous findings on the existence of a causality between income and health. Furthermore, I present a more detailed analysis of the related research by van Doorslaer et al. (2004). In section 3, the source and more information on the data set is lined out, as well as the methodology of this paper. In section 4, I present the findings of this research with its tables and figures. Finally, in section 5, I conclude this paper by answering the research question, providing the limitations of this study and proposing future research.
2. Previous Findings
Substantial evidence exists on a positive correlation between income and health (see e.g. Mackenbach et al., 2004, Semyonov et al., 2013, and Larrimore, 2011.). However, the
established correlation does not directly imply any causal relation between income and health. Therefore, numerous research is carried out with the aim of finding, if any, causality and its direction between income and health. This section provides an extensive overview of previous literature on the relationship between income and health. Former research is analyzed and
compared to find the strengths and limitations of those models and samples, which will be taken into account in the model of this thesis. Firstly, the findings regarding a causal effect of income on health are summarized respectively for positive evidence and insignificant or no evidence. Secondly, a research by van Doorslaer et al. (2004) that is similar to this study is lined out. Finally, I clarify my contribution to the debate on income having a causal effect on health.
2.1 Positive findings
Important evidence was found by Ettner (1996), who explored the causal effect of income on diverse health proxies: self-assessed health status (SAH), limitations on the ability to work for pay, limitations on the ability to perform daily activities, depressive symptoms, bed days, average daily alcohol consumption and alcoholic behaviours. Ettner, making use of three data sets on households, income and health in the US population, estimated two different effects. Firstly, she performed ordinary regressions of income on health, while assuming the
exogeneity of income. Results show that income is positively associated with SAH and negatively associated with work limitations, functional limitations, depressive symptoms and bed days, implying income has a causal effect on these health proxies (1996, p. 78). However, these results were likely to be biased from reverse causality, which causes correlation of income with the residual in the health equation. Secondly, she performed a two-stage IV regression to take away the possible simultaneous causality that would overstate positive effects of income on health (1996, p. 69). Surprisingly, Ettner found the extent of the income effects on health in the ordinary regressions being magnified with the IV regression (1996, p. 79), thus confirming causality running from income to health.
Secondly, a cohort study by van Lenthe et al. (2013) in the Netherlands throughout 1991 to 2013 on socioeconomic factors influencing health, proved that material (income), behavioural and psycho-social factors explain inequalities in health. Their analysis used household income and SAH as respective measures for income and health. The findings conclude that income plays a dominant role in health inequalities, implying a causality running from income to health (van Lenthe et al., 2013).
Correspondingly, Fritzell et al. (2004) find household income having a strong causal effect on SAH in Sweden, even after controlling for education, employment status and social class. Furthermore, the authors revealed a curvilinear, thus concave, association between income and health, which signifies that an income increase has a larger health impact the poorer you are in the first place (Fritzell et al., 2004, p. 32).
Further research in the shape of the relationship between household income and SAH was executed by Mackenbach et al. (2004), who conducted a cross-national investigation over seven European countries. Utilizing data from nationally representative health and income surveys, they examined men and women aged 25 years and older from Belgium, Denmark, England, Finland, France, Norway and the Netherlands. In all countries a higher household equivalent income is associated with better SAH, particularly in the middle-income range (Mackenbach et al., 2004). In the higher income ranges, curvilinearity was observed in all countries. However in the low income ranges, curvilinearity was only observed in four countries, of which one was the Netherlands. Mackenbach et al. (2004, p. 290) examined if this latter finding was due to the income measure used, however for the Netherlands the curvilinearity was observed for other income measures as well. Nevertheless, measurement biases cannot be excluded. Very small relative insignificant differences were observed between men and women (Mackenbach et al., 2004, p. 291).
Fifthly, Smith (1999) proposed another perspective of causality: the impact that health has on economic resources (p. 145). Smith studied the dual relation between health and economic status, which we already encountered in the above analyzed research. However, all of the authors only estimated the possible causality running from income to health, whereas Smith (1999) thus estimates the causal effect of income on health and, in particular, the causal effect of health on income. Thus, he considers the possibility of a good health implying a better job and a higher income. He finds clear evidence for causality running from health to income in middle and older ages. Then he stated: ‘While economic resources also appear to impact health outcomes, this may be most acute during childhood and early adulthood when health levels and trajectories are being established.’ (1999, p. 165).
Similarly, Rahkonen et al. (2000), who studied income inequalities in health in Britain and Finland, assume the possibility of both a causal association and reverse causality between income and health. They found that manual workers reported poorer health than non.-manual workers, that the unemployed reported poorer health than the full-time employed and that the level of education mattered: the lower the educational level is, the poorer SAH is (2000, pp. 34-35). These findings contribute to a causality between income and health, however, the direction could not be disentangled.
2.2 Negative findings
A recent study by Larrimore (2011) follows a similar instrumental variable approach to Ettner (see section 2.1), however produces opposing results. Although he observes a positive
correlation between income and self-reported health, testing for improvements in self-reported health statuses after increasing income did not yield any positive results. Thus, no causality was detected, which might have been due to the data sets used. Larrimore (2011) stated in the concluding part that his data sets follow people for only a relatively short time, wherefore he was not able to observe the effects of higher income over a lifetime. Thus, although this study found only very limited evidence in the short run, the possibility of larger long-term effects should not be ruled out. The data was obtained from the SIPP set, following US individuals from 1992 through 2005, and the CPS set, following US individuals from 1996 through 2009 (Larrimore, 2011).
Similarly, Semyonov et al. (2013) found very limited support for causality running from income to health. They utilized data for 16 national samples, including the Netherlands, to find evidence for a wealth-health gradient, with wealth consisting of wealth and income. The cross-national study used household income as a measure of income and a ‘Physical Health Index’ as a measure of health (Semyonov et al., 2013). ‘The Physical Health Index is constructed as a sum of health impairments from a list that pertains to limitations with activities of daily living, limitations in mobility, problems with arm and fine motor function, chronic diseases and illness symptoms.’ (Semyonov et al., 2013, p. 12). Of the 14 European countries examined, only Austria did not derive a positive correlation between income and health. Israel and the US also showed a positive correlation between income and health. However, where significant support is established for a causal relation between wealth and health, there is only very limited support for an income-health gradient (Semyonov et al., 2013). These two findings might have been caused by this research’ sample, in which at least one household member should have been 50 years or over. Respondents at that age had accumulated wealth and did not consider current income as important any more.
Furthermore, Frijters et al. (2005) tried to add to the debate by providing new evidence on the causal effect of income on health, using a panel of East and West Germans in the years following the reunification (2005, p. 1015). The causality was observed, however the
quantitative size of the effect was small.
2.3 A related study
Van Doorslaer et al. (2004) examined the extent to which health care use is unequally distributed by (individual) income. This cross-national study investigates the use of GP and specialist services in 12 European countries, exploiting data collected in the 1996 wave of the European Community Household Panel. Van Doorslaer et al. (2004) state that it is important
to model medical care use in nonlinear models, as this is well-known to be most appropriate. They find that in all European countries the need and use of the GP are more concentrated among the lower income population, although often the use distribution is even more pro-poor than the need distribution. The need distribution is the distribution that represents the
necessity of a certain medical care, whereas the use distribution represents the actual use of that medical care. According to van Doorslaer et al. (2004), the most essential variables that contribute to the more pro-poor use of the GP are not income, but rather other social
economic factors, as low education and unemployment. Thus, individuals who are low educated might be less informed on certain health problems, or have a less rational view on those health problems, which causes them to visit a GP sooner and more often. Moreover, the GP might influence the decision of repeated visits more expectedly rather than the patient himself. However, findings regarding specialist services are radically different: while the need for specialist services is greater for the lower income population, the higher income
population is more likely to report the use of a specialist (van Doorslaer et al., 2004). Thus individuals with a higher income utilize medical specialists sooner and more often than the lower income population, although the lower income population is in greater need for this medical care. This pro-rich distribution is weakest observed in the Netherlands, Belgium and Luxemburg. Furthermore, van Doorslaer et al. (2004) state that the decision of this use is more patient-initiated than driven by the GP and therefore the contribution of income is clearly more significant in explaining the inequalities in the use of specialist services. Thus, a causal effect of income on the use of medical care is considerably more observed in the case of medical specialist services than in the case of the GP.
2.4 Contribution
Summing the provided literature, we can conclude that a positive correlation between income and health is unquestionable, however causality and the direction of causality have not yet been clearly established. Furthermore, testing for causality running from income to health has yielded various results. Therefore, this research aims at supplying new evidence on the causal effect of income on health. It uses household income as the income measure, which has been used in most of the former research on this relationship. On the contrary of most of the
previous literature, this study does not make use of SAH as the proxy for health. According to Crossley and Kennedy (2002, p. 653), who extensively looked into SAH status, there is considerable measurement error in individuals self assessment of health, which can lead to unreliable output. Larrimore (2013, p. 702) agreed on this: ‘Self-reported health status is a
subjective measure of health, as it may vary in accordance with the respondent’s own
assessment of the scale. Therefore, the possibility exists that two people with the same health status could report different health ratings based on differences in their perceptions of good health.’ On the other hand, Fritzell et al. (2004, p. 7) indicated to be using SAH because it had been proven to be a robust and reliable measure of a person’s health status. Accordingly, there are two different perspectives on SAH as a health proxy, and as I agree on the unobjectivity of SAH, I introduce a proxy that has not been used much before. I measure health by the number of contacts with the GP per year, however I estimate my model based on SAH as well to capture the significance of the different proxies. The related study by van Doorslaer et al. (2004) in section 2.3 found little evidence of income contributing to the pro-poor distribution of GP use. They utilized the use of GP contacts as well, however the econometric models included individual income, rather than household income. I do replicate this research, however I take into account their findings and compare them with my own.
From the related literature I hypothesize the relation between income and health in the Netherlands to be causal, running from income to health.
3. Data and Methodology
This section provides information about the data and methodology of this paper and is divided into three subsections. The first subsection illustrates the source of the data sets used and the questionnaires involved, whereas the second subsection specifies the variables that are used in the research. The final subsection describes the methodology of the research.
3.1 Data sets and questionnaires
The data originate from the DNB Household Survey (DHS) that is part of the outstanding research institute CentERdata located at the Tilburg University. The purpose of the DHS is to study the economic and psychological determinants of the saving behaviour of households (CentERdata, n.d.). The first wave of the DHS was launched in 1993 and has been repeated every year since, resulting in a current number of 20 waves. This study only uses waves 2004-2012 for reasons provided in section 3.2. The DHS interviews all members aged 16 or over of 2000 Dutch households and consists of six questionnaires on different subjects:
1. General Information on the Household 2. Household and Work
3. Accommodation and Mortgages 4. Health and Income
5. Assets and Liabilities
6. Economic and Psychological Concepts.
For the purpose of this research the data of subjects 1 and 4 are utilized. The general information on the household data is the only set that contains every single household that completed any of the other questionnaires, and even includes the household members aged under 16. This resulted in a total number of observations of 39868. However, because the response on the health and income survey was lower (44.52%) and, in addition, only interviewed members aged 16 or over, the total of observations in this study diminished to 17750 when the data sets were merged.
Regarding the general information section, the household members are asked to specify their primary occupation, the degree of urbanization of the town/city of their
residence, the Dutch province of their residence, the composition of their household, the type of accommodation of their residence and the composition of their household. Since allowing for differences on all these variables and truly randomizing the participating households through the entire country, the surveys are representative samples for the Dutch population. Furthermore, in the general information section several questions are stated on the unique characteristics of the individual, like year of birth, gender, highest attended and completed education level and position within the household.
Concerning the health and income questionnaire, the household members are
questioned on several topics related to health and income. In the health section, they are asked to specify some individual physical characteristics, their SAH, their life expectancy, if he/she smokes, if he/she (on average) drinks more than 4 glasses of alcohol per day, the number of contacts with the GP and the number of days being ill for that year. In the income section, respondents are asked to indicate a variety of individual income measures, taxes, premiums and profits, household net income and if they have medical insurance. In the Netherlands, every single resident is obliged to have a (basic) medical insurance, thus for every respondent GP contacts are covered. Therefore, I do not need to include an insurance variable in my model. Both questionnaires can respectively be found in Appendices 1 and 2.
3.2 Variable specification
The dependent variable, the number of GP contacts per year, is constructed from a question in which the respondent is asked to specify how many times the respondent did contact their GP, by phone or personally, about their own health that year. In this research, it is assumed that
contacts with the GP by phone are entirely health-related, thus not to make an appointment. This is assumed to facilitate the data, as I can aggregate the number of GP contacts of three sub questions to find one absolute number per individual per year. Important to note is that the question regarding the GP contacts was only included in the questionnaire on health and income since 2003, wherefore it would be intuitive to use only the waves of 2003 until 2012. However, I also excluded wave 2003, finding that in the survey of 2004 a change was made concerning the most important question on measuring household income, which I illustrate more in the next paragraph. Furthermore, as has been stated in the literature review, SAH is a more common proxy for health and was used in many former studies. Though I think this proxy is unobjective, I also conduct the analysis using SAH as the dependent variable for the purpose of comparing both of my proxies and their results. Individuals are asked to indicate their SAH as excellent, good, fair, not so good or poor. Both health proxies are measured per individual in the analysis, as it is unintuitive and erroneous to assume health is equivalent for each household member.
Naturally, the main independent variable in this research is income in Euros. Income is measured by the net household income in, thus the income of all household members
combined after deduction of taxes and social security benefits. In the survey, respondents are asked to indicate the total net income of their household over that year in euros. However, if respondents answer ‘I don’t know’, they are referred to a next question that asks to indicate the range of the total net household income (see Appendix 2 for the specification of these income intervals). Since the income intervals were only defined in this form in 2004, whereas previous waves made use of different intervals, I decided to exclude 2003 and use data waves 2004-2012. In the research, I aggregate the two income questions by replacing the dummy value (1 to 11) with the average of the interval. The response rate of the question that asks to indicate the actual net household income was 57.1% against a rate of 42.9% for the question about the intervals. As a sufficient percentage of the sample stated the exact net household income number, I think it is valid to aggregate the two income measures. Furthermore, I control for household composition in the sense that the income measure will be extended to net equivalent household income. Thus, similarly to Larrimore (2011) and Mackenbach et al. (2004), I adjust income for household size by dividing by the square root of the number of household members.
In addition to the key variables, I include several control variables in the analysis. Firstly, I include education completed, distinguishing between a high and low level, where high equals HBO (high tertiary) or university, and low equals primary education, secondary
education, or MBO (low tertiary). Thus, I assume that individuals have a higher chance on a good job if they completed high tertiary education, as only primary or secondary education is not considered to be sufficient enough and the low tertiary education level really differs from the high level in the Netherlands. Secondly, I take account of the fact of living with a partner, of the age of the respondent, and of gender. Thirdly, I include an unemployment dummy, being equal to 1 if the individual is unemployed. I assume students to be unemployed, as, in general, students do not have a job contract. Fourthly, I control for several health-related characteristics of the individual, namely weight, if the individual smokes and if the individual drinks more than 4 glasses alcohol per day.
Finally, I add a variable that takes in to account if the respondent is suffering from a long illness, disorder, handicap or the consequences of an accident. These control variables might be accountable for explaining parts of the dependent variable and were for the same reason included in previous research as well (see e.g. Fritzell et al., (2004), Rahkonen et al., (2000), Larrimore, (2011), Semyonov et al., (2013) and van Doorslaer et al., (2004)). If I would not include them, I might experience omitted variable bias and/or endogeneity bias. In Table 1 an overview of all variables is given, including the number of observations, mean and standard deviation. In this table it is also made clear which variables are dummy variables and their values.
Concluding this subsection, assuming health and income are positively related, this research’ health proxy implies a negative relation with income. This follows from reasoning that the health proxy, the number of GP contacts per individual per year, is thought to be higher if health is worse. This means that if aiming at establishing a causality running from income to health, the econometric models should produce a negative coefficient of income. 3.3 Methodology
In this research I use two different questionnaires from nine waves (2004-2012). I apply a 1-to-1 merge on the number of the household and the number of the household members to aggregate the general information data with the income and health related data per individual. Then, I append the nine waves to find a combined data set, which provides the basis for my examination.
Firstly, I carry out a bivariate linear regression of health on income (see Table 2, Model 1). Then, I perform the regression, while controlling for respectively individual general characteristics only (see Table 2, Model 2) and individual health-related characteristics only (see Table 2, Model 3). Subsequently, I include both (see Table 2, Model 4). In Table 2 an overview of the different bivariate and multivariate econometric models is given.
TABLE 1
Descriptive statistics, variable specification
Variable
N % Mean Standard deviation
GP Contacts per year Total 15941 100 3.283 5.456
Self-Assessed Health Total 17750 100 2.142 0.711
1 if Excellent 2287 12.88 2 if Good 11614 65.43 3 if Fair 3040 17.13 4 if Not so good 654 3.68 5 if Poor 155 0.87
Net Equivalent Household Income in € Total 17750 100 19906.56 25703.02
Gender Total 17750 100 0.54 0.4998 1 if Male 9592 54.04 0 if Female 8158 45.96 Age Total 17750 100 56.12 15.938 Partner Total 17750 100 0.782 0.413 1 if Yes 13884 78.22 0 if No 3866 21.78
Completed Education Total 17722 100 1.364 0.481
1 if High Education 6459 36.45 0 if Low Education 11263 63.55 Unemployed Total 17750 100 0.476 0.499 1 if Yes 8454 47.63 0 if No 9296 52.37 Weight Total 17750 100 79.12 16.637
Smoking, every day Total 17750 100 0.166 0.373
1 if Yes 2995 16.87
0 if No 14795 83.35
Drinking >4 alcoholic drinks per day Total 17750 100 0.058 0.234
1 if Yes 1028 5.79
0 if No 16722 94.21
Long illness, disorder or handicap Total 17750 100 0.258 0.437
1 if Yes 4572 25.76
0 if No 13178 74.24
Secondly, I carry out a bivariate nonlinear regression, taking the logarithm of income (see Table 2, Model 5). Lokshin & Ravallion (2004), among others, prove that household income is nonlinear. Thus taking the logarithm of income in the same model, should increase the validity of the analysis. I test for this specification against the linear model by performing
the whole analysis again using the linear-logarithmic model.
Finally, regarding the dependent variable, I estimate for both the number of GP contacts per individual per year and the SAH. This is done for all the estimations in the remainder of this research. Thus the dependent variable GP contacts is replaced by SAH for every formula. Furthermore, as van Doorslaer et al. (2004) have stated in their paper, which is related to this research, medical care is well-known to be modeled most appropriately as a nonlinear variable. Therefore, I also take the logarithm of the number of GP contacts as a dependent variable to find more accurate results. These models are not added to Table 2, as the dependent variable GP contacts is simply replaced by its logarithm. In the remainder of this paper I will use the terms specification 1, 2 and 3 for respectively GP contacts, SAH and the logarithm of GP contacts as the dependent variable. All econometric models are estimated making use of the method of OLS.
4. Results
In this section an overview and analysis of the results are provided. To start with, I provide the correlations between income and the different dependent variables. Secondly, I analyze the bivariate linear models. Thirdly, I include the general and health-related controls and discuss their effects. Fourthly, I take the nonlinear specification including the logarithm of GP contacts as dependent variable in account. Fifthly, I consider the logarithm of net equivalent household income. Finally, I will evaluate the significance of the main independent variable in this study, trying to establish evidence for a causality. I extensively describe the test results of the research, referring to the output of the econometric models for the three different specifications, represented in Table 3, 4 and 5.
Firstly, it is important to look at the correlations between net equivalent household income and the different dependent variables: the number of GP contacts, SAH and the logarithm of GP contacts. Respectively the correlations are -0.0277, -0.0257 and -0.0393. As it can be observed, the relationships are all negatively signed and not of large quantitative size. This means that the direction of the relationship can be interpreted as follows: a very weak downhill linear relationship.
Secondly, looking at regression output of the bivariate linear models (see Table 3, 4), it becomes clear that net equivalent household income has a considerable, negatively signed and significant effect on both the number of GP contacts and SAH. This implies that if income increases, the number of GP contacts decreases as well as does the dummy SAH, which means that health improves. However, the models are likely to suffer from omitted
variable bias, as they both report very low R² values. Therefore, two multivariate linear models are obtained by adding controls for individual general and health-related
characteristics. TABLE 2 Econometric Models 1 2 3 4 5 6 7 8
The eight econometric models are also estimated for SAH and the logarithm of GP contacts as dependent variables, however are not included in this table. SAH (1 = excellent, 2 = good, 3 = fair, 4 = not so good, 5 = poor), gender (1 = male, 0 = female), partner (1 = partner, 0 = no partner), education (2 = high, 1 = low), unemployed (1 = unemployed, 0 = employed), disorder (1 = disorder, 0 = no disorder), smoking (1 = smoking, 0 = not smoking) and alcohol (1 = >4 glasses alcohol on average, 0 = ≤4 glasses alcohol on average) are dummy variables.
Thirdly, the effects of the individual general characteristics are analyzed for specification 1 and 2. In model (2) the quantitative effect of income on the number of GP contacts decreases compared to model (1), whereas the similar model estimating on SAH reports a slightly larger income effect compared to model (1). The significance of the main variable, net equivalent household income, in this specification, as well as in specification 2 and 3 will be discussed at the end of this section. In specification 1 the effects of being male, living with a partner, having completed high education and being employed are significant at
the 1% level in model (2), as well as in model (4), with the exception of being male, which is only significant at the 10% level. The effects are quite large, and indicate a better health. The effect of age is unexpectedly small in either model and only significant at the 10% level in model (4). Regarding specification 2, a conspicuous observation is made. Specification 2 reports p-values of 0.000 for every coefficient in every single model estimated on SAH. Although the R² of the models in specification 2 have a higher value than the R² of the models in specification 1 and 3, this specification 2 could be suffering from multicollinearity.
However, multicollinearity only has implications for individual estimators, not for the complete estimation, wherefore it is still possible to have a proper R². From another
perspective, it is possible that all the models are simply performing very well and represent a causality between income and health, even after controlling for general and health-related characteristics. To clarify this peculiarity, I tested the models for multicollinearity. However, no evidence for such bias was established. Thus, in the remainder of this paper, the models of specification 2 will be assumed to be working perfectly.
Furthermore, looking at the health-related controls, specification 1 and specification 2 are subsequently analyzed. In the first specification, weight, smoking, drinking more than 4 glasses of alcohol per day on average and suffering from a long illness, disorder or handicap are positively signed, implying to deteriorate health. However, drinking is only significant at a 10% level, where smoking appears to be insignificant at the 1% (model (3)) and 5% (model (4)) level in specification 1. Suffering from a long illness, disorder or handicap has an enormous and significant (at the 1% level) effect on the number of GP contacts per year, which is quite intuitive. If an individual is suffering from a long illness, disorder or handicap it is insightful to have more frequent contact with the GP. In specification 2, weight, smoking and drinking have similar quantitative effects on SAH as they had on the number of GP contacts, and are all significant at the 1% level. Suffering from a long illness, disorder or handicap has, among the other variables, the largest effect on SAH.
Subsequently, it is necessary to study the effects of specification 3 in Table 5, which takes the logarithm of GP contacts as dependent variable, in line with the study by van
Doorslaer et al. (2004). The foremost independent variable, net equivalent household income, has a similar considerable effect as in specifications 1 and 2. The other independent variables report the same significance levels as in specification 1, with the exception of smoking and drinking. Whereas smoking was significant at the 5% level in model (3) and at the 10% level in model (4), in specification 3 it is insignificant in both models. Drinking, however, is significant at the 1% level in model (3) and (4), which was not the case in specification 1. As
the R² of this specification has a higher value than specification 1, this model is not rejected. Nevertheless, the change in the R² is small as such, that I will not state to completely confirm the nonlinearity of medical care with van Doorslaer et al. (2004).
Moreover, I consider the models in the logarithmic form of net equivalent household income. The output is extremely comparable to the linear output, apart from the coefficient on income, as income now is measured in a logarithm. However, quantitative and significance effects are similar, implying that taking the logarithm of net equivalent household income does not add to the validity of the model. Comparing the R² of the linear and the linear-logarithmic models to validate this statement, I observe that the all the linear-linear-logarithmic models report either a lower or equal R². Therefore, this study does not confirm the
nonlinearity of household income that Lokshin & Ravallion (2004) found, however possibly due to bias in the data selection or model specification.
Finally, aiming at establishing a causality running from income to health, I analyze the effect of net equivalent household income and its p-values. In all three specifications, the effects are negatively signed, as expected, but substantial. This applies to either of the specifications and either of the econometric models. In specification 1, the effect diminishes with a small amount after controlling for all individual general and health-related variables (model (4)), whereas in specification 2 the effect of net equivalent household income increases when controlling for either individual general characteristics (model (2)) or for individual health-related characteristics (model (3)), however, slightly decreases when
controlling for both (model (4)). In specification 3, the net equivalent household income effect is lower in model (1) and model (3) compared to the effect in model (2) and model (4). Thus, controlling for health-related characteristics decreases the effect of net equivalent household income on health, which is intuitive. Additionally, examining the p-values belonging to net equivalent household income in every single econometric model, it can be concluded that the hypothesis of this paper should be accepted. In both specifications 1 and 3, when respectively GP contacts and the logarithm of GP contacts are used as dependent variables, income is reported to be significant at the 1% level in model (1) and model (3). Furthermore, in model (2), income is significant at the 5% level, however, in model (4) income is not significant at any level. These findings imply that there is a small, but causal effect of income on the number of GP visits. However, this statement should be questioned after controlling for all general and health-related characteristics, when income is insignificant. The probability exists that the models still suffer from omitted variable bias or endogeneity of income, as some of the literature stated, or some selection bias. However, the results are in line with the findings
of van Doorslaer et al. (2004): a small, but significant effect of a causality running from income to GP contacts.
TABLE 3
Effects variables on GP contacts
Dependent Variable Models
GP contacts (1) (2) (3) (4) (5) (6) (7) (8) Independent Variables Income -0.00000766*** -0.00000444** -0.00000863*** -0.00000345 - - - - (-0.00000219) (0.00000224) (0.00000213) (0.00000218) Log (income) - - - - 0.027 0.0652*** -0.0723*** 0.0284 (0.0171) (0.0205) (0.0182) (0.0201) Age - 0.0178*** 0.0204*** 0.0055* - 0.0104*** 0.0236*** 0.0018 (0.0031) (0.0027) (0.003) (0.0036) (0.003) (0.0035) Male - -0.9791*** -1.3451*** -1.1559*** - -0.9507*** -1.3743*** -1.1363*** (0.0875) (0.0898) (0.0906) (0.0879) (0.09) (0.0914)
Living with a partner - -1.1613*** - -0.8098*** - -1.1703*** - -0.8139***
(0.1037) (0.1019) (0.1037) (0.102) Education - -0.5214*** - -0.4423*** - -0.5927*** - -0.4858*** (0.0911) (0.0892) (0.0907) (0.889) Unemployed - 1.1798*** - 0.9429*** - 1.3407*** - 1.02*** (0.0984) (0.0972) (0.1062) (0.1044) Weight - - 0.0206*** 0.0231*** - - 0.022*** 0.0225*** (0.0026) (0.0026) (0.0027) (0.0027) Smoking - - 0.2474** 0.1915* - - 0.2828** 0.1874* (0.1135) (0.1141) (0.1136) (0.1142) Drinking - - 0.3469* 0.31* - - 0.3374* 0.2967 (0.1808) (0.1799) (0.1807) (0.1799) Disorder - - 2.945*** 2.7377*** - - 2.9444*** 2.739*** (0.0952) (0.0961) (0.0952) (0.0961) Constant 3.4357*** 3.4239*** 0.5208** 1.3762*** 3.0379*** 3.1038*** 0.7338*** 1.2706*** (0.0613) (0.1846) (0.2491) (0.2664) (0.1611) (0.2067) (0.2577) (0.2743) Observations 15941 15916 15941 15916 15941 15916 15941 15916 R² 0.0008 0.0414 0.0836 0.0945 0.0002 0.0418 0.084 0.0945
TABLE 4
Effects variables on SAH
Dependent Variable Models
SAH (1) (2) (3) (4) (5) (6) (7) (8) Independent Variables Income -0.000000711*** -0.000000869*** -0.000000976*** -0.000000668*** - - - - (0.000000208) (0.00000021) (0.000000183) (0.000000185) Log (income) - - - - 0.0173*** 0.0144*** -0.0051** 0.0072*** (0.0021) (0.0025) (0.002) (0.0022) Age - 0.0061*** 0.0047*** 0.0032*** - 0.0044*** 0.0049*** 0.0023*** (0.0004) (0.0003) (0.0003) (0.0004) (0.0003) (0.0004) Male - -0.0493*** -0.0907*** -0.0689*** - -0.0423*** -0.0928*** -0.0636*** (0.0107) (0.0101) (0.0102) (0.0107) (0.0102) (0.0103)
Living with a partner - -0.2028*** - -0.1144*** - -0.204*** - -0.1149***
(0.0126) (0.0115) (0.0126) (0.0115) Education - -0.0787*** - -0.0582*** - -0.094*** - -0.0677*** (0.0111) (0.01) (0.0111) (0.01) Unemployed - 0.1623*** - 0.1001*** - 0.1978*** - 0.1185*** (0.0119) (0.0108) (0.0123) (0.0117) Weight - - 0.0031*** 0.0034*** - - 0.0032*** 0.0032*** (0.0003) (0.0003) (0.0003) (0.0003) Smoking - - 0.1171*** 0.1069*** - - 0.1202*** 0.1059*** (0.0128) (0.0128) (0.0128) (0.0128) Drinking - - 0.1134*** 0.1086*** - - 0.112*** 0.1059*** (0.0204) (0.0203) (0.0205) (0.0203) Disorder - - 0.7086*** 0.6844*** - - 0.7093*** 0.6847*** (0.0109) (0.011) (0.012) (0.011) Constant 2.1565*** 1.9542*** 1.4908*** 1.6089*** 1.985*** 1.8857*** 1.5028*** 1.5839*** (0.0068) (0.0225) (0.0281) (0.03) (0.0195) (0.025) (0.029) (0.0308) Observations 17750 17722 17750 17722 17750 17722 17750 17722 R² 0.0007 0.0661 0.2359 0.2456 0.004 0.067 0.235 0.2455
TABLE 5
Effect variables on the logarithm of GP contacts
Dependent Variable Models
Log (GP contacts) (1) (2) (3) (4) (5) (6) (7) (8) Independent Variables Income -0.00000164*** -0.000000856** -0.00000153*** -0.000000626* - - - - (0.00000037) (0.000000357) (0.000000351) (0.000000358) Log (income) - - - - -0.0032*** 0.004 -0.0157*** -0.0009 (0.0031) (0.0035) (0.0031) (0.0034) Age - 0.0024*** 0.0033*** 0.0009* - 0.0019*** 0.004*** 0.0008 (0.0005) (0.0004) (0.0005) (0.0006) (0.0005) (0.0006) Male - -0.1771*** -0.2457*** -0.2139*** - -0.1759*** -0.2503*** -0.2141*** (0.0145) (0.0146) (0.0147) (0.0145) (0.0146) (0.0148)
Living with a partner - -0.1561*** - -0.0885*** - 0.157*** - -0.0883***
(0.0168) (0.0162) (0.0168) (0.0162) Education - -0.1106*** - -0.102*** - -0.1185*** - -0.1057*** (0.015) (0.0144) (0.0149) (0.0143) Unemployed - 0.197*** - 0.1371*** - 0.2086*** - 0.1388*** (0.0165) (0.016) (0.0175) (0.0169) Weight - - 0.0032*** 0.0034*** - - 0.0035*** 0.0034*** (0.0004) (0.0004) (0.0004) (0.0004) Smoking - - 0.0052 -0.0073 - - 0.0117 -0.0063 (0.0188) (0.0188) (0.0188) (0.0188) Drinking - - 0.0858*** 0.0805*** - - 0.0832*** 0.0788*** (0.03) (0.0299) (0.03) (0.03) Disorder - - 0.5271*** 0.4992*** - - 0.5263*** 0.4999*** (0.0146) (0.0148) (0.0146) (0.0148) Constant 1.0731*** 1.0661*** 0.5749*** 0.7102*** 1.0694*** 1.0415*** 0.6358*** 0.7112*** (0.0103) (0.031) (0.0412) (0.044) (0.0293) (0.0358) (0.0436) (0.046) Observations 12702 12683 12702 12683 12702 12683 12702 12683 R² 0.0015 0.052 0.1265 0.1375 0.0001 0.0517 0.1269 0.1373
5. Conclusion
In this paper, I aim to add to the debate on the existence of causality running from income to health. Using 9 data waves of the DNB Household Survey, I analyze household members in the Netherlands, trying to answer the research question:
Does net equivalent household income has a causal effect on health measured in the number of individual yearly contacts with the general practitioner in the Netherlands?
I utilize the data to estimate eight different models in three different specifications. Specification 1 uses GP contacts as dependent variable, specification 2 uses SAH and specification 3 uses the logarithm of GP contacts. The three specifications estimate eight different econometric models, consisting of four models taking net equivalent household income as a linear variable and four models taking net equivalent household income as a nonlinear variable. I perform a bivariate regression and multivariate regressions respectively controlling for age, gender, partnership, education and employment status, and controlling for weight, smoking, drinking and suffering from a long illness, disorder or handicap and a final regression controlling for all these variables.
Concluding from the analysis, net equivalent household income does have a causal effect on health measured in the number of yearly contacts with the general practitioner per individual in the Netherlands. Considerable, significant effects of net equivalent household income on the different health proxies are observed in model (1), (2) and (3). However, in specification 1 and 3 this significance disappears completely when controlling for general and health-related characteristics (model (4)). Thus, a causality is established in most of the models, except in the models that controls for both general and health-related characteristics. Therefore, more research that will clarify these results should be carried out in the future. In specification 2, significance is established in every model.
5.1 Limitations
In this study health is measured by the number of GP contacts per individual per year instead of SAH, for the purpose of unobjectivity of the latter. However, GP contacts itself is a quantitative rather than qualitative measure, as a visit to one general practitioner is not equal to a visit to another general practitioner. Therefore, this study might be suffering from measurement error.
Furthermore, in the Netherlands every single individual is obliged to be insured for basic medical care, which includes the GP, but excludes medical specialists. Therefore, the lower income population does not have a financial barrier when deciding to contact a GP. If,
for example, I would have examined the effect of net equivalent household income on the contacts with medical specialists, I might have found more evidence for the causality running from income to health.
Moreover, even after including several controls, this study might endure omitted variable bias and/or endogeneity of income. Several authors reasoned that income will have a causal effect on health, however, that additionally a better health could imply better chances on a good job and a high income.
5.2 Future research proposals
Aiming at further clarifying the findings of this paper, it is important to measure health as well in contacts with medical specialists. Especially in the Netherlands, because of the obliged insurance for the GP, this different health proxy might notably increase the evidence for a causality running from income to health. Furthermore, more health-related controls need to be added, like physical exercise, BMI, and stamina, among others. Possibly, an IV regression could be carried out, to test for the dual causality between income and health.
5.3 Acknowledgement
Appendix
1. General Information on the Household, complete questionnaire
1. Year of birth of the respondent 2. Sex of the respondent
1 male 2 female
3. The respondent’s position in the household 1 head of the household
2 spouse
3 permanent partner (not married) 4 parent (in law)
5 child living at home 6 housemate
7 family member or boarder
4. Highest level of education attended (regardless of certificate/diploma) 1 (Voortgezet) speciaal onderwijs / (continued) special education 2 Kleuter-, lager- of basisonderwijs / kindergarten/primary education
3 Voorbereidend middelbaar beroepsonderwijs (VMBO) / pre-vocational education 4 HAVO/VWO / pre-university education
5 MBO of het leerlingwezen / senior vocational training or training through apprentice system
6 HBO (eerste of tweede fase) / vocational colleges 7 Wetenschappelijk onderwijs WO / university education 8 Did not have education (yet)
9 other sort of education/training 5. Highest level of education completed
1 (Voortgezet) speciaal onderwijs / (continued) special education 2 Kleuter-, lager- of basisonderwijs / kindergarten/primary education
3 Voorbereidend middelbaar beroepsonderwijs (VMBO) / pre-vocational education 4 HAVO/VWO / pre-university education
5 MBO of het leerlingwezen / senior vocational training or training through apprentice system
6 HBO (eerste of tweede fase) / vocational colleges 7 Wetenschappelijk onderwijs WO / university education 8 Did not have education (yet)
9 other sort of education/training 6. Primary occupation of the respondent 1 employed on a contractual basis 2 works in own business
3 free profession, freelance, self-employed 4 looking for work after having lost job 5 looking for first-time work
6 student
7 works in own household
8 retired [pre-retired, AOW, VUT] 9 (partly) disabled
10 unpaid work, keeping benefit payments 11 works as a volunteer
12 other occupation
7. Number of household members 1 1 person 2 2 people 3 3 people 4 4 people 5 5 people 6 6 people 7 7 people 8 8 people 9 9 people or more
8. Number of children in the household 0 none 1 1 child 2 2 children 3 3 children 4 4 children 5 5 children 6 6 children 7 7 children 8 8 children 9 9 children or more
9. Degree of urbanization of the town/city of residence 1 very high degree of urbanization
2 high degree of urbanization 3 moderate degree of urbanization 4 low degree of urbanization 5 very low degree of urbanization 10. Region
1 Three largest cities 2 Other west 3 North 4 East 5 South 11. Province 20 Groningen 21 Friesland 22 Drenthe 23 Overijssel 24 Flevoland 25 Gelderland 26 Utrecht 27 Noord-Holland 28 Zuid-Holland 29 Zeeland 30 Noord-Brabant 31 Limburg
12. Composition of the household. The respondent: 1 is living by himself/herself
2 is living together with partner, no child(ren) living at home 3 is living together with partner, child(ren) living at home 4 is living without a partner, but with child(ren)
5 other
13. Are you the person who is most involved with the financial administration of the household? By financial administration we mean making the payments for
rent/mortgage, taking out loans, taking care of tax declarations, etc. 0 no
1 yes
14. Are you the main wage earner of the household?
The main wage earner is the person with the highest income. 0 no
1 yes
15. Is there a partner present in the household? 0 no 1 yes 16. Type of accommodation 1 owner-occupied property 2 rented house/flat 3 subrented house/flat 4 free accommodation 9 unknown
2. Questionnaire Health and Income, selected questions
The next questions concern two topics: health and income over 2011.
1. How much do you weigh, without clothes and shoes? Give your answer in whole kilos.
2. In general, would you say your health is: 1 excellent
2 good 3 fair
4 not so good 5 poor
3. Do you suffer from a long illness, disorder, or handicap; or do you suffer from the consequences of an accident?
1 yes 2 no
4. Do you smoke cigarettes at all? 1 yes, every now and then 2 yes, every day
3 no
5. About how many cigarettes do you smoke a day? 1 less than 20 cigarettes a day
2 at least 20 cigarettes a day
6. On average, do you have more than four alcoholic drinks a day? 1 yes
2 no
7. How many times did you contact your general practitioner about your own health in 2011?
1 contact by phone: x times
2 visit to your general practitioner: x times 3 visit of general practitioner to you: x times
8. What is the total net income for your household in 2011?
The total net income for your household is the net income of all household members combined. Net income means the income after deduction of taxes and social security benefits.
amount -9 don’t know
9. Please indicate about how much the total net income of your household was over the period 1 January 2011 through 31 December 2011.
1 less than 8.000 euro
2 between 8.000 euro and 9.500 euro 3 between 9.500 euro and 11.000 euro 4 between 11.000 euro and 13.000 euro 5 between 13.000 euro and 16.000 euro 6 between 16.000 euro and 20.000 euro 7 between 20.000 euro and 26.000 euro 8 between 26.000 euro and 38.000 euro 9 between 38.000 euro and 50.000 euro 10 between 50.000 euro and 75.000 euro 11 more than 75.000 euro
References
Crossley, T. F., & Kennedy, S. (2002). The reliability of self-assessed health status. Journal
of Health Economics, 21, 643-658.
DNB Household Survey CentERdata. Retrieved from: http://www.centerdata.nl/en/survey research/dnb-household-survey-dhs
Ettner, S. L. (1996). New evidence on the relationship between income and health. Journal of
Health Economics, 15, 67-85.
Frijters, P., Haisken-DeNew, J.P., & Shields, M.A. (2005). The causal effect of income on health: Evidence on the German reunification. Journal of Health Economics, 24(5), 997-1017.
Fritzell, J., Nermo, M., & Lundberg, O. (2004). The impact of income: assessing the relationship between income and health in Sweden. Scandinavian Journal of Public
Health, 32, 6-16.
Larrimore, J. (2011). Does a higher income have positive health effects? Using the earning income tax credit to explore the income-health gradient. The Milbank
Quarterly, 89(4), 694-727.
Loshkin, M., & Ravallion, M. (2004). Household dynamics in two transition economies.
Studies in Nonlinear Dynamics and Econometrics, 8(3), 1-33.
Mackenbach, J. P., Martikaien, P., Looman, C. W., Dalstra, A. E., Lahelma, E., & SEdHA. (2004). The shape of the relationship between income and self-assessed health: an international study. International Journal of Epidemiology, 34, 286-293.
Rahkonen, O., Arber, S., Lahelma, E., Martikainen, P., & Silventoinen, K. (2000).
Understanding income inequalities in health among men and women in Britain and Finland. International Journal of Health Services, 30(1), 27-47.
Semyonov, M., Lewin-Epstein, N., & Makileyson, D. (2013). Where wealth matters for more health: The wealth-health gradient in 16 countries. Social Science & Medicine, 81, 10-17.
Smith, J. P. (1999). Healthy bodies and thick wallets: the dual relation between health and economic status. Journal of Economics Perspectives, 13, 145-166.
Van Doorslaer, E., Koolman, X., & Jones, A. M. (2004). Explaining income-related inequalities in doctor utilisation in Europe. Health Economics, 13, 629-647. Van Lenthe, F. J., Kamphuis, C. B., Beenackers, M. A., Jansen, T., Looman, C. W.,
socioeconomic inequalities in health and health behaviours: A GLOBE study.
