Towards a right policy for good public health:
income as determinant of self-assessed health
Bachelor thesis
13 February 2014
Ilja Rensje Noordam
Faculty Economics and Business Student number: 10108416
Abstract
The government should ensure that every inhabitant of the Netherlands can lead a healthy life. If the government wants to establish a good health policy it is important to identify which factors actually affect people’s health. This paper investigates whether income has a causal effect on self-assessed health. In addition, this paper also examine which other factors have a causal effect on self-assessed health. Based on the results of the ordered probit model it can be concluded that income does not have a significant effect on self-assessed health. Other factors, like unemployment and smoking, do have a significant effect on health. The
consequence of these conclusions is that the government cannot affect people's health through income. The government can positively influence people’s health by reducing the
unemployment rate or by reducing the number of people who smoke.
Content
Abstract………. 1
1. Introduction……… 3
2. Literature review……… 6
3. Data and Method……… 8
3.1 Data……….. 8 3.2 Dependent variable……….. 8 3.3 Independent variable……… 9 3.3.1 Demographic ……… 9 3.3.2 Socioeconomic……….. 10 3.3.3 Health behavior………... 11 3.3.4 Network support……… 11 3.4 Method………. 13
4. Results and Analysis……….. 15
5. Conclusion and Discussion……….. 19
References……….. 21
Appendix………. 23
1. Introduction
In the universal declaration of human rights, signed by the Dutch government, article 25 states the following: ‘Everyone has the right to a standard of living adequate for the health and well-being of himself and of his family’. In addition, in the constitution of the Netherlands can be found in article 22, relating to the health of the population, housing and development: ‘The authorities shall take steps to promote the health of the population’. Based on the first
citation, it can be stated that every person has the right to lead a life in good health. And based on the above-mentioned law from the Dutch constitution the conclusion can be drawn that the Dutch government should ensure that the health of the Dutch population is good, so that every inhabitant of the Netherlands can lead a healthy life.
Health in all policies (HiAP) can be seen as a useful approach to positively influence people’s health.HiAP is a policy in which relevant sectors work together, both inside and outside the public health domain, in order to protect and improve the health of the society (Storm, Zoest & Broeder, 2007). The idea behind HiAP is that public health not only depends on activities in the health sector, but to a larger extent depends on the living environment and on social and economic factors. For that reason it makes sense to influence public health by policies and actions beyond the health sector (Ståhl, Wismar, Ollila, Lahtinen & Leppo, 2006). The report of the RIVM 1 named several factors that affect public health. The factors they mention include social and physical environmental factors, such as employment and the living environment.In addition, the economy and technological developments are mentioned as factors that affect health (Storm et al., 2007).
If the government wants to establish a good policy, it is important to identify which factors actually affect people's health. If they know these factors it can be determined what policies are needed to influence the determinants of health. So the first step of HiAP is to find out what are the determinants of health, thus what factors have a causal effect on health.
As noted before, the economy could have influence on health (Storm et al., 2007). One may wonder whether health is influenced by the level of income, or in other words, is there is a causal effect of income on health. If it turns out that this causal effect exists, then the government can improve health by influence people’s income. However, if it turns out that there is no causal effect running from income to health, the government should rely on other facets of HiAP to improve people’s health. In this case it is useful to investigate which other factors have a causal effect on health. An example of a potential factor is education: if it
1 National Institute for Public Health and Environment in the Netherlands
3
turns out that education has a positive causal effect on health, the government should
implement a policy in which education is stimulated. In this way a positive effect on health is exercised.
To come to a proper health policy based on economic factors it is essential to have a good understanding of the relationship between income and health. Correlation between income and health alone is not enough evidence to conclude that reducing income has a negative effect on people's health. ‘Standard economic theory predicts that persons in good health will have higher labor force participation rates and command higher wage rates, both of which lead to greater income’ (Ettner, 1996, p.68). This statement says that the causality is running from health to income. This is different from the leading question of this paper, which asked if income has a causal effect on people’s health. So it could be possible that the relationship between health and income works in both directions. In other words, there may be simultaneous causality.
The following example makes clear how this simultaneous causality occurs in the relationship between income and health. People with higher incomes often have better access to health care, whereas this is more difficult for people with lower incomes. This could be the reason why people with lower income, all else equal, are less healthy than people with higher income. But it could also be the other way around; people with poor health could be less able to work, and therefore generate less income than people with a good health status (Smith, 1999).
If both causal effects appear to be truth, then simultaneously causality bias occurs. This will be a problem when you want to do a normal regression, because the independent variable will be correlated with the error term. In a situation like this, the independent variable is called exogenous. One solution of the problem of exogenous variables is to use an
instrumental variable regression (Stock & Watson, 2012, p. 368).
When doing a regression, it is important to know which independent variables you should include in your model in order to minimize the bias of omitted variables. Omitted variable bias arise when there is a confounding variable, not included in the model, that is correlated with the independent variable and determines, in part, the dependent variable (Stock & Watson, 2012). This is also called a spurious relationship. Such a relationship is a false indication of causality: the effect of the confounding variable on the two other variables, the dependent and independent variable, leads to the unjustified conclusion that the two other variables are causally linked.
For example, when one performs a regression of income on health, a correlation can be found. But it could be that there is a third factor, for example education, not included in the model that is correlated with both health and income. This omitted variable is the reason why a correlation is found between income and health.
This bachelor thesis will focus on the question whether there exists a causal effect of income on health. In addition, but to a lesser extent, it will be examined what other factors affect health. Section 2 of this paper contains an overview of the theory about the income-health relationship. Section 3 describes what data and method is used in order to find an answer on the research question. Results and an analysis can be found in section 4 and a conclusion is drawn in section 5. In addition, section 5 contains some points for discussion and recommendations for further research.
2. Literature review
The main question of this study is whether there exists a causal effect of income on health and, in addition, what other possible factors affect health. There has been done a lot of research on this relationship in the recent decades.
In the early nineties, several studies from the United Kingdom and other Western European countries show that there is substantial evidence that higher income is correlated with higher life expectancy and good health (Hurrowitz, 1993). Doorslaer (2001) shows that the relationship between income and health for individuals is represented by a non-linear and concave function. A function like this means that every additional unit of income causes an improvement of health (all else equal), but as income increase there are diminishing returns of health. In other words, in low-income levels, the health benefits of the same income increase are higher than the health benefits in high-income levels. This concave relationship between income and health is also called the absolute income hypothesis (Lorgelly & J., 2007). The figure below shows this concave relationship (Wagstaff & Van Doorslaer, 2000).
Figure 1: Concave relationship between income and health Referrence: Wagstaff & Doorslaer, 2000
But can we interpret this relationship as a causal effect of income on health? In her research, Ettner (1996) tried to get to an answer on this question. The statistical research is made up of two different parts. In the first part it is assumed that income is an exogenous variable. This means that she assumes that the causal effect is running from income to health and that there is no possibility of simultaneous causality. In the second part of the study, the exogeneity assumption of income is released and estimates are made based on instrumental variable analysis. In both cases, the conclusion is drawn that income has a causal effect on health and
that there exists strong evidence that income has a big impact on people’s mental and physical health.
Frijters, Haisken-DeNew and Shields (2005) used an ordered logistic regression model to estimate the effect of the log household income on people’s health satisfaction. The data source was the German Socioeconomic Panel (GSEOP) between 1984 and 2002. In their research they found that income change has a small causal effect on health satisfaction.
Other evidence of the causal effect of income on health is found by Lindahl (2005). When he studied the effects of income on health he used monetary lottery prizes of
individuals as an exogenous variable of income. He estimated health as a function of income using the ordinary least square (OLS) method as well as the IV-estimation model. Using OLS, he found that a ten percent increase of income leads to an increase in general health of two percent of a standard deviation. Using IV-estimation, he found that a ten percent increase of income leads to an increase in general health of 4 to 5 percent of a standard deviation.
Lorgely and Lindy (2008) also found evidence that supports the absolute income hypotheses. They used the correlated random effects ordered probit model and found that income change had a small, positive association with self-assessed health for both men and for woman. They found a significant estimated value of 0.06 for men and 3.62 for woman.
But there are also studies that did not succeed in finding a causal effect of income on health. Meer, Millen and Rosen (2003) looked at the causal effect of wealth on health and they did not found a relationship in which there is a causal effect running from wealth to health in the short run. And also Contoyannis (2004) conclude that ‘it is not possible to separate a causal effect of long-term economic status on health and the correlation between mean income and the unobservable individual effect’ (p. 500).
Summarizing, it can be said that research shows that income and health are closely related. However, there is still no unambiguous answer to the question whether there exists a causal effect of income on health.
3. Data and Method
3.1 Data
The data used in this research is data from CentERdata. CentERdata is a Dutch research institute that is engaged in collecting and analyzing panel data. One manner that they use to collect data is through an internet panel called the CentERpanel. This internet panel consists of 2000 households and reflects the Dutch speaking population.The 2000 participating households answer questions about various topics such as: work, retirement, housing,
mortgages, income, health and personal characteristics.The data that is obtained on the basis of these questions can be used in studying both psychological and economic aspects of financial behavior. Because in this paper the relationship between income and health is
investigated, it is important that to look only at the participants with an income. Therefore, the sample includes only people aged sixteen years and older who have an income. For this paper data from the 2012 wave is used. The surveys are filled out in March 2013, but the data relates to the year 2012.
3.2 Dependent variable
Because the research question of this study is whether income has a causal effect on health, the dependent variable is health. A widely used measure of health status is self-assessed health (e.g. Ettner, 1996; Zimmer et all., 2000, Lorgely & Lindy, 2008). This measure is often used because it is easy to obtain and to process. Fayers and Hays (2005) conclude that ‘if one is allowed to use only one measure of health status, the rating of self-assessed health would be a good option’. It has been shown that self-assessed health is a good predictor of mortality and inequalities in mortality (e.g. van Doorslaer and Gerdtham, 2003).
A disadvantage of the subjective SAH measure is that there exists a possibility of a measurement error. This error occurs because different subgroups in the population use different cut point levels when reporting their SAH, while having the same level of ‘true’ health (Contoyannis et al., 2004). Research has shown that these cut point differences occur with respect to age and gender, but not with respect to education and income (Lindeboom & van Doorslaer, 2003).
In the household survey of CentERdata, the following question was posed: In general, would you say your health is: (1) 'excellent', (2) ‘good’ (3) ‘fair’ (4) ‘not so good’ or (5) 'poor'. In the model used to examine whether there exists a causal effect of income on health, the variable of self-assessed health is dichotomized in: (1) ‘excellent’, (2) ‘good’, (3) ‘fair’
and (4) ‘not so good’. The fourth category includes the people who answered ‘not so good’ and the people who answered ‘poor’.
3.3 Independent variables
In most of the studies about the relationship between income and health, self-assessed health is chosen as the variable that represents health. Some studies support the hypothesis that there exists a causal effect of income on self-assessed health, other studies rejects this hypothesis. However, income is unlikely to be the only determinant of self-assessed health. It is argued that ‘self assessments of health are likely to differ across populations characterized by
differences in actual health status as well as differences in other characteristics that have been shown to impact on health’ (Zimmer, Natividad, Lin & Chayovan, 2000, p. 467). These characteristics can be divided into different domains: a demographic domain, a domain concerning socioeconomic status, a domain about health behaviors and a domain concerning the existence of network support (Zimmer et al., 2000).
3.3.1 Demographic
Characteristics covered by the demographic domain are age, gender and the degree of
urbanization of the living environment. It is commonly known that, on average, when people become older their health deteriorates. When looking at self-assessed health, it is found that lower self-reported health is associated with an increasing age (Franks, Gold & Fiscella, 2003). The variable age is included in the model as a continuous.
Gender is added because of the difference in life expectancy between man and woman. In countries where the gross domestic product (GDP) per capita is less than $1,000, the life expectancy of woman is about four percent higher than for man. And for countries with a GDP per capita around $9000, the life expectancy of woman is about eleven percent higher than for man (Fuchs, 2004).
The last variable of the demographic domain is the degree of urbanization of the living environment. Previous research (Maas, Verheij, Groenewegen, Vries & Spreeuwenberg, 2006) showed that the percentage of green and nature in the living environment has a positive effect on people’s perceived health. It also shows a negative effect on perceived health when people live in an environment with a high level of urbanization. The variable in the model is constructed from five different scales: (1) ‘very high degree of urbanization’, (2) ‘high degree of urbanization’, (3) ‘moderate degree of urbanization’, (4)‘low degree of urbanization’ and (5) ‘very low degree of urbanization’.
3.3.2 Socioeconomic
Characteristics covered by the socioeconomic domain are income, education and employment status. The relationship between income and health is already discussed in paragraph 2 and it turns out to be very complex. In the model of this paper net household income of the
participant is used to represent income. Household income is divided into six different categories: (1) ‘less than 16000’, (2) ‘16000-26000’, (3)’26000-38000’, (4) ‘38000-50000’, (5) ‘50000-75000’ and (6) ‘more than 75000’.
According to Fuchs (2004), the education-health relationship appears to be less
complicated compared to the income-health relationship. First, the measurement for education is less problematic than for income. Second, the probability of simultaneous causality seems small because there is little empirical evidence that there is a causal relationship running from health to education (Fuchs, 2004).
Ross and Wu (1995) stated that the explanations for the positive association between education and health can be divided into three categories: (1) work and economic conditions, (2) social-psychological resources, and (3) health lifestyle. Research showed that, compared to poorly educated, highly educated respondents are less likely to be unemployed and get more satisfaction from their jobs. And based on the analyses, full-time and fulfilling work significantly improves health.Looking at the explanation that falls under social-psychological resource, it turns out that well educated respondents have more control over their lives. This sense of control is associated with good health. Besides that, the well educated have higher levels of social support, which is also associated with better health. In terms of health lifestyle the well educated generally do more sports, smoke less, and drink moderately, which are all associated with good health (Ross & Wu, 1995).
The variable schooling in the model is divided into four categories: (1) ‘primary school’, (2) ‘secondary school’, (3) ‘intermediate vocational education’ and (4) ‘higher vocational education/university’. The distribution is created based on the highest level of education completed by an individual.
The variable occupation is included in the model because the difference between whether you have a job or not can effect health. An unfavorable position in the labor market has indirect impact on health because it often causes stress, which has a negative effect on people’s health (Stronks & Hulshof, 2001). This variable is divided in 4 different categories: (1) ‘Employed’, (2) ‘unemployed’, (3) ‘retired’ and (4) ‘other’.
3.3.3 Health behavior
‘Behaviors can have direct influences on health outcomes and can also influence perception’ (Zimmer et al., 2000, p. 469).Examples of such behavior include smoking and the current consumption of alcohol, because these factors have a negative influence on people’s health. It should be said that drinking alcohol only has a negative effect on health when people
systematically consume too much alcohol (Hoeymans, Melse & Schoemaker, 2010).
Two variables included in the model are smoking and the consumption of alcohol drinks. Smoking behavior is dichotomized by (1) ‘never smoke’ and (2) ‘do smoke’. Drinking behavior is measured in two categories: (1) ‘drinks more than 4 alcoholic drinks a day’ and (2) ‘drinks less than 4 alcoholic drinks a day’.
Other factors included in the model, with respect to health behavior, are the number of doctor visits during the past year, whether people have an chronic disease, long illness,
handicap or accident and finally, if someone has obesity. Number of doctor visits is divided in two categories: (1) ‘more than 3 doctor visits’ and (2) ‘less than three doctor visits’.
3.3.4 Network support
This domain contains factors dealing with network characteristics. Psychosocial status, such as the existence of a strong and supportive network, will have effect on self-assessed health (Ross & Wu, 1995). Examples of different psychosocial statuses are: the number of living offspring and the presence of a partner (Zimmer et al., 2000).
The psychosocial factor marital status is called because research had determined that in every country, married men and woman are healthier than their unmarried peers (Fuchs, 2004, p. 659). There is a correlation between marital status and health. Wyke and Ford (1992) found material resources, less stress and perceived quality of social support as reason for the positive effect of marital status on health. Another explanation for the correlation is that healthier men and women are more desirable; therefore it is easier for them to find a wedding partner. In this explanation, the causality is running from health to marital status (Fuchs, 2004). The model includes a variable that makes a distinction between individuals with a partner in the household and individuals without a partner in the household.
The other factor that is included in the model is the number of children. The variable that indicated the number of children in the household is divided in three categories: (1) ‘no children’, (2) ‘one child’ (3) ‘two children’ and (4) ‘three children or more’. Table 1 gives an overview of the variables used in the model and the definition of these variables.
Table 1 Definitions of the variables
Variable Amount Percentage Definition
Self-assessed health, SAH N=1800
1: not so good 79 4.39 Five point scale of self-assessed health
2: fair 349 19.39
3: good 1134 63.00
4: excellent 238 13.22
Age mean = 58.82 Age of the participants
Gender
Female=0 782 56.56
Male=1 1018 43.44
Education
1: primary school 69 3.83 Highest level of education completed
2: secondary school 692 38.44
3: intermediate vocational education 295 16.39
4: higher vocational education/university 744 41.33
Income
inkomen1: less than 16000 299 16.61 Net level of household income
inkomen2: 16000 - 26000 382 21.22
inkomen3: 26000 - 38000 496 27.56
inkomen4: 38000 - 50000 376 20.89
inkomen5: 50000 - 75000 198 11.00
inkomen6: more than 75000 49 2.72
Occupation
occupation1: employed 849 49.61
occupation2: unemployed 120 6.66 Occupation of the participant
occupation3: retired 225 12.49
occupation4: other 563 31.24
Living environment
sted1: very high degree of urbanization 243 13.50
Five point scale of the degree of urbanization
sted2: high degree of urbanization 466 25.89
sted3: moderate degree of urbanization 389 21.61
sted4: low degree of urbanization 391 21.72
sted5: high degree of urbanization 311 17.28
Partner
partner_=1 having a partner 1400 77.78 Partner present in the household
partner_=0 no partner 400 22.22
Children
kids1: no children 1242 68.00 Number of children
kids2: 1 child 196 10.89
kids3: 2 children 231 12.83
kids4: 3 children or more 131 7.28
Smoking Dummy variable which makes a distinction
smoke=1 when smoke 318 82.33
between people who smoke and people who do not smoke.
smoke=0 when not smoke 1482 17.67
Drinking Dummy variable which makes a distinction
drinking=1 (>4 glasses) 87 4.83 between people consuming more than 4
drinking=0 (<4 glasses) 1713 95.17 alcoholic drinks a day and people who
consume less than 4 glasses a day. Huisa2, doctor visits
1: more than three visits 470 29.56 Number of doctor visits a year
0: less than three visits 1120 77.44
Overweight
1: BMI>30 261 14.50 BMI>30, someone has obesity
0: BMI<30 1539 85.50
Illness Whether someone has a long illnes, disorder
1: Yes 494 27.44 or a handicap; or suffers from the
0: No 1306 72.56 consequences of an accident.
3.4 Method
The outcome of the dependent variable SAH is an ordered ranking. One can say that ‘excellent’ is better than ‘good’, ‘good’ is better than ‘fair’ and ‘fair’ is better than ‘not so good’. The only information one has is the ranking of the categories. Even when the categories are coded as 1, 2, 3 and so on, the numbers do make sense. This is because one does not know how big the difference is between two categories. It could be that the difference between the first and the second outcome may not be the same as the difference between the second and the third outcome.
Because the dependent variable is an ordered variable, the ordered probit model will be used. The index model for a single latent variable 𝑦𝑦𝑖𝑖∗ is:
𝑦𝑦𝑖𝑖∗ = 𝛽𝛽𝑥𝑥′𝑖𝑖 + 𝜀𝜀𝑖𝑖 where 𝐸𝐸[𝑒𝑒𝑖𝑖] = 0
𝑥𝑥′𝑖𝑖 In the model is a set of independent variables and 𝛽𝛽 a set of coefficients to be estimated. A
latent variable is a variable that is unobservable; one can only observe the different categories. In the case of SAH one observers the different outcomes ‘excellent’, ‘good’, ‘fair’ and ‘not so good’, but not SAH itself.
The latent scale of 𝑦𝑦𝑖𝑖∗ corresponds to the observed outcomes 𝑦𝑦𝑖𝑖 as follows: 𝑦𝑦𝑖𝑖 = 𝑗𝑗 if 𝜏𝜏𝑗𝑗−1 < 𝑦𝑦𝑖𝑖∗ < 𝜏𝜏𝑗𝑗
This means that the observed variable is equal to 𝑗𝑗 if the underlying latent variable falls between the two unknown thresholds 𝜏𝜏𝑗𝑗−1 and 𝜏𝜏𝑖𝑖. In the model of SAH there are four different categories so three thresholds, where 𝜏𝜏1 < 𝜏𝜏2< 𝜏𝜏3.
𝑦𝑦𝑖𝑖 = 1 if 𝑦𝑦𝑖𝑖∗ < 𝜏𝜏1
𝑦𝑦𝑖𝑖 = 2 if 𝜏𝜏1 < 𝑦𝑦𝑖𝑖∗ < 𝜏𝜏2
𝑦𝑦𝑖𝑖 = 3 if 𝜏𝜏2 < 𝑦𝑦𝑖𝑖∗ < 𝜏𝜏3
𝑦𝑦𝑖𝑖 = 4 if 𝑦𝑦1∗ > 𝜏𝜏3
The probability that a person 𝑖𝑖 will be in the category 𝑗𝑗 is:
𝑝𝑝𝑖𝑖𝑗𝑗 ≡ 𝑝𝑝(𝑦𝑦𝑖𝑖 = 𝑗𝑗) ≡ 𝑝𝑝�𝜏𝜏𝑗𝑗−1< 𝑦𝑦𝑖𝑖∗ < 𝜏𝜏𝑗𝑗� =
𝑝𝑝�𝑦𝑦𝑖𝑖∗≤ 𝜏𝜏
𝑗𝑗� − 𝑝𝑝�𝑦𝑦𝑖𝑖∗ ≤ 𝜏𝜏𝑗𝑗−1� = 𝐹𝐹�𝜏𝜏𝑗𝑗 − 𝑥𝑥𝑖𝑖′𝛽𝛽� − 𝐹𝐹(𝜏𝜏𝑗𝑗−1− 𝑥𝑥𝑖𝑖′𝛽𝛽).
Where F is the standard normal cdf.
The estimated coefficients of the ordered probit model are difficult to interpret. The only information one can get from it, is that the sign of the coefficients shows whether latent variable 𝑦𝑦𝑖𝑖∗ increases or decreases with the independent variable. Another option is to look at the marginal effects: each unit increase in the independent variable increases or decreases the probability of selecting a specific outcome by the amount of the marginal effect. The formula of the marginal effect is:
𝜕𝜕𝑝𝑝𝑖𝑖𝑖𝑖
𝜕𝜕𝑥𝑥𝑖𝑖 = �𝐹𝐹
′�𝜏𝜏
𝑗𝑗−1− 𝑥𝑥𝑖𝑖′𝛽𝛽� − −𝐹𝐹′(𝜏𝜏𝑗𝑗− 𝑥𝑥𝑖𝑖′𝛽𝛽)�𝛽𝛽.
4. Results and Analysis
The analysis starts with model one in which demographic and socioeconomic variables are regressed on self-assessed health. The results of the ordered probit regression are shown in table 2. This table consists four columns. The first column listed the estimated coefficients of the variables. The second column contains the robust standard errors; this means that the standard error is adjusted for hetroskedasticity. The third and fourth column lists,
respectively, the z value and the p-value, which are indicators of the significance of a variable.
Table 2
In order to get an answer on the main question, whether income has a causal effect on SAH, it is important to look at the different categories of household income. From the table it can be seen that inkomen1, inkomen2, inkomen3 and inkomen4 are significant at 5%, because the p-value is below 0.05. This means that individuals in these income groups, compared to
individuals in the highest income bracket, are more likely to have smaller values of SAH. Other significant variables are: age, scholing2 and occupation2. Looking at age, it can be seen from the negative coefficient that when people become older they are less likely to be
Model 1. Demographic and socioeconomic variables
Self-assessed health Coefficient Robust Std. Err. z P>z gender .0919763 .0581795 1.58 0.114 age -.014515 .0028945 -5.01 0.000 scholing1 .0170553 .1516667 0.11 0.910 scholing2 -.1985873 .0643082 -3.09 0.002 scholing3 -.0683935 .0807138 -0.85 0.397 inkomen1 -.3708862 .145515 -2.55 0.011 inkomen2 -.3787902 .1398027 -2.71 0.007 inkomen3 -.3178389 .135365 -2.35 0.019 inkomen4 -.2731281 .1370368 -1.99 0.046 inkomen5 -.0825626 .1479949 -0.56 0.577 occupation2 -.8545806 .1204099 -7.10 0.000 occupation3 -.0478394 .0941305 -0.51 0.611 occupation4 .0197507 .0888593 0.22 0.824 s1 .0366518 .0973313 0.38 0.706 s2 -.0986764 .0843178 -1.17 0.242 s3 .0036828 .0899096 0.04 0.967 s4 -.0596459 .0879605 -0.68 0.498 /cut1 -309.915 .208906 /cut2 -2.027.688 .2020729 /cut3 -.0932838 .198869 Number of obs = 1800 Prob > chi2 = 0.0000 Pseudo R2 = 0.0467 15
in the higher categories of SAH. Occupation2 indicates the unemployed individuals. It can be seen that when individuals are unemployed they are less likely to be in the higher categories of SAH, compared to people who have a job.
The significance of scholing2 in this model means that individuals who only finished secondary school are less likely to be in the higher categories of SAH compared to people who finished university or higher vocational education. Remarkably, the model does not show a significant effect of the variable scholing1 on SAH. This is strange because one would expect that when only having a secondary school degree, compared to finishing university, has a negative effect on SAH, this must certainly be the case for people who have only finished primary school. The reason for this remarkable result could be that in or sample the group of people who only finished primary school is small, namely 3.83 percent. This low percentage can be explained by the fact that Dutch children between five and sixteen years old are required to go to school. Children finish primary school, on average, around their 12th and at that age are required to continue to go to school, so they go to secondary school.
In model two also network variables, having a partner and number of children, are included. The results of this regression are shown in table 3.
Table 3
Model 2. Demographic, socioeconomic and network variables
Self-assessed health Coefficient Robust Std. Err. z P>z gender .0685492 .0586308 1.17 0.242 age -.0131534 .0030715 -4.28 0.000 scholing1 -.0052397 .1548116 -0.03 0.973 scholing2 -.2162024 .065232 -3.31 0.001 scholing3 -.0917563 .0806707 -1.14 0.255 inkomen1 -.2974988 .1476011 -2.02 0.044 inkomen2 -.3238412 .1411514 -2.29 0.022 inkomen3 -.2829833 .1354595 -2.09 0.037 inkomen4 -.2575755 .1364132 -1.89 0.059 inkomen5 -.0718304 .1473486 -0.49 0.626 occupation2 -.8269972 .1210249 -6.83 0.000 occupation3 -.0742714 .0952957 -0.78 0.436 occupation4 -.0005091 .0905042 -0.01 0.996 s1 .1123295 .1015387 1.11 0.269 s2 -.0654098 .0854517 -0.77 0.444 s3 .0333064 .0909415 0.37 0.714 s4 -.0533696 .0876702 -0.61 0.543 kids1 -.1383465 .1130022 -1.22 0.221 kids2 -.2455905 .1306138 -1.88 0.060 kids3 -.0705345 .1227282 -0.57 0.565 partner .1758195 .0748596 2.35 0.019 /cut1 -2.989.253 .2339663 /cut2 -1.913.559 .2292533 /cut3 .0287235 .2276764 Number of obs = 1800 Prob > chi2 = 0.0000 Pseudo R2 = 0.0496 16
Looking at the results of table 3, it can be said that the variables age, inkomen1, inkomen2, inkomen3, occupation2 and partner are significant at 5%. The p-value of inkomen4 changed from 0.046 in model 1 to 0.059 in model 2 and is not significant any more. The variable partner is significant, which means that people who have a partner are more likely to be in the higher categories of SAH.
Model 3 includes demographic, socioeconomic, network and health behavior variables. The results of this regression are shown in table 4.
Table 4
Model 3. Including all variables
Self-assessed health Coefficient Robust Std. Err. z P>z gender -.0172452 .064613 -0.27 0.790 age -.0090022 .0034813 -2.59 0.010 scholing1 -.0732115 .16447 -0.45 0.656 scholing2 -.1715115 .0705936 -2.43 0.015 scholing3 -.0664947 .0906236 -0.73 0.463 inkomen1 -.2020572 .1623394 -1.24 0.213 inkomen2 -.1615819 .1569051 -1.03 0.303 inkomen3 -.1727343 .1519799 -1.14 0.256 inkomen4 -.1461346 .1506731 -0.97 0.332 inkomen5 .0316469 .1625695 0.19 0.846 occupation2 -.4932224 .1292221 -3.82 0.000 occupation3 -.019816 .10113 -0.20 0.845 occupation4 .0706171 .100041 0.71 0.480 s1 .0603385 .1129059 0.53 0.593 s2 -.1133381 .0949343 -1.19 0.233 s3 -.0326335 .1002581 -0.33 0.745 s4 -.1100194 .0961043 -1.14 0.252 kids1 -.0648847 .1301954 -0.50 0.618 kids2 -.1162295 .1522647 -0.76 0.445 kids3 .0364536 .141383 0.26 0.797 partner .1244675 .0809102 1.54 0.124 smoke -.2665572 .0815827 -3.27 0.001 drinking .0463205 .1374382 0.34 0.736 illness -.9767589 .0758128 -12.88 0.000 BMI -.3145863 .084645 -3.72 0.000 huisa2 -.5197193 .0686367 -7.57 0.000 /cut1 -3.439.842 .2645434 /cut2 -2.123.498 .2567636 /cut3 .1164888 .2533215 Number of obs. = 1590 Prob > chi2 = 0.0000 Pseudo R2 = 0.1539
It appears that almost all of the health behavioral variables added to the third model are significant. Only the variable drinking, with a p-value of 0.736, is not significant. It can be concluded that people who smoke are less likely to be in the higher categories of SAH. The
same is true for people with a long illness, a BMI larger than 30 and people who visit the doctor more than three times a year: people who are within one or more of these categories are less likely to be in the higher categories of SAH.
Looking at the other results of the third regression, it can be said that the variables age, scholing2 and occupation2 are significant, just as in the previous regressions. However, the income variables and the variable partner are not significant anymore. The possible reason for this result is that the health behavior variables are confounding variables. It could be that the health behavior variables are correlated with income and SAH. For example, this is could be the case for long illness: people with a long illness are unable to work and therefore have a lower income than people without a long illness and people with a long illness record a lower SAH. Leaving these confounding variables out of the regression can cause a case in which an unjustified causal effect is demonstrated, which probably is the case in model 1 and 2.
The fourth model only includes the significant variables of model 3. The output results of this regression can be found in table 5 in the appendix. The marginal effects are calculated based on this model. The output results of these marginal effects are also found in the
appendix, tables 7 to 10. The table below gives an overview of all the predicted marginal effects.
not so good fair good excellent
age 0.0004175 0.0024223 -0.13985 -0.0014414 scholing2 0.0086682 0.0482534 -0.0293714 -0.0275503 occuaption2 0.0390104 0.1473482 -0.1290806 -0.057278 smoke 0.0148086 0.0732856 -0.0515276 -0.0365666 illness 0.0695049 0.2485834 -0.2073365 -0.1107518 BMI 0.0172638 0.0826931 -0.0600624 -0.0398945 huisa2 0.0292796 0.1363942 -0.0982726 -0.0674012
Table 5. Marginal effects of the different self-assessed health outcomes
Looking at the variables that can be influenced by the government, such as unemployment, smoke and BMI, it can be seen that being unemployed (variable: occupation2) decrease the probability of being in the ‘good’ health status by 0.12908 and decrease the probability of being in the ‘excellent’ health status by 0.05728. Smoking increases the probability of being in the ‘not so good’ health status by 0.0148 and increase the probability of being in the ‘fair’ health status by 0.073286. And finally, people suffering from obesity are 0.06006 less likely to be in the ‘good’ health status and 0.039895 less likely to be in the ‘excellent’ health status.
5. Conclusion and Discussion
The Dutch government has the task to guarantee good health for all her citizens. One possible way to do this is to conduct health in all policies. If the government wants to achieve a good health policy it is important to identify the determinants of health. An example of such a determinant is income. If income has a causal effect on health, the government can positively influence health by using policies that effect income.
There is no clear answer to the question whether income has a causal effect on health. Several studies from the past decades show a causal effect, however, there are also studies where no causal effect is found. This study is based on data from a Dutch household survey conducted by CentERdata and an ordered probit regression is performed to determine whether a causal effect is running from income to self-assessed health.
The first model only includes the demographic and socioeconomic variables. The results of this regression show that income has a significant, positive effect on self-assessed health. It can be concluded that individuals from the lower income groups are less likely to be in the higher categories of self-assessed health, compared to the people in the highest income category. Besides the demographic and socioeconomic factors, network variables are also included in the second model. This model shows a positive significant effect of income on health as well.
In the third model, the health behavior variables are added. It turns out that all the health behavior variables are significant, except the variable ‘drinking’. However, even more striking is that there is no longer a significant effect running from income to self-assessed health. An explanation for the shift of significance is that there is omitted variable bias in model 1 and 2; the health behavior variables are confounding variables.
Based on the results of the third model, it can be concluded that income has no significant effect on self-assessed health. The government should use other determinants of self-assessed health to influence citizens’ health. For example, by reducing the unemployment rate, the government can improve people’s health. Another way to positively influence public health is to reduce the number of people who smoke. This can be achieved by raising the taxes on smoke or by warning about the dangers of smoking.
The models used in this study adopted many different variables from the demographic domain, socioeconomic domain, health behavior domain and the domain of network support. Because of this, there is a clear picture of which variables have a causal effect and which variables do not have a causal effect. However, it is possible that there exist other factors, not
included in this research, that have an effect on health. One reason for these missing variables is that the data was not available from the survey used in this paper. By proposing a more comprehensive questionnaire, it is possible to add more variables to the model. In addition, if more people than the 1800 of this research complete this questionnaire, the results of a new research will be more precise.
This research only used the survey wave of 2012. This means that this study only looked at the causal effect of income on health based on data from the year 2012. Further research may use more waves to get a better understanding of the relationship between health and income. When one looks at a relationship for an extended period, more precise estimates of the possible effects can be made.
The introduction of this thesis put forward the possibility of simultaneous causality between income and health. It was suggested that the causal effect between income and health could work in both directions. The instrumental variable regression was mentioned as a possible solution for this problem. In this research the income variable is assumed to be exogenous and the instrumental variable regression is not used. Further research can release this assumption of exogenous income and run an instrumental variable regression. When someone wants to use this method, it is important to find good instruments for the income variable. Ettner (1996) called as possible instruments work experience and the unemployment rate.
To summarize, it can be said that, based on the data and research method used in this paper, no evidence of a causal effect running from income to health can be found. However, when the recommendations mentioned above are used in future research, it could be possible that other conclusions can be drawn.
References
Contoyannis, P., Jones, A. M., & Rice, N. (2004). The dynamics of health in the British household panel survey. Journal of applied econometrics, 19(4), 473-503. Doorslaer, E. van (2001). Ongelijkheid in inkomen en gezondheid: “For richer, for
poorer, in sickness and in health”? In Stronks, K., & Hulshof, J.M. (Eds.), De kloof verkleinen (pp. 67-80). Assen: Koninklijke Van Gorcum BV.
Doorslaer, E. van, & Gerdtham, U-G. (2003). Does income inequality in self-assessed health predict inequality in survival by income? evidence from Swedish data. Social Science & Medicine, 57(9), 1621-1629.
Ettner, S., 1996. New evidence on the relationship between income and health. Journal of Health Economics, (15), 67–85.
Fayers, P. M., & Hays, R. D. (2005). Assessing quality of life in clinical trials: Methods and practice. (2nd ed.). Oxford: Oxford University Press.
Franks, P., Gold, M. R., & Fiscella, K. (2003). Sociodemographics, self-rated health, and mortality in the us. Social Science & Medicine, 56(12), 2505-2514.
Frijters, P., Haisken-DeNew, J. P., & Shields, M. A. (2005). The causal effect of income on health: evidence from German reunifiaction. Journal of Health Economics, 24(5), 997-1017.
Fuchs, V. R. (2004). Reflections on the socio-economic correlates of health. Journal of Health Economics, 23(4), 653-661.
Hoeymans, N., Melse, J. M., & Schoemaker, C. G. Rijksinstituut voor Volksgezondheid en Milieu, (2010). Gezondheid en Determinanten (270061006). Breda: Koninklijke Broese & Peereboom
Hurowitz, J.C., 1993, Towards a social policy for health, New England Journal of Medicine, 329(2), 130-133.
Lindahl, M. (2005). Estimating the effect of income on health and mortality using lottery prizes as an exogenous source of variation in income. The Journal of Human Resources, 40(1), 144-168.
Lindeboom, M., & Doorslaer, E. (2004). Cut-point shift and index shift in self-reported health. Journal of Health Economics, 23(6), 1083-1099.
Lorgelly, P.K., & Lindley, J. (2007). What is the relationship between income inequality and health? Evidence from the BHPS. Health Economics, 17(2), 249-265.
Maas, J., Verheij, R. A., Groenewegen, P. P., Vries de, S., & Spreeuwenberg, P. (2006). Green space, urbanity and health: how strong is the relation?. Journal of Epidemiol Community Health, 60, 587-592.
Mackenbach, J. P., Martikainen, P., Looman, C. W. N., Dalstra, J. A. A., Kunst, A. E., & Lahelma, E. (2004). The shape of the relationship between income and self-assessed health: an international study. International Journal of Epidemiology, 34, 286-293. Meer, J., Miller, D. L., & Rosen, H. S. (2003). Exploring the health-wealth nexus. Journal of
Health Economics, 22(5), 713-730.
Ross, C. E., & Wu, C-L. (1995). The links between education and health. American Sociological Review, 60(5), 719-745.
Smith, J. P., 1999, Healthy bodies and thick wallets: The dual relation between health and economic status. Journal of Economic Perspectives, 13(2), 145-166.
Stock, J. H., & Watson, M. W. (2012). Introduction to econometrics. (3nd ed., pp. 354-382). Harlow: Pearson.
Storm, I., Zoest, F. V., & Broeder, L. D. Rijksinstituut voor Volksgezondheid en Milieu, (2007). Integraal gezondheidsbeleid: theorie en toepassing (270851003). Retrieved from website: http://rivm.openrepository.com/rivm/handle/10029/13409.
Stronks, K., & Hulshof, J. M. (Eds.). (2001). De kloof verkleinen. (pp. 9-103). Assen: Koninklijke Van Gorcum BV.
Wagstaff, A., & Doorslaer, E. van (2000). Income Inequality and Health: What Does the Literature Tell Us? Annual Review of Public Health, 21, 543-67.
Wyke, S., & Ford, G. (1992). Competing explanations for association between marital status and health. Social science & Medicine, 34(5), 523-532.
Zimmer, Z., Natividad, J., Lin, H., & Chayovan, N. (2000). A cross-national examination of the determinants of self-assessed health. Journal of Health and Social Behavior, 41(4), 465-481.
Ståhl, T., Wismar, M., Ollila, E., Lahtinen, E., & Leppo, K. Ministry of Social Affairs and Health, Health department. (2006). Health in All Policies, prospects and potentials.
Appendix
Table 6. Model 4, including only the significant variables from model 3
Marginal effects
Table 7. Predicted probabilities of being in the 'not so good' health status
Table 8. Predicted probabilities of being in the 'fair' health status
Table 9. Predicted probability of being in the 'good' health status
Table 10. Predicted probability of being in the 'excellent' health status
