SES and Health: An Econometric
View on Socioeconomic Status and
the Number of Physician Visits.
EVA OUDEMANS2
TABLE OF CONTENTS
1. INTRODUCTION . . . 3
2. THEORY. . . 4
2.1 Causality bias . . . 5
2.2 Underpin the hypotheses . . . 5
2.3 Measurements of SES . . . 6
3. ECONOMETRIC MODEL . . . 7
3.1 The Poisson model . . . 7
3.2 A t-test for overdispersion . . . .8
3.3 An alternative model. . . .9
3.4 The hurdle model. . . .9
4. DATA . . . 10 4.1 LISS. . . .10 4.2 Choice of SES . . . 11 4.3 Descriptive statistics. . . 11 5. RESULTS . . . 13 5.1 Poisson model . . . 13
5.2 Negative binomial model. . . 14
5.3 Hurdle model. . . .15
6. CONCLUSION. . . 16
3
This thesis analyses the relationship between socioeconomic status and the number of physician visits, using the individual number of physician visits as outcome measure and data from the Longitudinal Internet Studies for the Social sciences of the survey from January 2012. A number of count data models were used to estimate the relationship. It appeared that persons with a low socioeconomic status visit the physician more often than persons with a high socioeconomic status. The results were not significant enough for robust conclusions, so further analysis could be done with the use of different models.
1. INTRODUCTION
The relationship between socioeconomic status (SES) and health is one that has often been examined in social science. Adler and Ostrove (1999) showed that the increase of interest in this subject started in the mid 80’s, where Dr. Alvin Tarlov organized a conference culminating with the publication of a volume ‘Pathways to health’. In this volume he bundled the theses of numerous researchers who suggested that the impact of SES factors on health was broader and more profound than the threshold model1 showed. Their research proved that the health effects of
SES were not only a result of the setbacks of extreme poverty, but remained at higher levels of SES.
Socioeconomic status is an economic and sociological combined total measure of an individual's or a family’s economic and social position in relation to others. Socioeconomic status is often broken into two categories: low SES and high SES. Although the classification differs per researcher which categorization of SES they employ in their model, the majority of researchers agree that income, education and occupation together best represent SES. Others researchers argue that changes in family structure should also be considered. Other SES indicators could include: marital status, ethnicity, whether the person lives in a good or bad neighborhood and dwelling.
Health can be measured in several ways. One measure of health might be expressed as probability of severe illnesses such as cancer or AIDS. Another means to determine a person's health is by measuring their life habits, such as alcohol and tobacco use. Although these observations are often only measurable for a small subset of a population and thus it is difficult to generalize for a random chosen population. Researchers as Winkelmann (2004) have tried to use the number of physician visits as a measure of health. For example, Winkelmann (2004) found that when a person is presented with a major health care reform, for instance an increase in co-payments (Germany 1997), the individual is less willing to visit the
1 Before 1985, the most frequent measure of SES was poverty status. Individuals were
characterized in terms of being either above or below the poverty line. The underlying assumption appeared to be a threshold model.
4
physician than before the health care reform. In this particular case, it can be derived that a decrease in income will cause a negative effect on the number of visits to a physician. This denotes that the number of physician visits seems to be a good indicator of health.
This thesis examines the relationship between SES and the number of physician visits, in particular the visits to a primary care physician. Just like Winkelmann (2004), Kiuila et al. (2007 p. 786) found with a long term survey that as income grows, the self-reported health status increases simultaneously. One who reports his own health status as being excellent or very good will intuitively not visit the physician as often as someone who reports himself as having a fair or poor health levels. Income proves to have a significant positive effect on the number of physician visits. Does this relationship exist for the other measurements of SES?
On one hand, an individual with low SES may be more susceptible to poor environmental circumstances and therefore will be more susceptible to poor health than a person with high SES. This correlation indicates that at-risk low SES individuals typically require more frequent physician visits than a high SES person. On the other hand, their low SES may result the inability to obtain proper medical care and delay detection of pertinent medical conditions. It may also reduce access to medical services or result in less effective treatment. Section two elaborate further on these hypotheses. These hypotheses will be examined using a data of Dutch households from the Longitudinal Internet Studies for the Social sciences (LISS) of the year 2012. It consists of 5000 households, covering 8000 individuals. The panel is based on a true probability sample of households drawn from the population register by Statistics Netherlands.
The relationship between SES and number of physician visits will be estimated through different methods, starting with a Poisson regression of the number of physician visits on the chosen variables of SES.
The next section of this thesis examines previous studies and provides background information concerning the relationship between health and SES. The third section gives an overview of the data set and the fourth section shows the methods employed in this thesis. Section five examines the results of the experiments are being presented and finally, in section six, a conclusion will be formed.
2. THEORY
This section provides background information on the relationship between SES and health. This allows for a summation of generalizations about the relationship in question. Previously, two hypotheses were presented. One stating that a low SES lead to more physician visits and the other one stated the opposite: that low SES
5
lead to less physician visits. The second part of this section tries to underpin these hypotheses. The last section contains information on the various measures of SES.
2.1 Causality bias
There is extensive literature explaining the relationship between SES and health, but not all studies succeed in finding direct causal paths. This might be due to the difficulty of quantifying health. For example, someone’s self-perception of illness influences his compliance with healthcare (Wichowski et al. 1997). Self-esteem as measure of self-perception of illness correlates positively with compliance. Wichowski (1997) showed that the lower the perception of illness, the lower the compliance with healthcare. Someone’s self-esteem is a variable which cannot be taken into account in this thesis because it is not directly measurable and not a valid indicator of someone’s SES.
When trying to find causal paths between SES and health, differing conclusions may be found from other studies. This may be due to measuring different data sets. Salas (2002) tries to find an explanation for the contradictory results between epidemiologic studies and economic studies about the relationship between poor health and low SES. Epidemiologists fail to find meaningful relationships between poor health and low SES, whereas economists consistently succeed in finding a positive relation. To exclude the possibility that this contradiction is because of different data sets, he used the same data for both experiments. Because the epidemiologic studies use mortality rates as indicators of health and economic studies use self-assessed health, he concludes that these health indicators produce different results because they probably relate to different ranges of the health variable. With this discovery he warns epidemiologists and economists not to make generalizations too quickly. By taking the results of Wichowski (1997) and Salas (2002) into consideration, the results of this thesis will be carefully analyzed and interpreted, even if results show significant effects.
2.2 Underpinning the hypotheses
Adams et al (2003) examined whether there are direct causal paths between SES and health using the Asset and Health Dynamics of the Oldest Old (AHEAD) panel, which consists of observations of individuals 70 years old and older. They express health in first and second class conditions; respectively acute, sudden health conditions such as cancer or strokes and chronic, mental and degenerative conditions such as diabetes or psychiatric conditions. They find no direct causal link from SES to first class health conditions, while the effect of second class conditions appear to have a direct causal relationship with SES due to common behavioral or genetic factors. These results suggest that there may be a SES gradient in seeking treatment for chronic, mental and degenerative conditions,
6
which may influence a person’s detection of illnesses in himself. He finds the possibility that the ability to pay medical services may be a causal factor in seeking and receiving treatment.
Whereas Adams et al. (2003) tried to find a causal relationship between SES and health with old individuals, Currie and Stabile (2003) looked at the relationship between SES and health for children. They find that the relationship gets stronger as the child ages. However, they find little evidence that the relationship differs between high SES and low SES children. Intuitively, this is an unexpected outcome, as someone’s SES is determined by inter alia education and income. Children do not have significant measurable data yet that covers those factors.
Another determining measure in a child’s SES is whether the child comes from a two-parent or single-parent family. Gorman et al. (2008) find that U.S. children from single-parent families, meaning a low SES, show a significant negative effect on health care utilization. Access and use of health care for children are facilitated by having two parents to carry out various roles. Earned income can enable parents to pay for treatments, co-payments and other medical care. Likewise, private health insurance is typically obtained for children through a parent’s employer. Private health insurance improves access to health care. In this thesis the results are obtained from data that consisting of Dutch families. As everyone in the Netherlands is required to have a healthcare insurance, private health insurance is not a relevant variable for use in this model.
The same conclusion as Gorman et al. (2008), about health care utilization, was drawn by Adams et al. (2003). Just as Currie and Stabile (2003) they did not seem to find a direct causal relationship between SES and health, but like Gorman et al. (2008) they did find a negative relationship between someone’s SES and access to or utilization of health care. Therefore, this substantiates the hypothesis that low SES leads to a diminished ability to seek medical care or reduce access to medical services.
2.3 Measurements of SES
As mentioned in section one, there are different variables in which socioeconomic status can be expressed. Fuchs (2004) reflects on these as correlated to health. He discusses the interactions among the variables, nonlinearities, causal conclusions, and potential mechanisms of action. He concludes that the relationship between income and health is the most complicated one; the causality can run in both directions and the correlation varies between extremely high and very low. Education, in comparison with income, appears to be less complicated. Causality can run in the opposite direction, but it is not likely and the relationship is almost always positive. Occupation appears to be a solid variable as well. Clearly, some occupations such as construction worker are less healthy and more hazardous than working at an office. Age proves to be a good
7
indicator of health; it is a biological fact that health degenerates as people age. In examining gender, women have longer life expectancy than men and a positive correlation with health. While marital status has not been shown to have a significant positive effect on health, marriage often correlates with having children which is assumed to make a positive contribution to the household production of health or to increase the demand for health. Finally, the ethnic differences in health rarely show convincing explanations.
Kemptner et al (2011) also find that education proves to have a significant positive effect on health and health-related behavior. They find that education raises efficiency in health production and that it has an indirect effect mediated through better dwelling, higher income and better environmental conditions.
Another correlate to health is the environment. The environment shapes health behaviors (Adler and Ostrove, 1999). Bad neighborhoods, which are often correlated with low-income families, contain more liquor stores and less access to nutritious foods. Also, because low-income families cannot afford it, there will be fewer opportunities for sport and physical exercise. These environmental circumstances will also increase the probability that a person with a low SES will use drugs, tobacco, or alcohol. Further they may engage in unprotected sex or other high risk behaviors as well and demonstrate a reduced potential for health promoting activities such as regular exercise and sound dietary practice.
The next sections show what models will be used in order to test for causality between number of physician visits and SES and which variables of SES are used in this thesis.
3. ECONOMETRIC MODEL
3.1 The Poisson Model
When studying the number of physician visits, one has to take into account that the number of physician visits is a count. In order to estimate count data, the standard probability distribution is the Poisson distribution. Poisson distributed data is inherently integer-valued, which is logical for count data.
( | ) ( ) [ | ] ( | )
The likelihood of this function for the whole sample y = (y1, . . . , yn ) will be given
8 ( | ) ∏ ( ) [ ∑ ( ) ] ∏[ ( )]
The mean depends on a vector of explanatory variables and is specified as ( ). Now the log-Likelihood is given by:
( ) ∑[ ( ) ( )]
The parameters will be estimated by maximizing the log-Likelihood, as a function of β.
The Poisson model will only work correctly under two assumptions. The first one states that the events are independent i.e. the assumption that one event has no effect on the likelihood of observing additional events in the same period. If this assumption is violated, there will be positive contagion. Positive contagion will, in effect, increase the variance of the observed counts. It is very likely that this assumption will be violated. A person will visit the physician with different probabilities. When a person visits the physician with a certain probability, the second time he visits the physician will most likely be with a different probability due to an unpleasant experience or an appointment for a check-up.
The second assumption that must be made is the assumption that rates are uniform within time periods i.e. that all micro-events have the same probability. When this assumption is violated, one assumes heterogeneity.
3.2 A t-test for overdispersion
Positive contagion and unobserved heterogeneity both lead to overdispersion: [ ] ( ). Since the Poisson model requires the variance and mean to be equal, a test for overdispersion has to be performed. The means of testing for overdispersion is to calculate if the squared errors have a variance that is statistically different from . This is a t-test for
H0: vs Ha : for the equation:
( | ) ̂ ( ) ( | ) ( ̂ )̂ ( ̂ ̂ )
̂ ( ) 2
2 Cameron, A. C., Trivedi, P. K., (2005), Ch. 20: MODELS OF COUNT DATA: Overdispersion. Microeconometrics: Methods and Applications, Cambridge University Press. (p. 670 – 671)
9
3.3 An alternative model
The Negative Binomial Model is used to deal with overdispersion. The Negative Binomial essentially adds some unobserved heterogeneity. In this model the mean remains equal to , but the variance grows by a factor . The density of the distribution is: ( | ) ( ) ( ) ( ) ( ) [ | ] ( | ) ( )
The likelihood of this function for the whole sample y = (y1, . . . , yn ) will be given
by: ( | ) ∏ ( ) ( ) ( ( )) ( ( ) ( ))
The parameters will be estimated by maximizing the log-Likelihood, as a function of α and β. Here β estimates the parameters and α indicates the level of overdispersion.
3.4 The hurdle model
With the Poisson model, once the mean is given, all other aspects of the distribution are determined as well. This single index structure is another possible criticism why the model could not work correctly. The most important generalization of this single index structure is the hurdle model. The hurdle model combines a binary model for the decision of visiting the physician with a truncated-at-one count data model for the degree of visiting the physician, given that the person visits the doctor. In the health literature, the hurdle model has been very popular. This is due to the fact that a structural interpretation can be given to the model that seems to be in good agreement with the dual decision-making structure of the dependent variable number of physician visits.
An attractive feature of the Negative Binomial model is that it has a closed probability function. However, it has been argued (Winkelmann, 2004) that a more suitable model for adopting unobserved heterogeneity is the Poisson-log-normal model. For the hurdle, a probit model is combined with a truncated Poisson-log-normal model. For the probit part of the distribution, is a latent indicator variable such that:
( ) and
10
The positive part of the distribution is given by:
| ( ) Now, the density of the distribution is:
( | ) ( ) [( ( )) ∫ [ ( ( ))( ( ))] [ ( ( ))] ( ) ] ( ) ( ) ( )
The parameters will be estimated by maximizing the log-Likelihood of the probit and Truncated Poisson-log-normal model, as a function of γ and β respectively.
The next section shows which variables are chosen for SES in order to estimate the relationship between SES and number of physician visits. Following that, that the results are shown.
4. DATA
4.1 LISS
The data used in this thesis stem from Longitudinal Internet Studies for the Social sciences (LISS): an ongoing monthly survey that began in 2007. These surveys cover different subjects, such as health, income, work and schooling.One member in the household provides the household data and updates this information at regular time intervals. For the purpose of this thesis, only observations from a survey in January 2012 are used. This survey relates to 11705 individuals. Not every individual took the survey on the same subject at the same time. This is why the missing observation in the dependent variable number of physician visits had to be deleted. After this correction, a total of 5629 observations remained. The number of observations had to be further reduced to 3659, due to missing information on income and other socio economic variables. The focus in this thesis will be on a family’s SES where only the observations of the household head will be included. This last adjustment in the data reduced the number of observations to 2118.
11
4.2 Choice of SES
In section 2.3, the different measures of SES were briefly addressed. In this thesis the choice of SES measures is: income, education, occupation, and partnership. These variables are chosen because they are highly correlated with each other. Income proves to be highly correlated with education. On one hand, education is serving as a substitute for long-term income. On the other hand, education’s favorable effect on health works partly due to higher income. Higher paid occupation comes often through higher levels of education, so occupation is correlating with education as well as income. Married men and women are healthier than unmarried men and women. Marital status is usually correlated with income and education.
4.3 Descriptive statistics
Table I describes the data for the sample used. Other than the SES variables occupation, education, income and partnership, the variables urban, gender, age and self rated health status were added in order to estimate the model(s). These are chosen because they are not likely to correlate with the SES variables, but do contribute in explaining the number of physician visits.
The variable Primary Physician visits is the dependent variable and a count variable which represents the number of times the individual has visited the primary physician in the past year. Net income represents the total net income of the individual’s household. The LISS data did not contain a variable of total household income, so this had to be computed. In this way, net income equal to zero was (almost) excluded from the data, which includes volunteer work and others. In LISS, retirement benefits or student loans are also seen contributing to an individual’s net income. The variable education denotes the highest achieved level of education, expressed in number of years of schooling. The lowest achieved level of schooling is primary school, which corresponds with eight years of schooling. When the individual finished high school, divided in VMBO and HAVOVWO, they respectively had twelve and fourteen years of education. If the individual has a college degree, divided in MBO and HBO, the individual had sixteen and seventeen years of education. And finally the highest achieved level of education is University, which corresponds with eighteen years of education. LISS
Table I. Descriptive statistics
Variable Mean
Primary Physician visits 2.37 Net income 2516.38 Education 14.88 Paid work 0.5434 Jobseeker 0.0349 Student 0.0104 Household 0.0212 Pensioner 0.3159 Urban 2.9259 Partner 0.6166 Gender 0.6997 Age 59.33 Age2 3520.28
12
did not contain a variable providing the answer to how many years of schooling the respondent had. The respondent only had to answer what his highest degree in education was. The number of years of schooling was imputed, according to the usual number of years that is necessary in order to achieve these levels. In this fashion, the variable contains little bias for long-term students. The higher an individual’s obtained degree in education, the higher an individual’s SES. The occupation category is divided into five dummy variables: paid work, jobseeker, student, household and pensioner. The description of these dummies is trivial. The omitted reference category is unemployed and represents a low SES. The variable Urban has to do with the urban character of the residence of the individual, divided into five levels from extremely urban (1) to not urban (5). When the individual indicates to live extremely urban, it means that his residence has a surrounding address density bigger than 2500 addresses p/km2. Amsterdam has for example a
surrounding address density of 6065 addresses per square kilometer3. People who
live in an urban area feel less healthy than people who live outside the city (Maas & Verheij, 2006). Green contributes positively to the perceived health of people. Urbanites use more 'formal' care than residents of less urban areas. The variable Partner is a dummy variable, which indicates whether the individual lives together with a partner (married or unmarried). Not having a partner correlates with a low SES. Although literature found that being married as appose to being unmarried contributes to a healthier life, the partner variable is deliberately chosen over for instance a married dummy variable. Nowadays, marriage is no longer required to live together or to have children, which formerly was quite a taboo. The number of unmarried couples living together has increased with more than 60% percent over the past decade, from 1 million couples to some over 1.6 million couples3. Having a
partner while unmarried, must contribute in the same way as being married, in explaining the number of physician visits. The dummy variable gender (male=1) proves to be a good indicator of health, since the life expectancy of women is consistently higher than men in The Netherlands. Looking at the mean of the Gender variable, the majority of the sample is male. This is because the head of the household is mostly male.
A measure of rated health status is provided by a subjective self-assessment in response to the question: ‘How would you describe your health, generally speaking?’ with responses from 1 ‘not at all satisfied’ until 5 ‘excellent’. It is likely that all self-reported health status suffer from some biases.
The next section shows the results of the estimated models with the sample used.
3 Bron: CBS Statline
13
5. RESULTS
5.1 Poisson model
The estimates for the basic Poisson model are displayed in table II. Most of the effects are significant in this particular model specification and are common with those found elsewhere in literature. The net income of the household does not have a significant effect on the number of physician visits. The outcome is inconsistent with literature, however this is not unexpected as physician visits are included within the basic health insurance in the Netherlands and in effect free for the average Dutch person. The higher the obtained level of education of the individual, the less they tend to visit the physician. Employed individuals tend to visit the physician less than unemployed individuals. Here, job seeking individuals visit the physician the least and students the most. Students tend to have a less healthy lifestyle than other demographics, so this outcome is not very surprising. Also jobseeker may be more worried about their immediate financial future than their long term health. The less urban the individual lives, the more they visits the physician which is a surprising result as urbanites are expected to receive more medical attention than residents of less urban areas due to physician density tending to be higher in in urban areas. Having a partner also decreases the demand for physician visits. Men visit the physician almost 0.7 times less as women. Physician utilization significantly increases with age, but with a small value. Younger individuals see the physician more. An obvious outcome is that individuals, who rate their own health status high, will see the physician less often than people who rate their own health status low.
Coefficient t-value Net income 0.0000 -1.0200 Education -0.0531 -11.1700 Paid work -1.1988 -27.5200 Jobseeker -1.2646 -13.2500 Student -0.7339 -4.1800 Household -1.1346 -13.4500 Pensioner -1.1289 -26.2100 Urban 0.1423 12.7000 Partner -0.1869 -5.3700 Gender -0.6860 -19.6900 Age 0.0389 5.1400 Age2 -0.0002 -2.7000
Self rated health status -0.1501 -7.5600 Test statistic p -value
Overdispersion test 38.622 0.000
14
5.2 Negative Binomial model
Although the Poisson estimates are quite accurate with literature, a test for overdispersion must be performed. The large test statistic of 38.622 and small p-value rejects the null hypothesis that . It indicates the Poisson model is inappropriate, so the alternative Negative Binomial regression was performed. The estimates of this regression are shown in Table III. The Maximum Likelihood Estimator for alpha again proves that the Poisson model is inappropriate. Its value is equal to 1.0132 and shows a small level of overdispersion. The results of the Negative Binomial regression are similar to those of the Poisson regression. The same negative and positive relationships remain. Though, more variables failed to show a significant relationship. The number of physician visits is here specified to be conditional on education, occupation, urban character, gender, and self-rated health status. In the Poisson model the number of physician visits is conditional on education, occupation, urban character, partnership, gender, age and self-rated health status (all variables, but net income). When comparing the effect of income, education, and occupation on number of physician visits between the two models, the negative binomial estimates show less significant effects. A somewhat counter-intuitive result is the insignificant relationship between partnership and number of physician visits. According to literature, having a partner (wed or unwed) should increase health. The variable age also loses its significant relationship in the Negative Binomial model. This result is surprising, since age has always proved to be a good indicator of health.
Table III. Negative Binomial Estimates Physician visits Coefficient t-value Net income 0.0000 -0.6800 Education -0.0311 -3.0600 Paid work -0.9862 -9.1800 Jobseeker -1.0627 -5.9900 Student -0.6299 -2.0000 Household -0.8940 -4.6200 Pensioner -0.9421 -8.2900 Urban 0.0569 2.5600 Partner -0.0928 -1.3700 Gender -0.6059 -8.3900 Age 0.0061 0.4400 Age2 0.0001 0.8900
Self rated health status -0.3400 -8.6800
15
5.3 Hurdle model
The presence of overdispersion in the Poisson model was not very great, so it is worthwhile to explore the hurdle model based on the Poisson-log-normal model. The parameter estimates of the probit-Poisson-log-normal model are reported in table IV. In this table, the first column gives the estimates for the hurdle portion: the decision to visit the physician or not. The second column provides estimates for the positive part. The coefficients generally should be of opposite sign, due to the parameterization of the model. The sign test shows a few deviations. In the case of the variables occupation (except student), gender and self-rated health status. This is possibly due to the fact that the Probit model does not show very significant results; the t-values are rather small. Just as with the Poisson model and the Negative Binomial model, the estimates of the hurdle model are similar. The negative and positive relationships remain. Just like in the Poisson model, as well as in the Negative Binomial model, the relationship between education and number of physician visits is small and almost insignificant, a degree obtained in higher education will cause a decrease of 0.06 in number of physician visits. Once the individual has chosen to see the physician, their self-rated health status does not influence the number of times that they see the physician. Also the decision whether or not to see the physician does not significantly depend on the health status that the individual rates themselves. The age variable also remains to show slightly significant relationships with the number of physician visits.
Table IV. Probit-Poisson-log-normal model
Coefficient t-value Coefficient t-value
Net income 0.000 0.060 0.000 -1.200 Education 0.001 0.130 -0.057 -11.240 Paid work -0.050 -0.390 -1.291 -27.640 Jobseeker -0.150 -0.780 -1.306 -12.200 Student 0.622 1.780 -1.128 -5.390 Household -0.239 -0.960 -1.136 -13.130 Pensioner -0.109 -0.760 -1.147 -25.360 Urban -0.004 -0.180 0.168 13.700 Partner 0.012 0.170 -0.254 -6.640 Gender -0.494 -6.400 -0.575 -15.160 Age -0.009 -0.560 0.045 5.350 Age2 0.000 1.830 0.000 -4.060
Self rated health status -0.298 -7.220 -0.003 -0.120 Probit (0/1+)
log-normal (1+) Truncated
Poisson-16
6. CONCLUSION
Table V shows the summary of the results presented in the previous section. It shows the significant relationship between SES and the number of physician visits in the two estimated models, where ↗ denotes a positive relationship and ↙ a negative relationship. The ‘-‘ sign shows that the variable had an insignificant relationship.
The three estimated models show similar relationships between SES and the number of physician visits. Due to the fact that every person in the Netherlands has health insurance, the relationship between income and number of physician visits is negligible. Obtained degrees in higher education will cause a decrease in physician utilization which is consistent according to literature. Degrees in higher education directly correspond to a higher SES and thus shows a negative relationship between high SES and number of physician visits. Employed individuals tend to visit the physician less often than unemployed individuals. Again, high SES and number of physician visits show a negative relationship. Having a partner (wed or unwed) causes a decrease in physician utilization. Partnership corresponds to a high SES and thus high SES and number of physician visits show a negative relationship.
The analysis revealed that a high SES person visits the physician less than a low SES person. It substantiates the hypotheses that people low SES is more susceptible to bad environmental circumstances and therewith will have a worse health than a person with high SES.
The Poisson model proved to be an inappropriate model in obtaining correct estimates. The two alternative models were shown to be more appropriate since they accounted for unobserved heterogeneity and the single index structure of the Poisson model. Although the two alternative models were more appropriate for estimating the relationship between SES and number of physician visits, the variables showed less significant effects. It is debatable whether the alternative models were chosen correctly. Instead of the probit-Poisson-log-normal model, one could for example explore the probit-Poisson-log normal model with correlated errors. This model may prove to be accurate as it relaxes the assumption of conditional independence between the probit model and the truncated Poisson model.
Table V. The relationship between SES
and physician visits
Poisson Model NB Model
Probit-Poisson- log-normal
High SES Low SES High SES Low SES High SES Low SES
Net income - - - -
Education ↙ ↗ ↙ ↗ ↙ ↗
Occupation ↙ ↗ ↙ ↗ ↙ ↗
17
7. REFERENCES
1. Adams, P., Hurd, M. D., McFadden, D., Merrill, A., Ribeiro, T. (2003). Healthy, wealthy and wise? Tests for direct causal paths between health and socioeconomic status. Journal of Econometrics, 2003(112), p. 3-56
2. Adler, N. E., Ostrove, J. M., (1999). Socioeconomic Status and Health: What
We Know and What We Don’t. St. Paul Minnesota: Macalester College,
Department of Psychology
3. Currie, J., Stabile, M. (2003). Socioeconomic Status and Child Health: Why is the Relationship Stronger for Older Children? The American Economic View,
2003 (Vol. 93, NO. 5)
4. Fuchs, V. R. (2004). Reflections on the socio-economic correlates of health.
Journal of Health Economics, 2004(23), p. 653-661
5. Gorman, K. B., Braverman. J. (2008). Family structure differences in health care utilization among U.S. children. Social Science & Medicine 2008 (67), p. 1766–1775
6. Kemptner, D, Jürges, H., Reinholds, S. (2011). Changes in compulsory schooling and the causal effect of education on health: Evidence from Germany. Journal of Health Economics, 2011 (30), p. 340–354
7. Kiula, O., Mieszkowski, P. (2007). The effects of income, education and age on health. Health Economics, 2007 (16), p. 781–798
8. Maas J., Verheij R. A., (2006), Green space, urbanity, and health: how strong is the relation? Journal of epidemiology and community health, 2006 (7), p. 587-592
9. Mulatu, M. S., Schooler, C. (2002). Causal Connections between Socio-economic Status and Health: Reciprocal Effects and Mediating Mechanisms.
Journal of Health and Social Behavior, 2002 (Vol. 43, No. 1), p. 22-41
10. Salas, C., (2002). On the empirical association between poor health and low socioeconomic status at old age. Health Economics, 2002 (11), p. 207-220 11. Wichowski, H. C., Kubsch, S. M., (1997). The relationship of self-perception
of illness and compliance with health care regimens. Joumal of Advanced
Nursing, 1997 (25), p. 548-553
12. Winkelmann, R. (2004). Health care reform and the number of physician visits – an econometric analysis. Journal of applied econometrics, 2004 (19), p. 455-472
