Obesity and the Demand for Canadian Physician Services

The objective of this study is to determine the role that obesity plays in how often Canadians visit their family doctors or general practitioners. Doctor visits are analyzed using mixtures of ordered probability models applied to sample survey data from the 2010 Canadian Community Health Survey. This procedure is shown to be superior in terms of likelihood criteria to the more usual one involving count models of doctor visits. The main result is that obesity is one of the leading causes of doctor visits. Obesity has become more important in the demand for physician services than smoking for all Canadians. Other factors including diabetes, the individual’s level of education, position in the income distribution, and drinking behavior are also important. The application of latent class’s ordered probability models by age-group and gender leads to results which are different from what others have found. While obesity is shown to be a serious problem in Canada, it has not yet reached the stage which some researchers have described as critical.


I
The weight of Canadians has increased enormously over the last three decades.For the thirty year period 1978-2008 a recent report on obesity produced by two Canadian health agencies 1 showed that the proportion of Canadians aged at least 18 who are obese has increased from 14% to 25%.This is one of the most remarkable and alarming changes in human physiology that has ever been recorded.Concerns arise because obesity is associated with the prevalence of diabetes for both children and adults, with heart disease, with some forms of cancer, stroke, and a large number of other ailments.The report mentioned above also noted that illnesses associated with obesity cost the Canadian economy an amount somewhere between 4.6 and 7.2 billion dollars per year in 2001.Longitudinal studies from several countries show high personal costs as well.At age forty, Peeters et al (2003) using data from the Framingham heart study showed that non-smoking males and females lost 5.8 and 7.1 years of life expectancy, respectively because of their obesity.In Canada, The Canadian Diabetes Association (2010) and the Heart And Stroke Foundation (2011) have both produced documents which stress the seriousness of what they call an 'obesity epidemic'.Similar concerns have been expressed by The World Health Organization and governmental agencies like the National Institute of Health (2009) in the United States and The European Commission (2007).
Most of the research on obesity in Canada is based on the Canadian Community Health Surveys which are now carried out every year.This is valuable series of large surveys which has made it possible to get a definitive picture of levels and trends in body weight over the last thirty years for representative samples of Canadians when they are combined with the earlier National Population Health Surveys and the Canada health Survey.These surveys also contain detailed information on the prevalence of diseases which afflict Canadians as well a battery of questions that elicit information on personal habits like smoking, the use of alcoholic beverages, illicit drug use, exercise, and dietary behaviour.In addition, these surveys contain information on the number times respondents visit their family doctor or a general practitioner, the number of visits to more specialized physicians and the number of nights that they spend in a hospital.
The purpose of this project is to analyze the determinants of doctor visits using data from the 2010 Canadian Community Health Survey.While some attention has been focused on the associations between obesity and heart disease and diabetes the effect of obesity on the actual utilization of the various services that are provided by Canada's health care system have not been examined by Canadian health researchers.
Unlike some of the studies that have been carried out using recent European and American data there are no studies of doctor visits for Canada and consequently there is no information on the impact of this country's obesity problems on the demand for access to family doctors and general practitioners.The same can be said about overnight visits to hospitals or the demand for specialized physician services.
It is important to have a clear picture of who visits their doctor, why they do this and how often this happens.Most individuals who utilize the health services begin by seeing a family doctor or visiting a health clinic so this is an important first step in the process where patients receive medical care.It is at this stage that trends in the type of user and the type of service can be identified.Latent class ordered probability models will be used to determine which factors are most important in determining how often an individual visits his or her family doctor or GP.The main focus is on the effects of obesity but other factors need to be considered in order to evaluate the effects of obesity relative to other factors like smoking, age and socioeconomic position.
To summarize the main results, for most male and all female age groups obesity is the most important explanatory variable in the determination of doctor visits.For both genders having a BMI of greater than thirty leads to significantly more doctor visits than smoking.Individuals who smoke, have diabetes and heart disease, are inactive, come from the lower part of the income distribution, or are poorly educated are also more likely to visit their doctor.Unobserved characteristics play an important role in determining who visits their doctor.The presence of genetic characteristics and various dimensions of behaviour and obesity history that are unobservable to the researcher require a statistical procedure do deal with them.When this is applied to the data a typology of respondents emerges where not all individuals are affected in the same way by how much they weigh.As a result, the role of obesity as a determinant in the demand for doctors' services is complex and depends on the evolution and duration of the individual's obesity status.
The paper is organized in the following way.The next section describes the data that is used in the analysis.Section 3 describes a set statistical models which can be used to explain doctor visits.The results are contained in section 4 and section 5 contains a discussion of the results.

D
The data comes from the 2010 Canadian Community Health Survey.This is a random sample of Canadians who have access to a telephone.The survey focuses on health issues but there is a very good selection of demographic variables which describes the respondents socioeconomic position as well as detailed information on smoking behaviour, alcohol and recreational drug use.Excluding respondents who failed to reveal their weight and height, the frequency of their doctor visits and those less than twenty years of age left a sample of 23755 males and 28400 females.The total sample survey size is 62909 and most of the exclusion are due to age.The variables used in the analysis which appear in Tables 3A and 3B are the natural logarithm of the respondent's body-mass index, BMI, marital status (married or common law), 5 smoking variables, starting with being a daily smoker, an occasional (former daily) smoker, etc. with the residual category being never smoked.There are two alcohol use variables: a regular drinker and an occasional drinker with the residual category being a non-drinker 2 .There are 4 educational categories: less than secondary school graduation, secondary school graduation, some post secondary education, and missing.The residual category is post secondary school graduation.Income is measured by the respondent's decile in the income distribution.Three diseases are included; these are diabetes, heart disease, and a group of other serious disease like cancer and stroke.The last variable is a physical activity index.Age was also included as a regressor but the survey uses five year age intervals and since all of the analysis involves ten year age intervals there is not enough variation in age within age groups to have any effect on doctor visits.Age group was never significant as a variable although doctor visits are much more frequent in the older age groups.
Much of the literature on obesity uses a set of discreet categories to represent the degree of obesity.Individuals are considered obese if their BMI is 30 or above.However, there are three categories of obesity.Category I is 30 < BM I ≤ 35; Category II, the severely obese, is 35 < BM I ≤ 40; and Category III, the morbidly obese is BMI > 40.Results involving cross tabulations were usually based on an aggregation of these categories because there were too few observations in category III.

P M D
V .
Most studies which analyze the number of times a respondent visits his or her family doctor model this behaviour as a count model.Often these models are augmented to take account of excessive zeros.This procedure is now well established and a thorough description of it may be found in Winkelmann (2009).However, an alternative method will be used to followed here.There are a number of reasons for this.First, the data for higher outcomes is not likely to be correct since some respondents appear to be unable to recall the exact number of visits when they report more than six or seven visits.This leads to spikes in the data at 10, 15, and 20 visits so that all of the commonly used count distributions will not fit the data very well.This problem can be circumvented by using a censored count distribution whose likelihood function is where d i is the number of visits by respondent i, k is the censoring point and f (d i ) is either the Poisson or negative binomial probability mass function.
As Winkelmann (2009) notes, there are situations where the data or the type of doctor requires that special treatment be given to the outcomes {d i = 0 or 1}.When this is needed two part models like the zip or the hurdle model can be employed.However, visits to Canadian family doctors or general practitioners do not display excessively large or small numbers of zeros as is clear from Table1.The event {d i = 1} is the most common for most male and female age groups but it does not stand out as a candidate for special treatment either.For males aged less than 50 the event {d i = 0} is the most likely but this event appears not to pose any problems for fitting statistical models to these age groups.
For reasons involving likelihood criteria and goodness of fit the procedure used here will treat the number of doctor visits in an ordered probability framework with a threshold parameter for each distinct number of visits rather than using a count distribution.In order to minimize the impact of guesswork in the determination of higher order outcomes all outcomes greater than 6 are grouped together as was the case for the censored count model.Unobserved respondent heterogeneity is treated by assuming that there are a finite number of latent classes or types who respond to their characteristics in a specific way.The respondent's characteristics and attributes are described by the vector Z i which is a set of respondent specific variables all of which have been normalized to have a zero mean an unit variance.This means that the size of the estimated coefficient indicates the importance of the regressor associated with it.
The the probability that respondent i has n visits is given by where Φ() is the cumulative normal distribution function with mean zero and variance one and {k n : n = 0, 1, ...6} is a set of increasing threshold points.γ j Z i is type specific morbidity or ill-health index.As this increases individuals have more doctor visits the more thresholds they cross.All respondents have the same threshold points but threshold points differ across types by allowing an intercept term, γ 0j , for j ≥ 2. To identify the model the intercept term for the first mixture is set equal to 0. The π j parameters are the type probabilities where π j is the probability of a respondent i being type j and π j = 1.
The sample ln-likelihood function for this model is, therefore where N is the sample size.The number of mixtures to be used is determined by the data.Mixtures are added until there is an increase in the Akaike index function or because the algorithm which maximizes the likelihood function fails to converge.The models that will be estimated here are more complicated than the simple probability model since it is necessary to allow for unobservable effects and for the possibility that not all respondents will react to variables which describe their situation in exactly the same way.Models which allow for these options are called latent class models and are described in Cameron and Trivedi (2009).Mixed ordered probability models always outperform mixed censored count models in terms of Vuong's (1989) non-nested criteria.Maximized likelihood function values are much higher for the mixed ordered probability models.They also fit the data better than mixed count models.The estimated individual outcome probabilities differ from the actual proportions only at the third or fourth decimal place so that the model fits the data extremely well for all age groups and each gender.
Parameter estimates for the most important variables for models using 4 or 5 mixtures appear in Table 3.At most 5 mixtures could be estimated as the maximum likelihood algorithm failed to converge for 6 mixtures.Akaike information criteria support 5 mixture models except for males aged 20-29 where 4 is the optimal number.This does not mean that there are only 5 latent classes.There may be more but the data is not rich enough to identify them.

R
As is clear from Tables 3A and 3B the logarithm of BMI is the most important and significant variable in the explanation of doctor visits for females of all age groups and for males over the age of 50. 3 .For males aged 30-49 it is the second most 3 Instead of the logarithm a quadratic could have been used.Both representations yield the similar results, except that the squared term is sufficiently negative to make the effect of large values of BMI negative with respect to doctor visits.However, using a set of dummy variables to represent a discreet set of BMI classes will lead to biased parameter estimates.See Breslaw and McIntosh (1999) for details.important.Only for males aged 20-29 is average ln(BMI) not significant, although even in this case there are two types for whom it matters a great deal.The message that Tables 3A and 3B convey is that obesity is the most important independent factor in determining how often individuals utilize the health care system through their family doctor or a general practice health clinic.And this is the case even when important diseases like diabetes, heart disease, and a large number of other diseases are allowed to play a part in explaining doctor visits.This confirms the cross tabulation results in Table 1 which are not altered when other factors are considered in the determination of doctor visits.
Various smoking and alcohol use variables are also significant as determinants of doctor visits.Both genders are significantly less likely to visit their doctor or a health clinic if they are regular drinkers.This is yet another confirmation of the beneficial effects of moderate alcohol use that have been found in other Canadian studies which look at the relation between alcohol intake and self-reported health, heart disease and diabetes 4 .Respondents who are better educated and higher up the income distribution have fewer doctor visits.
For both genders there is considerable variety across types with respect to the size and signs of the coefficients associated with ln(BMI).It is also the case that the probabilities of having at least one doctor visit are almost unrelated to the size of the associated ln(BMI) coefficient.For example, in the case males aged 20-29 type 2 has the largest ln(BMI) coefficient at 1.084 * * (0.260) but the probability of having at least one doctor visit is the second lowest at 0.611 † (0.388).But this is not typical of other age groups since there no consistent pattern in the relation between the effects of ln(BMI) and the probability of utilizing the services of a doctor.This is not what Datta-Gupta and Greve (2011) in their study of the effects of obesity on doctor visits which uses the same two class methodology as Deb andTrivedi (1997, 2002) versus unhealthy.This categorization scheme is well defined if more doctor visits are associated with a particular class.For example, Datta-Gupta and Greve (2011:57) in their two class model find small and insignificant parameters associated with being overweight or obese for their first latent class but much larger values of these two parameters for their second latent class leading to more doctor visits for the obese in the second latent class.This leads them to label the respondents in these two classes as infrequent and frequent users, respectively.In the model here BMI is continuous so that it is the size of the parameter that matters but there is no monotonic relation between the size of the BMI parameter and the frequency of doctor use as measured by the probability of a positive number of doctor visits.Why should the results reported here be different from what others have found?There are three potential causes.First, mixtures of ordered probability models instead of count models are being used.Secondly, four or five mixtures are being used instead of just two.Finally, rather than pooling all of the data, separate models are estimated on a two-way classification of age and gender.Akaike criteria always support more than two mixtures and likelihood ratio tests always reject pooling across gender and age.However, when the mispecified pooled two mixture model was run on the same type of sample as Dastta-Gupta and Greve (2011), males and females aged 25-60 with a gender dummy for males, there was a monotonic relation between ln(BMI) and the probability of positive doctor visits.In this case the two ln(BMI) coefficients are 0.178 * * (0.012) and 0.234 * * (0.016) and the two associated probabilities are 0.662 * * (0.154) and 0.915 (0.051), respectively.When this procedure was carried out on individual age groups using only two mixtures this relation was found for some but not all of the cases.From this one concludes that the two mixture pooled model leaves out important effects of age, gender, and those due to unobservables and this may produce misleading results.The effects of diabetes on doctor visits is a good illustration of this.For males, diabetes has a significant impact for age groups 30-59 but no significant impact for any of the female age groups.

D C
In order to get a better understanding of the results of the previous section it is necessary to focus on some of the more physiological and detailed characteristics of obesity that medical researchers have discovered.Matsuzawa (2010) describes obesity as an endocrine disease and Kyrou et al (2010: 214) characterize it as a 'chronic inflammatory state' involving white adipose tissue otherwise known as body fat.They observe that ".... the complex function of adipose tissue in obesity is the association of its expansion with the development of a low grade chronic inflammatory state...The proinflammatory nature of adipose tissue is heightened proportionally to the increased fat mass and shows consistently strong correlations with visceral adipose accumulation and BMI... Obesity has been shown to induce an unremitting proinflammatory response, which continues to evolve for as long as the weight gain is maintained.Recent evidence links this prolonged activation of inflammatory signalling pathways to the pathogenesis of both type 2 diabetes and atherosclerosis." This result helps to explain why individuals who are obese need to see their doctors more frequently than those whose weight is normal.In the severely or morbidly obese adipose tissue can make up more than 50% of body mass.This creates a serious health problem for these individuals which exacerbates medical conditions that are not so debilitating for the rest of the population.High levels of BMI lead to increased doctor visits over and above those which are associated with the diseases that are caused by obesity because the treatments of these diseases are usually not accompanied by weight losses sufficient to reverse the adverse effects on the individual's endocrine system.
The reasons why the importance of BMI differs across types are also better understood from a pathophysiological perspective.There are two factors that need to be considered here.First, not all fat is equally problematic for health.Bray (2007: 4) notes from earlier studies that 'patients with upper body fat were at greater risk for diabetes and heart disease than patients with lower body fat'.Location of fat deposits is important and waist circumference, which is a measure of visceral fat, has been shown to be associated with lipid abnormalities independently of BMI, Bray (2004: 2586).There is no data in the survey used in this study on the individual's fat distribution so that some of the variation in the coefficients associated with ln(BMI) may be due to omitted effects that arise from this missing information.
Perhaps more important is the lack of information on the individual's BMI history.Many research papers based on longitudinal data show that both the degree of the individual's overweight as well as its duration have an impact on the risks of insulin resistance, impaired glucose tolerance, and type 2 diabetes, Black et al (2005Black et al ( : 1199)).
Current BMI is a good representation of long term BMI for some respondents but not for others.Individuals whose increase in BMI was recent may not yet be paying the full price of being overweight or obese.Unfortunately, the obesity literature is largely silent on the importance of age at first obesity.It is also not clear whether the effects of obesity are the same for durations of the same length but start at different ages.Bray (2007: 73) notes that for a large American study no significant increases in plasma glucose levels were found for overweight or obese durations of less than ten years but for durations of more than ten years obesity did pose a serious health problem.This is, no doubt, part of the explanation of why ln(BMI) can have a significant negative coefficient for some of the types.Note, for example, from the first row of Table 3A, that type 1 males aged 30-39 have a highly significant coefficient of -0.515 * * (0.142).This negative coefficient is most probably picking up the negative association between current BMI and BMI ten or more years earlier for individuals who have had large weight gains over this period.Almost all changes in BMI are increases.The survey asks about weight loss and less than 1% of the respondents claimed to have lost any weight.The effects of obesity on health do not occur immediately; they appear with lags that can be quite long and possibly vary with the individual, Kopelman (2000: 635).
Prevalence rates of diabetes increase dramatically the higher the BMI class.For men these rates increase dramatically with age and reach a maximum of almost 40% between age 70 and 80 and then decline because of higher mortality rates for this group after age 70.Rates are a little lower for women and start to decline at earlier ages.For the survey data used in this study the simple correlation between diabetes and BMI for males is 0.21 and is highly significant.However, almost all of the respondents covered in this table who have diabetes acquired it at least ten years prior to the survey.While the present can not explain the past, present levels of BMI are likely to be highly correlated with BMI levels at or before the onset of diabetes.But, as already mentioned, BMI levels do change over time for part of the population and ten years provides ample opportunity for getting fat.Having a more complete picture of the individual's weight history would be very beneficial in improving our understanding of the role that obesity plays in determining who suffers from diabetes and other diseases whose risks are related to excess body weight.
So far the discussion has concentrated on the some of the more technical aspects of the results.Focus turns now to the broader question of the importance of overweight and obesity in terms of its impact on the health care system.Table 4 provides some simple comparisons of the effects of obesity with another major determinant of doctor visits: smoking.As was suggested in McCann et al (2007) the health impacts of the number of cigarettes smoked per day and BMI were qualitatively very similar in their significant adverse associations with self-reported health, coronary heart disease, and diabetes.The results in this table for doctor visits are even clearer and show that obesity is now significantly more important in determining the frequency of this type of contact with the health system than smoking.For both men and women, individuals who are obese, BM I ≥ 30, have more visits to their GP or family doctor than daily smokers.This is evident when daily smokers with BM I < 25 are compared to non-smokers with BM I ≥ 30.For 8 of the 10 age groups the number of doctor visits was more for respondents with a BMI at least 30 than for normal weight daily smokers.The two exceptions are males aged 20-29 and females aged 40-49.
Another way of assessing the impact of obesity on healthcare utilization rates is to ask what proportion of doctor visits is due to obesity.If all individuals were of normal weight then the total number of male doctor visits would be about 9.2% less than is actually observed.The same number for females is 10.6%.This is a substantial reduction in the use of family doctors that would occur in the hypothetical case where obesity were to disappear from Canadian society.Alternatively, one could also look at the case where everyone had a BMI of 30 or above.This would lead to an increase of about 7.5% in doctor visits.While both of these cases raise concerns about the seriousness of obesity the situation in Canada involving the use of medical services by individuals who are overweight or obese has not yet reached the stage that some of the literature describes as catastrophic or of epidemic proportion.Canadian health services are not going to be crushed by the weight of it users now or in the near future.Nor is the disappearance of the obese going to solve the current problems involving wait times and the shortage of GPs.
While the situation concerning the utilization of physician services may not be as critical as is sometime claimed the impact of obesity on the prevalence of diabetes is more pronounced.The application of these simple conceptual experiments to diabetes produces results which are much more worrisome.If obesity disappeared and every one was of normal weight there would be 43.3%fewer cases of diabetes in men and 53.8% fewer in women.On the other hand if everyone was obese this would lead to an 83.4% increase in male diabetes cases and a 123.2% increase in female diabetes cases.
In addition to the extra social costs that arise from the overuse of publicly funded services by individuals with excessive weight there are substantial private costs that have to be borne by these individuals.As was mentioned earlier, there are considerable reductions in life expectancy that are caused by being obese and the obese will eventually be more likely to suffer from diabetes.Moreover, the obese suffer stigma socially and in the work place and they are considerably less satisfied with their every day situation than is the case for the average citizen.BMI exerts a significant negative impact on the self-reported index of 'satisfaction with life in general' that the CCHS produces.
In summary, obesity is the most or second most important factor for both men and women in determining how often they seek the services of family doctors or general practitioners.Is is more important than being a daily or former smoker or having diabetes or heart disease if this is measured by the relative size of the average BMI coefficient in a mixture of ordered probability models which explains doctor visits.Drinking behaviour is also important as are the individual's level of education and position in the income distribution.Two counter-factual experiments were carried out to see how doctor visits would change if the weight characteristics of the population changed.Both revealed that the dramatic events like the disappearance of obesity or the arrival of universal obesity would lead to substantial but not overly dramatic changes in the demand for doctor services.Given the large personal costs that the obese pay in terms of shorter life expectancies, general ill health and restrictions on lifestyle the present situation, if not critical, is certainly very serious.

Acknowledgement
The author benefited from discussions with Peter R. Weldon, formerly professor , Jiménez-Martín et al (2002) and D'Uva (2006).The original motivation for using latent class models in the analysis of doctor visits in the Deb and Trivedi (1997) paper was to provide an alternative to the two part hurdle and zip models for dealing with excessive zeros.More recently the focus of attention has shifted from use versus non-use to infrequent use versus frequent use or healthy

Table 2
Distributions of Doctor Visits ByAge Group And Gender.Table 3A Average Maximum Likelihood Parameter Estimates For Mixed Probability Models of Doctor Visits For Males.

Table notes :
†, *, and ** indicate significant at the 10, 5, and 1 percent levels.The average effect is π j γ kj for variable k.

Table notes :
†, *, and ** indicate significant at the 10, 5, and 1 percent levels.The average effect is π j γ kj for variable k.

Table 4
Doctor Visits By Smoking and BMI Category:Males .