Adjusting for comorbidities in cost of illness studies

Abstract

Motivation:

Differences in cost of illness (COI) methodological approaches have led to disparate results. This analysis examines two sources of this variation: specification of comorbidities in the estimated cost models and assumed prevalence rates used for generating aggregate costs. The study provides guidance in determining which comorbidities are important to include and how to handle uncertainty in optimal model specification and prevalence rate assumptions.

Methods:

Comorbidities are categorized into four types. Type I comorbidities are those that increase the risk of the disease of interest; Type II comorbidities have no causal link to the disease of interest but are, nonetheless, highly correlated with that disease; Type III comorbidities are illnesses that the disease of interest may cause, and Type IV are comorbidities that have no causal link to the disease of interest and are only weakly correlated with that disease. Two-part models are used to estimate the direct costs of rheumatoid arthritis and diabetes mellitus using 2000–2007 Medical Expenditure Panel Survey data.

Results:

COI estimates are sensitive to the specification of comorbidities. The odds of incurring any expenses varies by 71% for diabetes mellitus and by 27% for rheumatoid arthritis, while conditional expenditures (e.g., expenditures among subjects incurring at least some expenditures) vary by 62% and 45%, respectively. Uncertainty in prevalence rates cause costs to vary. A sensitivity analysis estimated the COI for diabetes ranges from $131.7–$172.0 billion, while rheumatoid arthritis varies from $12.8–$26.2 billion.

Conclusions:

The decision to include Type II and Type III comorbidities is crucial in COI studies. Alternative models should be included with and without the Type III comorbidities to gauge the range of cost effects of the disease. In generating costs, alternative values for prevalence rates should be used and a sensitivity analysis should be performed.

Keywords::

• Cost of illness
• Rheumatoid arthritis
• Diabetes mellitus

Introduction

The number of cost of illness (COI) studies has risen substantially over time1. COI studies may be useful for a variety of reasons. They provide insight into the economic effects of medical treatments and technologies designed to alleviate diseases. From a policy perspective, they may inform private health insurance coverage decisions, the design of employer-sponsored healthcare benefits packages, and legislative decisions affecting the health insurance industry. Rice2 has further noted that COI studies are useful to

1. Define the magnitude of the disease or injury in dollar terms;

2. justify intervention programs;

3. assist in the allocation of research dollars on specific diseases;

4. provide a basis for policy and planning relative to prevention and control initiatives; and

5. provide an economic framework for program evaluation (p. 178).

COI studies have been used in Congressional testimony, official publications, reports, and public speeches2–6. Congress has expressed strong interest in using COI studies as one criterion for allocating research dollars among the National Institutes of Health2–6.

However, COI studies have been criticized for a number of conceptual reasons. First, it has been argued that these studies are inappropriate because they are not grounded in welfare economic theory7. A further concern is that COI studies, which typically measure healthcare expenditures, do not reflect the utility costs to the individual with the disease. In this view, willingness-to-pay (WTP) studies to avoid disease may provide more accurate measures1–8. Another criticism is that, unlike cost benefit and cost effectiveness analyses, COI studies do not compare costs to benefits and, therefore, have limited value to guide decision-making9–14.

In addition, questions have been raised about alternative empirical approaches to COI studies. The merits of using prospective vs retrospective data have been considered, as well as using epidemiological vs regression-based methods to estimate disease-attributable costs7^,15. There is great variability in methodologies and results among COI studies, even within specific diseases, which can limit the comparability of COI findings1^,10^,14. COI results may vary for a variety of reasons, including the perspective of the analysis, data sources, inclusion of specific cost items, statistical methods employed, and time frame of the analysis1^,14.

One challenging aspect of COI studies involves attributing costs to a specific disease. Rosen and Cutler6 stated:

Perhaps the biggest issue in disease-specific studies is the adding up constraint: it is not entirely clear what costs are associated with each disease, and how to ensure that all medical spending is allocated to one and only one disease (p. 88).

These concerns are well-founded, as failure to control adequately for comorbidities risks double-counting of disease costs and implausibly large aggregate cost estimates6. Lee et al.16 found that failing to control for comorbidities nearly doubled the estimated annual costs of osteoarthritis. Bloom et al.17 reported that total cost estimates for 80 diseases exceeded actual 1992 US healthcare expenditures.

Uncertainty regarding disease prevalence is another major concern for COI investigations. Nationally-representative healthcare databases such as the Medical Expenditure Panel Survey (MEPS) are often useful in estimating the incremental costs of disease. While these databases are rich in detail on expenditures, comorbidities and sociodemographic factors, self-reported estimates of disease prevalence may be unreliable, depending upon the disease studied. For conditions like osteoarthritis, individuals may be unaware that they have the condition, and only those with the most severe cases may report their disease. This would lead to under-estimates of the true disease prevalence. However, more symptomatic diseases, such as inflammatory bowel disease, may be less subject to this problem.

As an alternative to patient self-reports, determining disease prevalence based on clinician examination of patients has obvious appeal. However, studies that provide this level of detail often have dramatically smaller sample sizes than self-reported databases, and the resulting sample may not be representative of the population of interest.

To address the concerns of attributing costs to a specific disease and the uncertain prevalence of disease, this study was conducted. The study objectives are to examine (1) various econometric costs models for accurately attributing costs for specific diseases, (2) the potential biases surrounding controlling for disease comorbidities, and (3) different methodologies for estimating disease prevalence. This study describes the potential bias from controlling inappropriately for relevant comorbidities and provides guidance for proper specification of cost models. The study also considers the importance of alternative prevalence estimates and argues that any COI study should include a sensitivity analysis that takes into account uncertainty resulting from the challenges in specifying econometric cost models accurately and obtaining reliable estimates of disease prevalence to generate aggregate cost estimates1^,2.

For illustrative purposes, these issues are examined using a COI study for two chronic diseases, one that is quite common, diabetes mellitus, and one that is relatively rare, rheumatoid arthritis. Many COI studies have been conducted for diabetes mellitus15^,18. Diabetes mellitus is a common and costly illness that is both a risk factor for a number of other diseases, may itself be caused by other conditions such as obesity, and is correlated with other medical conditions15. Therefore, appropriately attributing costs and controlling for confounding comorbidities for this condition are important.

Rheumatoid arthritis is a less common chronic disease than diabetes mellitus, although it may be quite debilitating19–21. Moreover, although rheumatoid arthritis has a lower prevalence and lower healthcare expenditures than diabetes mellitus, annual per capita expenditures are higher than for individuals with diabetes mellitus22–26. Like diabetes mellitus, patients with rheumatoid arthritis report a number of comorbidities. Hence, the issue of controlling appropriately for confounding comorbidities is potentially quite important for both diseases.

Data, variables and methods

Data

Data from the 2000–2007 Medical Expenditure Panel Survey (MEPS), a nationally-representative database developed by the Agency for Healthcare Research and Quality that reports healthcare utilization and expenditures, health status, health insurance coverage, and socio-demographic and socioeconomic characteristics for the civilian, non-institutionalized population in the US were used27^,28. The MEPS database is a sub-sample of the National Health Interview Survey conducted by the National Center for Health Statistics (NCHS). MEPS utilizes a complex sampling design, which includes clustering and over-sampling of certain sub-groups such as minorities27. For MEPS, all interviews are conducted in person using a computer-assisted personal interview as the main data collection method27. The MEPS response rate is 75%27. Individuals with diabetes mellitus and rheumatoid arthritis were identified based on self-reporting. Individuals’ self-reported health conditions were mapped to the International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) diagnostic codes. Individuals were categorized as having diabetes mellitus and rheumatoid arthritis if they had an ICD-9-CM diagnostic code of 250 and 714, respectively. The study sample included adults who were at least 18 years of age.

Variables

Medical expenditures

The expenditure data included spending on physician visits, hospital and outpatient services, medications, diagnostic testing, and other medical services. We estimated the likelihood that a patient has positive medical expenditures and constructed conditional expenditure models for those subjects who did incur positive medical expenditures. To normalize the distribution, expenditures were log-transformed.

Clinical and demographic characteristics

The predictors of expenditures included major chronic diseases, sociodemographic characteristics, geographic region, and year. Chronic diseases were measured as binary variables and include diabetes mellitus and rheumatoid arthritis along with a large number of other major chronic illnesses and conditions. These conditions were chosen based on their prevalence and clinical considerations. A wide array of comorbidities were included given that previous research has shown that failure to control adequately for comorbidities may lead to substantial upward bias in the estimated expenditure impact of that disease16.

Sociodemographic and economic factors include age strata, education, race, income, marital status, and health insurance type. Race/ethnicity variables include African-American, Hispanic, and other non-Caucasian, with Caucasian serving as the reference group. Health insurance status was measured as a series of binary indicators of Medicaid, other public insurance, uninsured, private non-HMO insurance, and private HMO insurance (reference group). Geographic variables include Census Region (Midwest, South, West, and Northeast, with Northeast serving as the reference region) and whether a patient resides in a Metropolitan Statistical Area (MSA). Year is measured as a series of binary variables, with 2000 serving as the reference year.

Statistical methods

Empirical models

Two-part models were estimated in which the likelihood of incurring any expenditures and the natural logarithm of conditional expenditures were estimated separately by logistic and ordinary least squares (OLS) models, respectively29. The two-part model is frequently used in health economics research when many observations are clustered and the remaining observations are skewed to the right30^,31. The logistic regression models are specified as: (1) where Prob(expenditures) is a binary variable equal to 1 if medical expenses are positive and 0 otherwise; DM is a binary variable equal to 1 if the patient has diabetes mellitus and 0 otherwise; COMORBID are indicator variables for the presence or absence of other diseases; X is a vector of sociodemographic, economic, region, and year variables; α₀, α₁, β, and Θ are coefficients to be estimated; and ɛ is the error term. A similar equation is estimated for rheumatoid arthritis, using a variable, RA, which equals 1 if the patient has rheumatoid arthritis and 0 otherwise.

In the second stage, we use ordinary least squares (OLS) regression to estimate conditional expenditure models as: (2) where ln(expenditures) is the natural logarithm of healthcare expenditures and other terms are as defined above. Again, a similar equation is estimated for rheumatoid arthritis. Expenditures are adjusted to 2008 US dollars using the Medical Care component of the Consumer Price Index (CPI). All models are estimated using SAS version 9.1 (SAS, Cary, NC).

Specification of comorbidities

As noted above, proper control for comorbidities is important to obtaining reliable estimates for the incremental costs of the disease under study. Failure to include important comorbidities has been shown to increase the estimated costs of the study disease substantially16. However, it is unclear what comorbidities to include. To gain insight into this question, we identify four types of comorbidities, argue which of the types are most critical to include, and then use case studies of diabetes mellitus and rheumatoid arthritis to illustrate how the estimated cost of diabetes mellitus and rheumatoid arthritis changes as more of these comorbidity types are included in the model.

Type I comorbidities are those that can lead to the disease of interest but that the disease does not cause. How to analytically handle Type I comorbidities is conceptually challenging. To illustrate, suppose that a Type I comorbidity, call it disease X, increases the risk of the study disease. In this case, including disease X in the model would ‘rob’ the study disease of some of its effect on costs, because one would only be estimating the costs of the study disease that was not caused by disease X.

Type II comorbidities are those that are correlated with the study disease but have no casual connection to it. It is argued that Type II comorbidities should be included in cost models of the study disease. To illustrate the rationale for this, consider a variable such as patient’s age. Like Type II comorbidities, age, too, may be strongly positively correlated with the study disease but has no causal connection to it. Any reasonable cost model for the study disease would adjust for patient’s age, because failure to do so would lead to an over-estimate of the effects of that disease. The same rationale applies for including Type II comorbidities.

Type III comorbidities are conditions that the study disease is known to cause. As noted above, including these variables would lead one to under-estimate the cost effects of the study disease, since that disease is an important cause of these conditions. In the extreme, if every condition that the study disease causes were included in the model, one might find that it had little if any effect on cost. However, this would simply reflect that the model is estimating the cost effects of extremely benign cases of the study disease; namely, those that do not lead to any other diseases.

These observations appear to suggest that Type III variables should be excluded from the cost model. However, for some patients, Type III conditions may not have been caused by the study disease but are simply correlated with it. For example, many persons have diabetes mellitus and hypertension but there is not necessarily a causal link between the two in every case. Thus, for some patients, conditions that are seemingly Type III are in fact Type II, in which case they should be included in the cost models. Similarly, for some patients with comorbidities that seem to be of Type I, the relationship may not be a causal one but merely a correlation, in which case it is really a Type II comorbidity. For example, even though smoking is known to be a strong risk factor for lung cancer, some smokers who contract lung cancer may do so for reasons other than smoking. In most healthcare databases like MEPS, however, one cannot discern whether relationships between comorbidities for individual patients are causal or not. Given these uncertainties, we recommend estimating cost models both with and without Type I and Type III comorbidities.

Type IV comorbidities are those diseases that have no causal relationship with the study disease and little correlation to it. Whether these conditions are included or not should have little effect on the estimated cost of diabetes mellitus. However, to the extent that they are independent risk factors for cost, they should be included as part of any well-specified cost model. Figure 1 illustrates the four types of comorbidity situations discussed above.

Figure 1. Diagram of inclusion of comorbid conditions in cost of illness studies.

Disease prevalence rates

Uncertainty in disease prevalence rates may also lead to variation in COI estimates. Conditions are self-reported in the MEPS27. Depending upon the disease, some respondents may erroneously report whether or not they have the disease in question. For diseases such as osteoarthritis, prevalence rates based on self-reported data may be problematic, because many patients do not realize that they have the disease. Evidence also indicates that self-reports of heart failure do not align closely with medical records indicating the disease32. However, in the case of diabetes mellitus, hypertension, myocardial infarction, and stroke, patient self-reported data are quite accurate32.

The prevalence of diabetes mellitus from the MEPS data is 9.7% for adults aged 18 and above. This figure is close to evidence in the literature placing diabetes mellitus prevalence at 10.7% among adults aged 20 and above30. The baseline value for diabetes mellitus prevalence was 10%, and varied from 9–11% in calculating aggregate costs of diabetes mellitus.

The prevalence of rheumatoid arthritis is ∼0.5–1% in the US33^,34. Therefore, to address this source of uncertainty and to gauge the robustness of the results to alternative assumptions, sensitivity analyses were performed by varying the prevalence rate of rheumatoid arthritis from 0.5–1.0%, and used a baseline value of 0.75%.

Application of cost models to diabetes mellitus and rheumatoid arthritis

Alternative cost models were presented using the diseases diabetes mellitus and rheumatoid arthritis and using different combinations of comorbidity types to gauge the sensitivity of the results to these alternative specifications. In the diabetes mellitus case study, obesity serves as a Type I comorbidity. It is an increasingly important health condition that is known to strongly affect diabetes mellitus. Hyperlipidemia is a Type II comorbidity because, although highly correlated with diabetes, there is little evidence of any causal relationship between the two diseases. Type III comorbidities (e.g., those caused by diabetes) include heart disease, stroke, neuropathy, eye and kidney diseases, and depression35–37. As in the case of rheumatoid arthritis, a number of Type IV comorbidities will also be included in the Type IV cost models. Summary statistics for the variables used in these analyses, including the comorbidity types for rheumatoid arthritis and diabetes, are provided in Appendix A and B.

For rheumatoid arthritis, Type I or Type III comorbidities were not used, as there is an absence of empirical evidence establishing any strong causal linkages between rheumatoid arthritis and other diseases. In a preliminary analysis, it was hypothesized that lupus and multiple sclerosis might be risk factors for rheumatoid arthritis and that rheumatoid arthritis itself might lead to other auto-immune diseases, such as inflammatory bowel disease and psoriasis. However, the correlations between rheumatoid arthritis and these diseases were extremely low, suggesting that they are better treated as Type IV comorbidities with respect to rheumatoid arthritis. In contrast, rheumatoid arthritis has been shown to be correlated—although not causally linked—to a number of conditions, including cardiovascular disease, gastrointestinal problems, and depression. A number of Type IV comorbidities (e.g., those neither causally related to rheumatoid arthritis nor strongly correlated with it) will also be included in the cost models.

Results

Correlation analysis

Table 1 shows the correlations between diabetes mellitus and the comorbidities included in the diabetes model; corresponding results for rheumatoid arthritis are provided in Table 2. In the case of diabetes mellitus, these correlations are quite large and generally statistically significant for the Type I, Type II and Type III comorbidities, and far weaker for the Type IV comorbidities. Interestingly, although obesity (BMI ≥30 and <40) is strongly correlated with diabetes mellitus, there is no discernible relationship between diabetes and morbid obesity (e.g., BMI ≥40). For rheumatoid arthritis, Type I and Type III comorbidities were not included, for reasons given earlier. As anticipated, Type II variables show some non-trivial correlations with rheumatoid arthritis, while correlations are much lower for the Type IV comorbidities.

Table 1. Correlations between diabetes mellitus and other comorbid conditions.

	Correlation	p Value
TYPE I: cause diabetes mellitus
BMI ≥30 & BMI <40	0.10	0.00
BMI ≥40	0.00	0.91
TYPE II: complication
Hyperlipidemia	0.22	0.00
TYPE III: caused by diabetes mellitus
Hypertension	0.27	0.00
CAD	0.08	0.00
AMI	0.07	0.00
CHF	0.11	0.00
Stroke	0.03	0.00
Cardiac dysrhythmias	0.05	0.00
Angina	0.07	0.00
Diabetes caused kidney diseases	0.33	0.00
Diabetes caused eye problems	0.46	0.00
Depression	0.05	0.00
TYPE IV: other
Malignant neoplasm
Skin cancer	0.01	0.00
Colon cancer	0.02	0.00
Lung, bronchus, or trachea	0.01	0.00
Diseases of the digestive system
GERD	0.06	0.00
Gastritis	0.02	0.00
Gastric ulcer	0.02	0.00
Diverticulitis	0.02	0.00
Disease of genitourinary system
Kidney stones	0.01	0.00
Cystitis	0.01	0.00
Diseases of the musculoskeletal  System & connective tissue
Osteoarthritis	0.02	0.00
Lumbar disk disease	0.01	0.012
Diseases of the nervous system
Parkinson’s disease	0.02	0.00
Multiple sclerosis	0.00	0.42
Migraine	−0.03	0.00
Diseases of the respiratory system
Chronic bronchitis	0.02	0.00
Emphysema	0.04	0.00
Asthma	0.03	0.00
Chronic sinusitis	−0.02	0.00
Symptoms involving respiratory   system and other chest symptoms	0.05	0.00
Diseases of the skin
Eczema (dermatitis)	−0.01	0.00
Sebaceous gland diseases	−0.02	0.00
Metabolic diseases
Hypothyroidism	0.02	0.00

Table 2. Correlations between RA and other comorbid conditions.

	Correlation	p Value
TYPE I: cause RA—NONE
TYPE II: complication
Depression	0.03	0.00
Anxiety	0.03	0.00
Hyperlipidemia	0.03	0.00
Osteoarthritis	0.01	0.00
Osteoporosis	0.05	0.00
Diabetes mellitus	0.02	0.00
Hypertension	0.04	0.00
CAD	0.03	0.00
AMI	0.03	0.00
GERD	0.03	0.00
Chronic bronchitis	0.04	0.00
TYPE III: caused by RA—NONE
TYPE IV: other
Multiple sclerosis	0.00	0.26
Lupus	0.00	0.69
Psoriasis	0.01	0.00
IBD	0.00	0.48
CHF	0.01	0.00
Stroke	0.01	0.00
Cardiac dysrhythmias	0.00	0.00
Angina	0.01	0.00
Skin cancer	0.01	0.00
Colon cancer	0.01	0.02
Cancer of lung	0.01	0.01
Symptoms involving respiratory  system and other chest symptoms	0.01	0.00
Gastritis	0.01	0.00
Gastric ulcer	0.02	0.00
Diverticulitis	0.01	0.00
Kidney stones	0.00	0.88
Cystitis	0.01	0.00
Lumbar disk disease	0.01	0.03
Parkinson’s disease	0.01	0.02
Migraine	0.01	0.00
Emphysema	0.02	0.00
Asthma	0.02	0.00
Chronic sinusitis	0.01	0.00
Eczema	0.00	0.50
Sebaceous gland disease	0.00	0.82
Hypothyroidism	0.02	0.00

Multivariable results

Multivariable results for the probability of incurring any expenditures and for conditional expenditures are presented in Tables 3 and 4, respectively. In each case, alternative versions of the models are presented. These models differ in terms of what comorbidities types (e.g., Types I–IV) are included, but otherwise each model includes all of the explanatory variables listed in Table 1. For diabetes mellitus, models 1–4 successively add comorbidity types, starting with Type I comorbidities. Models 5–10 exclude Type I comorbidities. In the interest of brevity, Tables 3 and 4 present the estimated effects under the alternative models for the diabetes mellitus and rheumatoid arthritis variables only. However, full sets of regression results are presented for diabetes models 8–10 and for rheumatoid arthritis model 9 in Appendix C and D.

Table 3. Logistic regression results for alternative models.*.

	Diabetes		RA
Model type	Odds ratio	p Value	Odds ratio	p Value
Model 0: No comorbidities included	7.76	<0.01	2.47	<0.01
Model 1: Type I diseases only	7.73	<0.01
Model 2: Type I + Type II	6.46	<0.01
Model 3: Type I + Type II + Type III	4.58	<0.01
Model 4: Type I + Type II + Type III + Type IV	4.80	<0.01
Model 5: Type II	6.46	<0.01	2.01	<0.01
Model 6: Type II + Type III	4.55	<0.01
Model 7: Type II + Type III excluding eye & kidney problems due to diabetes	5.09	<0.01
Model 8: Type II + Type III + Type IV (lower bound)	4.76	<0.01
Model 9: Type II + Type IV (upper bound)	6.52	<0.01	1.95	<0.01
Model 10 Type II + Type III + Type IV excluding eye & kidney problems due to diabetes (baseline)	5.27	<0.01

*see Appendix for full set of regression results for Diabetes model and RA model.

Table 4. Ordinary least squares results for alternative models.*.

	Diabetes		RA
Model type	Coefficient	p Value	Coefficient	p Value
Model 0: No comorbidities included	0.81	<0.01	0.64	<0.01
Model 1: Type I diseases only	0.80	<0.01
Model 2: Type I + Type II	0.70	<0.01
Model 3: Type I + Type II + Type III	0.50	<0.01
Model 4: Type I + Type II + Type III + Type IV	0.50	<0.01
Model 5: Type II	0.71	<0.01	0.47	<0.01
Model 6: Type II + Type III	0.51	<0.01
Model 7: Type II + Type III excluding eye & kidney problems due to diabetes	0.58	<0.01
Model 8: Type II + Type III + Type IV (lower bound)	0.50	<0.01
Model 9: Type II + Type IV (upper bound)	0.69	<0.01	0.44	<0.01
Model 10 Type II + Type III + Type IV excluding eye & kidney problems due to diabetes (baseline)	0.57	<0.01

*see Appendix for full set of regression results for Diabetes model and RA model.

Turning first to the logistic regression results predicting the odds of incurring any expenses, the odds ratio for the effect of diabetes mellitus varies considerably across the models. The odds ratio ranges from 4.55 for model 6, which includes Type II and Type III comorbidities, to 7.76 for model 0, which includes no comorbidities other than diabetes. It is clear that how the comorbidity variables are specified has strong implications for the estimated effect of diabetes mellitus. However, whether Type I comorbidities are included or not appears to have little impact on the estimated effect of diabetes mellitus. Thus, comparing models 2 and 5, which differ only in that Type I comorbidities are included in model 2 but not in model 5, the odds ratio for diabetes mellitus is the same, 6.46. A consistent pattern emerges when models 3 and 6 and models 4 and 8 are compared. Similarly, including Type IV comorbidities has little effect on the results, as can be seen by comparing models 3 and 4 and models 6 and 8, respectively.

The real differences depend upon the Type II and Type III variables. If only Type II comorbidities are included (model 5), the odds ratio on diabetes is 6.46; this falls to 4.55—a decline of almost 30%—when Type III comorbidities are added. Models 8–10 are our preferred models, because they include Type II and Type IV comorbidities, and illustrate how the results vary as we successively include Type III comorbidities. Thus, model 8 includes Type II, Type III, and Type IV comorbidities, yielding a relatively low estimated odds ratio for diabetes mellitus of 4.76. Model 9 includes only Type II and Type IV comorbidities, leading to a higher odds ratio for diabetes mellitus of 6.52. Finally, model 10 includes Type II, Type III, and Type IV comorbidities, but excludes those Type III comorbidities (eye and kidney problems) that are known to have been caused by diabetes mellitus. Model 10 gives an intermediate odds ratio of 5.27. Based on these results, model 10 is our preferred model. This model is preferred because it is conservative by including many of the Type III comorbidities, but excludes those for which diabetes mellitus is clearly the cause. However, due to the uncertainty in whether other Type III variables should be included or not, models 8 and 9 are used to provide lower and upper bound estimates, respectively, in sensitivity analysis.

For rheumatoid arthritis, there is less variation across models, although failure to include any comorbidities leads to substantially higher odds ratios. This is particularly true with respect to the Type II comorbidities. Controlling for Type II comorbidities, the odds ratio for rheumatoid arthritis is 2.01, but this value increases to 2.47 without any Type II controls. In contrast, adding Type IV comorbidities has little effect on the results. While there is little difference between the rheumatoid arthritis models with or without the Type IV comorbidities, we prefer the model that includes these comorbidities, as it provides slightly more conservative estimates of the effects of rheumatoid arthritis and some of these additional comorbidities also affect the odds of incurring expenditures (see Appendix D).

The conditional expenditure results in Table 4 reveal a similar pattern—Type I and Type IV comorbidities have little impact on the estimated effects of diabetes mellitus, while including Type II and Type III comorbidities changes the results substantially. Again, our preferred specification is model 10, although models 8 and 9 may again be considered as providing, respectively, lower and upper bound estimates of the impact of diabetes. Similarly, the effects of rheumatoid arthritis also differ across models. Most importantly, including the Type II comorbidities substantially reduces the estimated effect of rheumatoid arthritis on conditional costs.

Aggregate cost estimates

Using the model 10 specification, and a baseline prevalence rate for diabetes mellitus of 10%, aggregate cost estimates for diabetes mellitus are generated. In addition, aggregate costs using our preferred model for rheumatoid arthritis (e.g., model 9) are estimated.

For diabetes mellitus, sensitivity analysis, using models 8 and 9, and varying diabetes prevalence from 9–11% was performed. Because the results for rheumatoid arthritis are so similar whether Type IV comorbidities are included or not, we only consider the model that does include these comorbidities. However, prevalence rates were allowed to vary from 0.5–1%, a range which is consistent with the literature33^,34.

The results are shown in Tables 5 and 6 for diabetes and rheumatoid arthritis, respectively. The baseline estimate indicates that diabetes increases aggregate annual healthcare expenditures by $147.80 billion. This ranges from a low of $131.69 billion to a high of $171.96 billion.

Table 5. Aggregate expenditures of diabetes mellitus and sensitivity analyses ($ billions per annum).

	Prevalence rates
Incremental effects	Low 9%	Medium 10%	High 11%
Lower bound model	131.69	146.32	160.96
Baseline	133.02	147.80	162.58
Upper bound model	140.69	156.33	171.96

To account for differences in the prevalence of disease based on the data source, the prevalence rates were varied. The prevalence of diabetes mellitus from the MEPS data is 9.7% for adults aged 18 and above. This figure is close to evidence in the literature placing diabetes mellitus prevalence at 10.7% among adults aged 20 and above30. Therefore, the diabetes mellitus prevalence was varied from 9–11%.

Population above 18 years old: 307,006,550*(1–24.3%) = 232,403,958.

Website: http://quickfacts.census.gov/qfd/states/00000.html

Table 6. Aggregate expenditures of RA and sensitivity analyses ($ billions per annum).

	Prevalence rates
Incremental effects	Low 0.5%	Medium 0.7%	High 2.0%
TYPE II + TYPE IV model	12.8	17.9	51.3

To account for differences in the prevalence of disease based on the data source, the prevalence rates were varied. The prevalence of rheumatoid arthritis is ∼0.5–1% in the US33^,34. Therefore, the RA prevalence was varied from 0.5–2%.

Population above 18 years old: 307,006,550*(1–24.3%) = 232,403,958.

Website: http://quickfacts.census.gov/qfd/states/00000.html

The lower bound estimate for the baseline prevalence of diabetes mellitus (10%) gives an estimated increase of $146.3 billion per annum, while the upper bound estimate yields an increase of $156.3 billion. Thus, alternative specification of comorbidities leads to a difference in annual costs of ∼$10 billion. For rheumatoid arthritis, costs range from a low of $12.8 billion to a high of $26.2 billion, with a mid-point estimate of $19.2 billion.

Discussion

Cost of illness estimates are sensitive to model specification. Proper inclusion of relevant comorbidities is particularly important. This study illustrates which comorbidity types are included in a model is more critical to the COI estimates than the number of comorbidities included. The choice of comorbidities to include in a COI model involves clinical judgement and may be limited by data availability. The present study has sought to provide some guidance on which comorbidities to include by categorizing comorbidities into four types.

Comorbidities that are clinically unrelated to the disease of interest but are highly correlated with it (Type II) warrant strong consideration for inclusion. If the goal is to provide conservative cost estimates, the case for including Type II comorbidities becomes stronger. In the case study of diabetes mellitus, the failure to include these comorbidities may result in significantly higher estimated effects of the study disease. Comparing model 1 (Type I only) and model 2 (Type I + Type II) in Table 3, the study reveals that excluding the Type II comorbidity (hyperlipidemia) would increase the estimated effect of diabetes mellitus on the probability of incurring expenditures by 20%; the estimated effect of diabetes on conditional costs would increase by 15% (see models 1 and 2 in Table 4). Thus, excluding one comorbidity that is positively correlated with diabetes mellitus, but that neither causes nor is caused by diabetes mellitus, results in substantially higher estimated effects of diabetes mellitus on costs.

The results for rheumatoid arthritis also highlight the importance of including Type II variables. Comparing model 0 (no comorbidities added) to model 5 (Type II included) for rheumatoid arthritis, the study reveals that excluding Type II comorbidities would increase the estimated effect of rheumatoid arthritis on the probability of incurring expenditures by 23%; the estimated effect of rheumatoid arthritis on conditional costs would rise by 36% (compare models 0 and 5 in Table 4).

Because hyperlipidemia bears no causal relationship with diabetes mellitus, including this comorbidity makes conceptual sense. Indeed, including all comorbidities that are significantly correlated with the study disease, but which bear no causal linkages with that disease, seems reasonable. In contrast, the conceptual argument for including Type I comorbidities seems less compelling. Dall et al.38 argue that failing to control for factors like obesity that increases the risk of both diabetes mellitus and other conditions will overstate the healthcare costs attributed to diabetes mellitus. However, including comorbidities that cause the disease in question allows one to estimate only the costs of the study disease that is not caused by any of these conditions. So controlling for obesity in a diabetes mellitus cost of illness study would yield estimates of the cost of diabetes mellitus that would exclude obesity-induced diabetes mellitus. In the extreme, if one were to consider a disease that is caused by a large number of other conditions, then including all of these comorbidities might lead one to estimate the cost effects of that disease only when its causes are idiopathic. Moreover, if there are other diseases that obesity causes but which are not causally linked to diabetes mellitus—such as hyperlipidemia—it would seem to make more sense to include these comorbidities directly than to use obesity as a proxy for them. We find that including or excluding Type I comorbidities (e.g., obesity) had little effect on the estimated impact of diabetes (e.g., compare models 4 and 8 in Tables 3 and 4, respectively), but we cannot be certain if this pattern would persist with other diseases.

It is also challenging to decide whether or not to include Type III variables—those that may be caused by the disease in question. When the causal connection is weak but the comorbidities are nonetheless highly correlated, the argument for inclusion is stronger. In this case, the comorbidity in question is much like a Type II comorbidity for which, as argued above, the case for inclusion is relatively strong. However, when the study disease is a very strong determinant of the comorbidity, including that comorbidity may rob the study disease of legitimate cost impacts. Given the uncertainties in handling Type III comorbidities, it seems prudent to perform sensitivity analysis, estimating cost impacts with and without these comorbidities.

Finally, Type IV comorbidities—those which are not causally related to the study disease and only weakly correlated with it—should have little effect on the estimated impact of the study disease, whether or not they are included into the model. Indeed, the study found this to be true for both diabetes mellitus and rheumatoid arthritis.

Although diabetes mellitus and rheumatoid arthritis were used as illustrative examples, these issues of comorbidity specification pertain to any other disease of interest as well. In sum, we recommend categorizing comorbidities into the four types as described above. Of particular importance is having as complete a list as possible of Type II and Type III comorbidities. Failure to include even one of these key comorbidities may lead to substantial bias in the estimated cost of the study disease.

These comorbidity categories also provide insight into why summing estimated costs of diseases leads to implausibly large aggregate cost estimates due to double-counting of costs, a phenomenon first noted by Trogden et al.42. To illustrate, consider the examples of the cost of diabetes mellitus and the cost of obesity. To estimate the cost of diabetes mellitus, suppose that one excludes the comorbidity of obesity since its inclusion in the calculation may rob diabetes mellitus of some of its legitimate costs. However, now consider estimating the cost of obesity. Diabetes mellitus is often caused by obesity, and thus we elect to exclude it from the obesity cost model as it may rob obesity of some of its legitimate costs. In these examples, however, the diabetes mellitus cost estimate includes the cost of obesity-induced diabetes mellitus, while the obesity cost estimate also includes this same cost because diabetes has not been controlled for. Summing costs across diseases may involve considerable double-counting of this type, leading to implausibly large aggregate estimates.

In addition to uncertainty in the estimated cost impacts of disease, uncertainty in disease prevalence is another important source of variation in aggregate cost estimates of a disease. Many disease prevalence estimates are based on self-reports. Given these uncertainties, alternative values for disease prevalence should be included as part of a sensitivity analysis in cost of illness studies.

The example of rheumatoid arthritis offers even stronger evidence of the importance of accounting for uncertainty in disease prevalence. Prevalence rates for rheumatoid arthritis in the US vary by 100%—from 0.5–0.1%. This uncertainty has a profound effect on the estimated costs associated with rheumatoid arthritis. Thus, while the estimated incremental effects of rheumatoid arthritis vary less across models than is true of diabetes mellitus, the aggregate estimates nonetheless vary by more because of greater uncertainty in prevalence rates.

Other types of economic evaluations, including cost effectiveness analysis, cost benefit analysis, and cost utility analysis, regularly employ sensitivity analyses to gauge the robustness and reliability of the findings. Therefore, it is surprising that sensitivity analysis has typically not been employed in COI studies. Since COI studies are subject to the same sources of uncertainty as these other evaluations, it seems reasonable to include sensitivity analysis as part of any COI study. In several recent COI studies, a broad range of comorbidities were included and sensitivity analyses were performed to address uncertainty in the estimated incremental cost effects and prevalence of the study disease39–41. This approach is essential for gauging the reliability and robustness of a COI study.

This study has some important limitations. Although particular attention was paid to the choice of appropriate comorbidities, the models may not have been able to control for all relevant factors that are related to healthcare expenditures. In particular, if there are additional Type II and Type III variables that were unable to be controlled for, the cost estimates may be biased. Second, this study was based on self-reported data collected as part of a national household survey. Thus, there is the potential for respondent recall bias, and the reported disease prevalence may not accurately reflect actual disease prevalence. This was addressed by performing a sensitivity analysis that varied the prevalence of diabetes mellitus. Evidence does indicate, however, that the vast majority of diagnosed adult diabetes cases—90–95%—are type 242^,43. In addition to self-reported data, this methodology may be performed in COI studies which utilize administrative claims data.

The two-part model methodology we employ is a widely accepted method of estimating expenditures in the econometrics literature because it adjusts for patients who have no expenditures30^,31. If this adjustment is not made, estimates of incremental costs associated with a disease are less reliable30^,31.

This analysis extends prior work performed on COI studies by providing a systematic evaluation of comorbidity types, providing guidance for selecting comorbidities to include in COI models, and highlighting the need for sensitivity analyses to address the inherent uncertainties in appropriate model specification and disease prevalence rates. We suggest that future COI studies consider carefully the choice of comorbidities to include in their analyses, and check the sensitivity of the results to alternative specifications. It is estimated that diabetes mellitus increases aggregate annual healthcare expenditures in the US by $147.8 billion. This is somewhat higher than a recent national estimate of $116 billion in 200738. Converting that estimate to 2008 dollars to make it more comparable to ours yields a figure of $120 billion per annum. An important difference between the two studies is that ours uses regression-based methods to estimate the incremental costs of diabetes, while the other study uses the attributable fraction method. The attributable fraction approach combines aggregate data with a population-attributable fraction to estimate attributable costs14. A study comparing these two methods also reports higher costs estimates for diabetes using the regression-based approach rather than the attributable fraction method15.

Conclusions

Cost of illness studies serve a variety of needs for policy-makers, third party payers, self-insuring employers, and other stakeholders. Proper estimation of the costs attributable to individual diseases is crucial, and requires careful attention to model specification issues, particularly the proper inclusion of comorbidities, and performing sensitivity analyses as part of any COI study to gauge the reliability and precision of these estimates. Categorizing comorbidities into the four types described here will assist the researchers in model specification and in performing sensitivity analysis. This methodology should lead to more reliable and robust COI estimates.

Transparency

Declaration of funding

This study was funded by Janssen Scientific Affairs, LLC.

Declaration of financial/other relationships

JL is an employee of Janssen Scientific Affairs, LLC. At the time of the study, AN was an employee of Janssen Scientific Affairs, LLC. CG is a consultant paid by Janssen Scientific Affairs, LLC. JR is a paid consultant to CTI Clinical Trial and Consulting. JME peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Acknowledgments

We would like to thank Amanda Teeple for her assistance in formatting and submitting the manuscript. Amanda Teeple is an employee of Kelly Services.

Appendix A

Table A1. Summary statistics and variable definitions of samples used in diabetes study (all population aged 18 and older, 2000–2007).

	People with diabetes mellitus (n = 14,116)		No diabetes mellitus (n = 130,915)
	Mean	SD	Mean	SD
Prob to have any exp	0.99	0.10	0.91	0.28
Total expenditure if any	11,346.99	19,620.84	4,779.709	12,204.84
Explanatory variable
4 types diseases related with diabetes
TYPE I: cause diabetes
BMI ≥30 & BMI <40	20%	40%	9%	29%
BMI ≥40	57%	50%	57%	50%
TYPE II: complication
hyperlipidemia	37%	48%	11%	32%
TYPE III: caused by diabetes
Hypertension	59%	49%	21%	41%
CAD	3%	17%	1%	8%
AMI	3%	16%	1%	8%
CHF	4%	20%	1%	8%
STROKE	1%	8%	0%	4%
Cardiac dysrhythmias	4%	20%	2%	13%
Angina	2%	15%	1%	7%
Diabetes caused kidney diseases	12%	32%	0%	0%
Diabetes caused eye problems	23%	42%	0%	0%
Depression	15%	36%	10%	30%
TYPE IV: other
Malignant neoplasms
Skin cancer	2%	14%	1%	12%
Colon cancer	1%	8%	0%	4%
Lung, bronchus, or trachea	0%	6%	0%	4%
Diseases of the digestive system
GERD	11%	31%	6%	23%
Gastritis	1%	11%	1%	8%
Gastric ulcer	1%	12%	1%	8%
Diverticulitis	1%	10%	1%	7%
Disease of genitourinary system
Kidney stones	1%	11%	1%	10%
Cystitis	2%	14%	2%	12%
Diseases of the musculoskeletal system & connective tissue
Osteoarthritis	1%	11%	1%	8%
Lumbar disk disease	11%	31%	10%	30%
Diseases of the nervous system
Parkinson’s disease	1%	8%	0%	5%
Multiple sclerosis	0%	5%	0%	5%
Migraine	2%	14%	3%	18%
Diseases of the respiratory system
Chronic bronchitis	1%	9%	0%	6%
Emphysema	2%	14%	1%	9%
Asthma	7%	26%	5%	22%
Chronic sinusitis	4%	20%	6%	24%
Symptoms involving respiratory system and other chest symptoms	8%	27%	4%	20%
Diseases of the skin
Eczema (dermatitis)	1%	11%	2%	13%
Sebaceous gland diseases	1%	11%	2%	14%
Metabolic diseases
Hypothyroidism	4%	20%	3%	16%
Gender
Female	56%	50%	57%	49%
Age
18–24	1%	11%	12%	33%
25–34	4%	19%	18%	38%
35–44	9%	29%	20%	40%
45–54	20%	40%	19%	40%
55–64	26%	44%	13%	34%
65–74	22%	42%	9%	28%
75 and older	17%	38%	8%	27%
Race/Ethnicity
Caucasian	51%	50%	62%	48%
Hispanic	23%	42%	19%	39%
African-American	17%	38%	12%	32%
Other race	8%	28%	7%	26%
Marital status
Married	55%	50%	52%	50%
Education
No high school degree	41%	49%	27%	44%
High school degree	29%	45%	29%	45%
Some college	12%	32%	15%	36%
College or advanced degree	18%	39%	28%	45%
Person’s income
<10 k	33%	47%	25%	43%
10–20 k	25%	43%	20%	40%
20–30 k	10%	31%	11%	31%
30–50 k	15%	36%	20%	40%
>50 k	12%	32%	20%	40%
Job variables
Unemployed	59%	49%	30%	46%
Health Insurance
Uninsured	9%	29%	14%	35%
Medicaid	22%	41%	12%	32%
Medicare	28%	45%	9%	29%
Other public health insurance	2%	12%	1%	9%
Private non-HMO plan	35%	48%	42%	49%
Private HMO plan	18%	38%	26%	44%
Location
MSA (urban)	77%	42%	80%	40%
US Census Regions
Northeast	15%	36%	16%	37%
Midwest	18%	39%	22%	41%
South	43%	49%	38%	48%
West	23%	42%	25%	43%
Year
2000	8%	27%	10%	30%
2001	11%	31%	13%	34%
2002	14%	34%	15%	36%
2003	12%	33%	13%	33%
2004	13%	33%	13%	33%
2005	13%	34%	13%	33%
2006	15%	36%	13%	33%
2007	14%	35%	11%	31%

Appendix B

Table A2. Summary statistics and variable definitions of samples used in RA study (all population aged 18 and older, 2000–2007).

	People with RA (n = 1208)		No RA (n = 144,069)
	Mean	SD	Mean	SD
Prob to have any exp	0.98	0.13	0.92	0.27
Total expenditure if any	12,351.12	21,502.98	5,417.61	13,232.47
Explanatory variable
4 types diseases related with diabetes
TYPE I: cause RA—NONE
TYPE II: complication
Depression	20%	40%	11%	31%
Anxiety	14%	35%	7%	25%
Hyperlipidemia	24%	43%	14%	35%
Osteoarthritis	2%	13%	1%	9%
Osteoporosis	13%	33%	3%	18%
Diabetes	16%	37%	10%	30%
Hypertension	44%	50%	24%	43%
CAD	4%	19%	1%	9%
AMI	4%	19%	1%	9%
GERD	15%	36%	6%	24%
Chronic bronchitis	3%	17%	0%	6%
TYPE III: caused by RA—NONE
TYPE IV: other
Multiple sclerosis	0%	6%	0%	5%
Lupus	1%	8%	1%	8%
Psoriasis	1%	11%	0%	7%
IBD	0%	6%	0%	5%
CHF	2%	15%	1%	10%
Stroke	1%	9%	0%	5%
Cardiac dysrhythmias	5%	23%	2%	14%
Angina	2%	13%	1%	8%
Skin cancer	3%	16%	2%	12%
Colon cancer	1%	8%	0%	5%
Cancer of lung	0%	7%	0%	4%
Symptoms involving respiratory system and other chest symptoms	7%	25%	5%	21%
Gastritis	1%	12%	1%	9%
Gastric ulcer	2%	15%	1%	9%
Diverticulitis	1%	11%	1%	7%
Kidney stones	1%	10%	1%	10%
Cystitis	3%	16%	2%	12%
Lumbar disk disease	12%	32%	10%	30%
Parkinson’s disease	1%	8%	0%	5%
Migraine	5%	22%	3%	18%
Emphysema	3%	18%	1%	9%
Asthma	11%	31%	5%	22%
Chronic sinusitis	8%	27%	6%	24%
Eczema	2%	14%	2%	13%
Sebaceous gland disease	2%	14%	2%	14%
Hypothyroidism	6%	24%	3%	16%
Gender
Female	73%	44%	57%	49%
Age
18–24	2%	14%	11%	32%
25–34	5%	22%	17%	37%
35–44	10%	30%	19%	40%
45–54	24%	43%	20%	40%
55–64	25%	43%	14%	35%
65–74	17%	37%	10%	30%
75 and older	17%	37%	9%	28%
Race/Ethnicity
Caucasian	67%	47%	61%	49%
Hispanic	14%	34%	20%	40%
African-American	10%	30%	12%	33%
Other race	10%	30%	7%	26%
Marital status
Married	51%	50%	52%	50%
Education
No high school degree	32%	47%	28%	45%
High school degree	34%	47%	29%	45%
Some college	14%	35%	15%	35%
College or advanced degree	19%	40%	28%	45%
Person’s income
<10 k	34%	48%	26%	44%
10–20 k	27%	44%	20%	40%
20–30 k	10%	30%	11%	31%
30–50 k	14%	35%	19%	40%
>50 k	11%	32%	20%	40%
Job variables
Unemployed	63%	48%	32%	47%
Health Insurance
Uninsured	8%	27%	14%	34%
Medicaid	20%	40%	13%	33%
Medicare	26%	44%	11%	31%
Other public health insurance	2%	13%	1%	9%
Private non-HMO plan	39%	49%	41%	49%
Private HMO plan	16%	37%	25%	43%
Location
MSA (urban)	76%	43%	80%	40%
US Census Regions
Northeast	14%	34%	16%	37%
Midwest	22%	42%	21%	41%
South	40%	49%	38%	49%
West	24%	43%	25%	43%
Year
2000	8%	27%	9%	29%
2001	10%	30%	13%	34%
2002	12%	32%	15%	36%
2003	10%	30%	13%	33%
2004	11%	31%	13%	33%
2005	12%	32%	13%	33%
2006	11%	32%	13%	34%
2007	26%	44%	11%	32%

Appendix C

Table A3. Full set of regression results for diabetes (model 8–model 10).

	LOGIT: any expenditure (n = 145,031)						OLS: LN (total expenditure) (n = 133,541)
	Model 8		Model 9		Model 10		Model 8		Model 9		Model 10
	OR	p	OR	p	OR	p	coef	p	coef	p	coef	p
Diabetes	4.76	0.00	6.52	0.00	5.27	0.00	0.50	0.00	0.69	0.00	0.57	0.00
Hyperlipidemia	4.22	0.00	5.67	0.00	4.22	0.00	0.34	0.00	0.46	0.00	0.34	0.00
Hypertension	5.09	0.00			5.09	0.00	0.34	0.00			0.34	0.00
CAD	11.48	0.00			11.53	0.00	0.25	0.00			0.25	0.00
AMI	2.84	0.01			2.85	0.01	0.91	0.00			0.92	0.00
CHF	15.72	0.00			15.81	0.00	0.66	0.00			0.69	0.00
Stroke	4.41	0.04			4.40	0.04	0.57	0.00			0.57	0.00
Cardiac dysrhythmias	6.02	0.00			6.03	0.00	0.48	0.00			0.48	0.00
Angina	2.18	0.21			2.18	0.21	0.25	0.00			0.26	0.00
Kidney diseases	1.18	0.66					0.26	0.00
Eye problems	2.14	0.03					0.23	0.00
Depression	1.36	0.00			1.36	0.00	0.48	0.00			0.48	0.00
Skin cancer	4.38	0.00	4.31	0.00	4.38	0.00	0.33	0.00	0.35	0.00	0.32	0.00
Colon cancer	14.17	0.01	14.06	0.01	14.16	0.01	0.92	0.00	1.00	0.00	0.91	0.00
Lung, bronchus, or trachea	15.47	0.01	16.32	0.01	15.52	0.01	1.16	0.00	1.20	0.00	1.16	0.00
GERD	7.11	0.00	7.69	0.00	7.11	0.00	0.55	0.00	0.61	0.00	0.55	0.00
Gastritis	4.41	0.00	4.23	0.00	4.42	0.00	0.39	0.00	0.44	0.00	0.40	0.00
Gastric ulcer	1.31	0.32	1.37	0.24	1.31	0.32	0.40	0.00	0.46	0.00	0.40	0.00
Diverticulitis	7.32	0.00	7.68	0.00	7.33	0.00	0.47	0.00	0.52	0.00	0.47	0.00
Kidney stones	3.91	0.00	3.96	0.00	3.91	0.00	0.68	0.00	0.70	0.00	0.68	0.00
Cystitis	8.51	0.00	8.43	0.00	8.52	0.00	0.31	0.00	0.32	0.00	0.32	0.00
Osteoarthritis	1.76	0.01	1.77	0.01	1.76	0.01	0.33	0.00	0.38	0.00	0.33	0.00
Lumbar disk disease	1.35	0.00	1.34	0.00	1.35	0.00	0.23	0.00	0.25	0.00	0.23	0.00
Parkinson’s disease	27.44	0.00	27.58	0.00	27.45	0.00	0.68	0.00	0.76	0.00	0.69	0.00
Multiple sclerosis	2.73	0.06	2.56	0.07	2.73	0.06	1.13	0.00	1.17	0.00	1.12	0.00
Migraine	1.57	0.00	1.59	0.00	1.57	0.00	0.31	0.00	0.35	0.00	0.31	0.00
Chronic bronchitis	1.98	0.08	2.10	0.06	1.98	0.08	0.14	0.03	0.28	0.00	0.14	0.03
Emphysema	2.27	0.02	2.25	0.02	2.27	0.02	0.44	0.00	0.53	0.00	0.44	0.00
Asthma	2.27	0.00	2.34	0.00	2.27	0.00	0.37	0.00	0.42	0.00	0.37	0.00
Chronic sinusitis	3.03	0.00	3.01	0.00	3.03	0.00	0.17	0.00	0.16	0.00	0.17	0.00
Symptoms involving respiratory system and other chest symptoms	1.71	0.00	1.81	0.00	1.71	0.00	0.41	0.00	0.46	0.00	0.41	0.00
Eczema (dermatitis)	2.60	0.00	2.54	0.00	2.60	0.00	0.03	0.25	0.02	0.41	0.03	0.28
Sebaceous gland diseases	8.90	0.00	8.69	0.00	8.91	0.00	0.36	0.00	0.36	0.00	0.36	0.00
Hypothyroidism	52.43	0.00	54.17	0.00	52.42	0.00	0.19	0.00	0.23	0.00	0.19	0.00
Female	2.58	0.00	2.57	0.00	2.58	0.00	0.30	0.00	0.31	0.00	0.30	0.00
age 18–24	1.05	0.32	0.86	0.01	1.05	0.32	−0.40	0.00	−0.52	0.00	−0.40	0.00
age 25–34	0.84	0.00	0.71	0.00	0.84	0.00	−0.28	0.00	−0.36	0.00	−0.28	0.00
age 35–44	0.80	0.00	0.71	0.00	0.80	0.00	−0.22	0.00	−0.26	0.00	−0.22	0.00
age 55–64	1.13	0.04	1.28	0.00	1.13	0.04	0.18	0.00	0.21	0.00	0.18	0.00
age 65–74	1.55	0.00	1.95	0.00	1.55	0.00	0.18	0.00	0.21	0.00	0.17	0.00
age 75+	2.08	0.00	2.84	0.00	2.08	0.00	0.32	0.00	0.40	0.00	0.31	0.00
Hispanic	0.64	0.00	0.64	0.00	0.64	0.00	−0.25	0.00	−0.28	0.00	−0.25	0.00
African American	0.66	0.00	0.71	0.00	0.66	0.00	−0.19	0.00	−0.19	0.00	−0.19	0.00
Other race	0.63	0.00	0.65	0.00	0.63	0.00	−0.24	0.00	−0.26	0.00	−0.24	0.00
Married	1.27	0.00	1.26	0.00	1.26	0.00	0.04	0.00	0.01	0.29	0.04	0.00
No high school degree	0.89	0.00	0.89	0.00	0.89	0.00	−0.05	0.00	−0.04	0.00	−0.05	0.00
Some college	1.22	0.00	1.21	0.00	1.22	0.00	0.06	0.00	0.06	0.00	0.06	0.00
College or more	1.64	0.00	1.60	0.00	1.64	0.00	0.13	0.00	0.12	0.00	0.13	0.00
Income 10–20 k	0.98	0.67	0.99	0.84	0.98	0.68	0.01	0.54	0.02	0.14	0.01	0.49
Income 20–30 k	1.11	0.03	1.13	0.01	1.11	0.03	0.05	0.01	0.05	0.01	0.05	0.01
Income 30–50 k	1.34	0.00	1.35	0.00	1.34	0.00	0.07	0.00	0.07	0.00	0.07	0.00
Income 50 k+	1.68	0.00	1.68	0.00	1.68	0.00	0.12	0.00	0.11	0.00	0.12	0.00
Unemployed	1.31	0.00	1.37	0.00	1.31	0.00	0.35	0.00	0.40	0.00	0.35	0.00
Uninsured	0.29	0.00	0.29	0.00	0.29	0.00	−0.65	0.00	−0.65	0.00	−0.65	0.00
Medicaid	1.27	0.00	1.29	0.00	1.27	0.00	0.23	0.00	0.28	0.00	0.24	0.00
Medicare	0.85	0.06	0.88	0.14	0.85	0.06	0.02	0.37	0.04	0.06	0.02	0.28
Other public plan	1.33	0.20	1.35	0.18	1.33	0.20	0.05	0.38	0.08	0.12	0.05	0.37
Private non HMO	0.92	0.03	0.91	0.02	0.92	0.03	0.06	0.00	0.06	0.00	0.06	0.00
Urban	0.92	0.05	0.90	0.02	0.92	0.05	0.03	0.13	0.02	0.18	0.02	0.13
Midwest	0.88	0.01	0.88	0.02	0.88	0.01	0.04	0.06	0.05	0.02	0.04	0.06
South	0.81	0.00	0.83	0.00	0.81	0.00	−0.04	0.04	−0.02	0.25	−0.04	0.04
West	0.77	0.00	0.76	0.00	0.77	0.00	−0.06	0.00	−0.06	0.01	−0.06	0.00
Yr2001	0.98	0.73	0.98	0.70	0.98	0.73	0.10	0.00	0.11	0.00	0.10	0.00
Yr2002	0.94	0.34	0.95	0.38	0.94	0.34	0.10	0.00	0.12	0.00	0.10	0.00
Yr2003	0.90	0.10	0.90	0.12	0.90	0.10	0.12	0.00	0.14	0.00	0.12	0.00
Yr2004	0.81	0.00	0.82	0.00	0.81	0.00	0.11	0.00	0.13	0.00	0.11	0.00
Yr2005	0.84	0.01	0.86	0.02	0.84	0.01	0.09	0.00	0.11	0.00	0.09	0.00
Yr2006	0.83	0.00	0.85	0.01	0.83	0.00	0.05	0.01	0.08	0.00	0.05	0.01
Yr2007	0.86	0.02	0.89	0.09	0.85	0.02	0.00	0.83	0.03	0.13	−0.01	0.75
Constant							6.69	0.00	6.80	0.00	6.69	0.00
R^2	0.23		0.22		0.23		0.30		0.27		0.29

Appendix D

Table A4. Full set of regression results for rheumatoid arthritis (Type II + Type IV model).

	LOGIT: Any expenditure (n = 145,227)		OLS: LN (total expenditure) (n = 133,785)
	OR	p	Coef	p
Rheumatoid arthritis	1.95	0.01	0.44	0.00
Depression	1.27	0.00	0.44	0.00
Anxiety	2.42	0.00	0.40	0.00
Hyperlipidemia	4.11	0.00	0.33	0.00
Osteoarthritis	1.63	0.03	0.30	0.00
Osteoporosis	3.72	0.00	0.28	0.00
Diabetes	5.29	0.00	0.58	0.00
Hypertension	5.07	0.00	0.34	0.00
CAD	11.83	0.00	0.25	0.00
AMI	2.86	0.01	0.91	0.00
GERD	6.80	0.00	0.52	0.00
Chronic bronchitis	2.02	0.07	0.10	0.11
Multiple sclerosis	2.77	0.05	1.11	0.00
Lupus	25.81	0.00	0.45	0.00
Psoriasis	2.65	0.00	0.28	0.00
IBD	5.65	0.00	0.83	0.00
CHF	15.71	0.00	0.68	0.00
Stroke	4.19	0.04	0.56	0.00
Cardiac dysrhythmias	5.99	0.00	0.46	0.00
Angina	2.20	0.20	0.26	0.00
Skin cancer	4.31	0.00	0.31	0.00
Colon cancer	14.45	0.01	0.90	0.00
Cancer of lung	15.93	0.01	1.16	0.00
Symptoms involving respiratory system and other chest symptoms	1.69	0.00	0.40	0.00
Gastritis	4.36	0.00	0.38	0.00
Gastric ulcer	1.28	0.36	0.39	0.00
Diverticulitis	7.25	0.00	0.46	0.00
Kidney stones	3.95	0.00	0.69	0.00
Cystitis	8.60	0.00	0.31	0.00
Lumbar disk disease	1.34	0.00	0.23	0.00
Parkinson’s disease	26.51	0.00	0.66	0.00
Migraine	1.54	0.00	0.30	0.00
Emphysema	2.19	0.03	0.40	0.00
Asthma	2.25	0.00	0.36	0.00
Chronic sinusitis	3.03	0.00	0.16	0.00
Eczema	2.60	0.00	0.03	0.33
Sebaceous gland disease	8.90	0.00	0.36	0.00
Hypothyroidism	51.93	0.00	0.18	0.00
Female	2.54	0.00	0.29	0.00
Age 18–24	1.08	0.13	−0.38	0.00
Age 25–34	0.85	0.00	−0.27	0.00
Age 35–44	0.80	0.00	−0.21	0.00
Age 55–64	1.11	0.06	0.17	0.00
Age 65–74	1.49	0.00	0.17	0.00
Age 75+	2.04	0.00	0.31	0.00
Hispanic	0.66	0.00	−0.23	0.00
Black	0.68	0.00	−0.16	0.00
Other race	0.64	0.00	−0.22	0.00
Married	1.28	0.00	0.05	0.00
No high school degree	0.90	0.00	−0.04	0.00
Some college	1.22	0.00	0.06	0.00
College or more	1.64	0.00	0.12	0.00
Income 10–20 k	0.98	0.67	0.01	0.69
Income 20–30 k	1.12	0.02	0.04	0.01
Income 30–50 k	1.34	0.00	0.07	0.00
Income 50 k+	1.69	0.00	0.12	0.00
Unemployed	1.28	0.00	0.33	0.00
Uninsured	0.29	0.00	−0.64	0.00
Medicaid	1.26	0.00	0.22	0.00
Medicare	0.83	0.04	0.01	0.45
Other public plan	1.32	0.21	0.04	0.45
Private non HMO	0.92	0.03	0.06	0.00
Urban	0.91	0.03	0.02	0.23
Midwest	0.88	0.01	0.04	0.05
South	0.81	0.00	−0.04	0.04
West	0.77	0.00	−0.06	0.00
Yr2001	0.97	0.65	0.10	0.00
Yr2002	0.93	0.27	0.09	0.00
Yr2003	0.89	0.07	0.11	0.00
Yr2004	0.79	0.00	0.10	0.00
Yr2005	0.82	0.00	0.07	0.00
Yr2006	0.81	0.00	0.03	0.09
Yr2007	0.84	0.01	−0.03	0.23
_cons			6.68	0.00
R^2	0.24		0.30