When a mental health condition, such as depression, is combined with a chronic medical diagnosis, the costs of medical care can be substantially inflated (
1). It has been suggested that the added costs of treating a chronic illness, such as diabetes mellitus, in combination with a lower mental health score, are significantly higher than the costs of treating each condition independently. The cost-multiplier estimates range from about 1.5 to 4.5 (
2–
6), suggesting that the costs of treating the combination of conditions are 150% to 450% higher than the expected costs of treating each condition separately. However, other analyses have indicated that the costs of treating the two conditions follow an additive model—that is, the overall costs equal the sum of the costs of treating each condition. For example, using aggregation of Medicaid claims from across states, Kronick and colleagues (
7) found that the cost of treating the combination of depression and diabetes was approximately equal to the sum of the costs of treating each condition separately.
The literature on the effects of comorbid mental illness on cost is difficult to interpret because studies apply different definitions of mental health conditions and use different outcome measures. Although a few studies have had large samples, most have been limited to observations of only a few hundred individuals. Furthermore, most studies have focused on convenience samples, and few have been generalizable to the larger population that receives health care. Studies using claims data have confounded diagnosis with care access and care seeking. Population-based surveys avoid these biases.
Using data from the Medical Expenditure Panel Survey (MEPS), a large-scale survey of individuals representative of the U.S. population, medical providers, and employers, we examined the effects on total health care expenditures of treating eight chronic general medical conditions among persons who had or did not have comorbid mental health problems. Mental health status was estimated from the mental component summary (MCS) score of the 12-item Short Form Health Survey (SF-12) (
9). Most previous studies have assessed mental health status on the basis of a formal diagnosis in the medical record. However, diagnosis is not the most reliable estimate of mental well-being. Some patients avoid discussing mental health conditions with their providers, and some providers are reluctant to enter a diagnosis of a mental health condition in the patient’s record. Use of the mental health measure provided by the MCS score has the advantage of assessing mental health status independent of provider judgment. Another advantage of the MEPS data is that the MCS score is a continuous variable. Many previous studies have considered only patients with serious mental illness.
Results
The numbers of persons with each of the eight general medical conditions are shown in
Table 1, which also summarizes the analysis-of-variance results for each diagnosis. The table estimates the effect of depression on cost, the effect of diagnosis of a chronic disease on cost, and the interaction between diagnosis and depression. The influence of statistically controlling for PCS is also shown. The results, including estimated costs, are also summarized in
Figure 1.
As the data in
Figure 1 suggest, the presence of a chronic disease diagnosis had a very strong effect on health care expenditure. The main effects for mental health status and the presence of each of the eight general medical conditions were highly significant. In addition, the effect of the PCS score on cost of medical care was highly significant for each condition.
Figure 1 shows that respondents with any of the eight diagnoses consumed significantly more resources, compared with those without these diagnoses. Health care services for respondents with a low level of mental health functioning had significantly higher costs, compared with those with high mental health functioning, even after the analysis controlled for physical health status using the PCS score.
The central focus of this analysis was the interaction between mental health functioning and chronic disease diagnosis. In
Figure 1, an interaction is indicated by deviation from parallel lines in the two-dimensional graphs. The interaction terms for three of the conditions (diabetes, COPD, and cancer) were statistically significant. However, in each of these cases, the effect sizes tended to be small. Eta
2 (η
2) is an approximate index of the proportion of total variance in cost attributable to each effect in the model. The interaction effect of η
2 was 0.1% in cancer, 0.06% in diabetes, and 0.004% in COPD. In contrast, averaged across the eight general medical conditions, the η
2 for the PCS score was 7%. Visual inspection of
Figure 1 suggests that the effect of a mental health condition on cost was similar in magnitude for those who did or did not have each of the eight diagnoses.
Psychiatric diagnoses are more likely to occur in groups with lower SES, and some chronic conditions, such as diabetes, are also more likely to occur in these groups. To address this issue, we applied additional statistical adjustment for SES. Education had a strong effect on health outcome for all eight general medical conditions. Across the eight conditions, adjustment for education had slight effects on the mental health × condition interaction. However, adjustment for education changed the statistical significance only in relation to asthma (F value changed from 2.43 to 3.54; p value changed from 0.08 to 0.03). For cancer, adjustment for education reduced the p value for the interaction (from <.001 to <.03). Although education and SES have been shown to have strong effects on cost and health outcome, these effects appeared independent of the interaction between mental health status and chronic disease diagnosis.
To explore the robustness of the results, we performed a series of sensitivity analyses. One concern was that medical care costs were not normally distributed. A few individuals have very high costs, whereas most have low costs. To address this issue, we transformed the cost variable using a log10 transformation. Considering only the interaction terms, the log10 transformation resulted in a nonsignificant interaction for diabetes (p=0.25), whereas the interaction was significant without the transformation (p<0.001). Conversely, the interaction terms for arthritis and for high cholesterol became statistically significant after the transformation while they had not been when raw units were used. The interaction for cancer remained statistically significant but the p level declined from 0.001 to 0.02. Overall, the effect size for the interactions remained low (all η2<0.1%) in relation to the effect size for PCS score and for diagnosis. (A table in an online supplement to this article presents the results of this analysis.)
We also completed a reanalysis of the interaction effects using both natural and log10 units for cost but without adjustment for the PCS score (see online supplement). When the data were analyzed in natural units, there was a significant interaction only for diabetes, arthritis, and cancer. When log10 units were used, the only significant interaction effect was observed for arthritis (p=0.009). Once again, the effect sizes for all the interactions were very small (<0.1%) in relation to the effects of PCS score and the effects of a diagnosis of a general medical condition.
The third component of the sensitivity analysis considered the concern that the mental health status variable was divided into three categories. We addressed this issue by using the MCS score as a continuous variable. Using multiple regression, we estimated the effect of MCS score, diagnosis, and a term representing the product of diagnosis × the continuous MCS on total health care costs. The analysis indicated significant interaction effects for diabetes, arthritis, COPD, and cancer (see online supplement). There were significant effects of mental health status for all diagnoses except coronary heart disease and stroke. Although the interaction effects were significant for four of the eight diagnoses, all effect sizes accounted for less than 0.1% of the variance.
Next, we calculated the 10% trimmed mean for expenditures. Although 90% of the respondents had medical expenses less than $10,000 per year, a few respondents had extremely high expenditures, ranging to over $8 million. To address this issue, we assigned a value of $10,000 to any expenditure greater than $10,000 (see online supplement). As in the other analyses, four of the eight interactions terms were statistically significant, but none accounted for more than 0.1% of the variance. Overall, the truncation did not affect the conclusions.
Finally, we considered whether dividing the MEPS respondents into tertiles was too broad a stroke. Estimates suggest that about 6% of the U.S. population has a major depressive disorder. By including the lowest-scoring third of the MEPS sample in the highest depression level, we may have mixed those with a major depressive disorder and those with subclinical depression and other mental health problems. Several studies have suggested that a multiplicative effect occurs only with persons in the category of major depressive disorder. To address this problem, we identified 2,130 MEPS participants with MCS scores less than 37.95, and they formed a group representing the lowest 6% of MCS scores. For comparison, we selected the 12% of the MEPS sample who had MCS scores in the middle of the distribution, between 51.73 and 56.95.
Table 2 shows the significance levels for the interaction term (condition × mental health status) derived from analyses of the eight chronic disease categories among respondents with the lowest 6% versus lowest 12% of MCS scores. The three significant interactions from the main analysis (diabetes, COPD, and cancer) were replicated in this subgroup analysis that focused on individuals at the extreme of the distribution of MCS scores. In addition, there was a significant interaction for arthritis.
Discussion
We did not find strong evidence that poor mental health functioning acts synergistically with other chronic disease diagnoses to inflate health care expenditures.
Methods for statistical testing of hypotheses were originally developed for studies with relatively small samples. In textbooks, most of the tables that give p levels for F ratios are for samples of up to only 1,000. Because the MEPS sample is more than 20,000, it was unusual to find a comparison that was not statistically significant. Although the interaction effect was statistically significant for three or four (depending on the analysis) of the chronic disease categories, the effect size was consistently very small. For example, the significant interaction between mental health status and diabetes accounted for 0.06% of the variance in cost. In contrast, treatment of the diagnosis of diabetes accounted for 0.4% of the cost variance (about seven times more than the interaction between mental health status and diabetes), and being in the lowest third of the distribution on MCS accounted for 0.3%. Physical health functioning (as measured by the PCS score) accounted for a full 6% of the variance (about 100 times more than the interaction between comorbid mental problems and diabetes). The observation of limited synergistic effects on treatment cost of comorbid mental problems and chronic disease is in contrast to various studies that have suggested that mental illness in combination with chronic general medical illness has a strong synergistic effect on health care costs (
8). In our study, the main findings were statistically adjusted for physical health status (PCS score) and did not change substantially by further adjustment for educational attainment.
Our results should in no way be interpreted as justification to restrict funding for mental health services. Mental health care improves quality of life and may extend life expectancy (
19). Justification for treating mental health problems is no different than that for any other health problem, and evidence suggests that individuals with mental health challenges require more rather than less attention. Among persons with chronic general medical conditions, studies suggest that those with mental health conditions receive less care than those without mental health conditions. For example, Druss and colleagues (
20) reported that patients with any comorbid mental health condition were less likely to receive costly but potentially effective cardiovascular interventions. Underdetection of depression, in particular, perpetuates a substantial shortfall in health care’s potential yield (
21).
Our study differed from previous contributions in several ways. First, we used a continuous score for mental health status rather than a psychiatric diagnosis. Epidemiologic studies suggest that major depression occurs in about 6% of the population (
22). Many previous studies have focused on patients with a diagnosis of serious mental illness (
23). However, focusing on the top 6% of the MEPS respondents who might be more likely to have major depressive disorder or another serious diagnosis did not alter the results, which suggested that the effects of comorbid mental challenges on costs are independent of the effects of other chronic conditions (
Table 2). To be clear, the interactions shown in
Table 2 were statistically significant for four of the eight general medical conditions. However, the effect sizes for all interactions, with the possible exception of cancer, were very small. Although the MEPS survey includes a nationally representative sample, it excludes individuals living in institutions. As a result, we likely excluded an important segment of people with serious mental health conditions. This difference in samples may explain why our results deviated from some those of previous studies.
One concern might be that because the SF-12, from which the MCS score was derived, is a short instrument, it may not be sensitive to important variations in mental health. We recognize the shortcomings of the MCS score as an important limitation of the study. On the other hand, as in many other studies, persons whose MCS score indicated poorer mental health functioning used significantly more health care, compared with those whose MCS score indicated better functioning, even after the analysis controlled for physical health status (
24). If the MCS score were capturing only error variance, we would not have expected to see this systematic variation.
Another concern is that most previous studies focused on depression rather than on mental wellness generically. The MCS includes a variety of questions, some of which address mental health issues other than depression. Thus the MCS score may not be a reasonable proxy for an estimate of depression. However, other studies have shown that MCS scores are highly correlated with depression measures, such as the CES-D screening tool (
14). More work is necessary to determine whether the MCS score captures meaningful variation in depression.
In addition to possible error in the assessment of mental health status, we recognize that assessments of the eight general medical conditions were based on self-report. Thus we cannot say with certainty that the classification into each of these chronic disease categories was accurate. Furthermore, total costs were estimated for each chronic condition separately. It was difficult to adjust for the effect of other comorbid chronic conditions, but it is known that some chronic conditions tend to occur together (e.g., coronary heart disease, diabetes, and high cholesterol). As a result, the estimated total costs may have been confounded by a third chronic condition.