The Role of Health and Mental Health Status - NCBI

Determining Provider Choice for the Treatment of Mental Disorder: The Role of Health and Mental Health Status Richard G. Frank and Mark S. Kamlet This article specifies and estimates a model of provider choice for mental health services. Three types ofproviders are identified: specialty mental health providers, general medical providers, and informal providers. Specific attention is paid to the role ofhealth and mental health status in determining provider choice. The model is estimated using a multinomial logit approach applied to a sample of 2,800 respondents to the Baltimore Epidemiological Catchment Area Survey. The results are largely consistent with the previous work of Wells et al. (1982), suggesting that health and mental health status play an important role in the decision to seek care but have little effect on the type ofprovider chosen. The results also reveal that 22 percent of individuals obtaining mental health care did so through the informal care sector. One exemplary benefit design simulation is performed using the estimation results.

Individuals seeking care for mental disorders receive treatment in a variety of settings and from a variety of providers. The settings include the specialty mental health sector, the general health sector, and informal settings. Treatment in the specialty mental health sector is generThis research was supported by grant MH42338 from the National Institute of Mental Health. Address correspondence and requests for reprints to Richard G. Frank, Ph.D., Associate Professor, Health Service R and D Center, Johns Hopkins University, 624 North Broadway, Baltimore, MD 21205. Mark S. Kamlet, Ph.D. is Professor of Economics, School of Public Affairs, Carnegie Mellon University, Pittsburgh.

84

HSR: Health Services Research 24:1 (April 1989)

ally provided by psychiatrists, psychologists, psychiatric social workers, and mental health counselors. Treatment in the general health sector tends to be provided by nonpsychiatric physicians, nurses, and allied health professionals. Treatment through informal care is provided by individuals such as clergy, family, friends, and crisis centers. Little is known about the factors that influence the choice of provider for mental health care. We are aware of no studies that have added utilization of the informal mental health care sector by the mentally ill to an examination of their choice between specialty and general medical care. Indeed, we are aware of only one work, Wells et al. (1982), that examines influences on the choice between the general and specialty mental health sector for mental disorders. A better understanding of what leads individuals to seek treatment for mental disorders and what determines their choice of provider is important for several reasons. First, utilization patterns differ dramatically between patients treated in the specialty mental health and those treated in the general medical care sectors. In the RAND Health Insurance study, Wells and his colleagues (1982) report that mental health care from specialty providers led to median expenditures of $280 per year. In contrast, mental health care from general medical providers resulted in median mental health expenditures of $14 per year. Data from the National Institute of Mental Health (NIMH)-sponsored Epidemiological Catchment Area (ECA) program in Baltimore reflect a similar pattern. Average use of ambulatory mental health services obtained from a specialty provider was 10.46 visits per year. Average use of ambulatory mental health services delivered by a general medical provider was 2.60 visits per year. These observations lead one to question the extent to which utilization differences can be explained by differences in the health and mental health status of patients being treated in each sector. Clearly, an important policy issue would be clarified by determining whether resource use differences are caused by the levels of illness of patients, the variety of provider practice styles, or other factors. A second reason to examine factors influencing choice of provider in mental health care is related to evaluating the desirability of prospective payment mechanisms for ambulatory mental health care treatment. The desirability of such mechanisms depends, in part, on the nature of patients treated in the different settings. Suppose patient selfselection leads to treatment of more severely ill patients in the specialty mental health sector than in the general health sector. In that situation, providing prospective payments that do not vary as a function of setting will systematically discriminate against providers in the specialty

Provider Choice for Mental Disorder

85

mental health sector and may lead to lower-quality care in that sector. As we discuss in more detail further on, other, comparable issues in mental health benefit design exist, for which information on factors that influence the type of provider plays an important role. Over the past decade, much research has examined the use of health and mental health services by individuals suffering from emotional disorders and mental illness. Such research has not typically been able to analyze factors influencing an individual's choice of treatment setting. A significant portion of this work has examined in aggregate terms the extent to which services are provided in the general health versus specialty mental health sector (Shapiro et al. 1984). Results show that over half of those seeking mental health care receive treatment only in the general health care sector and that women are greater users of ambulatory mental health care than men, although men who seek such treatment are more likely to be seen by a mental health specialist. This research also indicates that the aged have particularly low rates of utilization. This article, using data gathered from the NIMH-sponsored ECA program in Baltimore, analyzes factors that lead an individual to seek mental health treatment and to seek a particular type of treatment source. As discussed in more detail later, the NIMH ECA study interviewed noninstitutionalized individuals age 18 and over during 1981 and 1982, concerning utilization of health and mental health services. The Baltimore ECA survey followed a multistage sampling design and was implemented in the eastern part of Baltimore city. The sample includes a higher proportion of minorities and low-income individuals than the U.S. population in general. This is important because these groups typically display a greater risk of mental problems than the general population. It should be noted that we limit our attention to individuals age 18 and over living in the community. This article, therefore, does not allow for analysis of use of mental health care by chronically mentally ill individuals who are institutionalized. Nor does it allow for an analysis of mental health care for those under 18 years of age. A brief overview of the data used in this article appears in the next section. The third section presents a model as well as estimation procedures for determining factors that influence the choice of provider. Subsequently, we discuss the measurement of variables in the analysis, leading to a presentation of the results of the empirical analysis. The final section presents estimation results from one exemplary simulation and provides some brief concluding remarks.

86


DATA USED IN THIS ANALYSIS The Baltimore ECA survey is one of five in the Epidemiological Catchment Area program (Eaton et al. 1981). The study, which was principally aimed at establishing prevalence and incidence estimates for mental disorders in the adult population of east Baltimore, collected detailed information on health and mental health care utilization. The survey was based on a longitudinal multistage probability design that induded two face-to-face interviews one year apart and a telephone interview between personal contacts (for a total of three waves of data collection). The geographic area covered by the Baltimore ECA is the eastern third of Baltimore city, an area with a population of 241,000, which is 38 percent black and where 19 percent of the population is Medicaid eligible. Formal service providers indude 3 short-term general hospitals, 11 small health centers and prepaid plans, and 1 freestanding community mental health center. There were approximately 0.36 primary care physicians per 1,000 population in 1981. The field survey was designed to obtain household interviews with one randomly selected person from those 18 to 64 years of age, and with every household member 65 or older; proxy respondents were accepted for the 2.7 percent of subjects who were ill or had language problems. The overall response rate in the baseline survey was 78 percent, which resulted in 3,481 completed interviews. The follow-up six-month telephone survey and the one-year household survey had response rates of 83 percent and 81 percent, respectively. The longitudinal sample was compared to the baseline sample for evidence of systematic selection by respondents (Shapiro et al. 1984). Only minor differences were found between the two samples (which should not influence analyses of utilization). The questionnaire used in the household survey contained a battery of mental health and health status measures. At the core of the ECA data base is the Diagnostic Interview Schedule (DIS), a structured questionnaire administered by lay interviewers that indudes data for computer-generated diagnoses according to the criteria specified in the Diagnostic and Statistical Manual, Third Edition (DSM-III) of the American Psychiatric Association. Other mental health status measures indude: a score on the Mini-Mental State Examination (MiniMental), a question asking if a disability day occurred due to an emotional problem, and a 20-item version of the General Health Questionnaire (GHQ) (Goldberg 1972). The GHQ identifies current symptoms of distress and demoralization, and is designed to identify individuals at high risk of emotional distress. While the scale provides a

Provider Choicefor Mental Disorder

87

numerical index of symptomatology (the properties of which are not clear), a score of 4 or more has been shown by Goldberg to designate individuals at high risk of having a diagnosable mental disorder (which implies nonlinearity in the numerical scale). The health status measures indude a question that asks how an individual would rate his or her health status relative to people of the same age (excellent, good, fair, or poor); a count of the number of chronic medical conditions from which an individual suffers; and a detailed review of a person's specific health conditions (e.g., cancer and heart disease). All three waves of the ECA collected detailed health care utilization information. Thus, a complete picture of utilization based on three consecutive six-month recall periods is possible. The complete health and mental health status battery was administered only during the two face-to-face interviews. Nevertheless, the panel design allows us to study the effect of health and mental health status on utilization while avoiding difficulties associated with "post-diction" often found in crosssectional studies (Manning et al. 1981).1 Therefore, we make use of baseline health and mental health status measures and utilization that occurred subsequent to baseline. The ECA surveys also collected information on a variety of social, demographic, and economic characteristics of the individuals questioned. This information included: income, employment status, living arrangements, type of health insurance, education, and other demographics. The sample of individuals used here indudes those who answered the baseline questionnaire and continued in the study. After eliminating cases with values missing for key variables to be used in the analysis, we obtained a usable sample for analysis of 2,802 individuals. The loss of cases did not alter the age, sex, race, and service utilization distributions of the sample. All analyses reported in this article are weighted to take into account differential sampling probabilities, response rates, and the 1980 decennial census of population counts by age, race, and sex for Baltimore city. Variable names, short definitions, and descriptive statistics appear in Table 1.

SPECIFICATION AND ESTIMATION ISSUES The choice of whether or not to seek treatment for a mental disorder and the choice of provider if treatment is sought are discrete decisions. A powerful way to model such choices is through the use of multinomial logit (MNL). Consistent with MNL, we presume that an individ-

88


Table 1: Variable Descriptions Variabk Name

Description

Mean (S.D.)

Age in years

Age

44.7 (18.8) 1 if female; 0 otherwise Sex 0.61 (0.49) 1 if nonwhite; 0 otherwise Race 0.37 (0.48) 1 if married; 0 otherwise MARITAL ST 0.43 (0.49) Years of schooling EDUCATION 10.81 (3.07) 1 if employed; 0 otherwise EMPLOYED 0.46 (0.50) 1 if unemployed but in labor force; 0 otherwise 0.03 LAIDOFF (0.17) 15,350 Household income in dollars INCOME (12,000) 1 if enrolled in Medicaid; 0 otherwise 0.16 MEDICAID (0.36) 1 if at least two signs and symptoms (DIS, GHQ, 0.10 MHNEED (0.09) or self-reported disability); 0 otherwise 0.86 Number of chronic health conditions Number chronic (1.01) 0.73 1 if self-reported health is excellent or good; Self-health 0 otherwise (0.44) 0.55 Ambulatory insurance 1 if person has private ambulatory insurance (0.50) coverage; 0 otherwise 0.69 1 if person has private hospital insurance Hospital insurance (0.46) coverage; 0 otherwise 1 if person is covered by Medicare; 0 otherwise 0.21 Medicare (0.41) 0.06 1 if there is a private office-based health care Private provider provider in home census tract; 0 otherwise (0.23) 0.05 1 if there is an alcohol, drug, or mental health ALC/DRUG (0.22) clinic in home census tract; 0 otherwise 1 if there is a CMHC in home or adjacent 0.28 CMHC (0.43) census tract; 0 otherwise

choose among four options: to seek no treatment; to seek treatment from an informal treatment provider; to seek treatment from the general medical sector; or to seek treatment from the specialty mental health sector. In choosing whether to obtain mental health treatment and from whom, individuals consider a variety of characteristics of the care provider options, such as proximity and specialization. The decision is

ual

can

Provider Choicefor Mental Disorder

89

also affected by the individual's health and mental health status and the specific characteristics of any illness being experienced. Other particularities of the individual may also contribute to the manner in which various alternatives are viewed (education, insurance coverage, family support, or random variation in tastes). We thus assume that utility for each treatment provider depends on a set of observed characteristics of: (1) the provider, (2) individual health and mental status, and (3) the individual. We differentiate individual characteristics from characteristics of health and mental health status in order to allow individuals to make different utility-maximizing selections as the particularities of a health problem vary. The MNL framework is consistent with the standard model of consumer choice, in which an individual chooses among options based on the utility of each alternative. More specifically, we can posit that the utility of choice optionj to individual i, Uij, is:

U,

=

U(MjA, Zi, Hi,

el)

(1)

where M is a vector of provider characteristics. Z is a vector of other individual characteristics. H is a vector of individual health and mental health status descriptors. e is a vector of unobserved characteristics. j denotes provider choice alternative (0 = no care, 1 = informal, 2 = general medical, 3 = specialty mental health). Following McFadden (1981, 1982), we separate the utility function into its observed and unobserved components:

Uij = V# (M., Zi, Hi) + Eij (2) where V(M, Z, H) represents utility determined by observed data, while e is the unobserved component, which will be treated as a random variable. Utility-maximizing behavior implies that an individual i will only choose a particular alternative j if U., > Uik for all k not equal toj. Because e is assumed to be a random variable, the situation U# > Uik is also random. The probability of any given alternative j being chosen by an individual can be expressed as:

90


P.,= PU4> UiU) for all k * j =

P[(ei4 eik) > Vi (M,Z,H) for all k * j -

-

V

(M, Z, H)]

(3) Computation of Equation 3 requires making assumptions about the distribution of the e's. Letting X4 = (Mj, Zi, Hi) and assuming V() to be a linear function of components of X, we operationalize Equation 2 as: U4j = gi X4 + sj4(4) where ,j is a vector of coefficient values indicating the effect of the various X,;s on individual i's utility for option j. Note that i3j is subscripted by the choice indexj. This means that in our analysis a given (e.g., age) is allowed to "interact" with each choice option. For X4 example, age may have one effect on the utility of treatment in the general health sector and another effect on the utility of treatment in the specialty mental health sector. MNL assumes that e4 is distributed according to the Type 1 extreme-value distribution. This distribution has a cumulative distribution function J(cE- < e) = exp[-exp(-E)] and a probability density function F (e,-) = exp[-c-4 - exp (-Es)]. It is assumed that the family C-4 is independent, identically distributed, and independent of the X4 variables. Under the above assumptions: 4 (5) Prob(Option ijlx) = exp(fljX4)1EeXP(fkXk k = 1 The parameters of this model can be estimated straightforwardly using maximum-likelihood methods, with asymptotic standard errors given by the square root of elements of the inverse of the information matrix.2 The sampling design and the nonresponse weighting make the Baltimore ECA data a so-called complex sample. This means it is inappropriate to use estimation techniques that assume a simple random sample. In particular, autocorrelation arising from the situation where individuals are drawn from the same geographic duster may lead to biased variance estimates. We obtain measures of the design effect on the t-statistics by using a weighted fixed-effects model for the simple logit model. That is, each geographic cluster used for sampling is accounted for by a fixed effect. The results reported in the second column of Table 2 are design effects: the ratio of the 1-statistic from the weighted fixed-effects model to that from the simple logit. Thus, the reader can multiply the design effect by the reported t-statistic to obtain some guidance regarding the effect of the complex sample design for


91

statistical inference. We refer to the product of the design effects and reported t-statistics as the adjusted t-test.3 The discussion of results is in terms of adjusted t-tests.

DEFINITION AND MEASUREMENT OF KEY VARIABLES MEASURES OF MENTAL HEALTH SERVICE USE

Use of mental health services is a complex phenomenon that does not lend itself to exact definitions. Like previous studies (Wells et al. 1982), we operationalize our setting choice variables by creating classification rules. We begin by defining the providers included under each class of settings. Informal providers include clergy, family or social service agency, self-help group, crisis center, natural therapist, and other providers. General medical providers consist of non-psychiatrist physicians, outpatient departments of general hospitals, and hospital emergency rooms. Specialty mental health providers are defined to include mental health clinics, community mental health centers, psychiatric outpatient departments, alcohol clinics, and office-based mental health providers (including psychiatrists, psychologists, and social workers). Cases where only a single source of mental health care was reported were both the most common and the simplest with which to deal. Single-source users were, of course, assigned to the category in which their provider fit. The difficult situations occurred when a patient consulted providers of mental health services from multiple categories. In cases where both a mental health provider and a general medical provider saw a patient, we classified that patient as a specialtysector mental health patient (this occurred with eight patients). In cases where a patient saw an informal provider and a specialty mental health provider, we classified that patient as a mental health patient (this happened with two patients). Finally, when an informal and a general medical provider were both reported as sources of mental health care we assigned the patient to the sector where the most visits occurred (16 cases). This was largely due to the lack of a clear referral hierarchy among these sectors. The resulting distribution of cases was as follows: nonusers (2,533, or 90 percent), informal users (60, or 2.1 percent), general medical users (141, or 5 percent), and specialty users (68, or 2.4 percent). This distribution of mental health service utilization indicates

92


that, consistent with much of the literature on the treatment of the mentally ill, more individuals receive mental health treatment from the general medical sector than from the specialty mental health sector. Moreover, a substantial number of individuals receive mental health care through the informal sector. Roughly 22 percent of the individuals receiving mental health care did so through the informal mental health care sector, almost as many as received care through the specialty mental health care sector. MEASURING MENTAL HEALTH AND HEALTH STATUS

Measuring mental health status is a difficult problem in which any solution is subject to important caveats (Ware et al. 1983). The approach taken here draws on the previous work of Shapiro et al. (1985) and Frank (1987). The ECA data provide a rich source of information on mental health status; the DIS (including the MiniMental), GHQ, and self-report questions provide three separate and somewhat different assessments of signs, symptoms, and distress associated with mental health problems. Each of the indicators of mental health status is limited in various ways. The DIS and standardized diagnoses made by clinicians have very uneven rates of agreement across psychiatric diagnoses (Anthony et al. 1985). The GHQ was originally designed as an aid to primary care clinicians in evaluating mental health status. Its stability has been questioned in previous research (Shapiro et al. 1985). Self-reported disability due to a mental disorder is generally viewed as a strict criterion requiring rather severe distress. The approach taken in this research to measuring mental distress is based on the Shapiro et al. (1985) measure of need. Our measure makes use of all three measures of mental health status: a DSM-III diagnosis, as measured by the DIS and the MiniMental (score of less than 18); a GHQ score indicating at least four symptoms of psychiatric distress; and a self-reported disability day due to a mental or emotional problem. Because we are reluctant to rely on any single indicator, we set a criterion identifying a subject in need as an individual suffering from mental or emotional problems by at least two of the three indicators. In order to make the indicators consistent with one another we used the two-week version of the DIS and the self-report questions. Our measure of mental distress is thereby an indicator of the dustering of signs and symptoms of a mental disorder. In this study we use a simple dichotomous variable equal to one when at least two indications are present, and zero otherwise. In a previous analysis Frank (1987) used this indicator both as a dichotomous variable and as an ordinal


03

variable that could take on a value from zero to three. In that analysis he found that little added information resulted from use of the ordinal variable. Thus, we hypothesize that the dichotomous measure of mental distress will be related to a relatively high probability of use of services and especially those services of mental health providers. Two indicators of general health status were also included in the multinomial logit models. These were a self-reported health status measure based on a question that asked how a subject rated his or her health relative to that of others the same age (excellent, good, fair, or poor). While the limitations of this measure have been pointed out by Manning et al. (1981), Taubman and Rosen (1981), among others, have had reasonable success with it. We used a dichotomous indicator of good health, which takes a value of one if the response to the question was excellent or good, and zero otherwise. A second indicator of health status was based on a count of chronic conditions for each subject. The value of the variable ranged from zero to nine. Finally, in order to study the confluence of mental and physical problems, we created an interaction between our indicator of mental distress and the number of chronic conditions. We expected that poor health increases the likelihood of mental health care use but is also more likely to make the general medical provider the provider of choice (Eastwood 1975; Lipowski 1975). OTHER MEASURED INFLUENCES ON PROVIDER CHOICE

A number of sociodemographic factors given on Table 1 were included in the model: age (as a quadratic), sex, race, marital status, education, and household income. Three other sets of variables -insurance descriptors, employment status, and measures of the proximity of health care providers -are of particular interest to this research. The insurance variables consist of four dichotomous indicators: (1) whether or not a subject had Medicaid coverage; (2) whether or not a subject was covered only by Medicare; (3) whether or not a subject had private inpatient coverage; and (4) whether or not a patient had outpatient coverage. The Medicaid coverage for mental disorders in Maryland is reasonably generous although the fee schedule for office-based providers is quite low, discouraging supply from these providers. The presence of outpatient coverage means that the individual is certain to be covered by the mental health mandate in Maryland that specifies coverage for ambulatory mental health services (50 percent coinsurance and 30 visits per year). We therefore expect Medicaid and private

94


outpatient coverage to encourage use of services in general and specialty services in particular relative to the no-coverage reference group. Employment status is measured by two dichotomous variables. The first is equal to one if an individual is currently employed, and zero otherwise. The second is equal to one if an individual is in the labor force but unemployed (looking for work). The reference group is made up of those not in the labor force (homemakers, retirees, and others). The two employment variables may reflect both the relatively high opportunity cost of time for the employed and the effect of the stigma associated with mental illness in the workplace. Thus, we expect the effect of both variables on use to be negative and most strongly negative for specialty mental health providers. This prediction is made because specialty services tend to require more time-intensive visits and are more likely to generate stigma. Three measures of availability of specific providers are induded. These are the presence of a private physician in a subject's home census tract; the presence of an alcohol, drug, or specialty mental health clinic in a subject's census tract; and the presence of a community mental health center in a subject's or adjacent census tract. Theoretically we would expect these measures to encourage use and specific types of use. However, since east Baltimore is a confined area that is relatively well serviced, there may in fact be little effective variation in availability.

EMPIRICAL RESULTS Table 2 reports results for two sets of logit estimates. The first column of Table 2 reports a dichotomous logit for the probability of any use of mental health services. These results appear primarily to orient the reader for interpreting the multinomial logit results presented in the second, third, and fourth column of Table 2. Because the multinomial logit results normalize the coefficients for the no-use category to zero, the estimated coefficient for the three treatment modes can be viewed as the effect of a particular variable on the probability that a given treatment mode will be chosen relative to the probability of no use. Therefore, the differences between estimated coefficients for a specific variable across equations indicate the effect of that variable on the probability that one treatment mode will be chosen over another. The pseudo R2 statistic reported at the bottom of the first column indicates that roughly 11 percent of the variation in the probability of use is explained by the logit model.4


95

Table 2: Multivariate Results* Variable Intercept

Age

SEX RACE MARITAL ST

EDUCATION EMPLOYED

LAIDOFF INCOME MEDICAID MHNEED Number chronic

Self-health Ambulatory insurance Hospital insurance

Logit Probability of Use -3.32 (4.21) 0.11 (3.80) 0.07 (0.45) -0.84 (4.98) -0.37 (2.31) 0.0003 (0.12) -0.31 (1.74) -0.03 (0.08) -0.31e°5 (0.39) 0.45 (2.00) 1.34 (9.40) 0.35 (4.40) -0.14 (0.77) -0.02 (0.07) -0.06

Design Informal

0.79

0.51

-4.75 (3.19) 0.047 (0.86) 0.28

0.99

-0.45

-0.98

0.71

0.04 (0.70) -0.13 (0.38) -0.66 (0.63) 0.Sle°5

-0.03 (0.90) -0.14 (0.57) -0.23 (0.44) -0.21e°5

0.83

-0.34

0.98

0.98 0.75

1.20 0.88

0.98

Private provider

ALC/DRG CMHC

AGESQ N Log of likelihood function R2 (pseudo)

0.30 (1.15) -0.19 (0.54) -0.04 (0.13) -0.16 (0.99) -0. 15e02 (4.50) 2,802 -759.736

(0.91) (1.43) -0.39 (1.29)

(0.36) (0.76) 1.26 (4.48)

(3.41) 0.09 (2.42) 0.33

-6.68 (4.20) 0.27 (4.12) -0.63

(1.53) (4.34) -0.21 (0.99)

(2.21) (3.05) -0.67 (2.12)

(0.19) 0.54 (1.91) 1.33 (6.95)

(0.97) 0.98 (2.20) 1.55 (5.27)

-3.56

-0.96

0.04 (0.86) -0.88 (2.62) -0.21 (0.32) -0.15e°4

0.34 (2.18) -0.05

1.06

-0.25

0.36 (3.51) -0.14

1.15

(0.74) 0.04 (0.10)

-0.33

(0.60) 0.23 (0.72)

-0.46

(0.16) 0.49 (1.32)

(0.68)

(1.29)

(2.26)

-0.18 (0.33) -0.58

0.33 (0.97) -0.19

0.64 (1.31) 0.06

(0.78) 0.13 (0.23)

(0.38) 0.007 (0.01)

-0.29

1.00

0.34

(2.23)

1.23

(0.22) Medicare

Multinomial Logit Mental Medical

Effectt

1.16 1.02 1.21

0.77 1.00

0.11

Asymptotic t-statistic in parentheses. tDesign effect - t weighted fixed effects/t simple logit.

0.008 (0.03) -0.67e°3 (1.21)

-0.31 (1.41) -0.21e°2 (2.97) 2,802 -1,011.99

1.11

(0.11) (0.52)

-0.71 (0.23)

-0.35e02 (4.64)

96


The empirical work presented is aimed primarily at explaining patterns of utilization. Thus, the multinomial logit results will be the primary focus of the discussion. However, there are a number of variables for which effect on the probability of use is of research and policy interest. For these factors we will present results from the simple logit model. We begin the review of results by examining the variables describing coverage by third party payers. The coefficient estimates for the Medicaid variable indicate that Medicaid leads to increased probabilities of use of mental health services delivered by both general medical providers and specialty mental health providers. The Medicaid coefficient in the simple logit model suggests an odds ratio of 1.56 for Medicaid, which is significant (adjusted t = 1.65) at the .05 level (using a one-tailed t-test).5 This means that individuals covered under the Medicaid program are 1.56 times more likely to use mental health services than are otherwise similar, uninsured individuals. Taube and Rupp (1986) estimated a model of use of mental health services using data from a subset of the National Medical Care Utilization and Expenditure Survey (NMCUES). Their estimate of the relative odds of utilization for the Medicaid population relative to the general population was 2.09, which is larger than that found here. In a separate analysis on the full NMCUES data set, Taube et al. (1986) estimated that the relative odds of mental health care use for Medicaid enrollees was 1.69 that of the general population. This is very close to our results. The results point weakly to some response in the type of provider chosen as coverage becomes complete (i.e., out-of-pocket costs under Medicaid are zero). The odds ratios for Medicaid in the general medical and specialty mental health equations were 1.71 and 2.64, respectively. This implies, controlling for the other variables, that Medicaid recipients are 1. 7 times more likely to receive mental health care from a general medical provider than similar individuals not on Medicaid. Similarly, Medicaid increases the chances, by a factor of 2.6, of receiving some mental health care from a specialty provider. The results suggest that Medicaid enrollment increases the likelihood that a user of mental health services will choose a specialty mental health provider over a general health care modality by a factor of 1.55- which is not significantly different from 1 (adjusted t = 0.75). The Medicaid coefficient in the informal provider equation was negative and not significantly different from zero at conventional levels. This is a sensible finding since Medicaid does not as a rule reimburse informal providers. These results are consistent with a price response in a decision about whether or not to use services.


97

The results for the other insurance variables indicate essentially no effect of coverage on either use of any services or the pattern of use. While for the most part the coefficients had the expected signs, they were not significant. Several anomalous results, however, should be noted. The most puzzling of the insurance results is the coefficient estimate for the ambulatory coverage variable, which was negative with an adjusted t-statistic of 1.51 for specialty mental health in the multinomial results. One possible explanation is that those with private ambulatory coverage are more likely to have an office-based nonpsychiatrist physician as their usual source of care, and that those physicians may treat all but the most severe mental and emotional problems themselves (Carey and Kogan 1971). The employment variable estimates reveal that, controlling for other factors including mental health status, labor force participants are in general less likely to make use of any type of mental health services. For example, the simple logit results for the EMPLOYED variable imply an odds ratio of 0.73, meaning that, controlling for other factors, employed individuals are 27 percent less likely to use mental health services than those not in the labor force (significant at the .05 level, adjusted t = 2.09). The coefficient for EMPLOYED in the specialty mental health equation suggests that these individuals are especially averse to using the specialty mental health sector. That is, if employed individuals use mental health services, they are least likely to obtain them from a specialty provider. The LAIDOFF variable is most often negative but not significant at conventional levels. These results are consistent with the opportunity cost of time as well as the stigma arguments posed earlier. Unfortunately, we cannot clearly differentiate between the two explanations for the observed results. The results for the indicator of mental distress (MHNEED) imply that use of mental health services from all sources is significantly related to our measure of distress. This can be seen most clearly in the odds ratio derived from the simple logit model. The relative odds of use of any care for those meeting our clustering criteria is 3.82. The results from the multinomial logit model suggest that MHNEED has a monotonically increasing effect on probability of use as one moves from informal providers to specialty providers. For example, the relative odds of using a specialty versus a general medical provider for someone in "need" is 1.24. The differences in the effect of MHNEED on the likelihood of use of each type of provider are not significant (t = 1 for the largest difference). Thus, mental distress does appear to explain who uses services but does not serve to differentiate to any large extent

98


among sources of care. This finding is in accord with the work reported by Wells et al. (1982). The estimates for the variable that measures the number of chronic health conditions indicates a significantly positive effect on the probability of any use.6 The effect of the number of chronic conditions on the source of mental health treatment is constant across alternatives. This means that, although physical health status serves to explain probability of use, it does not allow one to differentiate among the modes of care used. Finally, the self-reported positive health status measure (Self-Health) was consistently negative, as expected. However, the estimates were very imprecise (t-statistics always < 0.75). In order to probe further the results for health and mental health status, we reestimated the multinomial logit model with an interaction between MHNEED and the number of chronic conditions. This interaction variable was never significant and always had a negative sign. This indicates that if there is any interaction effect, it is to decrease the likelihood of any mental health care use. The magnitude of the interaction coefficient was largest in the specialty mental health equationalthough still not significant at conventional levels (t = -1.21). Proximity to various providers never seems to influence either use of services or the pattern of care. (We suspected that we might obtain such a result because the geographic area from which data were collected is relatively compact and well served.) Among the demographic factors, the results for the age variables are particularly notable. Age was specified as having a quadratic relation to probability of use. The results from the simple logit suggest that the probability of use increases until age 37, then drops at a rate of about 0.003 per year. This means that, holding other factors constant, persons age 67 years will have utilization rates for mental health services that are roughly 9 percent below those of 37-year-old individuals. The pattern for the age effect differs slightly across treatment modes. For example, the probability of use of specialty mental health services peaks at roughly 39 years, whereas use of informal providers peaks at 36 years. These results imply lower rates of use among the elderly for all types of services. The use of the specialty sector appears to drop off at a somewhat slower rate than use of the other two sectors. Examination of the results for the informal care sector suggests that, except for variables related to third party coverage of health expenses, the qualitative results are similar to those for the medical and specialty mental health sectors. The coefficient estimates for the health and mental health status indicators, in particular, are notable due to their close correspondence across sectors. The difference in odds of

Provide Choicefor Mental Disorder

99

obtaining care in the specialty versus informal care sector is 1.19 for those with MHNEED equal to one. This estimate is not significantly different from one. Thus, our indicator of a clustering of signs and symptoms drives people equally into both types of care. The estimated effects for employment status and marital status also suggest that the use of informal care appears to be driven by forces similar to those determining use of other classes of providers. DISCUSSION OF RESULTS

Perhaps the most notable results reported here are those that link health and mental health status to choice of setting. While health problems measured by chronic conditions and a clustering of signs and symptoms of mental distress both significantly increase the probability of using some kind of mental health service, they do not strongly differentiate among those sources of care. Given that utilization of specialty mental health care services is five times that of use of mental health care services from the general medical sector, one would expect to observe differences in illness. Several points need to be raised in connection with these results. We do not measure outcomes stemming from treatment in each sector-which may differ substantially. Differences in outcome associated with higher use could explain differences in use across treatment modes. Another explanation that has been suggested is that patients with a mental health problem begin in the general medical sector and end up being treated for their mental illness in the specialty sector. Since we defined as specialty patients those who sought care from both the general medical sector and the specialty mental health sector, this explanation is largely precluded in our results. Finally, there is some weak evidence of some differences in mental health status across settings. The monotonically increasing coefficient estimates for the MHNEED variable suggest a greater tendency among those in distress to seek care from the specialty sector ceterus paribus. A second set of results of interest for policy purposes includes those for the Medicaid variable. Those results indicate that a subject's eligibility for the Maryland Medicaid program substantially increases the probability-relative to being uninsured-that he or she will use mental health care delivered in both the general medical and specialty mental health sectors. Moreover, the results suggest that Medicaid's effect may be especially pronounced for specialty mental health services. This result is only suggestive, since the differences between the general medical and specialty effects were not significant at conven-

100


tional levels. Given the high rates of mental morbidity among the poor and near-poor, expansion of Medicaid may be a particularly effective approach to expanding access to "appropriate care" for this population. We take up this issue in a bit more detail in the simulations reported in the next section. A third result of particular interest is the estimate related to the EMPLOYED variable. The strong negative-coefficient estimate points to lower use of all mental health services by employed individuals, holding constant such factors as mental health status. Moreover, the EMPLOYED coefficient was significantly more negative for specialty services than for the other two alternatives. As mentioned earlier, this is consistent with both the opportunity cost and stigma hypotheses for lower use among the employed. We have made some preliminary attempts to tabulate responses to attitudinal questions about mental health with employment status. Our initial analyses indicate that a somewhat higher proportion of the employed appear to be concerned about what others will think if they are treated for a mental problem. However, the differences in proportions are not sufficient to explain the large negative coefficient. This raises a need to investigate further the opportunity cost hypothesis and to attempt to incorporate stigma indicators more workably into demand models such as the one presented here.

POLICY SIMULATIONS In this section we present one simulation as an example of the type of benefit design changes currently being debated among policymakers and clinicians. We simulate an expansion of the Medicaid program to all individuals who live in households below 125 percent of the poverty line. This is particularly relevant due to recent concerns about problems of health care financing for the indigent. States such as Florida, South Carolina, and Virginia, among others, have started making use of the Medicaid program as a means of funding health care for the uninsured poor. The problem of medical indigency is particularly serious in the mental health area. For example, the percentage of psychiatric inpatients who are uninsured is twice that of the population of nonpsychiatric admissions (Hospital Discharge Survey 1985). The results are summarized in Table 3. As is evident, use of mental health services provided by either general medical providers or specialty mental health providers would expand substantially with


Table 3: Simulation Results Before

After

.904 .021 .050 .024

.893 .020 .058 .029

Medicaid Simulation No use Informal Medical Mental

101

Medicaid coverage. Utilization rates are shown to increase from 9.6 percent to 10.7 percent, an increase of almost 10 percent in the number seeking care. Use of informal services declines slightly and is shifted into reimburseable modes (2.1 percent versus 2.0 percent). The increases in use are substantial for both the services delivered in the medical care sector (5.0-5.8 percent) and those in the specialty mental health sector (2.4-2.9 percent). The change in use of specialty services amounts to about a 25 percent increase in use; the increase would come largely from the population of poor and distressed (MHNEED = 1). If one is concerned about use of services by those not seriously distressed, the results suggest that expansion of Medicaid will largely encourage use among those who exhibit a clustering of signs and symptoms. This would indicate a tendency toward "appropriate use of services." This simulation result is consistent with points raised by Shapiro et al. (1985) and Taube and Rupp (1986). Both of those studies point to substantial barriers to treatment for the uninsured poor and near-poor. The result of those barriers is relatively high levels of unmet need. The simulation results suggest that the change of the income-eligibility criterion would effectively lead to more needs being met. In summary, this article has shown that use of informal care accounts for a significant portion of all mental health care use (21 percent). Use of informal care responds to demographic and health status factors in a manner akin to medical and specialty sector utilization. However, factors related to reimbursement have little effect on informal use. Health and mental health status are important predictors of use per se, but do not predict the type of use well. Medicaid coverage is an important determinant of use and tends to encourage use of specialty services. This is clearly reflected in the simulation results that traced the consequence of an expansion in Medicaid eligibility. Finally, future work should be concerned with developing an understanding of why treatment costs differ so dramatically from sector to sector, while health and mental health status do not.

102


ACKNOWLEDGMENTS The authors are grateful to Becky Clark for computing assistance, and to Sam Shapiro, Will Manning, Gordon DeFriese, and two anonymous referees for useful comments.

NOTES 1. Post-diction occurs when one uses a health status measure made subsequent to reported utilization to predict utilization. 2. The MNL model is often associated with the "independence of irrelevant alternatives" assumption. To relax this assumption, models are sometimes estimated as nested or sequential. In this application - since all right-handside variables are individual characteristics - nesting of models produces the same results as nonnested models (McFadden 1982, 10-11). To verify this contention, we nested the choice of treatment alternatives conditional on use of any service. The resulting odds ratios were identical to that of the

nonnested model. 3. The software technology for developing complex survey-variance estimates for the MNL model has not been developed. Thus, the design effects presented should serve only as a general point of reference. 4. This was calculated as R2 =1 - (Lw/LT)21x where Lw is the likelihood function when maximized for all parameters, LT is the likelihood function when it is maximized with only a constant, and n is the sample size. 5. Apart from complex sample design effects on the reported asymptotic tstatistics (and the significance levels they imply), the reader should bear in mind that statements about statistical significance are based on ex post multiple comparisons. While this is conventional practice, it reemphasizes the advisability of taking significance levels as general guides rather than interpreting them literally. 6. This is consistent with the literature on somatization and work on the health/mental health interface (see Lipowski 1975).

REFERENCES Anthony, J. C., M. Folstein, A. J. Romanoski, et al. "Comparison of the Lay DIS and a Standardized Psychiatric Diagnosis." Archives of General Psychiatry 42 (July 1985):667-75. Carey, K., and N. S. Kogan. "Exploration of Factors Influencing Physician Decisions to Refer Patients for Mental Health Services." Medical Care 9 (January/February 197 1):55-66. Eaton, W. W., D. A. Regier, B. Z. Locke, and C. A. Taube. "The Epidemiological Catchment Area Program of the National Institute of Mental Health." Public Health Reports 96, no. 4 (August 1981):1218-25.


103

Eastwood, M. R. The Relation between Physical and Mental Illness: The Physical Status of Psychiatric Patients at a Multiphasic Screening. Toronto: University of Toronto Press, 1975. Frank, R. G. "Use of Mental Health Services and Persistence of Emotional Distress: An Exploratory Analysis." Working Paper, Health Services Research and Development Center, Johns Hopkins University, 1987. Goldberg, D. P. The Detection of Psychiatric Illness by Questionnaire. London:

Oxford University Press, 1972. Lipowski, Z. "Psychiatry of Somatic Diseases: Epidemiology, Pathegenesis Classification." Comprehensive Psychiatry 16, no. 2 (1975): 105-24. Manning, W. G., J. P. Newhouse, and J. E. Ware. "The Status of Health in Demand Estimation; or Beyond Excellent, Good, Fair, Poor." In Economic Aspects of Health. Edited by V. Fuchs. Chicago: University of Chicago Press, 1981. McFadden, D. "Econometric Models of Probabilistic Choice." In Structural Analysis of Discrete Data. Edited by C. Marski and D. McFadden. Cambridge, MA: MIT Press, 1981. . "Qualitative Response Models." In Advances in Econometrics. Edited by W. Hildenbrand. Cambridge, UK: Cambridge University Press, 1982. National Center for Health Statistics. Hospital Discharge Survey. Washington, DC: Government Printing Office, 1985. Shapiro, S., et al. "Utilization of Health and Mental Health Services." Archives of General Psychiatry 41 (October 1984):971-78. . "Measuring Need for Mental Health Services in a General Population." Medical Care 23, no. 9 (September 1985):1033-43. Taube, C. A., and A. Rupp. "The Effect of Access to Ambulatory Mental Health Care for the Poor and Near Poor Under 65." Medical Care 24, no. 8 (1986):677-86. Taube, C. A., et al. "Estimating the Probability and Level of Ambulatory Mental Health Use." Health Services Research 21, no. 2 (1986):241-66. Taubman, P., and S. Rosen. "Healthiness, Education and Marital Status." In Economic Aspects of Health. Edited by V. Fuchs. Chicago: University of Chicago Press, 1981. Ware, J. E., et al. "Health Status and Use of Outpatient Mental Health Services." American Psychologist 39, no. 10 (1983):1090-100. Wells, K., et al. Cost Sharing and the Demandfor Ambulatory Mental Health Services. RAND Report R-2960-HHS. September, 1982.