The report of the President's New Freedom Commission called for the U.S. mental health care system to be transformed into a consumer-centered system that is focused on recovery and delivers excellent care without disparities (1). Such a transformation will require large-scale workforce development efforts, such as training and redistribution initiatives, that are informed by a national perspective on county-level need. Although there is evidence of widespread unmet need in this country (2,3), we lack a current and accurate assessment of need for mental health professionals on a county-specific basis.

In the absence of national minimum standards for mental health services (based on mental disorder, symptoms, and functional limitations, for example), need estimation is challenging for at least four reasons. First, the prevalence of mental disorders (and therefore the appropriate level of utilization) varies across demographic and socioeconomic groups (4,5,6,7). Second, not everyone needs mental health services, and among those who do, the level of need varies greatly (8,9,10). For example, people with diagnosed mental disorders are more likely to require mental health services than are those without (10). Third, some need for mental health services is met not by mental health professionals but by primary care physicians, who see about half of those with mental health needs and cover about 21% of all mental health visits (11). This complicates the measurement of utilization. Fourth, there is often a mismatch between the level of need and the amount of services received. This may be related to the fact that perceived need correlates very weakly with measured morbidity (12). Among people with serious mental illness, 54% do not receive timely care (2,3,13). On the other hand, about half of those who receive mental health care do not have serious mental illness, in part because decreased stigma among the more affluent segments of the population has turned mental health services into something of a consumer good (3,14,15).

The goal of this study was to formulate the best estimates possible, given current data, to aid in workforce planning and to call attention to the need for better data. We address the lack of a national picture of local need and the challenges of utilization-based need estimation by using a synthetic estimation technique combined with U.S. census data to develop county-level estimates of the prevalence of serious mental illness throughout the United States. This method was developed in response to an invitation by the Health Resources and Services Administration (HRSA) to update the process for designating shortages in the supply of mental health professionals (16); the process is used for resource allocation, including the placement of specific types of mental health professionals throughout the country by the National Health Service Corps in response to the needs of local communities. Assessments of health workforce requirements typically use one of three methods: need estimation based on disease prevalence and the time required for treatment, economic forecasting of demand for services, or benchmarking to a known population (for example, a health maintenance organization) (17). Our approach was to combine the need and demand methods. Estimates of disease prevalence and current utilization were generated from nationally representative survey data, but the utilization estimates for adults with serious mental illness were adjusted in order to include people who need treatment and are not receiving it. A full report describes the method and results in more detail (18).

Estimating county prevalence of serious mental illness

Following several earlier estimates of the prevalence of serious mental illness (19,20), we generated county-level prevalence estimates by applying the predicted probabilities from demographic models to cross-tabulations of census population data. The predictive models were based on the most recent nationally representative data on the distribution of psychiatric disorders in the general population: the 2001 National Comorbidity Survey Replication (NCS-R) (21). The NCS-R (N=9,282) contains information about diagnoses, functional limitations, mental health service utilization, and demographic descriptors.

Defining serious mental illness operationally is a complex task involving diagnosis, functional impairment, and duration (22). Although a conceptual definition is published (23), no commonly accepted computational algorithm exists (24). We classified NCS-R respondents as having serious mental illness if they met three requirements: first, Composite International Diagnostic Interview diagnosis of bipolar I, bipolar II, mania, major depressive disorder, agoraphobia, generalized anxiety disorder, hypomania, panic disorder, posttraumatic stress disorder, social phobia, or specific phobia; second, a high level of disability as indicated by either the inability to carry out normal activities as a result of mental health problems for at least 120 days in the past year or a mean self-rated impairment level of 7 or higher on a 10-point scale across four dimensions (home, work, relationships, and social life); and third, age at onset at least two years less than the respondent's age (in order to exclude those whose disorders had lasted less than 12 months).

Our diagnosis, disability, and duration criteria were designed to identify the group of people of primary concern to the mental health service system—those who have significant mental health service needs. Therefore, although we included a relatively broad range of diagnoses, these criteria reflect a higher level of impairment than does the traditional definition of serious mental illness. Because we planned to compare our need estimates with workforce estimates, and because data on the substance abuse treatment workforce are quite limited, our criteria did not include substance use disorders. We also excluded some typically less severe mental disorders: adult separation anxiety disorder, attention-deficit disorder, subthreshold bipolar symptoms, conduct disorder, dysthymia, intermittent explosive disorder, oppositional defiant disorder, and panic attack. Unfortunately we were unable to incorporate schizophrenia and other psychotic disorders into the inclusion criteria because the NCS-R does not elicit enough information to diagnose psychotic disorders, its psychosis screening questions are subject to false positives, and the psychosis section does not include questions about functional impairment. Analyses of clinical data not in the public release files suggest that 79% of people with nonaffective psychosis would be identified through other diagnoses as having serious mental illness (25,26).

NCS-R data were used to model the probability of having serious mental illness in relation to demographic predictors. We used a two-stage logit model to generate the predicted probabilities, because poverty level, an important predictor of serious mental illness, was available only for the 61% of NCS-R respondents who completed part 2 of the survey instrument, and participation in part 2 was strongly associated with having serious mental illness. (Among NCS-R respondents with serious mental illness, 99% responded to part 2.) In order to maximize use of the available information, we predicted the probability of part 2 participation as a function of age, sex, race, marital status, and education level. Then, given part 2 participation, we predicted the probability of having serious mental illness as a function of poverty level along with the other demographic predictors. We combined the predicted probabilities from the two models to yield overall predicted probabilities of serious mental illness (based on all six predictors) that were not conditional on part 2 participation. The set of predictors was limited to demographic variables in order to avoid reinforcing any disparities related to other predictors (such as region). The small cell sizes for many combinations of predictor values did not support interaction terms in the models.

Using synthetic estimation procedures (27,28), we then applied the predicted probabilities to county subpopulations, defined by all the permutations of the same demographic variables used in the logit models. We used demographic data from the U.S. census for 2006 (29,30) to yield estimates of the number of people with serious mental illness for each county subpopulation. The county subpopulation estimates were then aggregated to obtain an estimate of the number of people with serious mental illness in each county. Because these prevalence estimates were derived from NCS-R sample data rather than measured in the entire U.S. household population, they are subject to sampling error. Taking into account the NCS-R sampling design, we used balanced repeated replication to assess the effect of sampling variation on the prevalence estimates. Prevalence estimates were generated from 88 randomly selected half-samples from the NCS-R data, and a 95% confidence interval was constructed around the whole-sample prevalence estimate for each county and state.

Estimating need for mental health services

We estimated the need for mental health professionals separately for individuals with and without serious mental illness based on each group's use of mental health services. For people with serious mental illness, we chose the NCS-R as the most appropriate data source; for those without serious mental illness, we used the 2000 Medical Expenditure Panel Survey (MEPS). Both data sets provide detailed information on service use and types of mental health providers.

A total of 377 persons met the above three criteria for serious mental illness. Using data on outpatient services from the 356 individuals in the NCS-R sample with adequate data, we estimated mental health service need. We assumed that everyone with serious mental illness should receive outpatient mental health services at some time over the course of a year and estimated their need for mental health services as the mean number of visit minutes among service users. Using provider categories defined in the NCS-R instrument, we generated separate estimates of need for prescribers (psychiatrists, general practitioners, or family physicians) and for nonprescribers (psychologists, social workers, counselors, or other mental health professionals such as psychotherapists or mental health nurses). Visits to primary care providers for mental health care were included in order to estimate the full extent of need for mental health services. The role of primary care providers is addressed below.

Because the NCS-R survey obtained utilization data only from individuals who met diagnostic criteria, need among people without serious mental illness was estimated with the 2000 MEPS (31), which has a large sample of noninstitutionalized civilians. Individuals who appeared to have serious mental illness (that is, those with ICD-9 codes 295–301 or 312 who rated their conditions as "serious" in any data collection round; weighted proportion 2.1% of respondents) were excluded to avoid duplication with the NCS-R sample, yielding a very large sample of persons without serious mental illness (N=16,418).

Based on respondents' self-report data, we defined mental health visits as outpatient visits involving a possible mental health provider (physician, nurse, nurse practitioner, psychologist, social worker, or other) and one of the following: psychotherapy, a psychotherapeutic drug, or diagnosis of a mental disorder. Because it was not appropriate to assume that every individual without serious mental illness should receive outpatient mental health services over the course of a year, the estimate of need for providers was the mean number of visit minutes overall (not only among service users). Again we generated separate estimates of need for prescribers and need for nonprescribers. As with the population with serious mental illness, we included primary care providers because the goal was to estimate the total need for mental health services. Our estimates exclude inpatient services, general health care providers other than physicians and nurses, hotlines, religious or spiritual advisors, support groups, self-help, and complementary and alternative medical professionals.

We converted minutes of services needed for the populations with and without serious mental illness to full-time-equivalent (FTE) estimates of prescribers and nonprescribers, using practice pattern data from various sources (32; also unpublished data: American Psychiatric Association's National Survey of Psychiatric Practice, 2002; Center for Substance Abuse Treatment's Practitioner Services Network II, 2003). Conversion factors reflected that 71% of nonprescribers' time (range 64%–79% across professions) and 60% of psychiatrists' time is spent in direct contact with patients. (These factors imply that psychiatrists typically provide 1,208 hours of direct patient care per year, whereas other mental health professionals typically provide 1,410 hours.) For each provider category (prescribers and nonprescribers), the sum of the need estimates for people with and without serious mental illness was used as a preliminary county-level estimate of total need.

To account for the portion of need that is met by primary care providers, we adjusted the preliminary need estimates. Primary care providers account for about 21% of all mental health visits; however, the scope of their mental health practice is constrained by their mental health training and the physical health issues competing for their limited time (11,33,34,35,36,37). Also, primary care providers tend to see a mix of patients with less severe mental health problems than the problems of clients seen by mental health professionals (38,39). In order to acknowledge the role of primary care providers in addressing mental health need while estimating conservatively the appropriate size of their contribution, the need estimate was reduced by 15% in counties where there is a sufficient primary care workforce (in other words, there is no shortage). We chose 15% as the factor because the metric of need here was visit minutes rather than number of visits; because we found no empirical guide in choosing a cutoff point, the choice was arbitrary. This percentage was prorated on the basis of the proportion of the county's primary care need that is met, according to the shortage score proposed by Ricketts and colleagues (40) (calculated with 1998 data).

The University of North Carolina's Public Health Institutional Review Board determined that this study did not require board approval.

Prevalence of serious mental illness

The results of the two-stage logit model to predict the probability of having serious mental illness in relation to demographic predictors are presented in Table 1. The two-stage logit model had 65% sensitivity and 64% specificity with a predicted probability threshold of .04. The area under the receiver operating characteristic curve was .71, which indicates acceptable fit (41).

Our method yielded a prevalence estimate of serious mental illness in the NCS-R sample of 3.9% (model-based prevalence 3.7%), which is in line with national estimates derived from other sources (13,42,43,44,45) (Table 2). The highest estimates counted persons with any mental disorder in the past 12 months; measures with shorter recall periods or narrower frequency and severity criteria yielded lower estimates. Our criteria were among the strictest in that they did not include substance use disorders or psychosis screen information and required significant functional impairment. Thus we attempted to identify a group of people who have both serious disorders and a level of functional impairment that necessitates significant service use.

Synthetic estimation resulted in state-level prevalence estimates (N=51) that ranged from 3.2% to 4.5% (mean±SD=3.8%±.3%); county-level estimates (N=3,140) of prevalence of serious mental illness ranged from 2.5% to 9.0% (4.0%±.5%). The width of the 95% confidence interval (CI) around the prevalence estimate for each county or state derived from the half-sample replications gives an indication of the effect of sampling variation on the estimates. At the state level, the width of the CI ranged from .6 to 1.8, with a mean and median of .9. At the county level, it ranged from .5 to 4.1, with a mean and median of 1.2. (The mean relative standard errors were .06±.01 and .08±.01 at the state and county levels respectively.)

Service use and requirements for health professionals

We calculated the proportion of respondents who were service users and the mean total hours of visits along with CIs. About half of adults with serious mental illness used services; they typically spent 10.54 hours per year (CI=5.46–15.63) with nonprescriber mental health professionals and 4.38 hours per year (CI=3.40–5.37) with primary care physicians or prescriber mental health professionals. Less than 10% of adults without serious mental illness used specialized mental health services. Overall, adults without serious mental illness spent only 7.8 minutes (CI=5.4–9.6) with nonprescriber mental health professionals and 12.6 minutes (CI=10.8–14.4) with primary care physicians or prescriber mental health professionals in the reference year.

Table 3 displays our estimates with CIs of the mental health professional FTEs required to treat the U.S. adult household population (calculated as described above). For 2006 we estimated that the U.S. adult household population with serious mental illness was 8,138,223, and another 210,106,179 adults without serious mental illness required a much lower amount of mental health services. County-level adjustments for the contribution of primary care physicians reduced the national estimates of mental health professional FTEs needed by 14.5%. Under our assumptions (detailed above), approximately 56,462 FTE prescribers and 68,581 FTE nonprescribers are needed to provide services to the U.S. adult household population. Figure 1 shows the distribution of need among counties after considering prevalence and primary care availability. The total number of FTE mental health professionals needed per county has a wide range, from near zero to 4,000 (40±124, median=12). [The map can be viewed in closer detail as an online supplement to this article at ps.psychiatryonline.org.]

Clearly, the quality of our estimate of the national level of need for mental health services is only as good as the data and assumptions we used. It is important to note, for example, that the estimates presented here do not reflect the needs of children or the needs of adults who are homeless, in the military, or living in institutions. Because of the lack of data for people living in institutions and the difficulty of distinguishing long and short hospital stays, the estimates presented here are limited to need for outpatient visits. They do not reflect local neighborhood, community, or personal factors that affect individual need, such as stressors and environmental stigma (46,47), nor do they speak to the most appropriate mix of prescribers and nonprescribers. Although we believe that we treated race and ethnicity appropriately, variation in the prevalence of serious mental illness associated with race, ethnicity, and linguistic isolation is not well understood. Further, although we believe that the county is the appropriate unit of analysis given the available data and the goals of the effort, there may be within-county variation in need not reflected here.

Because our focus was on licensed mental health professionals, our estimates do not reflect need for providers such as registered counselors or registered hypnotherapists, nor do they address need for frontline workers such as hospital- or community-based aides, who are an important foundation of mental health services (48). Although these estimates were adjusted to account for the role of primary care providers, the validity of our adjustment factor is untested.

Furthermore, there are three major sources of error associated with our estimates of need. First, there is statistical error associated with modeling serious mental illness status in the NCS-R. Our two-stage logit model had acceptable fit but was nonetheless subject to prediction error. Second, although our county-level prevalence estimates had relative standard errors that met the usual reliability criterion of less than .30, there was sampling error associated with estimating the prevalence of serious mental illness by applying the NCS-R model to census data. Our NCS-R and MEPS utilization estimates were subject to sampling error as well. (These estimates also had relative standard errors less than .30.) Third, there is error associated with collecting census data. To the extent that the census data were inaccurate (for example, undercounting people who are likely to have serious mental illness or incorrectly assessing low income, transient residence, or household composition), our prevalence estimates will be imprecise. Estimating need in stages compounds the error from these three sources.

We are aware of no contemporary standard by which to validate our need estimates. However, it is worth noting that our estimates of need for psychiatrists, which translated to about 25.9 psychiatrists per 100,000 adult population (corrected for primary care substitution), substantially exceed the need-based standard developed almost 30 years ago by the Graduate Medical Education National Advisory Committee (49) of 15.4 psychiatrists per 100,000 population. This may not be surprising, given changes in mental health treatment patterns over the past several decades. For example, about 38% of the mental health professional FTEs in our estimate would be required to treat adults with serious mental illness who now live in the community, although such individuals might have been long-stay hospital patients in the 1970s.

This article presents a method to estimate the need for mental health professionals in individual counties throughout the United States and reports preliminary national estimates of need. Our estimates are probably most useful when taken as an expression of relative rather than absolute need. Our continuous measure of county-level need could be used in conjunction with other information, such as mental health professional supply or met need (50, this issue), in order to target resources based on a given threshold or mental health professional shortage measure (51, this issue) to guide policy decisions. Assessing local need in absolute terms would require a more detailed classification of levels of need, specific population estimates for different levels of functioning, and comprehensive standards for level of service use. We view our preliminary estimates of the national need for mental health professionals as a starting point that mental health planners, educators, and workforce analysts can improve on in future work.

This work was supported by contract HHSH-230200532038C from the Health Resources and Services Administration. The authors acknowledge the help of the project officer, Andy Jordan, M.S.P.H.; their advisory board, which included Michael Almog, Ph.D., David Bergman, J.D., Tim Dall, M.S., Sheron R. Finister, Ph.D., John C. Fortney, Ph.D., Nancy P. Hanrahan, Ph.D., R.N., Sharon M. Jackson, M.S.W., L.C.S.W., Nina Gail Levitt, Ed.D., Ronald W. Manderscheid, Ph.D., Noel A. Mazade, Ph.D., Bradley K. Powers, Psy.D., Richard M. Scheffler, Ph.D., Laura Schopp, Ph.D., Lynn Spector, M.P.A., Marvin S. Swartz, M.D., and Joshua E. Wilk, Ph.D.; and the following individuals: Rick Harwood, Marlene Wicherski, Jessica Kohout, Ph.D., Lynn Bufka, Ph.D., Olivia Silber Ashley, Dr.P.H., Bob Bray, Ph.D., J. Valley Rachal, Ph.D., Mark Holmes, Ph.D., Edward Norton, Ph.D., and Gary Koch, Ph.D. The views expressed in this report do not necessarily reflect the official policies of the U.S. Department of Health and Human Services, nor does mention of organizations imply endorsement by the U.S. Government.

The authors report no competing interests.