A new way to estimate neurologic disease prevalence in the United States

Regional MS prevalence studies

A recent systematic review⁴ identified 9 population-based MS prevalence studies conducted in small regions of the United States between 1985 and 2011 (table e-1; https://doi.org/10.5061/dryad.t1k42p8). MS point prevalence was lowest in the 19 counties surrounding Lubbock, Texas (age-standardized prevalence 39.9/100,000; 95% confidence interval [CI] 34.0–45.7)⁵ and highest in Olmstead County, Minnesota (age-standardized prevalence 191.2/100,000; 95% CI 165.6–216.8).³ A noteworthy advantage of such regional studies is that the relatively small population size enables investigators to use many overlapping methods to ascertain possible MS cases in the study population (case ascertainment) and apply standardized criteria to determine which cases have MS (case definition). Regional MS prevalence studies have used a variety of methods to maximize the completeness of case ascertainment, with the most common sources being specialty clinics or academic MS centers, community neurologists or other physicians, patient service associations, nursing homes or long-term care facilities, and death certificates.⁴ Case definitions are usually established using medical chart reviews abstracted to apply standard diagnostic criteria.⁶ However, because MS prevalence can have considerable geographic variation,⁴ projecting estimates from one or a few regions to the entire United States is of uncertain validity.

National MS prevalence studies in the United States

For a country the size of the United States (∼310 million population in 2010), it is not feasible to use the intensive case ascertainment methods that are routine in smaller geographic regions. In the last 4 decades, 4 studies have estimated the prevalence of MS in the United States by using methods intended to provide representative samples of the US population (table 1). The first study was conducted in 1976 using a probability sample of 725 acute short-term general hospitals and 8,800 physicians in the contiguous United States and surveying them to estimate the number of persons with MS with whom they had contact.⁷ MS prevalence was estimated to be 58 per 100,000 population, corresponding to 123,000 cases in the United States, but Anderson et al.⁸ believed that the prevalence of MS was underestimated due to several methodologic factors. A primary source of underascertainment was relying on physician recall to identify MS cases, as at the time these studies were conducted, computerized medical records systems were rare. Another source of underascertainment was the single-year surveillance window, which would have missed patients with MS who did not seek medical attention for MS during that year. Evidence for a third source of underascertainment came from a prevalence survey conducted in the Weld and Larimer Counties of northern Colorado,⁹ which used multiple case ascertainment methods and found that 21% of all prevalent cases identified came exclusively from the records of 3 MS service organizations in Colorado, a source that was not used in the 1976 study.⁷ Anderson et al.⁸ used several adjustments for these sources of underascertainment to provide an updated estimate of MS prevalence in the United States in 1990, ranging from 250,000 to 350,000 persons.

The next 2 national MS prevalence studies were conducted using the National Health Interview Survey (NHIS), an annual probability sample of the noninstitutionalized US population. During 1982–1996, one-sixth of the NHIS sample was surveyed about MS and other health conditions resulting in limitation of activity. Collins¹⁰ used NHIS data for a 3-year period (1990–1992) to estimate national MS prevalence because a single year of data did not yield enough persons with MS for stable estimation. MS prevalence was estimated to be 70 per 100,000, corresponding to an estimated count of 180,000 cases. A second study also used NHIS data, over 1982–1996.¹¹ MS prevalence was estimated to be 85 per 100,000 population, corresponding to an estimated count of 211,000 cases in the United States (95% CI 191,000–231,000). A major weakness of the NHIS studies is that survey data were collected from one family member within each sampled household who reported the medical conditions for the entire household. This method would underestimate MS prevalence if household members underreport the condition but could also overestimate MS prevalence if self-reported data resulted in the inclusion of provisional cases (or noncases) who would not meet formal case definition criteria.

More recently, a national MS prevalence study was carried out using the Household Component of the Medical Expenditure Panel Survey (MEPS-HC), a nationally representative survey of the US civilian noninstitutionalized population drawn from NHIS that collects data on health status, health utilization, health costs, and health insurance coverage.¹² MS cases were identified on the basis of having one or more medically coded diagnoses of MS (ICD-9 code 340). Between 1998 and 2009, 526 MEPS-HC participants met this criterion, and the annual average number of persons with MS was estimated to be 572,312 (95% CI 397,004–747,619). Requiring only 1 ICD-9 code during the study year could have overestimated MS prevalence if false-positives were included because individuals who had a single ICD-9 340 code were ultimately not diagnosed with MS or the code was entered erroneously. The MEPS-HC probability sample only captured noninstitutionalized persons with MS, which could have resulted in underascertainment because it excludes the estimated 5%–10% of persons with MS residing in nursing homes.^13,14

Two studies have used administrative health claims (AHC) data to estimate the number of persons with MS. Both studies assessed subpopulations of the United States and then used this information to estimate the total number of persons with MS in the United States as a whole (table 2). Pope et al.¹⁵ obtained administrative claims data from privately insured individuals, those aged into Medicare, and Medicaid blind or disabled populations to determine the prevalence of MS in 6 states for 1991–1997. MS prevalence (per 100,000 population) was estimated to be 240 in the privately insured, 180 among Medicare age 65 and above, and 710 among Medicaid disabled populations. These estimates are not strictly comparable due to the different age distributions and the expectation that MS prevalence would be much higher in the Medicaid disabled population. In addition, no algorithm was developed or used for MS cases. Dilokthornsakul et al.¹⁶ estimated MS prevalence for 2008–2012 in the privately insured US population by applying a case definition algorithm using the PharMetrics Plus Health Plan Claims data, a large US private insurance claims database that contained claims data for approximately 42 million enrollees in 2011. The case definition algorithm required at least 2 inpatient claims with ICD-9 code 340, at least 3 ICD-9 code 340 outpatient claims, or at least one MS-specific disease-modifying treatment (DMT) claim. The estimated MS prevalence for 2012 was 149 per 100,000 (95% CI 147.6–150.9), and when extrapolated to the United States as a whole, yielded an estimate of 403,630 individual cases (95% CI 387,445–419,833).

Population-based MS registries

Longstanding MS registries have been successfully used to estimate MS prevalence in 4 countries: Denmark,^17,18 Sweden,¹⁹ Norway,²⁰ and Poland.²¹ However, a similar US population-based MS registry does not exist. In the United States, there are several voluntary MS registries for conducting outcomes research; however, none of these registries attempts the near-complete case ascertainment in a defined geographic region that would enable the estimation of MS prevalence. The benefits and pitfalls of population-based disease registries were recently reviewed²² and are summarized in table 3. A nationwide population-based registry would be costly to implement; however, it is possible that a geographically representative system of regional registries could be implemented, similar to the approach used for cancer surveillance.²³ In the United States, only one neurologic disease registry has been implemented. The National Amyotrophic Lateral Sclerosis Registry was established by a Congressional mandate and began estimating national amyotrophic lateral sclerosis prevalence in 2010.^24,25

National probability health surveys

The primary method used to estimate national MS prevalence in the past has been probability health surveys. However, almost none of the national probability surveys conducted by the US government are sufficiently large enough to contain enough people with MS to enable precise estimation of MS prevalence, much less prevalence within demographic or regional subgroups. Moreover, the national survey that once allowed an overall estimate of MS prevalence (NHIS) no longer includes questions about MS and other disabling conditions. National probability survey data rely on patient or family member reports of physician-diagnosed MS,^10,12 risking the inclusion of false-positive MS cases or provisional cases that would not meet formal case definition criteria. Only one national survey, the Behavioral Risk Factor Surveillance Survey (BRFSS), which surveys more than 500,000 individuals each year, has a sample size that would potentially yield precise estimates of MS prevalence. Feasibility work would be needed to determine if this BRFSS could be used for this purpose.

Administrative health claims data

In other countries with universally funded health care systems such as Canada and many European nations, AHC databases have been used to provide a population-based alternative to the traditional intensive methods of case ascertainment used in regional studies.¹ In the United States, AHC data are maintained by government (public) insurance programs, including Medicare and Medicaid, and by private (commercial) insurance providers. AHC databases are a tremendous resource for studying neurologic disease incidence and prevalence because of their rich prospective data on patients from defined populations and their use of the standardized international disease-coding framework (using the ICD system).²⁶ However, AHC databases are not created for research purposes, making it necessary to demonstrate the validity and the accuracy of case ascertainment algorithm prior to using them for estimating disease prevalence. These challenges exist for estimating prevalence of all other common chronic neurologic conditions.²⁶ AHC data sources have potential strengths and limitations when applied for the purpose of estimating disease prevalence (table 5). The information from AHC databases begins with enrollment into health insurance plans and provides a standardized source of longitudinal patient health care utilization. These data contain near complete capture of demographic information (sex, date of birth, race/ethnicity, zip code, or state of residence) and health care utilization data, including diagnostic codes and pharmacy records. They can be used for research without specific consent because the databases are de-identified. Identifying persons with MS through pharmacy records is an excellent case ascertainment strategy given that many of the DMTs are prescribed only for MS, thus, the persons with MS using such medications are highly likely to be true MS cases. However, over 40% of persons with MS in the United States do not take DMTs, necessitating additional case-finding methods. Because clinical details are often not available, such as whether a diagnostic code was entered by a neurologist, physician, or nonphysician, most claims database studies have relied on case-finding algorithms that have not been judged against case definitions or even neurologist diagnoses for accuracy. Two studies used a case definition that required only a single health care encounter for MS,^12,15 which is likely to be highly nonspecific, risking an unknown number of false-positives. The study by Dilokthornsakul et al.¹⁶ that used private insurance claims data was strengthened by conducting sensitivity analyses that examined the performance of 2 alternative algorithms, one of which produced estimates very similar to the primary algorithm.