Determinants of specialist physician ambulatory visits: a neurology example

Abstract

Background: Economic theory argues that specialization in medicine improves efficiency. Current literature suggests that access to and utilization of specialist care vary widely based on many determinants. Thus, understanding the determinants of specialist physician ambulatory care utilization is integral to healthcare policy.

Objectives: The objective is to investigate the individual and community determinants of specialist ambulatory care utilization—specifically neurologists. The aim was to find predictors of specialist utilization and to identify the particular determinants that can be modified by regulatory or legislative action.

Methods: A large claims database, Truven Health Analytics™ Marketscan data, was used from 2007–2010 as the sample. These data are supplemented with data from the American Academy of Neurology (for geographic distribution of neurologists) and the US Census American FactFinder (for community demographic factors). Multivariate regression analysis was run to test the hypotheses. Several robustness tests of our models were included.

Results: Most importantly, neurologists per capita has a meaningful impact on utilization. Additionally, the difference in neurologist usage by neurological condition is an important factor. It was also found that union status, age, comorbidities, and diagnosis are significant individual level determinants, and that the percentage of Hispanic residents and median income are significant community level determinants.

Conclusions: There are two predictors believed to be the most important. The first is the unique neurologists per 1,000 capita variable, which shows a small increase in the number of neurologists would be correlated with a small increase in the probability of seeing a neurologist. We suggest that this is within policymakers’ control, and policymakers should consider this action in the face of the predicted shortage. The second is what appears to be possible sorting by neurologists of patients based on diagnosis – the large difference in the fraction of patients seeing a neurologist by disease.

Keywords:

• Specialist physicians
• determinants of medical care utilization
• neurology
• ambulatory care

JEL CLASSIFICATION CODES:

• I1
• I12
• I14

Background and theory

Background

This paper provides insight into the determinants of specialist physician ambulatory visit utilization¹ in the US. In this study we chose neurology as our example specialty, and four commonly encountered neurological conditions: (1) dementia, (2) epilepsy, (3) multiple sclerosis (MS), and (4) Parkinson’s disease (PD).

We chose neurology because neurology represents specialized knowledge with relatively few unique procedures or therapies. The function of specialization in neurology is acquisition of knowledge on the human nervous system function and pathology, as any generalist medical practitioner would be able to prescribe medications or order tests that neurologists typically use (e.g. magnetic resonance imaging of the brain). This would be in contrast to orthopedic surgeon utilization for knee repair, where the surgeon utilization is the only way to affect the procedure (knee repair). We argue that patients should seek specialist care (more medical specialization) if they believe specialist care will increase their utility².

Our study follows several limited studies on the determinants of neurologist utilization for particular disease states. Willis et al.¹ found male gender and white race more often led to neurologist care in a Medicare cohort with Parkinson’s disease. Minden et al.² noted access to neurologists among multiple sclerosis sufferers was diminished in rural areas, African Americans, and patients with mobility issues. Mattsson et al.³ equated urban living and higher parental income with increased use of child neurologists in Swedish pediatric epilepsy patients. Our study adds to the existing literature on the determinants of neurologist utilization by focusing on office visit utilization and by employing unique data on the availability of US neurologists.

There are numerous studies on individual and community level determinants of utilization for specific diagnostic, therapeutic, and communication interventions in healthcare. Demographic predictors of cancer screening have been investigated⁴. Braun et al.⁵ looked at demographic predictors of gambling treatment. Ketterer et al.⁶ examined the determinants of patient portal enrollment in primary care pediatrics. Studies on the use of ambulatory care for specialist physicians are rare in comparison.

If specialization in medicine improves economic efficiency,³ then understanding the determinants of specialist physician ambulatory care utilization is integral to healthcare policy. This research aims not only to find predictors of utilization generally, but to identify the particular determinants that can be modified by regulatory or legislative action. We further demonstrate methodology in evaluation of determinants of ambulatory care use that could be applied to non-neurologist specialty medical disciplines.

Some basic economics of medical specialization

Why does specialist utilization matter? Smith⁷ argued that the division of labor and specialization leads to greater productivity—a more efficient production process. He also argued that specialization leads to greater skill in the area of specialization⁷. However, there are limits to specialization based on the market^7,8. If neurologists improve efficiency, have greater skill in their area of specialization, and are not bound by economies of scale or scope in the market for healthcare, then patients will utilize neurologists.

Each patient has a health production function⁹: $$H=f(M,\mathrm{\hspace{0.17em}}X)$$

where $H$ is health or health capital, $M$ is medical care, and $X$ represents all of the other goods, services, and activities in which a person can engage to produce health—also, where $\frac{\delta H}{\delta M}>0$ and $\frac{{\delta }^{2}H}{\delta M^2}<0.$ Once a patient receives a specific diagnosis, the patient may wish to seek specialized care in that disease area in order to produce health capital more efficiently: $$H=f(M_s,M_p,\mathrm{\hspace{0.17em}}X)$$

where $H$ is health or health capital, and $X$ represents all of the other goods, services, and activities a person can engage in to produce health, but M has been divided into to two parts: (1) $M_s$ or specialist medical care and (2) $M_p$ or primary care physician (PCP) medical care. The patient may believe this division of labor will lead to the more efficient production of health. Thus, the patient seeks specialist care.

Smith’s⁷ observations about specialization and the division of labor may not be directly applicable to health and healthcare, as they came from a pin factory. Much of the benefit of specialization, in the pin factory context, was a function of the homogeneity of the inputs and the output. While “healthcare factories” can and do exist (e.g. Mayo Clinics and Cleveland Clinics), each patient is different at admission and discharge, and not only are the treatments different for each patient, but the goals of the treatment process are different for each patient. Specialization in the practice of medicine is more complex.

Following from the theory, the literature is rich with discussions on specialist impact on costs and outcomes. Baicker and Chandra¹⁰ looked at specialists within Medicare and found increased costs and no difference in outcomes, but did provide ample evidence that patients seek and use specialist care. In 2012, the Journal of the American Medical Association featured a series of letters on specialization that discuss whether the US healthcare system has become too specialized and costs have increased without a clear improvement in outcomes¹¹, but provide an argument that specialization is beneficial to patients if well-coordinated¹².

Specifically in neurology, benefits of specialization have been demonstrated. Minden et al.² show that neurologist care increases the probability with which MS patients receive disease-modifying drugs—which should improve outcomes. Willis et al.¹ show neurologist care reduces the probability of Parkinson’s disese patients being placed in a skilled nursing facility. Aspinal et al.¹³ demonstrate that specialization occuring in non-physicians, neurology nursing in this case, improves continuity of care. Ney et al.¹⁴ found that patients who saw neurologists were more likely to be placed on immunologic therapies for MS, dopaminergic medications for Parkinsons, and anticoagulation after atrial fibrillation-associated stroke. Between the theory and empirical evidence, it is clear why patients seek specialist care.

Given that healthcare specialists improve the efficiency of the production of health (or health outcomes), the goal of this research is to investigate the determinants of specialist utilization, and in our study specifically neurologist office visits, as greater than 70% of care by neurologists is administered in the ambulatory care setting (Nuwer et al.¹⁵). Dall et al.¹⁶ calculate that a full-time neurologist will have 2,840 office visits in a year. We hypothesize that the number of neurologists in a given geography will be positively correlated with the number of neurologist ambulatory openings, visits, possible appointments, etc., which in turn would be positively correlated with a patient having an ambulatory visit. We also hypothesize that the diagnosis (dementia, epilepsy, MS, or PD) will be an important determining factor in neurologist ambulatory care visit utilization. For example, dementia is a relatively common neurological condition, and internists and geriatricians are likely to regularly see patients with dementia. Therefore, specialized care may be less of a priority for the patient; the patient feels the care from the PCP is sufficiently efficient. Parkinson’s disease (PD), on the other hand, is a less common condition, and patients may feel specialized care will measurably improve the production of health.

Data

Our dataset comes from three sources: (1) the micro-level data from Truven Health Analytics MarketScan (THMS) database 2007–2010¹⁷, (2) the 3-digit ZIP code area characteristics from the US Census American FactFinder (AFF)¹⁸, and (3) the number of neurologists per 3-digit ZIP code come from the American Academy of Neurology (AAN)¹⁹. The AAN membership data is confirmed with the AMA Masterfile to represent 95% of practicing neurologists in the US. Summary statistics for individual level variables and the 3-digit ZIP code characteristics including neurologists per 1,000 capita and neurologists per square mile stratified by disease of interest are shown in Table 1. The same information, stratified by physician (the dependent variable), is displayed in Table 2.

Table 1. Summary statistics by disease.

	DEM	EPI	MS	PD
Neurologist	19%	51%	59%	57%
Neuro/square mile	0.21	0.22	0.21	0.21
	(1.28)	(1.38)	(1.36)	(1.36)
Neurologist/1,000 capita	0.04	0.04	0.04	0.04
	(0.04)	(0.04)	(0.04)	(0.04)
Union	24%	11%	10%	20%
Female	62%	56%	76%	42%
Age	79.43	45.93	48.37	71.04
	(12.59)	(17.12)	(12.37)	(12.58)
Charlson	1.47	0.60	0.44	0.88
	(1.45)	(1.11)	(0.91)	(1.24)
Medicare	87%	13%	8%	67%
White	75%	74%	75%	76%
African-America	13%	13%	13%	12%
Hispanic	10%	11%	10%	11%
HS or more	86%	86%	86%	86%
Median Income ($10,000)	5.38	5.45	5.58	5.44
Married	50%	50%	50%	50%

Table 2. Summary statistics by physician group

Variables	No Neuro	Neuro	Neither	PCP only	N-only	Both
Neurologist	0%	100%	0%	0%	100%	100%
Neurologist/square mile	0.19	0.23	0.18	0.19	0.29	0.22
	(1.19)	(1.46)	(1.13)	(1.22)	(1.69)	(1.41)
Neurologist/1,000 capita	0.04	0.04	0.04	0.04	0.05	0.04
	(0.04)	(0.04)	(0.04)	(0.04)	(0.04)	(0.04)
Union	20%	14%	11%	24%	10%	15%
Female	59%	59%	58%	60%	56%	59%
Age	67.19	55.55	60.18	70.11	47.74	57.18
	(20.17)	(19.00)	(21.05)	(19.04)	(17.83)	(18.83)
Charlson	1.08	0.83	0.77	1.21	0.35	0.92
	(1.36)	(1.28)	(1.19)	(1.41)	(0.83)	(1.33)
Medicare	57%	32%	41%	64%	16%	35%
White	76%	74%	76%	75%	73%	74%
African-American	13%	13%	12%	13%	13%	13%
Hispanic	10%	11%	10%	11%	12%	11%
HS or more	86%	86%	86%	86%	86%	86%
Median Income ($10,000)	5.33	5.54	5.44	5.27	5.58	5.53
Married	50%	50%	50%	50%	50%	50%
Dementia (DEM)	54%	19%	37%	61%	9%	19%
Epilepsy (EPI)	27%	43%	37%	24%	52%	41%
Multiple Sclerosis (MS)	12%	24%	17%	8%	28%	23%
Parkinson's disease (PD)	7%	13%	8%	6%	11%	14%
Observations	470,438	341,185	143,211	327,227	66,494	274,691

Mean (SD), except where noted. “No Neuro” is a combination of Neither and PCP only. Neuro is a combination of “N-only” and “Both”.

THMS is a large insurance claims database which includes information about diagnoses, ambulatory care visits, patient characteristics (excluding race and ethnicity), and insurance characteristics (for example, insurance from union employment vs commercial insurance vs Medicare). THMS only includes commercially insured and Medicare insured individuals. The AFF includes data on geographic area characteristics including land area, population, percentages of racial and ethnic groups, median income, education level, and marital status. The AFF data in combination with the AAN data determines the number of neurologists per 1,000 capita and the number of neurologists per square mile.

Our sample size is 944,571⁴ person-years across all four diseases—dementia, epilepsy, multiple sclerosis, and Parkinson’s disease. The smallest disease sample size is 89,211 for Parkinson’s disease and the largest disease sample is dementia with 376,579 observations. The inclusion criteria for our sample included that the patient must have one of the disease or condition-related ICD-9 CM codes.⁵ For all diseases the ICD-9 CM code could appear in an inpatient or outpatient setting. The patient had to be an adult between 18 and 100 years old (inclusive). Additionally, the patient had to be uniquely identified according to THMS criteria. Finally, the observations could not have missing geographic information—neither 3-digit ZIP nor region.

After a patient is identified in THMS, all claims matching that patient’s unique identifier are extracted from THMS. Patient data are extracted from the year of diagnosis forward, and any patient with data in more than 1 year has a single year of her/his data randomly selected for inclusion in the final dataset. The dataset is constructed as a repeated cross-section.

To identify ambulatory care visits for the disease of interest, we use the same ICD-9-CM code list for each disease. To identify physician type, we use a field to identify if the physician is a PCP or a specialist. For this study, the only specialty considered is neurology. PCPs include both MDs and DOs in family practice, general internal medicine, pediatrics, primary care, and geriatrics. If the patient has a visit that is coded by both the diagnosis codes and the physician type code, then it is considered either a disease-specific specialist visit or disease-specific PCP visit accordingly.

THMS is also used to calculate the Charlson Comorbidity Index (CCI) for each patient. The CCI has been used extensively for risk-adjustment of health in large observational datasets based on patient comorbid medical conditions (D’Hoore, Bouckaert, and Tilquin²⁰; Austin et al.²¹). It is coded from outpatient diagnosis codes. We used the first two outpatient diagnosis codes for consistency; only two are available for 2007 and 2008. Also, we used outpatient information in order to capture long-term or chronic conditions as opposed to conditions that were diagnosed and treated on an acute basis in an inpatient setting. We do not have within-disease severity so the CCI serves as our control for how sick a patient is—in any dimension of illness.

Because THMS does not include individual level data on race, ethnicity, income, or marital status, we use 3-digit ZIP code level data from the AFF. The AFF provides all of the listed information at the 5-digit ZIP code level and is converted to the 3-digit level using weighted means. The median income data is a weighted median. AFF data is not available for each year. The 2011 5-year average sample data is used; the 3-digit ZIP code level data does not change over time.

The number of neurologists per 1,000 capita and the number of neurologists per square mile is calculated by combining the aggregated AAN membership information and the AFF data. For each 5-digit ZIP, the AFF provides land area and population. That data converted to 3-digit ZIP code information for use in this study. These two variables are used in separate specifications as proxies for neurologist availability. Neurologists per square mile is a proxy for distance or travel time, while neurologists per 1,000 capita is a proxy for the availability of an appointment.

Empirical model and methods

Variable and disease selection

Anderson and Newman¹⁷ provide a framework for the empirical study of the determinants of utilization. This paper provided the framework for our research. The authors divided the individual level determinants into three general groups, each containing sub-groups of influencing factors, and each sub-group having several factors (the groups and sub-groups are shown below).

1. Predisposing

   1. Demographic

   2. Social structure

   3. Beliefs

2. Enabling

   1. Family

   2. Community

3. Illness level

   1. Perceived

   2. Evaluated

We included as many of the factors as were available in our data—see Data section above. We were able to include demographic and social structure variables, but not beliefs, as claims data does not contain any information about beliefs. We included one proxy for family, marital status, and several community variables.

One of the unique aspects of our study is that we have a measure for one of the specific enabling community factors suggested by Anderson and Newman²², “Ratios of health personnel and facilities to population”. We have unique data on the number of neurologists at the community level (in our case “a community” is a 3-digit ZIP code area and our use of term community does not imply any other definition)—see Table 4 for more details. Also, based on current utilization and trends in both the number of neurologists and the population, we expect to see a reduction in the number of neurologist office visits^6,23; there is a projected shortage of neurologists with a reduction in the ratio of neurologists to population, particularly in the population over 65 (Dall et al.¹⁶).

Table 3. Descriptive statistics for 3-digit ZIP codes.

Variable	5th Percentile	Median	95th Percentile	Mean	SD
Land Area (square miles)	41	1,727	9,806	2,931	5,562
Population	68,992	253,093	1,193,455	391,724	389,259
Neurologists	<10	<10	65	17	28
Neurologists/square mile	<0.01	<0.01	23	0.17	1.38
Neurologists/1,000 capita	<0.01	0.03	0.63	0.04	0.04
Observations (# of 3-digit ZIPs)	756
Total Neurologists (2011)	12,391

Additionally, we have individual level data on diagnoses and general health status which represent elements of 3 b above. We cannot know the individual’s perceived health state from our claims database, only the “evaluated” information.

Disease selection

Based on expert consensus⁷, we have selected four neurological conditions to study: (1) dementia, (2) epilepsy, (3) multiple sclerosis (MS), and (4) Parkinson’s disease (PD). These conditions were selected specifically because they represent the wide range of diseases types that neurologists treat in ambulatory patient visits. Based on the characteristics of the diseases, we expect to see differential use of specialist care by disease.

Model

For the purposes of this study, utilization is defined as having at least one disease-specific office visit in a given calendar year; therefore, in our analysis, we employ a logistic regression and multinomial logistic regression: $${\mathrm{Y}}_{\mathit{ijrt}}=\mathrm{\alpha }+{{\mathrm{\beta }}_0{\mathrm{N}}_j+\mathrm{X}}_{\mathit{it}}^{\mathrm{\text{'}}}{\mathrm{\beta }}_1+{\mathrm{Z}}_j^{\mathrm{\text{'}}}{\mathrm{\beta }}_2+{\delta }_r+{\theta }_t+{\epsilon }_{\mathit{ijrt}}$$

In Model 1, our dependent variable is dichotomous; ${\mathrm{Y}}_{\mathit{ijrt}}=1$ if person $i,$ in 3-digit ZIP code area $j,$ in region $r,$ in year $t$ has seen a neurologist, otherwise ${\mathrm{Y}}_{\mathit{ijrt}}=0.$ Where our unique variable, the measurement of neurologist availability, is ${\mathrm{N}}_j,$ ${\mathrm{X}}_{\mathit{it}}$ are the individual level covariates, ${\mathrm{Z}}_j\mathrm{ }$are the 3-digit ZIP code level covariates, ${\delta }_r$ are regional dummies, and ${\theta }_t$ are year dummies.

For the multinomial model, the setup is identical: $${\mathrm{Y}}_{\mathit{ijrt}}=\mathrm{\alpha }+{{\mathrm{\beta }}_0{\mathrm{N}}_j+\mathrm{X}}_{\mathit{it}}^{\mathrm{\text{'}}}{\mathrm{\beta }}_1+{\mathrm{Z}}_j^{\mathrm{\text{'}}}{\mathrm{\beta }}_2+{\delta }_r+{\theta }_t+{\epsilon }_{\mathit{ijrt}}$$

except now ${\mathrm{Y}}_{\mathit{ijrt}}$ can be one of four categories. There are four patient groups determined by the types of physicians utilized in any given calendar year: (1) Neither, indicating the patient has seen neither a PCP or neurologist in that year for the disease of interest (the reference group, $j_0$), (2) Neurologist only, indicating a patient has seen a neurologist without a PCP in that year for the disease of interest ($j_1),$ (3) PCP only, indicating a patient has only seen a PCP without seeing a neurologist in that year for the disease of interest ($j_2$), and (4) Both indicating a patient has only seen a neurologist and a PCP in that year for the disease of interest ($j_3$). All other subscripts are the same as the logistic model. Each model is run 4-times—once for each disease individually. All the regressions are repeated for the second measure of neurologist availability.

Results

Descriptive statistics

Table 1 displays the summary statistics by disease. We found the patterns generally follow our prior beliefs. The oldest average patients are those with dementia, while epilepsy is the youngest group, followed closely by MS. Both the fraction with Medicare and the CCI follow similar patterns as expected.

Table 2 provides an overview of right-hand side variables used in the models. The columns represent the different control and treatment groups used in the analysis.

Table 3 contains the descriptive statistics of our unique neurologist density data. The large differences in the geographic and population distribution of neurologists are exploited in our regression analysis below.

Table 4. Logistic model results for neurologist/square mile.

Variables	DEM	EPI	MS	PD
	Neurologist	Neurologist	Neurologist	Neurologist
Neurologist/square mile	1.004	1.031***	1.037***	1.036***
	[0.008]	[0.008]	[0.008]	[0.010]
Union	0.907***	1.090***	1.143***	0.817***
	[0.033]	[0.034]	[0.043]	[0.038]
Female	0.860***	1.119***	1.074***	0.916***
	[0.008]	[0.009]	[0.014]	[0.014]
Age	1.187***	1.004***	1.056***	1.183***
	[0.005]	[0.002]	[0.003]	[0.006]
Age^2	0.999***	1.000***	0.999***	0.999***
	[0.000]	[0.000]	[0.000]	[0.000]
Charlson	1.078***	1.011	1.016	0.950***
	[0.009]	[0.007]	[0.012]	[0.012]
Charlson^2	0.998	0.998	0.998	1.005*
	[0.001]	[0.001]	[0.003]	[0.003]
Medicare	1.219***	1.198***	0.954	1.149***
	[0.042]	[0.034]	[0.035]	[0.051]
Percentage white	0.999	1.001	1.000	1.001
	[0.003]	[0.004]	[0.005]	[0.004]
Percentage African-American	1.005	1.005	1.004	1.006
	[0.004]	[0.004]	[0.005]	[0.005]
Percentage Hispanic	1.007**	1.008***	1.009***	1.006***
	[0.003]	[0.001]	[0.002]	[0.002]
Percentage HS or more	0.986**	1.010**	1.009	0.993
	[0.006]	[0.005]	[0.007]	[0.006]
Median income ($10,000)	1.152***	1.077***	1.037	1.134***
	[0.028]	[0.024]	[0.030]	[0.030]
Percentage married	1.002	1.002	1.006	0.999
	[0.006]	[0.005]	[0.007]	[0.007]
Observations	376,579	321,388	157,383	89,221

Notes: Odds ratio [SE]; SE clustered on patient 3-digit ZIP in brackets; year and region dummies included in all models.

***p < 0.01,

**p < 0.05,

*p < 0.1.

Logistic model with neurologist/mile²

The first row of Table 4 shows the relationship between neurologists per square mile and the probability a patient sees a neurologist. As the number of neurologists per square mile increases, the probability of seeing a neurologist increases. We see the expected positive relationship, but the magnitude is very small. Because the odds-ratio is interpreted as relative to a one-unit change, and the mean land size for a 3-digit ZIP code is 2,247 miles², it should be interpreted only for sign and significance.

In the second row, the odds ratios (OR) for Union (insurance is provided through union employment) are inconsistent in direction, but significant in each individual disease specification—but not in the all disease specification. The third row, the OR for female, also does not have a consistent sign and is significant in each specification except for the all disease regression. Age and Age^2 (rows 4 and 5) have the expected directions—increasing at a decreasing rate or increasing linearly. The CCI coefficients (row 6) are not as expected. All but one is greater than 1; the coefficient in the Parkinson’s disease regression is less than one. The CCI^2 coefficients are as expected—less than one—except the coefficient for Parkinson’s, which is 1.005 and marginally significant. The coefficient in the all disease regressions is positive, but is of a relatively small magnitude (1.001) and not significant. The coefficients on Medicare (row 8) indicate that generally Medicare patients are more likely to be seeing a neurologist (between 19.9% (PD) and 46.3% (ALL)) with the exception of MS, where the odds-ratio is 0.954 and not significant.

Rows 9–14 show the odds ratios for the 3-digit ZIP code level data. Looking over all of these rows, only two of the 3-digit ZIP code level variables are consistently statistically significant—percentage Hispanic and (weighted average) and Median Income. The odds-ratios for Hispanic are all positive, between 1.005 and 1.009, meaning a 1% increase in the percentage of the Hispanic population in a 3-digit ZIP code area correlates with an increased probability of seeing a neurologist of 0.5% and 0.9%. The median income odds ratios range from 1.037 (not significant) to 1.152.

Logistic model with neurologist/1,000 capita

The first row of Table 5 shows the relationship between neurologists per 1,000 people and the probability a patient sees a neurologist. As the number of neurologists per 1,000 people increases, the probability of seeing a neurologist increases. We see the expected positive relationship and the odds ratios are all relatively large, although two are not significant and three are only marginally significant.

Table 5. Logistic model results for neurologist/1,000 capita.

Variables	DEM	EPI	MS	PD
	Neurologist	Neurologist	Neurologist	Neurologist
Neurologist/1,000 capita	2.356*	1.834	2.674	3.897*
	[1.179]	[0.940]	[1.896]	[3.107]
Union	0.908***	1.089***	1.142***	0.817***
	[0.033]	[0.034]	[0.043]	[0.038]
Female	0.860***	1.119***	1.074***	0.916***
	[0.008]	[0.009]	[0.014]	[0.014]
Age	1.187***	1.005***	1.056***	1.183***
	[0.005]	[0.002]	[0.003]	[0.006]
Age^2	0.999***	1.000***	0.999***	0.999***
	[0.000]	[0.000]	[0.000]	[0.000]
Charlson	1.079***	1.011	1.016	0.951***
	[0.009]	[0.007]	[0.012]	[0.012]
Charlson^2	0.998	0.998	0.999	1.004*
	[0.001]	[0.001]	[0.003]	[0.003]
Medicare	1.219***	1.201***	0.953	1.151***
	[0.042]	[0.034]	[0.035]	[0.051]
Percentage white	0.998	1.001	0.999	1.000
	[0.003]	[0.004]	[0.005]	[0.004]
Percentage African-American	1.005	1.004	1.003	1.005
	[0.004]	[0.004]	[0.005]	[0.005]
Percentage Hispanic	1.007**	1.008***	1.008***	1.005**
	[0.003]	[0.001]	[0.002]	[0.002]
Percentage HS or more	0.985**	1.008*	1.007	0.990
	[0.006]	[0.005]	[0.006]	[0.006]
Median income ($10,000)	1.140***	1.077***	1.034	1.123***
	[0.030]	[0.026]	[0.032]	[0.033]
Percentage married	1.007	1.002	1.007	1.002
	[0.006]	[0.006]	[0.008]	[0.008]
Observations	376,472	321,192	157,284	89,180

Notes: Odds ratio [SE]; SE clustered on patient 3-digit ZIP in brackets; year and region dummies included in all models.

***p < 0.01,

**p < 0.05,

*p < 0.1.

The balance of Table 5 looks almost identical to Table 4. This indicates the results are not sensitive to the choice of the neurologist availability measure we chose.

Multinomial model with neurologist/mile²

Table 6 shows the complete results for the multinomial model with neurologist per square mile. The omitted group is Neither—patients not seeing any physician for the disease of interest. All coefficients are reported as relative risk (RR). As the number of neurologists increases per square mile, the RR of seeing a physician is increased. The geographic density of neurologists is positively related to seeing PCPs as well as neurologists. This is true for each disease.

Table 6. Multinomial logistic model results for neurologist/square mile.

	Dementia			Epilepsy
Variables	PCP only	N-only	Both	PCP only	N-only	Both
Neurologist/square mile	1.060**	1.080***	1.050*	1.063***	1.072***	1.067***
	[0.028]	[0.028]	[0.030]	[0.017]	[0.017]	[0.018]
Union	1.937***	1.639***	1.547***	2.548***	1.774***	2.131***
	[0.142]	[0.160]	[0.124]	[0.141]	[0.105]	[0.136]
Female	1.102***	0.835***	0.938***	1.046***	0.895***	1.232***
	[0.018]	[0.024]	[0.015]	[0.011]	[0.013]	[0.014]
Age	0.998	1.186***	1.185***	1.004*	1.014***	1.007***
	[0.003]	[0.011]	[0.006]	[0.002]	[0.003]	[0.002]
Age^2	1.000***	0.998***	0.999***	1.000***	1.000***	1.000
	[0.000]	[0.000]	[0.000]	[0.000]	[0.000]	[0.000]
Charlson	1.377***	0.825***	1.463***	1.196***	0.661***	1.258***
	[0.021]	[0.024]	[0.024]	[0.015]	[0.012]	[0.014]
Charlson^2	0.969***	1.006	0.969***	0.982***	1.039***	0.975***
	[0.002]	[0.005]	[0.002]	[0.002]	[0.003]	[0.002]
Medicare	1.302***	1.307***	1.508***	1.109*	1.334***	1.283***
	[0.089]	[0.102]	[0.115]	[0.067]	[0.086]	[0.083]
Percentage white	1.005	0.990	1.005	1.002	0.997	1.005
	[0.008]	[0.007]	[0.009]	[0.007]	[0.006]	[0.008]
Percentage African-American	1.009	1.007	1.014	1.009	1.005	1.012
	[0.012]	[0.011]	[0.013]	[0.007]	[0.006]	[0.008]
Percentage Hispanic	1.018***	1.016***	1.023***	1.019***	1.023***	1.020***
	[0.006]	[0.005]	[0.005]	[0.004]	[0.003]	[0.004]
Percentage HS or more	0.999	0.982	0.986	1.021**	1.026***	1.023**
	[0.012]	[0.012]	[0.011]	[0.010]	[0.008]	[0.010]
Median income ($10,000)	1.018	1.192***	1.164**	0.962	1.068*	1.044
	[0.058]	[0.061]	[0.069]	[0.041]	[0.039]	[0.048]
Percentage married	1.012	1.011	1.013	1.017*	1.006	1.015
	[0.013]	[0.012]	[0.014]	[0.010]	[0.009]	[0.011]
Observations	376,579	376,579	376,579	321,388	321,388	321,388

Table

	Multiple sclerosis			Parkinson’s disease
Neurologist/square mile	1.057***	1.077***	1.064***	1.101***	1.109***	1.106***
	[0.015]	[0.015]	[0.016]	[0.031]	[0.030]	[0.033]
Union	2.521***	1.919***	2.050***	2.992***	1.675***	1.797***
	[0.156]	[0.137]	[0.140]	[0.237]	[0.176]	[0.156]
Female	1.026	0.910***	1.145***	1.183***	0.840***	1.050**
	[0.019]	[0.020]	[0.020]	[0.028]	[0.028]	[0.024]
Age	0.978***	1.049***	1.040***	0.931***	1.126***	1.130***
	[0.004]	[0.006]	[0.005]	[0.006]	[0.011]	[0.008]
Age^2	1.000***	0.999***	1.000***	1.001***	0.999***	0.999***
	[0.000]	[0.000]	[0.000]	[0.000]	[0.000]	[0.000]
Charlson	1.270***	0.658***	1.307***	1.248***	0.638***	1.178***
	[0.025]	[0.018]	[0.023]	[0.028]	[0.022]	[0.026]
Charlson^2	0.975***	1.053***	0.969***	0.982***	1.040***	0.985***
	[0.005]	[0.006]	[0.004]	[0.004]	[0.007]	[0.004]
Medicare	1.066	0.953	1.011	1.245***	0.972	1.329***
	[0.073]	[0.069]	[0.071]	[0.095]	[0.079]	[0.109]
Percentage white	1.006	0.995	1.006	1.004	0.992	1.007
	[0.007]	[0.006]	[0.008]	[0.009]	[0.007]	[0.009]
Percentage African-American	1.016**	1.004	1.015*	1.018*	1.014	1.020*
	[0.008]	[0.007]	[0.009]	[0.011]	[0.009]	[0.012]
Percentage Hispanic	1.017***	1.021***	1.018***	1.021***	1.015***	1.022***
	[0.004]	[0.003]	[0.005]	[0.006]	[0.005]	[0.006]
Percentage HS or more	1.013	1.016*	1.018	1.000	0.990	0.995
	[0.010]	[0.010]	[0.012]	[0.013]	[0.012]	[0.014]
Median income ($10,000)	1.005	1.059	1.030	0.984	1.167***	1.110*
	[0.045]	[0.046]	[0.053]	[0.058]	[0.060]	[0.068]
Percentage married	1.015	1.011	1.015	1.031**	1.019	1.019
	[0.010]	[0.010]	[0.011]	[0.013]	[0.013]	[0.014]
Observations	157,383	157,383	157,383	89,221	89,221	89,221

Notes: Relative Risk [SE]; SE clustered on patient 3-digit ZIP in brackets; year and region dummies included in all models. Omitted group—not seeing any physician.

***p < 0.01,

**p < 0.05,

*p < 0.1.

People who work in unions are more likely to be seeing a physician than not. This is true across all the regressions. Females are more likely to be in the PCP only group or the Both group than not seeing a physician, but less likely to be in the Neurologist only group. Again, these results are consistent across all regressions. Age and Age^2 are generally as expected—increasing at a decreasing rate or increasing linearly. The CCI relative risks reveal a unique pattern in these models. The relative risk of only seeing a neurologist is decreasing in CCI. Also, it is decreasing at a decreasing rate. The more comorbidities a patient has the less likely she/he is to be using a neurologist exclusively for care. The Medicare RRs are greater than 1 for dementia and epilepsy, while for MS and Parkinson’s disease the RR for N-only group is less and not significant.

Again, most of the 3-digit ZIP code level variables are not significant in these models. The Hispanic RR is greater than 1 and significant in every model. The RRs range between 1.015 and 1.022. The only other RRs that have any sort of pattern are those for median income. The neurologist only group have RRs greater than 1 in every model, but the MS RR for neurologist is not significant. Generally, an increase in median income is associated with an increased RR of seeing a physician.

Multinomial model with neurologist/1,000 capita

Table 7 displays the results for the multinomial model that includes neurologists per 1,000 capita. The results for the individual level characteristics are almost identical (in some cases identical) to the results for neurologist per square mile model. Union, female, age, CCI, and Medicare are essentially the same.

Table 7. Multinomial logistic model results for neurologist/1,000 capita.

	Dementia			Epilepsy
Variables	PCP only	N-only	Both	PCP only	N-only	Both
Neurologist/1,000 capita	3.633	8.390	6.659	0.862	2.000	1.722
	[6.426]	[13.372]	[11.493]	[0.869]	[1.869]	[1.760]
Union	1.939***	1.641***	1.551***	2.543***	1.769***	2.127***
	[0.142]	[0.161]	[0.124]	[0.141]	[0.105]	[0.136]
Female	1.102***	0.834***	0.938***	1.046***	0.895***	1.232***
	[0.018]	[0.024]	[0.015]	[0.011]	[0.013]	[0.014]
Age	0.998	1.186***	1.185***	1.004*	1.014***	1.007***
	[0.003]	[0.011]	[0.006]	[0.002]	[0.003]	[0.002]
Age^2	1.000***	0.998***	0.999***	1.000***	1.000***	1.000
	[0.000]	[0.000]	[0.000]	[0.000]	[0.000]	[0.000]
Charlson	1.377***	0.825***	1.464***	1.195***	0.661***	1.257***
	[0.021]	[0.024]	[0.024]	[0.015]	[0.012]	[0.014]
Charlson^2	0.969***	1.006	0.969***	0.982***	1.039***	0.975***
	[0.002]	[0.005]	[0.002]	[0.002]	[0.004]	[0.002]
Medicare	1.302***	1.310***	1.508***	1.113*	1.340***	1.289***
	[0.089]	[0.102]	[0.115]	[0.067]	[0.087]	[0.083]
Percentage white	1.004	0.989*	1.004	1.002	0.996	1.004
	[0.008]	[0.007]	[0.009]	[0.007]	[0.006]	[0.008]
Percentage African-American	1.008	1.004	1.013	1.006	1.002	1.009
	[0.012]	[0.010]	[0.012]	[0.007]	[0.006]	[0.008]
Percentage Hispanic	1.018***	1.015***	1.023***	1.018***	1.023***	1.019***
	[0.006]	[0.005]	[0.005]	[0.004]	[0.003]	[0.004]
Percentage HS or more	0.996	0.978*	0.983	1.020**	1.024***	1.021**
	[0.012]	[0.012]	[0.011]	[0.009]	[0.008]	[0.010]
Median income ($10,000)	1.014	1.181***	1.147**	0.978	1.077*	1.053
	[0.063]	[0.067]	[0.076]	[0.044]	[0.043]	[0.052]
Percentage married	1.013	1.013	1.018	1.008	1.000	1.009
	[0.015]	[0.014]	[0.016]	[0.011]	[0.011]	[0.012]
Observations	376,472	376,472	376,472	321,192	321,192	321,192

Table

	Multiple sclerosis			Parkinson's disease
Neurologist/1,000 capita	0.912	2.969	2.519	1.061	1.671	5.111
	[1.090]	[3.320]	[3.103]	[1.870]	[2.849]	[9.158]
Union	2.518***	1.917***	2.048***	2.987***	1.671***	1.796***
	[0.156]	[0.137]	[0.140]	[0.237]	[0.176]	[0.155]
Female	1.025	0.910***	1.144***	1.183***	0.840***	1.051**
	[0.019]	[0.020]	[0.020]	[0.028]	[0.028]	[0.024]
Age	0.978***	1.049***	1.040***	0.930***	1.126***	1.130***
	[0.004]	[0.006]	[0.005]	[0.006]	[0.011]	[0.008]
Age^2	1.000***	0.999***	1.000***	1.001***	0.999***	0.999***
	[0.000]	[0.000]	[0.000]	[0.000]	[0.000]	[0.000]
Charlson	1.269***	0.658***	1.306***	1.249***	0.639***	1.180***
	[0.025]	[0.018]	[0.023]	[0.028]	[0.022]	[0.026]
Charlson^2	0.975***	1.053***	0.970***	0.982***	1.039***	0.985***
	[0.005]	[0.006]	[0.004]	[0.004]	[0.007]	[0.004]
Medicare	1.067	0.953	1.012	1.249***	0.975	1.334***
	[0.073]	[0.069]	[0.071]	[0.095]	[0.079]	[0.109]
Percentage white	1.005	0.993	1.005	1.003	0.991	1.005
	[0.007]	[0.006]	[0.008]	[0.009]	[0.007]	[0.009]
Percentage African-American	1.012	1.001	1.012	1.014	1.009	1.016
	[0.008]	[0.007]	[0.009]	[0.011]	[0.009]	[0.012]
Percentage Hispanic	1.016***	1.020***	1.018***	1.020***	1.014***	1.021***
	[0.004]	[0.004]	[0.005]	[0.006]	[0.005]	[0.006]
Percentage HS or more	1.012	1.013	1.016	0.998	0.987	0.991
	[0.010]	[0.010]	[0.012]	[0.013]	[0.012]	[0.013]
Median income ($10,000)	1.018	1.064	1.034	1.002	1.186***	1.110
	[0.049]	[0.051]	[0.057]	[0.066]	[0.069]	[0.076]
Percentage married	1.008	1.007	1.013	1.020	1.009	1.017
	[0.011]	[0.012]	[0.013]	[0.016]	[0.015]	[0.017]
Observations	157,284	157,284	157,284	152,407	152,407	152,407

Notes: Relative Risk [SE]; SE clustered on patient 3-digit ZIP in brackets; year and region dummies included in all models. Omitted group—not seeing any physician.

***p < 0.01,

**p < 0.05,

*p < 0.1.

The RRs for the 3-digit ZIP code level variables are also basically the same as the previous specification. The percentage white, percentage African-American, the percentage Hispanic, the percentage high school educated or higher, the percentage married, and the median income are all the same patterns observed previously, and in some cases the exact same RRs.

Discussion and conclusion

Variables of interest: neurologist availability

In the logistic regressions, we see that availability of neurologists is correlated with seeing a neurologist. The logistic regressions in Table 4 show that increasing the number of neurologists per square mile is correlated with between a 0.04% and 3.7% increase in the probability of seeing a neurologist, but the number of neurologists needed to affect that marginal change is in the thousands for each 3-digit ZIP code area.

The odds ratios on neurologists per 1,000 capita (Table 5) in the logistic regression are relatively large, but none are statistically significant. Converting the OR to neurologists per 1,000,000 capita, increasing the number of neurologists by one is correlated with a 0.54% increase in the probability of seeing a neurologist.⁸ There are ∼16,000 neurologists in the US, and about 316 million people. Increasing the number of neurologists by ∼2% (320 new neurologists) would be associated with an ∼0.25% increase in the probability of seeing a neurologist for the four neurological conditions analyzed in this paper.⁹

The multinomial RRs tell a similar story as the ORs on the logistic regressions when we look at our measures of neurologist availability. The RRs for neurologists per square mile are between 1.050–1.124% (see Table 7, columns N-only and Both), representing a positive correlation between neurologists per square mile and the probability of seeing a neurologist. However, as before, this requires thousands of new neurologists in order to affect that level of change.

In the multinomial specification, the neurologists per 1,000 capita RRs are not statistically significant. Converting the result to neurologists per 1,000,000 capita, we get an association of almost a 1% change in the probability of seeing a neurologist. Again, the correlated change would require 316 new neurologists.

Union, age, CCI, and Medicare

Union-based insurance is a significant predictor of seeing a neurologist in all models. The OR estimates are consistent between Tables 5 and 6, but across diseases the effect varies. In the multinomial specifications the RRs for Union are always positive. Getting coverage through a union (compared to non-union) always increases the probability of seeing a physician relative to not seeing a physician at all, but is not consistently correlated with seeing a neurologist. This could be related to the utilization of neurologists to establish and evaluate workman’s compensation claims for job-related conditions such as back pain or carpal tunnel syndrome. It is also possible that spouses (dependent coverage) of union members have differential access to neurologists, and those spouses are more likely to be female.

Age and CCI are important drivers of neurologist office visit utilization. The ORs and RRs on age for seeing a neurologist are positive in every specification (all ORs and the RRs on N-only and Both). This is expected.

The CCI ORs are generally increasing at a decreasing rate. People seek more care as they get sicker, but on the margin the investment in health capital (seeing another specialist, in this case) is getting smaller as the level of sickness increases. This story holds for every disease except PD.

The ORs and RRs for PD present an interesting story. As comorbidities increase, patients with PD are less likely to see a neurologist, or they are less likely to be seeing a neurologist exclusively. This relationship is also decreasing at a decreasing rate, so there is a turnaround where, with enough comorbidities, the probability of seeing a neurologist is increasing and will eventually be greater than the probability of not seeing a neurologist. The turnaround happens around a CCI of 5 and becomes positive again around a CCI of 11. A person with a CCI of 5 has a predicted 10-year survival rate of ∼21%²⁴.

The Medicare ORs and RRs are almost all positive and significant. Where the Medicare ORs and RRs are not consistently positive is for MS. MS is a disease with average age of onset of ∼34. We believe distinguishing between private commercial coverage and public coverage is important but, because our public plan primarily covers patients 65 and over, so these ORs and RRs may be equivalent to an age dummy for patients 65 and over.¹⁰

3-digit ZIP code variables

Only two of the community-level variables were consistently significant—percentage Hispanic and median income. While percentage Hispanic is positively correlated with seeing a neurologist, all of the ORs and RRs are relatively small in magnitude. It is unclear what the mechanism is for this result. Median income is consistently positively correlated with the probability of seeing a neurologist, as expected.

Limitations

This study has limitations. Our micro-level data does not have individual race, ethnicity, or income data, all of which are suggested by Anderson and Newman²². Instead we rely on 3-digit ZIP code level data—because that is the smallest geographic region we can identify in our claims data. Also, we do not have within-disease severity, which would also be informative to our analysis. We assume one calendar year is the correct frequency for visits for our diseases of interest. The geographic distribution of neurologist data were unavailable for the same years THMS data were available; we are reasonably confident that neurologist distribution has not changed dramatically over this period of time. Another limitation is that neurologist availability is endogenous in our reduced form equation. Future research will investigate ways to deal with the endogeneity issue, although we are not convinced this particular issue is a major concern in estimating determinants of neurologist ambulatory care; this research is not estimating a demand curve or a demand elasticity.

Conclusions

Despite the above limitations, we are able to provide insight into the determinants of neurologist ambulatory care visit utilization. Individual characteristics, like age and CCI, are generally positively correlated with utilization. Plan characteristics, i.e. union vs public plan, are not as helpful in policymaking, as they differ by disease. Community level characteristics like median income level also have some predictive power, but are generally outside the control of policymakers.

There are two predictors we believe to be the most important. The first is our unique neurologists per 1,000 capita variable, which shows a small increase in the number of neurologists would be correlated with a small increase in the probability of seeing a neurologist. We suggest that this is within policymakers’ control, and policymakers should consider this action in the face of the predicted shortage²³.

It is also possible that redistributing neurologists would increase the probability of seeing a neurologist; there may exist regional shortages, but not a shortage overall. The difference between the highest density and lowest density 3-digit ZIP codes is quite large, so moving neurologists without harming the highest density areas may be possible. One possible downside of redistribution is that areas of high density may be centers of teaching and/or research, and moving neurologists from those activities to the clinic may decrease teaching and/or research output and, in turn, impact quality of care. Another concern is that each market may have a different point at which increasing the density of neurologists yields zero marginal benefit or that the heterogeneity of neurologists may make choosing which neurologist to redistribute impossible. Use and reimbursement of telemedicine is an emerging solution to potentially address the geographic imbalances of neurologist distribution.

Also, our analysis shows that much of the work (ambulatory care visits) in dementia is being done by PCPs. If PCPs had improved training in the confirmation of diagnosis of dementia and initiating treatment for dementia, the number of neurologists may not need to be increased nor redistributed.

In addition to redistributing neurologists, there may be other methods of increasing access such as telemedicine, decreasing the need for face-to-face time with the neurologist as part of a care team, and using advance practice providers to extend the range of one specialist. The second is what appears to be possible sorting by neurologists of patients based on diagnosis—the large difference in the fraction of patients seeing a neurologist by disease—see Tables 1, 4, and 5. If neurologists have already started prioritizing patients based on diagnosis, the predicted shortage may actually be larger than already expected.

Transparency

Declaration of funding

This research was funded in part by the American Academy of Neurology “2013 Value of Neurologic Care” program.

Declaration of financial/other interests

JN is an employee of the federal government; this manuscript is not a term of his employment. He is a consultant for SpecialtyCare, Ceribell, BioVarRXinfusion, and JEM Research institute. He has received honoraria from the American Academy of Neurology, American Society for Neuromonitoring, serves on the HSR1 VA Scientific Merit Review Board, and is on the editorial Board of Neurology Clinical Practice.

DVDG has been a consultant for PRMA Consulting, Fleet UK. LG is a director of VeriTech Consulting, Inc. JME peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Acknowledgements

The authors thank the UNM Department of Economics, the Combined BA/MD program at UNM, and the Robert Wood Johnson Foundation Center for Health Policy at UNM. The authors would also like to thank two anonymous referees for their valuable time and energy.

Notes

Notes

1 Utilization is defined as the use of healthcare services.

2 Utility is the satisfaction received from the consumption of goods and services.

3 Economic efficiency includes, but is not limited to Pareto efficiency, productive efficiency, and allocative efficiency.

4 In the analysis using number of neurologists per 1,000 capita, 499 observations are lost due to a 0 population for one of the 3-digit ZIP codes; the dependent variable cannot be calculated. This represents a loss of less than 0.05% of the sample due to missing data.

5 The ICD-9-CM code lists are available upon request.

6 David N. van der Goes confirmed this potential shortage using different data sources and analysis techniques as part of a presentation at the AAN Annual Meetings in San Diego, 2013.

7 The panel consisted of nine neurologists on the American Academy of Neurology Value of Neurology subcommittee.

8 This number comes from converting the OR of 5.396-times as likely to 539.6% more likely and then moving the decimal place to the left from per 1,000 to per 1,000,000.

9 See previous footnote for calculation.

10 We ran regressions including an age dummy for 65 and over, and the results of our main variables of interest did not change. Also, the Age and Age² results looked almost identical. We also ran the model without either the 65+ age dummy or the Medicare dummy; all of the results were robust to the exclusion of those variables as well.