*Equal contribution. Author names listed alphabetically.
Introduction
Mental Health Awareness and the Need for Mental Health
Professionals
Since 2020, there has been a significant surge in mental health
awareness in the United States, with a growing emphasis on understanding
and addressing emotional well-being. This heightened awareness has led
to an increased demand for mental health professionals, fueled by
reduced stigma surrounding seeking care. To effectively tackle these
issues, it becomes crucial to improve access to mental health
professionals, as doing so can help lower the number of untreated mental
health cases while simultaneously promoting greater access to mental
health treatments.
In this report, we analyze the relationship between access to
mental health care, as measured by the number of available
mental health professionals per capita, and people’s mental
health status, as measured by their self-reported number of
poor mental health days. In doing so, we also focus on two crucial
factors that may impact the relationship.
Impact of Socioeconomic Status
According to Social Psychiatry and Psychiatric Epidemiology,
instances of poor mental health have been associated with socioeconomic
status (Boe et al., 2012). Socioeconomic status is defined as “the
position of an individual or group determined by a combination of social
and economic factors, such as income, amount and kind of education, type
and prestige of occupation, place of residence, and—in some societies or
parts of society—ethnic origin or religious background” (American
Psychology Association, 2010). Studies have reported association between
lower socioeconomic status and increased amounts of negative mental
health events. It is also known that “exposure to stressful life events
is related to mental health problems including depression, anxiety and
substance abuse. The total number of stressors experienced by an
individual may have a direct impact on mental health” (Businelle et al.,
2014).
In this study, we specifically focus on two socioeconomic factors:
median household income and the education level, as
measured by the percentage of college graduates,
aggregated over each county in the U.S. Other socioeconomic factors,
such as median wages, housing prices, racial and gender composition,
unemployment rate, and income inequality, may further impact the mental
health variables; we leave these variables for future work.
Impact of the Affordable Care Act
In 2010, the Affordable Care Act (ACA) was
implemented in the United States. The ACA has three primary goals: “make
affordable health insurance available to more people, expand the
medicaid program, and support innovative medical care delivery methods
designed to lower the costs of health care generally”
(HealthCare.gov).
The most notable changes the ACA made to healthcare was its expansion
of coverage to decrease the amount of the uninsured in the
United States. Its expansion also included additional
mental health services, trying to address the precedent of
siloed mental health care versus physical health care coverage.
According to Baumgartner et al. (2020), “the ACA guarantees access to
mental health services within individual, small-group (fully insured),
and Medicaid expansion plans by mandating that they cover 10 essential
health benefits, including mental health and prescription drugs.”
Furthermore, the ACA requires all plans (including large-group) to cover
preventive services like mental health screenings at no cost. Finally,
insurance companies cannot deny coverage for preexisting medical
conditions, cannot cancel insurance in the event of sickness, and cannot
cap the amount of care received annually or during a patient’s
lifetime.
The ACA has been considered controversial in American politics. While
the implementation of the ACA is at the federal level, individual states
can choose to reject the additional coverage that the ACA offers its
residents. According to Kantarjian et al. (2014), at least 30 million of
the 50 million previously uninsured US citizens were set to gain health
insurance coverage, although those living in states that rejected the
ACA did not ultimately get the coverage.
As of 2023, a total of 12 states have not implemented (either
rejected or not implemented) the ACA. Wisconsin, Wyoming,
Kansas, Tennessee, Texas, Mississippi, Alabama, Georgia, South Carolina,
and Florida rejected the ACA; South Dakota and North Carolina have
adopted but not implemented it (Kaiser Family Foundation, 2023). In our
analysis, we focus on the binary variable of whether each state has
implemented the ACA.
Topic Question and Hypothesis
The key questions and hypotheses in this report are summarized as
follows.
- Main Question
- How does the number of mental health providers affect the number of
self-reported poor mental health days?
- Impact of Socioeconomic Status
- How does socioeconomic status impact the number of poor mental
health days?
- How does socioeconomic status impact the relationship between mental
health providers and the number of poor mental health days?
- Impact of the Affordable Care Act
- Is there a relationship between ACA implementation and access to
mental health care, in terms of both the availability of mental health
providers and whether or not people are insured?
- How does ACA implementation affect the relationship between the
number of mental health providers per population and the number of poor
mental health days?
We answer these questions with a combination of exploratory analysis,
linear modeling, and predictive modeling.
Data and Key Variables
Throughout our analysis, we use the 2023 version of the county
health rankings data for the purpose of assessing the overall health
status of counties across the United States. This dataset collects
various health and socieconomic variables for each county in the United
States and report aggregate statistics per county. This means
that all of our analysis uses data aggregated at the county
level, rather than the individual or state level. The following
is a glossary of the key variables we consider from the dataset.
- Mental Health Variables
- Poor Mental Health Days: the average
number of self-reported poor mental health days, per 30 days,
in each county. This is our primary outcome variable that
assesses the people’s mental health status. While this variable is known
to be a reliable measure of people’s mental health status, its
consistency as a measure of mental health status may be influenced by
individual’s personal and cultural backgrounds. The mean and median poor
mental health days across counties are around 4.8 out of 30 days.
- Mental Health Providers (per 100K
population): the number of mental health providers in a
county, normalized by 100,000 people. This is our primary
predictor variable that measures the availability of mental
health care.
- Socioeconomic Factors
- % College Graduates: the percentage of
college graduates (at least a two-year program) in the county. This is
one of our main measures of socioeconomic status (education).
- Median Household Income: the median
household income in the county. This is one of our main measures of
socioeconomic status (income).
- ACA-Related Factors
- ACA Implementation: whether the county
belongs to one of the 38 states that implemented the ACA.
- % Uninsured: the percentage of people
without insurance in the county.
Exploratory Data Analysis
Each of our exploratory data analysis (EDA) result corresponds to a
question.
Main Question
How does the number of mental health providers affect the number of
self-reported poor mental health days?
We show the relationship between poor mental health days versus the
number of mental health providers as a scatterplot, along with a linear
fit. Each point in the plot represents one of 3,082 counties in the
U.S.
The plot demonstrates an inversely proportional relationship
between the amount of health providers per capita and the amount of
self-recorded poor mental health days. This is expected, as better
access to mental health care should ideally decrease poor mental health
days, although the relationship also appears quite weak.
Impact of Socioeconomic Status
How does socioeconomic status impact the number of poor mental
health days?
We show a county-level representation of the relationship between
poor mental health days and the percentage of college graduates. As
before, each point in the plot represents a county in the U.S. The color
of each point further indicates the percentage of current college
graduates.
We see that there is a clear inversely proportional relationship
between the two variables. As the mean education level increases, we
expect people to experience less number of poor mental health days
experiences, and the plot confirms this hypothesis. One explanation for
this is that the education level correlates with many factors related to
the quality of living (income; access to a stable job, housing, and
healthcare; and more). This trend is expected to be similar if we
substitute the education level with the median household
income.
How does socioeconomic status impact the relationship between mental
health providers and the number of poor mental health days?
In this plot, we redraw the relationship between poor mental health
days and the number of mental health providers at the county level, but
we additionally show the percentage of college graduates as the colors
of each point.
There is an observed shift in gradient color representing the
increase in percentage of college graduates as mental health providers
per 100,000 people increases.
Impact of the Affordable Care Act
Is there a relationship between ACA implementation and access to
mental health providers (in select states)?
We plot four histograms on the distribution of mental health
providers per 100K population, across two ACA and two non-ACA states.
ACA states are shown in Blue (New York and Louisiana), while non-ACA
states are shown in Red (Florida and Alabama).
This shows that non-ACA states have more counties that have few
mental health providers per capita.
Is there a relationship between ACA implementation and access to
mental health providers (across all states)?
We now plot histograms on the distribution of mental health providers
per 100K population, across all counties in ACA and non-ACA states.
ACA-implements states are shown in Blue (New York and Louisiana), while
non-implemented states are shown in Red (Florida and Alabama).
This plot shows that non-ACA states as a whole have an overall
lower amount of mental health providers than ACA states.
Is there a relationship between ACA implementation and the
percentage of uninsured people?
Here, we draw violin and box plots displaying the percentage of
uninsured people in ACA vs non-ACA states.
This visualization shows that there are fewer uninsured
individuals in ACA states than non-ACA states. The majority of ACA
states have less than 10% uninsured individuals while non-ACA states are
closer to 15% of its populations being uninsured. The outlier
represented in the non-ACA states is Presidio county (TX), having about
37% of its population being uninsured, possibly due to outside factors
not accounted for in this research.
How does ACA implementation affect the relationship between the
number of mental health providers per population and the number of poor
mental health days (in select states)?
We show the number of mental health providers per 100K population
versus poor mental health days, specifically in the two ACA states (New
York and Louisiana) and non-ACA (Florida and Alabama) states. We include
a linear fit corresponding to mental health professionals versus poor
mental health days in these states.
We see that, in non-ACA states, the number of available mental
health providers per 100,000 people has a clear negative association
with the number of poor mental health days. For ACA states, the
relationship is more nuanced. In New York, the number of poor mental
health days decreases as the number of mental health providers
increases; however, in Louisiana, the number of poor mental health days
seems to remain constant regardless of the number of mental health
providers per population.
How does ACA implementation affect the relationship between the
number of mental health providers per population and the number of poor
mental health days (in all states)?
We now show the number of mental health providers per 100K population
versus poor mental health days, across all counties in ACA and non-ACA
states. As before, we include a linear fit corresponding to mental
health professionals versus poor mental health days in these states.
Again, we see that, in non-ACA states, the number of available
mental health providers per 100,000 people has a clear negative
association with the number of poor mental health days. Having more
mental health providers in non-ACA states may lead to an improved mental
health status of the residents.
For ACA states, somewhat surprisingly, the relationship appears
insignificant. One possible explanation is that a majority of residents
in these states do have some access to mental health care, such that it
is not as crucial of a factor than other quality-of-life
variables.
Linear Modeling
In this section, we formally examine the linear relationships of the
predictors and the outcomes examined in our EDA. We use a simple linear
regression model that makes a rather restrictive assumption that the
underlying relationship is linear. While far from perfect, the model
allows certain inferential conclusions to be made regarding the
relationships.
Basic Linear Model
##
## Call:
## lm(formula = PoorMentalHealthDays ~ LogMentalHealthProviders,
## data = county_health_data_clean)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.83806 -0.40267 0.01791 0.40873 2.13766
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 5.05318 0.05442 92.847 < 2e-16 ***
## LogMentalHealthProviders -0.10514 0.02559 -4.108 4.1e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.6105 on 2892 degrees of freedom
## (188 observations deleted due to missingness)
## Multiple R-squared: 0.005801, Adjusted R-squared: 0.005458
## F-statistic: 16.88 on 1 and 2892 DF, p-value: 4.102e-05
This model examines the relationship between the number of mental
health providers and the average number of recorded poor mental health
days on a county based level. According to the model, the relationship
is statistically significant (p-value < 0.001). The model states that
a 10-fold increase in the number of available mental health care
providers would result in the average number of poor mental health days
recorded to decrease by .105 days. (Note that we use data from 2,892
U.S. counties, after deleting 188 counties due to lack of
data.)
Poor Mental Health Days vs. (log)Mental Health Providers,
controlling for socioeconomic factors
##
## Call:
## lm(formula = PoorMentalHealthDays ~ LogMentalHealthProviders +
## SomeCollege + MedianHouseholdIncome, data = county_health_data_clean)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.82331 -0.31256 0.02943 0.33186 1.53220
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 6.268e+00 5.791e-02 108.236 < 2e-16 ***
## LogMentalHealthProviders 1.565e-01 2.293e-02 6.824 1.08e-11 ***
## SomeCollege -2.019e+00 1.114e-01 -18.131 < 2e-16 ***
## MedianHouseholdIncome -9.740e-06 8.019e-07 -12.146 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.5131 on 2890 degrees of freedom
## (188 observations deleted due to missingness)
## Multiple R-squared: 0.2981, Adjusted R-squared: 0.2974
## F-statistic: 409.2 on 3 and 2890 DF, p-value: < 2.2e-16
This model examines the relationship between the (log) number of
mental health providers and the average number of recorded poor mental
health days on a county, while controlling for the college graduation
percentage and the median household income. The model says that a
10-fold increase in the number of available mental health care providers
would actually result in the average number of poor mental health days
recorded to increase by 1.56 days (p-value < 0.001). This is somewhat
surprising, and it reveals how the relationship may be more complex than
a simple linear fit.
Poor Mental Health Days vs. (log)Mental Health Providers,
controlling for ACA implementation
Here, we use a linear model with an interaction term for ACA
implementation: \[
\mathrm{PoorMentalHealthDays} = \alpha_0 + \alpha_1 \cdot \mathrm{ACA} +
\beta_0 \log_{10}(\mathrm{MentalHealthProviders}) + \beta_1 \cdot
\log_{10}(\mathrm{MentalHealthProviders}) \cdot \mathrm{ACA}.
\]
In the case where \(ACA=0\) (not
implemented), \(\alpha_0\) is the
intercept and \(\beta_0\) is the slope.
* In the case where \(ACA=1\)
(implemented), \(\alpha_0 + \alpha_1\)
is the intercept and \(\beta_0 +
\beta_1\) is the slope.
##
## Call:
## lm(formula = PoorMentalHealthDays ~ LogMentalHealthProviders *
## ACA, data = county_health_data_clean)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.87124 -0.39080 0.02236 0.39040 2.09571
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 5.49504 0.08585 64.009 < 2e-16
## LogMentalHealthProviders -0.36292 0.04414 -8.222 2.99e-16
## ACAImplemented -0.64756 0.11341 -5.710 1.24e-08
## LogMentalHealthProviders:ACAImplemented 0.36395 0.05535 6.576 5.72e-11
##
## (Intercept) ***
## LogMentalHealthProviders ***
## ACAImplemented ***
## LogMentalHealthProviders:ACAImplemented ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.6051 on 2890 degrees of freedom
## (188 observations deleted due to missingness)
## Multiple R-squared: 0.02396, Adjusted R-squared: 0.02295
## F-statistic: 23.65 on 3 and 2890 DF, p-value: 4.038e-15
This model examines the relationship between the number of mental
health providers and the average number of recorded poor mental health
days considering whether or not a state has implemented the ACA. The
model displays an increased need for mental health providers in non-ACA
states, as there is a significant relationship between the two
variables: a 10-fold increase of mental healthcare providers may lead to
a decrease in poor mental health days recorded by 0.363 days. The
interaction term also models how the relationship becomes essentially
uncorrelated for counties in ACA states.
Poor Mental Health Days vs. (log)Mental Health Providers,
controlling for socioeconomic factors and only considering Non-ACA
Implemented States
##
## Call:
## lm(formula = PoorMentalHealthDays ~ LogMentalHealthProviders +
## SomeCollege + MedianHouseholdIncome, data = county_health_data_non_aca)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.71314 -0.26054 0.03652 0.29359 1.16945
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 6.221e+00 7.784e-02 79.924 < 2e-16 ***
## LogMentalHealthProviders -1.128e-01 3.606e-02 -3.129 0.0018 **
## SomeCollege -1.532e+00 1.628e-01 -9.408 < 2e-16 ***
## MedianHouseholdIncome -6.107e-06 1.301e-06 -4.694 3.05e-06 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.432 on 1016 degrees of freedom
## (100 observations deleted due to missingness)
## Multiple R-squared: 0.2666, Adjusted R-squared: 0.2644
## F-statistic: 123.1 on 3 and 1016 DF, p-value: < 2.2e-16
Given that the relationship is significant only with non-ACA
states, we will now further examine the relationship while controlling
for the socioeconomic variables. This model examines the relationship
between the number of mental health providers and the average number of
recorded poor mental health days in non-ACA states only, while
controlling for the college graduation percentage and the median
household income. This model shows that a 10-fold increase on the number
of mental health providers results in a decrease in the number of poor
mental health days recorded by 1.13 days, emphasizing that in non-ACA
states an increase in available mental health care providers may result
in the overall decrease in recorded poor mental health days.
Predictive Modeling via Random Forests
We now fit a random forest model using several predictors in the
dataset. We first remove counties that have any missing entry in any of
the predictors. The following shows the list of predictors used as well
as the number of missing entries for each.
## PoorMentalHealthDays LifeExpectancy
## 1 17
## LongCommuteDrivingAlone DrivingAlonetoWork
## 0 0
## SevereHousingProblems SocialAssociations
## 0 0
## ChildreninSingleParentHouseholds IncomeInequality
## 0 6
## Unemployment SomeCollege
## 0 0
## HighSchoolCompletion MentalHealthProviders
## 0 187
## Uninsured TeenBirths
## 0 129
## SexuallyTransmittedInfections ExcessiveDrinking
## 89 1
## AccesstoExerciseOpportunities FoodEnvironmentIndex
## 50 32
## State PoorPhysicalHealthDays
## 0 1
## ACA MedianHouseholdIncome
## 0 1
After removing the missing rows, we are left with the following
number of counties.
## [1] 2749
##
## Call:
## randomForest(formula = PoorMentalHealthDays ~ LongCommuteDrivingAlone + SevereHousingProblems + SocialAssociations + ChildreninSingleParentHouseholds + IncomeInequality + Unemployment + SomeCollege + MentalHealthProviders + Uninsured + TeenBirths + ExcessiveDrinking + AccesstoExerciseOpportunities + FoodEnvironmentIndex + ACA + MedianHouseholdIncome, data = county_health_data_RF, importance = TRUE, ntree = 500)
## Type of random forest: regression
## Number of trees: 500
## No. of variables tried at each split: 5
##
## Mean of squared residuals: 0.1554119
## % Var explained: 57.34
Note that we did not include predictors capturing the physical health
status. This would increase the % variance explained further. In our
models, we focus on variables that are directly related to mental health
and/or socioeconomic status.
#variable importance
vip(rf_model) + theme_minimal() +
labs(title = "Variable Importance Plot, Random Forest",
y = "Importance (% increase in mean squared error)")
The number of mental health providers per capita has the following
variable importance.
## [1] 22.78494
The variable is not considered more important than some of the other
variables we considered. This makes sense because variables such as
excessive drinking, unemployment, and socioeconomic status can be more
direct predictors of mental health problems than the availability of
mental health providers. Nevertheless, the number of mental health
providers is still useful for accurately predicting the number of poor
mental health days, having a substantial variable importance as measured
by the increase in MSE.
One of the variables that was expected to be seen in this variable
importance plot is the uninsured variable. Being uninsured is likely
correlated with low socioeconomic status; furthermore, we previously saw
that states that did not implement ACA tend to have higher number of
poor mental health days. This suggest that some individuals can be
mentally unhealthy and would like to report their mental health, however
they are uninsured. In such cases, the implementation of more affordable
health insurance may be beneficial. This also suggests that, even though
ACA is not the most important variable considered, it may still be
necessary.
Discussion
Conclusion
- Both our exploratory analysis and our simple linear model analysis
demonstrate that there is a significant relationship between the number
of mental health providers and the average number of poor mental health
days in counties.
- As the number of mental health providers per 100,000 people
increases, the number of poor mental health days decreases.
- We find that counties with higher percentage of college graduates
have more mental heal providers per 100K people and less reported poor
mental health days.
- Counties in states that implemented ACA tend to have a higher number
of mental health providers per 100,000 people and a lower percentage of
uninsured people.
- In states that did not implement ACA, the number of poor mental
health days decreases as the number of mental health providers per 100K
people increases, even after controlling for socioeconomic variables
(college graduate % and median household income).
Limitations
- There were 187 counties that did not input data for the number of
poor mental health providers present, while other counties left multiple
variables uncompleted.
- Race and gender were two variables that we were unable to explore
and examine how they had an effect on the data used.
- Because the number of poor mental health days is self-reported, the
outcome measure may not be as consistent as needed, as the number can be
affected by personal and cultural backgrounds.
- The linear model analyses rely on strong assumptions, and they
should be interpreted with extra caution.
Future Work
- A longitudinal study following up on the impact of access to mental
health care, especially among states that implement alternatives to ACA
(if any), can may further clarify the answer to our main question.
- Having other relevant variables, such as the number of individuals
living in a single household, could also help predict improvements that
should be made to the ACA public policy.
- For socioeconomic factors, the analysis may benefit from see more
socioeconomic variables be included, such as race and gender. These
variables may allow us to give more accurate predictions along with
being able to build and explore other models.