Introduction

Telemedicine has varying definitions, but can be broadly described as healthcare occurring virtually, with a patient in a different geographic location than the physician or healthcare worker. This includes live-video patient exams, remote patient monitoring, and instant communication between healthcare professionals. With the emergence of telemedicine, 35 states have implemented legislation that mandates private payers cover telemedicine at the same level of in-person care, while other states have lagged behind (Center for Connected Health Policy, 2017). By mandating that private payers reimburse healthcare costs involving telemedicine at the same level of in-person care, these states should have a rise in the use of telemedicine.

The innovative healthcare practices involved with telemedicine are seen as potential cost-cutters and tools for an overall more efficient healthcare system, thus many states are continually moving towards mandating coverage (Sameer 2013). A study found that by utilizing telemedicine, the Veterans Health Administration saved an estimated $6,500 per patient compared to patients with no telemedicine participation (American Hospital Association, 2016). With 17% of the current GDP being spent on healthcare in 2014 and an estimated 20% of GDP to be spent on healthcare in 2020 (World Bank Data, 2015), the promise of more efficient healthcare processes such as telemedicine is alluring to policymakers. Six states already have laws pending that mandate private payer reimbursement of telemedicine care (Center for Connected Health Policy, 2017). This makes 2018 an important time to study the impact of telemedicine, as differences in health status of those with and without mandated Private Payer Reimbursement (PPR) can be more easily studied.

There have been limited studies of the resulting health outcomes when telemedicine is used, especially clinical trials that involve experimental controls and treatment (Heinzelmann 2005). This causes telemedicine effectiveness to be difficult to measure. However, with the current data, 56% of published research gathered on telemedicine has determined that telemedicine has a favorable impact on health outcomes (Heinzelmann 2005). The lack of research is most likely due to the fact that telemedicine is a new concept that is continually changing as technology is improved and the difficulty of measuring a distinct impact, since health outcomes are dependent on many factors such as lifestyle and compliance with a medical regiment (Heinzelmann 2005).

The ultimate goal of this study is to determine whether the use of telemedicine rather than conventional healthcare practices improves or worsens health by measuring the impact PPR has on the number of bad health days of insured persons. Differences in certain age and income groups may be present, so this data will be included as well. The data used in this analysis was gathered from the 2016 Behavioral Risk Factor Surveillance System (BRFFS), a survey regarding health risks and conditions of U.S. residents.

Methods

Data

#----------------------------------------------------------------------------------------------#
#Project: Data Challenge 1                                                                      
#Author: Sarah Ziolkowski
#Program Name: 1_dataprep
#Data Used: BRFSS16_DC1.dta
#Created: 1/31/2018
#Last Revised & Notes: 2/07/2018 to keep sampling weightt finalwt. 2/26/2018 to add back in recode as.factor variables

#Reminders: Set working directory.

#Contents: This file converts original data from Stata to R format and recodes variables for analysis
#----------------------------------------------------------------------------------------------#

#----------------------------------------------------------------------------------------------#
#Set Up and load data. Remember that after the first run you can comment out install.packages lines
#----------------------------------------------------------------------------------------------#
#Set working directory
setwd("U:/ECO307/DC1")

#Install packages
#install.packages("foreign")
#install.packages("tidyverse")
#install.packages("stargazer")

#Library packages
library(foreign)
library(tidyverse)

#Import Data from Stata format to R
library(foreign)
BRFSS16 <- read.dta("BRFSS2016_DC1.dta")

#Look at your data
#View(BRFSS16)

#-----------------------------------------------------------------------------------------------#
#Rename variables that start with _ because R hates them. My example here with _state will work for any other
#variables that start with _ too. 
#-----------------------------------------------------------------------------------------------#
names(BRFSS16)[names(BRFSS16) == "_state"] <- "fips"
names(BRFSS16)[names(BRFSS16) == "_ageg5yr"] <- "ageg5yr"
names(BRFSS16)[names(BRFSS16) == "_finalwt"] <- "finalwt"
names(BRFSS16)[names(BRFSS16) == "_finalwt"] <- "finalwt"
#-----------------------------------------------------------------------------------------------#
#Select the Variables Relevant for your Analysis. Don't worry, you can always add more later.
#I chose to put mine in a new dataframe called "ltcdata" ltc stands for long term care.
#-----------------------------------------------------------------------------------------------#
tmdata <- select(BRFSS16, fips, physhlth, hlthpln1, ageg5yr, income2, finalwt)
#View(tmdata)

#-----------------------------------------------------------------------------------------------#
#Merge in your state policy data
#-----------------------------------------------------------------------------------------------#
#First open it
statepol <- read.csv("PolicyData_Ziolkowski.csv")

#Look at it (Does it look right?)
#View(statepol)

#We need to name the variable that contains the fips code the same as in our BRFSS file
names(statepol)[names(statepol)== "X_state"] <- "fips"

#Merge it with your analytic data set (mine is lctdata - what did you call yours above?)
tmdata_statepol <- left_join(tmdata, statepol, by="fips")

#Remove the other dataframes from your workspace to conserve memory
rm(BRFSS16, tmdata, statepol)

#-----------------------------------------------------------------------------------------------#
#Recode variables and label values for categorical variables
#When we looked at the data we could see there were lots of weird values. For example 88's and 99's
#Use your BRFSS codebook to decide what values those should be recoded to.Then adapt the code below.
#For variables where numbers are used to represent categories, we can prepare the variables as factors.
#-----------------------------------------------------------------------------------------------#
#The recode verb in dplyr is used here for a numeric variable
#First look at the original values
#table(tmdata_statepol$physhlth)


#Recode 88's numeric to numeric for physhlth
tmdata_statepol$physhlth <- recode(tmdata_statepol$physhlth, '88' = 0L)


#Recode 99's and 77's to NA numeric to character for physhlth
tmdata_statepol$physhlth <- as.numeric(recode(as.character(tmdata_statepol$physhlth), "77" = "NA", "99" = "NA"))

#Check to see values 0 to 30 and nothing else. 
#table(tmdata_statepol$physhlth)


#Now hlthpln1
#Recode 2's, 7's, and 9's to character
tmdata_statepol$hlthpln1 <- as.numeric(recode(as.character(tmdata_statepol$hlthpln1), "2" = "0", "7" = "NA", "9" = "NA"))

#Now ageg5yr
#recode 14 to NA. 1-13 are age groups to be interpreted later
tmdata_statepol$ageg5yr <- as.numeric(recode(as.character(tmdata_statepol$ageg5yr), "14" = "NA"))

#Recode ageg5yr to a factor to fit codebook
tmdata_statepol$ageg5yr <- as.factor(tmdata_statepol$ageg5yr)

#Define levels of ageg5yr as in codebook
levels(tmdata_statepol$ageg5yr) <- c("18 to 24",
                                     "25 to 29",
                                     "30 to 34",
                                     "35 to 39",
                                     "40 to 44",
                                     "45 to 49",
                                     "50 to 54",
                                     "55 to 59",
                                     "60 to 64",
                                     "65 to 69",
                                     "70 to 74",
                                     "75 to 80",
                                     "80 and older")

#Now income2
tmdata_statepol$income2 <-as.factor(tmdata_statepol$income2)

#define levels for factor variable income
levels(tmdata_statepol$income2) <- c("Less than $10K", 
                                     "$10K to <$15K", 
                                     "$15K to <$20K", 
                                     "$20K to <$25K", 
                                     "$25K to <$35K",
                                     "$35K to <$50K",
                                     "$50K to <$75K",
                                     "$75K or more",
                                     "Don't Know",
                                     "Refused")

#Now PPR
tmdata_statepol$PPR <- as.factor(tmdata_statepol$PPR)
levels(tmdata_statepol$PPR) <- c("No", "Yes")
                               
#-----------------------------------------------------------------------------------------------#
#Apply your sample selection criteria using the filter verb from DataCamp
#-----------------------------------------------------------------------------------------------#


#filter to only people with insurance coverage and those under 65 (medicare)
tmdata_statepol <- tmdata_statepol %>% filter(hlthpln1 == 1)
tmdata_statepol <- tmdata_statepol %>% filter(ageg5yr == "18 to 24"|ageg5yr == "25 to 29"|ageg5yr == "30 to 34"|ageg5yr == "35 to 39"|ageg5yr == "40 to 44"|ageg5yr == "45 to 49"|ageg5yr == "30 to 34"
                                              |ageg5yr == "35 to 39"|ageg5yr == "40 to 44"|ageg5yr == "45 to 49"|ageg5yr == "50 to 54"|ageg5yr == "55 to 59"|ageg5yr == "60 to 64")


#Drop observations with missing values on any of your variables by replacing my dataframe name with yours
#in the code below.
#Pay attention - if a ton of observations are dropped, you may have a problem here
tmdata_statepol <- tmdata_statepol %>% filter(complete.cases(.))

##################################################################################################
#Analytic sample is prepared. Proceed to 2_eda
##################################################################################################

The Behavioral Risk Factor Surveillance System and the Center for Connected Health Policy (CCHP) were two resources used to gather my data. States with and without PPR legislation were recorded with data collected by the CCHP and used as my key independent variable. PPR is coded as a binary variable: 0 is attributed to states without the policy and 1 is attributed to states with the policy.

Figure 1 States Participating in PPR

Figure 1 States Participating in PPR

The physical health outcome, PHYSHLTH, and income and age variance of a random, national sample was provided by BRFSS data. BRFFS is a nation-wide telephone survey that collects data on health conditions and risks of U.S. residents, most recently done in 2016. The dependent variable, PHYSHLTH, was measured with the survey question: “Now thinking about your physical health, which includes physical illness and injury, for how many days during the past 30 days was your physical health not good?” Larger values to this question indicate worse health outcomes. Income ranged from below $10,000 per year to income above $75,000 per year. Age intervals ranged from 18-24 years old to 80 years old and above.

The BRFFS also collects data on insurance coverage of individuals, so I filtered out any respondents without insurance and anyone over the age 65 since people over the age of 65 are most likely on Medicare and therefore aren’t affected by private payer policies. This should create a regression with data that is better able to measure the impact PPR has on health. Those without insurance would not be affected by a policy which mandates private payers cover telemedicine at the same level of conventional care. After this filter was applied, 261,973 observations remained. Unweighted statistics of a comparison of PPR and average physical health are shown in Table 1 below. Before conducting any regressions, the average days of poor health are almost exactly the same for PPR and non-PPR states. The regressions explained below will attempt to control for different variables that could be causing the actual of PPR to be hidden, but for now, the effect of PPR is neglible.

tmdata_statepol %>% 
    group_by(PPR) %>%
      summarize(Average.physhlth=round(weighted.mean(physhlth, finalwt),3))
PPR Average.physhlth
No 3.580
Yes 3.578

Empirical Strategy

The regression models used to estimate the effect of PPR on the health of participating and non-participating states are shown below. Model 1 is a simple regression that compares PPR to physhlth through Ordinary Least Squares (OLS) and weighted least squares. The BRFFS data is weighted to account for any sampling discrepancies that could alter my results.

Model 1 Simple Regression

\[physhlth_{i} = \beta_{0} + \beta_{1}PPR_{i}+\epsilon_{i}\]

The model 1 regression equation uses physhlth as the dependent variable. The binary variable, PPR, is the independent variable. I would expect \(\beta_{1}\) to be a negative number since a state with PPR should see a decrease in the number of poor health days per month. This model excludes numerous factors that could alter the physical health outcome between states such as pre-existing health conditions, exposure to pollution, diet, and in particular, age and income. The model 2 regression below addresses this issue and controls for age and income to hopefully get a better estimate of PPR’s impact.

Model 2 Multiple Regresssion Controlling for Income and Age

\[physhlth_{i} = \beta_{0} + \beta_{1}PPR_{i} + age_{i}\theta_1 + income_{i}\theta_2 + \epsilon_{i}\]

To include the factors income and age, model 2 adds them as independent variables. Income and age are recorded as a scale variables with ten and fourteen categories, respectively. The omitted categories for age are 18-24 and for income are below 10K. By controlling for age and income, \(\beta_{1}\) should better reflect the impact PPR has on health. As age increases the number of days of poor health increases. States with more of an older population may also be more likely to implement PPR since healthcare legislation would be a priority for these states. If age was excluded from the regression equation, there would be a positive bias on the PPR coefficient and it would show PPR to be more likely to increase days of poor health.

The second variable added to the simple regression equation is income. It is known that as income increases, days of poor health is estimated to decrease. Additionally, states with higher incomes may be less likely participate in PPR. This would cause PPR to be positively biased, which would also show PPR to be more likely to increase days of poor health if income were excluded.

Model 3 Alternative Functional Form Regression with PPR and Age Interaction

\[physhlth_{i} = \beta_{0} + \beta_{1}PPR_{i} + age_{i}\theta_{1} + income_{i}\theta_{2} + PPR_{i}*age_{i}\theta_{3} + \epsilon_{i}\]

To further account for the effect of age on PPR and physhlth, it is added as an interaction term with PPR in the model 3 regression. This interaction term allows for the difference between certain age groups with PPR and certain age groups without PPR to be measured. In other words, the effect of PPR can change depending on age.

Results

The results of the study are summarized well in the graph below. There is no estimated impact of PPR on health, even when considering the differences in age and resulting PPR effects. Poor health days increases as age increases, as was predicted, however, the difference between those in a PPR participating state and a non-PPR participating state are too small to be significant. The unit of measurement is in days, so for a real impact to be present, there should be a difference of one or more whole days for PPR and non-PPR states .

#----------------------------------------------------------------------------------------------#
#Now create a scatter plot showing changes in physhlth by ageg5yr in PPR and non PPR states
#----------------------------------------------------------------------------------------------#
#First prep the plot data
plot_data2 <- tmdata_statepol %>%
  group_by(PPR, ageg5yr) %>%
  summarize(Average.physhlth=weighted.mean(physhlth, finalwt))

#----------------------------------------------------------------------------------------------#
#Create a scatter plot showing changes in physhlth by income2 in PPR and non PPR states
#----------------------------------------------------------------------------------------------#
ggplot(plot_data2, aes(x = ageg5yr, y = Average.physhlth, color = as.factor(PPR))) + geom_point() +
#Format Graph
  labs(x="Age", y="Number of Bad Health Days", title = "States With PPR vs. States Without: Average Number of Bad Health Days by Age") +
  theme_minimal() +
  theme(axis.text.x=element_text(angle=90)) +
  scale_color_discrete(name="PPR")

The simple regression results are listed in Table 2. The estimates for OLS and weighted least squares have a large enough difference to indicate that the final weights method should be used for the rest of the regression models. The small, negative value given to the PPR variable in the weighted least squares method indicates that states with PPR are estimated to have .002 fewer days of poor health in a month. This estimated effect is so small that there seems to be no actual impact of PPR on health from this estimate.

model1_ols <- lm(physhlth ~ PPR, data = tmdata_statepol)
model1_wt <- lm(physhlth ~ PPR, data = tmdata_statepol, weights = finalwt)
stargazer(model1_ols, model1_wt, header=FALSE, title = "Table 2 Results for Model 1 Simple Regression", column.labels = c("OLS", "Weighted"), covariate.labels = c("PPR"), type = "html")
Table 2 Results for Model 1 Simple Regression
Dependent variable:
physhlth
OLS Weighted
(1) (2)
PPR -0.091** -0.002
(0.036) (0.035)
Constant 4.052*** 3.580***
(0.030) (0.029)
Observations 261,973 261,973
R2 0.00002 0.000
Adjusted R2 0.00002 -0.00000
Residual Std. Error (df = 261971) 8.465 199.561
F Statistic (df = 1; 261971) 6.316** 0.002
Note: p<0.1; p<0.05; p<0.01

Table 3, on the other hand, includes the formatted results of the multiple regressions, Model 2 and 3 listed above. By controlling for age and income in the multiple regression, the estimate of PPR’s impact on health increases, but in the opposite direction desired. The 0.022 coefficient for PPR in Model 2 is saying that those in PPR states have around 0.022 more days of poor health. If income and age had positive biases if left out of the regression equation, the coefficient for PPR should have decreased, instead it increased to become a positive number. So this regression could be showing that states with PPR have more days of poor health when controlling for age and income, except the coefficient is incredibly small and statistically insignificant.

Column 2 of table 3 shows Model 3, which includes an interaction term between PPR and age. PPR’s coefficient now represents PPR’s effect on bad health days for those 18 to 24 years old and income below $10K. The value of the coefficient is a negative value as predicted, however it is still too small to indicate any real effect PPR could have on this category of the population. 0.08 less days of poor health would not indicate an improved health outcome. The coefficient for PPR is also not statistically significant in Model 3 regression results. With the interaction term between age and PPR, the difference between certain age groups with and without PPR can be estimated. At first glance it looks like PPR may actually have an impact on those in a PPR state in the 25 to 29 age category with a -0.128 coefficient, and with the PPR coefficient this turns into 0.210 fewer days of poor health for this category. However, this is still too small to indicate a true effect and it is also statistically insignificant, as well as the rest of the interaction coefficients, except for the 30 to 34 age group with a small, positive coefficient. Altogether, model 3 is not able to show any effect PPR may have on physical health.

model2 <- lm(physhlth ~ PPR + ageg5yr + income2, data = tmdata_statepol, weights = finalwt)
model3 <- lm(physhlth ~ PPR + ageg5yr + income2 + PPR*ageg5yr, data = tmdata_statepol, weights = finalwt)
stargazer(model2, model3, column.labels = c("Model 2", "Model 3"), covariate.labels = c("PPR", "Age 25 to 29", "Age 30 to 34", "Age 35 to 39", "Age 40 to 44", "Age 45 to 49", "Age 50 to 54", " Age 55 to 59", "Age 60 to 64", "Income $10K to less than $15K", "$15K to less than $20K", "$20K to less than $25K", "$25K to less than $35K", "$35K to less than $50K", "$50K to less than $75K", "$75K or more", "Don't Know Income", "Refused to Provide Income", "PPR * Age 25 to 29", "PPR * Age 30 to 34", "PPR * Age 35 to 39", "PPR * Age 40 to 44", "PPR * Age 45 to 49", "PPR * Age 50 to 54", "PPR * Age 55 to 59", "PPR * Age 60 to 64"), header=FALSE, title = "Table 3 Results for Model 2 and 3 Multiple Regressions", type = "html")
Table 3 Results for Model 2 and 3 Multiple Regressions
Dependent variable:
physhlth
Model 2 Model 3
(1) (2)
PPR 0.022 -0.082
(0.033) (0.084)
Age 25 to 29 0.691*** 0.786***
(0.061) (0.116)
Age 30 to 34 1.405*** 1.082***
(0.059) (0.111)
Age 35 to 39 1.902*** 2.001***
(0.062) (0.114)
Age 40 to 44 2.415*** 2.318***
(0.062) (0.115)
Age 45 to 49 3.019*** 2.787***
(0.063) (0.116)
Age 50 to 54 3.551*** 3.368***
(0.059) (0.109)
Age 55 to 59 3.994*** 4.021***
(0.060) (0.111)
Age 60 to 64 4.194*** 4.128***
(0.059) (0.109)
Income 10K to less than 15K -0.117 -0.115
(0.108) (0.108)
15K to less than 20K -1.918*** -1.917***
(0.097) (0.097)
20K to less than 25K -3.332*** -3.331***
(0.093) (0.093)
25K to less than 35K -4.211*** -4.212***
(0.090) (0.090)
35K to less than 50K -5.170*** -5.170***
(0.085) (0.085)
50K to less than 75K -5.926*** -5.926***
(0.083) (0.083)
75K or more -6.811*** -6.810***
(0.077) (0.077)
Don’t Know Income -3.042*** -3.041***
(0.092) (0.092)
Refused to Provide Income -5.895*** -5.895***
(0.094) (0.094)
PPR * Age 25 to 29 -0.128
(0.136)
PPR * Age 30 to 34 0.445***
(0.130)
PPR * Age 35 to 39 -0.137
(0.134)
PPR * Age 40 to 44 0.136
(0.135)
PPR * Age 45 to 49 0.324**
(0.136)
PPR * Age 50 to 54 0.255**
(0.127)
PPR * Age 55 to 59 -0.039
(0.130)
PPR * Age 60 to 64 0.091
(0.128)
Constant 6.280*** 6.354***
(0.083) (0.099)
Observations 261,973 261,973
R2 0.088 0.088
Adjusted R2 0.088 0.088
Residual Std. Error 190.625 (df = 261954) 190.616 (df = 261946)
F Statistic 1,397.430*** (df = 18; 261954) 968.778*** (df = 26; 261946)
Note: p<0.1; p<0.05; p<0.01

Conclusion

This study focuses on the impact that the policy PPR has on health. A quantifiable result could help reveal the effect of using telemedicine over conventional healthcare on health to policy makers as they decide whether to implement legislation that mandates private payer coverage of telemedicine. Telemedicine includes various healthcare practices but is overall viewed as an innovative type of care that could become the future of healthcare. Prior studies on this topic are limited, but on average, have come to positive or neutral conclusions. A study on the effects of telemedicine care used on asthma patients shows that those with the treatment (utilizing telemedicine-based care such as video-exams) provided the same health outcome compared to traditional healthcare practices like in-person physical exams (Zhao, 2015). This specific study is not necessarily an important finding, but it acknowledges the fact that telemedicine is so heterogeneous that measuring its impact is difficult, especially when the many factors that can affect health are considered.

This study found no relationship between states that participate in PPR and an improved or worsened health outcome. Furthermore, a difference is not found for PPR participants and non-participants of different age groups. Altogether, PPR is estimated to have no effect on physical health. This could be because of the issues mentioned above regarding the many forms of telemedicine, or because of limiting factors that are specific to this study. This could include the fact that the data used for this study is cross sectional, and does not revisit the sample to survey them again to measure a before-and-after impact. This study also lacks the accuracy of a clinical experiment with control and treatment groups. Even though this study found no evidence of PPR having an impact on health, further research should still be a priority for this new type of healthcare. Many legislators are executing new policies regarding telemedicine insurance coverage, specific telemedicine licensure, online prescription use, and more, in hopes of increased telemedicine use as healthcare costs rise. While cost effectiveness is crucial, quality of care cannot be ignored and should continue to be studied.

Works Cited

Greenwald, Peter, Michael Ethan Stern, Sunday Clark, and Rahul Sharma. “Older Adults and Technology: In Telehealth, They May Not Be Who You Think They Are.” International Journal of Emergency Medicine 11, no. 1 (December 2018). https://doi.org/10.1186/s12245-017-0162-7.

“Health Expenditure, Total (% of GDP).” Chart. 2015. World Health Organization Global Health Expenditure Database

Heinzelmann, Paul J., Christy M. Williams, Nancy E. Lugn, and Joseph C. Kvedar. “Clinical Outcomes Associated with Telemedicine/Telehealth.” Telemedicine and E-Health 11, no. 3 (June 2005): 329-47. https://doi.org/10.1089/tmj.2005.11.329.

Sameer, Kumar, and Blair John T. “U.S. Healthcare Fix: Leveraging the Lessons from the Food Supply Chain.” Technology and Health Care, no. 2 (2013): 125-41. https://doi.org/10.3233/THC-130715.

“State Telehealth Law and Reimbursement Policies Report” Center for Connected Health Policy, November 6, 2017, http://www.cchpca.org/state-laws-and-reimbursement-policies

“Telehealth: Helping Hospitals Deliver Cost-Effective Care” American Hospital Association, April 22, 2016, https://www.aha.org/system/files/content/16/16telehealthissuebrief.pdf

Zhao, Jie, Yun-kai Zhai, Wei-jun Zhu, and Dong-xu Sun. “Effectiveness of Telemedicine for Controlling Asthma Symptoms: A Systematic Review and Meta-Analysis.” Telemedicine and E-Health 21, no. 6 (June 2015): 484-92. https://doi.org/10.1089/tmj.2014.0119.