2 United Healthcare’s Participation Decisions in 2014 and 2015

United Healthcare is the largest health insurance company in the United States by number of covered individuals. Prior to the first year of marketplaces in 2014, United Healthcare had a large number of enrollees in Medicare Advantage program. As the ACA substantially cut federal reimbursements to this program, the company conceded to reporters in 2013 that this change is likely to have an adverse effect on profits. As a result, the company took a conservative approach to the marketplaces in 2014; United Healthcare did not offer any insurance plans in the 34 states with FFMs, while offering plans in only four state-run marketplaces.

In July 2014, while discussing the company’s second-quarter earnings report, chief executive officer Stephen J. Hemsley stated that the company would offer health insurance plans in the marketplaces under the ACA ‘to as many as two dozen’ states in 2015. Accordingly, United Healthcare offered plans in 16 of the 34 FFMs, as well as additional state-run marketplaces. There was a wide range in the scale of United Healthcare’s participation in these 16 FFMs. In Alabama, Georgia and Indiana, United Healthcare offered plans in every rating area. However, it offered plans in only 1 out of 13 rating areas of Illinois. Among the most populated of 16 FFMs, United Healthcare participated in 65 out of 67 rating areas in Florida, 7 out of 26 rating areas in Texas, 5 out of 17 rating areas in Ohio and three out of nine rating areas in Pennsylvania.

United Healthcare’s entry can have two effects on 2LPS premiums. First, the company could itself have offered one of the two lowest-priced silver premiums in a rating area. Second, its participation can lead rival insurance companies to strategically lower premiums in hopes of competing more successfully with United Healthcare. We expect both effects to be more pronounced in areas where United Healthcare offered a plan and where limited competition existed before United Healthcare’s entry. The consequences that result from United Healthcare’s entry into FFMs is due to the competition levels existing prior to entry and how the competitors react afterward. Such changes in the industry’s competitive landscape would create variation in competition across rating areas. One concern is that United Healthcare’s entry may have discouraged other companies from entering. If this is the case, our estimates would be downward-biased. However, due to the long application process to enter the FFMs, we expect this to be a marginal bias.

3 Literature review

Some earlier studies focused on how premiums are set in the health insurance markets. Zweifel and Breuer (Reference Zweifel and Breuer2006) find that risk-independent insurance premiums are not optimal; they advocate risk-based premiums combined with means-tested, tax-financed transfers. Their research supports the need to control for an estimate of the risk pool in determining the marketplace premiums and the omitted variable bias that arises due to the lack of data on this estimate. McGuire (Reference McGuire2006) argues that risk-based premiums would enable the gains from insurer competition to be realized, granted that the information was perfect or at least adequate. In the absence of adequate information, he concludes that risk-based premiums would not be feasible. Herring and Pauly (Reference Herring and Pauly2001) study the relationship between expected medical expense and actual paid premiums. They conclude that premiums vary less than proportionally with expected expense and they are related to only certain risk characteristics.

Before the ACA, the individual insurance market was a small market with limited competition. Therefore, earlier literature on the effects of competition on premiums was constrained by a lack of data. Wholey et al. (Reference Wholey, Feldman and Christianson1995) examine the effects of Health Maintenance Organization (HMO) market structure on premiums from 1988 to 1991. Managed care organizations are designed to increase competition among health care plans, thereby reducing costs and premiums. They find that an increased number of HMOs in the market area is associated with reduced HMO premiums. Robinson (Reference Robinson2004) highlights the dominance of few insurance firms in each state and the associated increase in premiums and profits. He argues that the industry must undergo rejuvenation through the addition of new firms and products, or the market will face increased regulatory oversight from the public sector. Dafny (Reference Dafny2010) analyzes the effects of imperfect competition on premiums and market power of health care insurers. He finds that health care insurers are exercising market power in an increasing number of geographic markets. Due to lack of data availability, none of these studies quantify the effects of competition on premiums. By using data from the first 2 years of ACA’s implementation, we are able to quantify this effect, which is important as increased competition and reduced premiums can help millions of uninsured individuals obtain health insurance.

Two recent papers analyze the premiums in ACA’s marketplaces during 2014 and 2015. Gabel et al. (Reference Gabel, Whitmore, Green, Stromberg, Weinstein and Oran2015) initially use descriptive statistics to demonstrate the changes in premiums from 2014 to 2015 in both FFMs and state-run marketplaces. They find a 1.6% decrease in premiums from 2014 to 2015. In an OLS regression analysis of pooled 2014 and 2015 data, they find that one more insurer is associated with a 2.2% decline in the 2LPS premium. In this analysis, they control for the plan year, whether the state had an FFM or ran its own marketplace, plan type, carrier category, whether the area was urban, suburban or rural, market concentration in the individual insurance market prior to the ACA, number of insurers competing in the rating area, median family income, percentage of the uninsured population, and a state identifier. Jacobs et al. (Reference Jacobs, Banthin and Trachtman2015) expands upon this analysis by removing omitted time-invariant factors with a FD multivariate regression analysis of 2014 and 2015 data from the FFMs. They find that one more insurer is associated with a 1.2% reduction in the 2LPS premium, after controlling for the 2012 US census population estimate, the percentage of population under the age of 18 years and over the age of 64 years, the percentage of rural population, median household income, the percentage of uninsured adults, and the percentage of population with incomes below poverty. Both studies use an OLS method, which would yield biased results in the presence of omitted time-variant factors and reverse causality. We identified three other studies that use instrumental variables to address these concerns.

Dafny et al. (Reference Dafny, Gruber and Ody2012) uses an instrumental variables approach for the 1998 merger between Aetna and Prudential, two large insurers. This merger had differential effects on the level of competition across markets, which helps them identify the most affected markets. They find an average increase in premiums of 7% between 1998 and 2006, following the merger. We focus on more recent data from the ACA. In addition, we do a threshold analysis to investigate how many insurers create enough competition to affect the premiums.

Guardado et al. (Reference Guardado, Emmons and Kane2013) uses another large merger between United Health Group and Sierra Health Services. This merger changed the market concentration in Nevada’s health insurance markets. Using a difference-in-difference approach, they find premiums in most affected Nevada markets increased by 13.7% following the merger. Various factors in Nevada may have affected this increase, aside from a change in market concentration. Therefore, results from a single state may not be generalizable to other states. We include competition and ACA premiums data from 34 states, which represent diverse regions of the United States and the results may provide a more holistic outlook on the effects of competition on premiums.

Dafny et al. (Reference Dafny, Duggan and Ramanarayanan2015) exploits United Healthcare’s decision not to participate in any of the marketplaces of 2014. They find that the 2LPS premium would have decreased by 5.4%, on average, had United Healthcare participated. If all insurers active in 2011 also participated in all rating areas of 2014, they estimate that this premium would be 11.1% lower and 2014 federal subsidies would be reduced by $1.7 billion. A powerful aspect of Dafny et al. (Reference Dafny, Duggan and Ramanarayanan2015) is their use of the exogenous variation introduced by United Healthcare’s decision not to participate. We find a larger effect on premiums when an additional insurer competes in a rating area, due to using a different instrument and our inclusion of 2015 data. In construction of their instrument, Dafny et al. (Reference Dafny, Duggan and Ramanarayanan2015) use market concentration in the individual insurance market of the same area prior to the ACA and the change in this concentration due to United Healthcare’s non-participation in 2014. Therefore, they include data from both 2014 and the time period before 2014. Instead of using the variation in competition due to United Healthcare’s non-participation in 2014, we use the variation caused by United Healthcare’s participation in 2015 as an instrument for competition. As the ACA came with substantial insurance policy reforms and led to the creation of online marketplaces, premiums and levels of competition may not be comparable to pre-ACA premiums and competition levels. Accordingly, this study analyzes only post-ACA data created by the marketplaces in 2014, and expands the analysis to include 2015 data.

5 Methodology and Results

This section presents our methodology and main empirical results. We begin our analysis by employing a simple OLS regression of 2LPS on competition and other control variables in Section 5.1. We examine the correlation between premiums and competition separately for 2014 and 2015. In Section 5.2, we use a FD regression model, with which we aim to eliminate sources of bias that arise in the OLS regressions. Next, we use United Healthcare’s decision to enter the marketplaces in 2015 as an instrument for competition in Section 5.3. This section begins with an analysis of the necessary assumptions required for a good instrumental variable. A 2SLS method is then used to estimate the casual effects of competition on premiums. In Section 5.4, we run a threshold analysis to allow for differential casual effects depending on the initial level of competition in a rating area in 2014. This threshold analysis enables us to observe whether one more insurer entering a rating area has different effects on premiums at varying levels of competition. We examine the robustness of our results in Section 5.5.

5.1 Are premiums correlated with competition?

We begin our analysis by first examining the correlation between 2LPS and competition. In Figure 3, we use a scatter plot to observe the relationship between 2LPS and the number of insurers in a rating area for 2014. Figure 4 does the same for 2015. In both 2014 and 2015, a majority of rating areas had five or fewer insurers in competition. We also observe an inverse relationship between the number of insurers, that is, the level of competition and premiums; premiums decline as competition increases. These scatter plots are useful because they preview industry dynamics. However, a bivariate analysis does not control for the competition characteristics, market structure or rating area demographics. Each of these factors may also influence the premiums charged by insurers in a rating area. Therefore, a multivariate analysis with these control variables would provide more meaningful insight into the effects of competition on premiums.

Figure 3

Scatter plot 2014 (number of issuers and premiums)

Figure 4

Scatter plot 2015 (number of issuers and premiums)

We use a simple OLS model to estimate the following multivariate regression model:

(1) $$\eqalignno{ ln(2LPS)_{{rt}} {\equals}&\,\beta _{0} {\plus}\,\beta _{1} Insurers_{{rt}} \,{\plus}\,\beta _{2} Non{\hbox-}profits_{{rt}} {\plus}\beta _{3} COOPs_{{rt}} \cr & {\plus}\beta _{4} {\rm ln}(HospitalPrice)_{{r(t{\minus}1)}} {\plus} \beta _{5} PoorHealth_{{r(t{\minus}1)}} {\plus}\beta _{6} SickDays_{{(t{\minus}1)}} \cr & {\plus}\beta _{7} {\rm ln}(Population)_{{r(t{\minus}1)}} {\plus}\beta _{8} {\rm ln}(PcIncome)_{{r(t{\minus}1)}} {\plus} \beta _{9} Black_{{r(t{\minus}1)}} \cr & {\plus}\beta _{{10}} Hispanic_{{r(t{\minus}1)}} {\plus}\beta _{{11}} Uninsured_{{r(t{\minus}1)}} {\plus}\beta _{{12}} Poverty_{{r(t{\minus}1)}}\cr & {\plus} \beta _{{13}} PopUnder18Over65_{{r(t{\minus}1)}} {\plus}{\varepsilon}_{{rt}} , $$

where the variable 2LPS _rt represents the 2LPS premium for rating area r in year t; Insurers _rt is the number of insurance companies offering plans; Non-profits _rt the percentage of non-profit insurers; COOPs _rt a binary variable indicating the existence of at least 1 CO-OP insurer in a rating area; Hospital Price _r the average hospital price in the rating area; PoorHealth _r and $$ SickDays_{{r } $$ represent the percentage of population reporting poor or fair health status and the average number of sick days reported in that rating area, respectively; Population _r the rating area population; PcIncome _r the per capita income of the rating area; Black _r and Hispanic _r represent the percentage of black and Hispanic population; Uninsured _r the percentage of the uninsured; Poverty _r the poverty rate of the rating area; PopUnder180Over65 the percentage of population under the age of 18 years or over the age of 65 years. We run this model separately for 2014 and 2015. As insurance companies can only observe previous year’s rating area demographic characteristics, we include a lagged value of all control variables except for Non-profits and CO-OPs. These two variables were only available after the implementation of ACA in 2014 and they have the potential to directly reduce premiums without a lag. The model includes an error term (ε _rt ) we assume to be uncorrelated with other unobserved determinants of premiums. As the premiums may be more volatile in markets with more insurers, we use robust standard errors to correct for potential heteroskedasticity. Furthermore, rating areas in the same state may have correlated standard errors due to state-specific factors. Therefore, we cluster standard errors by state in all our regressions.

Table 2 displays the results from the OLS model utilizing equation (1), with the first two columns reporting 2014 results and the last two columns reporting 2015 results. Columns 1 and 3 display results that do not control for other variables that may affect the premiums, whereas columns 2 and 4 display results that account for control variables. Although not statistically significant across both specifications and in both years, we find that an additional insurer in a rating area is associated with lower 2LPS. The magnitude of this effect ranges from 1.9% in 2014 to 2.9% in 2015, across all four specifications. This effect is similar to the range of 1.2–3.5% decline in premiums found in Jacobs et al. (Reference Jacobs, Banthin and Trachtman2015) and Gabel et al. (Reference Gabel, Whitmore, Green, Stromberg, Weinstein and Oran2015). As the OLS model is likely to suffer from both simultaneity and omitted variable biases, as we describe in Section 1, we do not impose a causal interpretation to the results in Table 2.

Table 2

Ordinary least squares estimates of the effects of competition on premiums in the federally facilitated marketplaces

Note: Standard errors are clustered by state and given in parentheses.

*p<0.10, **p<0.005, ***p<0.01.

5.2 FD estimates

One way to mitigate the endogeneity concerns is to use a FD regression analysis. With this approach, variables now represent changes between 2 years. For 2LPS, the number of insurers, share of non-profits, and existence of CO-OPs, we use the change from 2014 to 2015. All other controls use changes between 2013 and 2014. The rationale behind this decision intuitively makes sense as only the previous year’s data will be available and relevant to insurance companies as they set their premiums for the following year. First-differencing all variables eliminates the time-invariant omitted factors that were previously captured within the error term. If these factors were correlated with the level of competition in a rating area, then the endogeneity bias caused by this correlation will no longer be a concern. Therefore, we run the following regression of changes-on-changes:

(2) $$\eqalignno{ ln(\Delta 2LPS_{r} )\,{\equals}\,&#x0026;\,\beta _{0} {\plus}\,\beta _{1} \Delta Insurers_{r} \,{\plus}\,\beta _{2} \Delta Nonprofits_{r} {\plus}\beta _{3} \Delta COOPs_{r} \cr &#x0026;{\plus}\beta _{4} {\rm ln}(\Delta HospitalPrice)_{r} {\plus} \beta _{5} \Delta PoorHealth_{r} {\plus}\beta _{6} \Delta SickDays_{r} \cr &#x0026;{\plus}\beta _{7} {\rm ln}(\Delta Population)_{r} {\plus}\beta _{8} {\rm ln}(\Delta PcIncome)_{r} {\plus}\beta _{9} \Delta Black_{r} \cr &#x0026;{\plus} \beta _{{10}} \Delta Hispanic_{r} {\plus}\beta _{{11}} \Delta Uninsured_{r} {\plus}\beta _{{12}} \Delta Poverty_{r} \cr &#x0026;{\plus}\beta _{{13}} \Delta PopUnder18Over65_{r} {\plus}{\varepsilon}_{r} . $$

In Table 3, we report the results. In the first column, we do not control for any variables, whereas the second column includes all additional controls. The first column suggests that if a rating area were to gain one more insurer from 2014 to 2015, the 2LPS premium will be 4.13% lower. In the second column, one additional insurer results in 4.47% lower premium increases from 2014 to 2015 per rating area. Both effects are statistically significant at 5% level, unlike the endogenous regression in (1). Magnitudes of both effects are also greater. However, even though this analysis eliminates the bias caused by omitted time-invariant factors, simultaneity bias due to reverse causality between premiums and competition remains unaccounted for. An instrumental variables method can drastically improve our understanding of the effects of competition on premiums.

Table 3

First-differenced (FD) estimates of the effects of competition on premiums in the federally facilitated marketplaces

Note: 2LPS=second-lowest-priced silver plan.

Standard errors are clustered by rating area and given in parentheses.

*p<0.10, **p<0.005, ***p<0.01.

5.3 Instrumental variable: United Healthcare’s 2015 entry

In order to examine whether there is a casual relationship between premiums and competition, we use United Healthcare’s participation in 2015 as an instrument for the change in number of insurers from 2014 to 2015 in a rating area. United Healthcare began offering insurance plans in 169 rating areas throughout 16 states. If this participation is correlated with the change in competition and if the United Healthcare’s choice of rating areas was exogenous, then it can serve as an effective instrumental variable to eliminate both simultaneity and omitted variable biases. In the following subsections, we first ask whether United Healthcare’s participation is correlated with the change in competition. Next, we justify our reason to believe that United Healthcare’s choice of rating areas was exogenously determined. In the last two subsections, we estimate the reduced form regression and 2SLS to find unbiased estimates of competition’s effect on premiums.

First-stage regression

We estimate the following model to see if United Healthcare’s participation in 2015 is correlated with the change in the number of insurers in a rating area from 2014 to 2015:

(3) $$\Delta Insurers_{r} \,{\equals}\,\beta _{o} {\plus}\beta _{1} UnitedHealthcare_{r} {\plus}\Delta X_{r} \delta {\plus}{\varepsilon}_{r} ,$$

where UnitedHealthcare _rt serves as a dummy variable and takes a value of 1 if United Healthcare participated in rating area r in year t. In 2014, the first year of the marketplace, this variable takes the value of 0 for all rating areas, while taking the value of 1 throughout the 169 rating areas of 16 states in 2015. X _rt is a vector of all other control variables given in equation (2).

We present our results in the third and fourth columns of Table 4, with and without other control variables, respectively. In both specifications, coefficient estimates for United Healthcare’s participation are statistically significant and as high as 0.691 with full set of controls. The change in the share of non-profits, the change in the existence of a CO-OP, the change in the black percentage of the population and the change in the uninsured are also statistically significant.

Table 4

Instrumental variables estimates of the effect of change in competition on change in premiums

Note: OLS=ordinary least squares; 2SLS=two-stage least squares; 2LPS=second-lowest-priced silver plan.

Standard errors are clustered by rating area and given in parentheses.

* p<0.10, ** p<0.005, *** p<0.01.

Exclusion restriction

If United Healthcare based its entry decision on an omitted factor that affects premiums, the resulting correlation with the error term would negate its ability to serve as an effective instrumental variable in our analysis. Although this is not testable due to data on omitted factors not being available, we can still compare rating areas where United Healthcare entered in 2015 to rating areas that it did not in terms of both the premiums and the included control variables. If United Healthcare did not target certain areas, then we can make a stronger argument that United Healthcare’s entry decision was exogenous in 2015, thereby rendering its entry as a viable instrumental variable.

If the United Healthcare based its entry decisions on expected premiums, then this reverse causality of premiums affecting the entry decision would make our instrumental variable endogeneous. Therefore, in the first row of Table 5, we report the 2LPS across rating areas that United Healthcare entered in 2015 and across those that it did not. In 2014, the areas that United Healthcare subsequently entered the following year had an average 2LPS of $241.04 and the areas that it did not had an average of $235.42. However, this difference is not statistically significant at 5% significance level. Similarly, the average 2LPS in entry-areas was $246.28 in 2015, while the non-entry areas had a 2LPS of $241.27. We failed to reject the null hypothesis that the two areas had significantly different premiums. We do not have enough evidence to conclude that the United Healthcare might have based its entry decisions on expected premiums.

Table 5

Comparison of rating areas by United Healthcare’s entry

Note: 2LPS=second-lowest-priced silver plan.

* p<0.10.

We report the means and standard deviations of all our included variables by United Healthcare’s entry decision in Table 5. The first three columns compare the rating areas in 2014, before United Healthcare’s entry. The last three columns compare rating areas post United Healthcare entry. In the third and sixth columns, we report the p-value for the null hypothesis that two means are equal in 2014 and 2015, respectively. All p-values are above 5%, indicating a failure to reject the null hypothesis. In other words, we do not have sufficient evidence to claim that United Healthcare selectively entered certain rating areas. For three select variables, their respective p-values are only marginally significant in 2014 at a 10% significance level: hospital price, percent uninsured, and percent under the age of 18 years or over the age of 64 years. This does not constitute a strong pattern in United Healthcare’s entry. Therefore, we feel confident that United Healthcare’s entry is exogenous in 2015 and unlikely to be correlated with omitted factors that affect the premiums.

Another factor that may have affected the entry decisions of United Healthcare into each rating area is the share of the urban population in that rating area. US Census Bureau considers areas with populations greater than 50,000 as ‘urbanized’ areas. We report the share of urban population in Table 5. The average rating area where United Healthcare entered had a somewhat larger share of urban population both in 2014 and 2015 than the areas where it did not. However, the wide variation in urban share of the population across ratings areas makes this difference statistically insignificant. Alternatively, we divide rating area populations by the rating area’s geographic size, provided by the US Census Bureau, to find population per square mile of each rating area. This variable does not demonstrate a noticeable difference between the two types of rating areas in 2014 or 2015.

Finally, United Healthcare may have based its entry decisions on counties per rating area, which may affect both the population and population density of rating areas. We report the average number of rating areas by year and the type of rating area in Table 5 along with its standard deviation. The rating areas that the United Healthcare entered had a lower number of counties per rating area in both years. However, standard deviations are quite high, making the differences statistically significant. Therefore, we fail to reject the null hypothesis that the number of counties is the same between two types of rating areas.

Another way to explore the factors that may determine how United Healthcare made the decision to enter certain rating areas is to run a simple multivariate OLS regression with a binary dependent variable that takes value 1 if United Healthcare entered a certain rating area in 2015. In this regression, we use 2015 values of all our independent variables. As United Healthcare did not know the 2015 data when they made a decision to enter certain rating areas, this regression would be using future values to predict United Healthcare’s decision. In this regression, we found three independent variables with marginally significant coefficients at 10% significance level: population, per capita income and poverty rate. United Healthcare was slightly more likely to enter rating areas with greater number of people and slightly less likely to enter poorer rating areas. None of these coefficients were significant at 5% significance level after controlling for all independent variables.

Reduced form

We regress the natural logarithm of the change in premiums from 2014 to 2015 on our instrument and other exogenous control variables in the reduced-form model:

(4) $$ln(\Delta 2LPS_{r} )\,{\equals}\,\beta _{o} {\plus}\beta _{1} UnitedHealthcare_{r} {\plus}X_{r} \delta {\plus}{\varepsilon}_{r} .$$

We display our results in the fifth and sixth columns of Table 4. All the coefficients have the expected signs. Without other controls, United Healthcare’s entry reduces the change in 2LPS by 6.26%. This effect is 6.31% when we include all other control variables. As the premiums increased from 2014 to 2015, average premium increase would be 6.31% lower in rating areas where United Healthcare entered as opposed to the rating areas that it did not.

2SLS estimates

As our final step, we estimate the 2SLS regression with United Healthcare’s entry in 2015 serving as the instrumental variable:

(5) $$ln(\Delta 2LPS_{r} )\,{\equals}\,\beta _{o} {\plus}\beta _{1} \Delta Insurers_{r} {\plus}\Delta X_{r} \delta {\plus}{\varepsilon}_{r} .$$

The binary variable above indicates United Healthcare’s 2015 entry for rating area r in 2015 as the instrument for ΔInsurers r. We present the results in the final two columns of Table 4. The coefficients indicate that competition has a significant effect on premiums. In magnitude, the coefficients are slightly higher than those from reduced-form estimates, confirming the positive correlation in the first-stage estimates. Compared with the OLS results, we find significantly higher effects of competition on premiums, thereby validating the initial endogeneity concerns.

To interpret the magnitude of the results, we consider the positive increase in 2LPS premium from 2014 to 2015 in a rating area. Based on our preferred specification in column 8, we find this premium increase would be 6.97% lower due to United Healthcare’s entry in a rating area. This effect is significant even at 1% significance level and higher than the 5.4% lower premiums with United Healthcare’s entry that Dafny et al. (Reference Dafny, Gruber and Ody2015) finds with fewer control variables.

We also find three other control variables to have statistically significant effects on the change in premiums. If there is at least one CO-OP in a rating area or if the share of non-profits increase, we observe a negative effect on premiums. CO-OPs and non-profit insurers have lower administrative costs; therefore, these plans are likely to offer lower premiums. We also find a negative effect on premiums as the percentage of Hispanics increase in a rating area. Hispanics are more likely to be uninsured and insurance companies may be offering lower premiums to incentivize and increase enrollment.

5.4 Threshold analysis

Insurance companies may lack adequate incentive to change their premiums if significant competition already exists in a rating area. We postulate whether the effect of competition on premiums varies by the initial level of competition. To validate this postulate, we run equation (5) for only those rating areas with a certain number of insurers in 2014. We identify this effect by the separate subsample analyses.

Table 6 displays the results of this threshold analysis, where each row corresponds to a separate regression. We only report the main coefficient of interest, β ₁. Each individual regression includes only the rating areas with a specific number of insurers in 2014 and a full set of control variables. Coefficients are negative and significant for rating areas with 1, 2 or 3 insurers in 2014. This indicates that a rating area with four or more insurers in 2014 does not display a significant change in premiums as the number of insurers increase in 2015. For rating areas with only one insurer in 2014, one additional insurer in 2015 reduces the increase in 2LPS premium by 7.51%. The same effects are 7.03 and 6.81% for rating areas with two and three insurers in 2014, respectively. When a rating area has four or more insurers, we do not find any statistically significant effects on premiums as more insurers enter the rating area.

Table 6

Threshold analysis of the effects of competition on premiums in the federally facilitated marketplaces

Note: 2SLS=two-stage least squares; 2LPS=second-lowest-priced silver plan.

Standard errors are clustered by rating area and given in parentheses.

*p<0.10, **p<0.005, ***p<0.01.

5.5 Robustness

We examine the robustness of our results to alternative measures of premiums, non-linear specification of the regression model and using county-level data instead of rating areas. We present our results in the first two columns of Table 7 for alternative measures of premiums. The first column displays the results that we obtain by using the mean premium across all silver plans offered in a ratings area, instead of the 2LPS premium. In the second column, we do the same for the median premium across the silver plans. For both alternative measures of premiums, the coefficients have the expected signs and the significance levels do not change. The magnitudes are slightly greater than the coefficients for the same IV regression with 2LPS in the column 8 of Table 4. Therefore, we conclude that our results are robust to whether we use the 2LPS premium, mean premiums or median premiums as our dependent variable.

Table 7

Robustness to alternative measures of premiums, non-linear specification, county-level analysis and additional controls

Note: Standard errors are clustered by state and given in parentheses.

*p<0.10, **p<0.005, ***p<0.01.

In column 3 of Table 7, we use a non-linear regression specification that includes the squared number of insurers. With this squared term, we allow for a non-linear relationship between competition and premiums. The coefficient of the squared term is small and not statistically significant. Therefore, we do not have evidence to claim that the results are driven by the omission of non-linearity in our regression models. In column 4 of Table 7, we use a non-aggregated version of our data set so that each observation represents a county, instead of a ratings area. With this version, we have 2462 counties in 34 states. The same IV method yields similar results to our ratings area-level analysis. Coefficients are similar in magnitude and significance levels. Therefore, our results are robust to a county-level analysis of competition’s effects on premiums in 2014 and 2015.

In the last column of Table 7, we include two additional control variables: share of urban population and number of counties per rating area. With these two additional control variables, magnitudes of the coefficients slightly change. However, the signs and significance levels are the same as all other 2SLS regressions. Our results are robust to the inclusion of these two variables in our econometric model.