Supplemental Digital Content
Technical Appendix 1: Definition of Integration
I defined integration using a strategy developed by Neprash and colleagues (2015). They utilize the place of service codes found in Medicare claims. During this paper’s study period, when a hospital acquired a physician, the physician became eligible to bill under the Hospital Outpatient Department (HOPD) place of service code. The incentives to make this billing change were strong since reimbursement under an HOPD designation was often higher than under an office designation. I created an indicator for each physician in each year to indicate integration status by counting the number of line items billed in the Medicare Carrier (physician/supplier) claims under an office code and under an HOPD code. When 75 percent or more of these line items were billed under an HOPD code, I classified the physician as integrated. Results were not sensitive to a stricter definition of integrated: the effect estimate with 75 percent was .034 (.016, . 052) and with 100 percent was .033 (.015, .051). To account for the possibility that some
practices may have a lag in updating their place of service codes, I used MD-PPAS data to supplement the claims-based approach. After applying the claims definition, if an unintegrated physician’s tax identifier legal name contained the keywords “medical center,” “hospital,”
“system,” “health science center,” “health sciences center,” or “med ctr,” I identified them as integrated. About 80 percent of integrated physicians were identified using claims and the remainder with the keyword search. While the effect estimates were small in all specifications, they were smaller in magnitude when using only the claims-based definition (.014 compared to . 034), although still statistically significant at p = .001.
Technical Appendix 2: Estimating Equations and Specifications
In the main analysis, I estimated a linear probability model among unintegrated physicians with a physician-level dataset:
(1)Integratedi=B0+B1FacilityFeei+B2Xi+ei
Where Integrated was a binary variable that takes the value of 1 if physician i was hospital- integrated for at least one year from 2014-2016. FacilityFee is a continuous variable that measures the change in the facility fee value of physician i’s office visits that resulted from the 2014 reform; it can be positive or negative depending on the effects of the reform for the
composition of services delivered by physician i in 2013. X includes a state fixed effect (state of physician residence) and covariates: physicians' 2013 values for number of Medicare
beneficiaries, number of office visits, number of total services, percentage of female patients, average patient age, percentages of patients by race, average patient severity score, average patient number of months dually enrolled in Medicaid, and area characteristics of education and income. e is a physician-level error term. Standard errors were clustered by hospital referral region. This is a cross-sectional estimation that examines the association of current-year (2013) unintegrated physician characteristics (particularly the variable of interest, FacilityFee) on subsequent integration.
To test whether the association between changes in facility fee potential and integration was stronger among physicians who were especially exposed to the 2014 reform (i.e., physicians whose facility fee potential experienced large gains or large losses), I estimated:
(2)Integratedi=B0+B1FacilityFeei+B2Xi+B3FacilityFeei∗LargeGaini+B4FacilityFeei∗LargeLossi+B5LargeGaini+B6LargeLossi+ei
In which dummy variables for large-gain and large-loss physicians are interacted with the facility fee variable to flexibly allow for a distinct effect among these groups. In these cases, the
coefficients are not reported; rather, results reported are the total marginal effects of the facility fee variable over the values of the dummy variable when equal to 1. That is, I report the total marginal effect of treatment among large-gain and large-loss physicians rather than the effect over and above some reference category. I estimate Model (2) and use a post-estimation
hypothesis test for B3 = B4. The F-statistic is 6.66 (p = 0.0103). I conclude that the effect of the reform is not equal for large-loss and large-gain physicians.
To test whether the association between changes in facility fee potential and integration was stronger within markets that were consolidated, I estimated:
(3)Integratedi=B0+B1FacilityFeei+B2Xi+B3FacilityFeei∗Concentratedi+B4Concentratedi+ei
In which a dummy variable for concentrated hospital markets was interacted with the facility fee variable to flexibly allow for a distinct effect among this group. Given the presence of interaction terms, I report the marginal effects of the facility fee variable rather than the coefficients as in (1). Results were that the marginal effect within concentrated markets was statistically
insignificant (.027, p = 0.226). From this, I concluded that changes in facility fee potential associated with the 2014 reform did not significantly affect integration among primary care physicians in concentrated markets. The marginal effect among non-concentrated markets was . 037 (p < .001). Using a test of differences in average marginal effects between concentrated and non-concentrated markets (i.e. testing for significant differences between .027 and .037) yielded an F-test of 0.13 (p = 0.713). This implies that the differences in marginal effects were not estimated precisely enough to identify a statistically significant difference between them (i.e.
between .027 and .037; results allow me to reject that .037 is equal to 0, but I do not reject that . 027 is equal to 0, nor do I reject that .037 - .027 = 0).
I also examined the role of rural areas with respect to the facility fee reform and hospital- physician integration. I constructed two definitions of rurality (there are many different
approaches to measuring rurality; since this is a secondary or tertiary aim of the paper, I limited it to two). First, I appended a rural-urban continuum code from the Area Health Resources File.
Secondly, I identified the areas that CMS qualifies as rural and appended that information to my dataset. I added each of these rural dummy variables to the main analysis in turn. Using either definition (i.e., a county rural-urban continuum code of 4 or greater; or a CMS-designated rural zip code), rurality was a statistically significant, and large, predictor of integration.
(4)Integratedi=B0+B1FacilityFeei+B2Xi+B3Rurali+ei
Technical Appendix Table TA1. Rural estimation with rural indicator as a covariate
Specification Treatment effect
estimate of policy change
Coefficient on rural indicator
(1) Main .034*** N/A
(2) Main, plus rural continuum code
>=4 as a predictor
.032** 3.10***
(3) Main, plus CMS rural zip as a predictor
.033*** 2.32**
***p < .001
** p < .01
The main treatment effect (of the policy change) was nearly unchanged when including a rural definition, implying that rurality exerts a strong independent effect but does not confound the main estimate. I next estimated an interacted model analogous to those of the concentrated hospital markets; the results of this model are reported in the body of the paper. For this analysis, I used the rural continuum code definition. I estimated:
(5)Integratedi=B0+B1FacilityFeei+B2Xi+B3Rurali+B4Rurali∗FacilityFeei+ei
Using the full sample, I calculated the total marginal effect of the treatment over rural and non- rural areas. The marginal effect among rural areas was not significant (.018, p = 0.56). Among non-rural areas, the effect was significant (.035, p < .0001). From this, I concluded that changes in facility fee potential associated with the 2014 reform did not significantly affect integration among primary care physicians in rural areas. Using a test of differences in average marginal effects between rural and non-rural areas (i.e. testing for significant differences between .018 and .035) yielded an F-test of 0.28 (p = 0.599). This implies that the differences in marginal effects were not estimated precisely enough to identify a statistically significant difference between them (i.e. between .035 and .018; results allow me to reject that .035 is equal to 0, but I do not reject that .018 is equal to 0, nor do I reject that .035 - .018 = 0). Last, I also estimated a version of (5) in which I added hospital HHI to the vector of covariates found in X in case the apparent effect of rurality was simply a function of concentrated rural hospital markets; it was not.
Technical Appendix 3. Robustness and Falsification Tests
I checked for potential problems with, and considered alternatives to, the above specifications.
One potentially important confounder in this study is that physicians whose facility fee potential increased had a unique composition of office visit levels that may have been correlated to other characteristics that appealed to hospitals, which would overstate the role of the change in facility fees; in other words, an omitted variable could have been correlated to both the facility fee variable and the error term, causing omitted variable bias.
I estimated a model using pre-reform data to test whether composition of office visit levels was associated with integration when the reform was not in play. If this composition were a strong predictor of subsequent integration even in pre-reform times, it would suggest omitted variable bias. I identified the set of office codes which received facility fee increases (99201, 99202, 99211, 99212, 99213) and calculated, for each physician, the proportion of their office visits that fell into these categories in 2011. I then estimated:
(6)Integratedi=B0+B1CompositionofVisitsi+B2Xi+ei
Where Integrated was a binary variable that takes the value of 1 if physician i was hospital- integrated for at least one year from 2012-2013. CompositionofVisits measured the proportion of a physician’s office visits that would, in 2014, receive facility fee increases. I used the same control variables and state fixed effects as in (1). I found that B1 was not statistically significant (-.015, p=0.92) and showed virtually no correlation to subsequent integration. I also examined the correlation between the composition of office visits and patient Hierarchical Condition Category score (a measure of patient clinical severity) and found a weak relationship (r = -.068).
These findings lend credibility to the analysis.
I also tested a physician fixed effects model:
(5)Integratedi ,t=B0+B1FacilityFeei ,t+B2Xi , t+μi+ei ,t
This model included physicians from 2011-2015 who were unintegrated in 2011. Integrated is a time-varying binary variable that takes the value of 1 in each year in which a physician is integrated. FacilityFee is a time-varying measure of the facility fee value of a physician’s office visits in each year. The covariates are the same as in model 1, with the difference that they are time-varying; the model includes state fixed effects, a physician fixed effect μi , and standard errors clustered at the physician level. The results of the physician fixed effects analysis yielded an estimate of B1=¿ 0.00039 (p-value < 0.001). While this still implies a positive association between facility fee increases and the probability of integration, it is smaller than the main estimate. I have comparatively little confidence in the validity of the fixed effects model. The main advantage of a physician fixed effects model is that it controls for any potential time- invariant unobserved confounders: if the main analysis omitted an important characteristic of a physician that was correlated to integration and the facility fee variable (e.g., the school in which a physician completed his or her medical training), the fixed effects model removes this bias. In this setting, a physician fixed effects model has at least three serious drawbacks: first, the overwhelming majority of physicians are thrown out of the model (93 percent): identification comes only from physicians with variation in their outcome variable. Second, in this setting, once a hospital integrates a physician, it is likely to change aspects of the working environment, quite possibly including billing protocol; for the fixed effects analysis, this is a problem, since I rely on billing data to construct the key variable (facility fee potential). The measure of facility fee potential among physicians who integrate could be contaminated by such employment setting changes, leading to biased estimates. The main analysis does not suffer from this problem
because it does not include facility fee potential calculations from years in which physicians were integrated. Third, I focus on the 2014 reform because it provides plausibly exogenous changes in facility fee potential; I cannot say with certainty that changes in office visit facility fees in the other years were as exogenously determined. For all these reasons, I prefer the main model, and recommend that the fixed effects analysis be viewed as exploratory. If anything, however, it further supports the paper’s evidence that the changes in facility fee potential associated with this reform exerted only a modest effect on integration among hospitals and primary care physicians.
The analysis also depends on a reasonably high degree of continuity in the composition of office visits within a physician from year to year. That is, if physicians’ distribution of codes 99201- 99215 were highly variable from year to year, hospitals would not make acquisition decisions on the basis of a physician’s set of codes in any given year. For this reason, I tested a linear
regression model in which the dependent variable was the current-year proportion of visits that fell into codes 99201, 99202, 99211, 99212, 99213 (i.e. the codes which received a facility fee increase in 2014). The independent variable was the one-year lag of this variable. The lag was highly predictive of current-year proportion (B = 0.65; t-statistic = 508, p < .001). This result indicated that a physician’s distribution of office visit code acuities persisted from year to year.
Appendix Figure A1. Changes in composition of services before and after CMS facility fee reform
Note: Annual proportions are adjusted for state and year fixed effects and shown by physicians who integrated and those who did not. High-level office visits are defined as those with Current Procedural Terminology code 99215; low-level office visits are those with code 99211. Facility fees increased for codes 99211, and decreased for codes 99215, beginning January 1, 2014. The composition of office visit levels was effectively unchanged from 2011 through 2016.
Appendix Figure A2. Net effects of Medicare’s office visit facility fee reform on physicians’
value-to-hospitals (i.e., facility fee potentials)
Notes: Each point in this histogram represents one primary care physician in 2013 (n = 98,884).
This graph displays the sum of the potential office visit facility fees under the 2014 pricing schedule minus the sum under the 2013 pricing schedule. A positive number implies that the physician’s facility fee potential increased from 2013 to 2014. 11 percent of physicians
experienced a net change of less than $500. The average facility fee value of a physician’s full set of services in 2013 was $106,399. Values Winsorized at 1st and 99th percentiles for display purposes.
Appendix Figure A3. Hospital-physician integration over time among primary care physicians
Appendix Table A1. Sample Flow
Inclusion Criteria
Resulting Sample Size (physicians) Valid National Provider Identifier with 2013 medical specialty listed
as non-hospitalist general practice, internal medicine, or family
practice 143375
Non-zero number of office visits 143281
Provider located in U.S. state 141687
Provider has at least 50 service line items billed to Medicare 140628
Covariate data present 123886
Exclude physicians integrated prior to 2014
98884
Final sample size
98884
Appendix Table A2. Marginal effects of facility fee reform using alternative thresholds for large fee-increase and large fee-decrease physicians
Threshold Mean change in facility fee potential among physicians in threshold
Total marginal effect among physicians in
threshold
1st percentile (increase) $72,700 0.102***
1st percentile (decrease) -$41,200 -0.038
2nd percentile (increase) $59,100 .095***
2nd percentile (increase) -$32,700 -0.026
3rd percentile (increase) $52,000 .094***
3rd percentile (decrease) -$28,300 -0.015
5th percentile (increase) $44,700 .098***
5th percentile (decrease) -$23,700 .015
10th percentile (increase) $34,600 0.091***
10th percentile (decrease) -$17,700 0.018
*** p < .0001
Source: Author’s calculations using Medicare claims data.
Note: Total marginal effects shown in percentage points for a $1,000 change in facility fee potential associated with CMS 2014 reform on the probability of a physician integrating with a hospital from 2014-2016. The thresholds vary the definition of large fee-increase and large fee- decrease physicians. Effects for physicians at the 1st, 2nd, 3rd, 5th, and 10th percentiles of fee- increase and fee-decrease physicians are shown. The regression is a linear probability model run separately for each threshold (i.e. Equation 2 of the Technical Appendix). The 10th percentile definition is used as the main analysis in the paper. In the main analysis, the marginal effect among the middle of the distribution (i.e. those not included in the large-gain or large-loss groups) was not statistically significant (marginal effect = -0.025, p = 0.177). The average marginal effect across the full sample was .034 (p < .001).
Appendix Table A3. Regression output from main models
(1) (2) (3) (4)
Full Sample
Estimate 95% CI Sample mean
Large-gain and large-loss
physicians Estimate
95% CI Sample mean
Concentrated Hospital Markets Estimate 95% CI Sample mean
Rural areas
Estimate 95% CI Sample mean Integrated
(dependent)
( - ) ( - ) ( - ) ( - )
4.931 4.931 4.931 4.931
Large-gain NA -2.128***
(-3.328 - -0.927) 0.0688
Large-loss NA -1.033
(-2.309 - 0.244) 0.0260 2014 Reform
(difference in office visit facility fees, continuous, $k)
0.0342*** -0.0254 0.0367*** 0.0346***
(0.0163 -
0.0522) (-0.0624 -
0.0116) (0.0171 -
0.0564) (0.0181 - 0.0510)
4.995 4.995 4.995 4.995
Large-gain * reform NA 0.116*** NA NA
(0.0692 - 0.164)
2.382
Large-loss * reform NA 0.0433 NA NA
(-0.0209 - 0.107) -0.459
Concentrated 1.759**
(0.316 - 3.201)
0.234 Concentrated *
reform
-0.00934 (-0.0594 -
0.0407)
1.365
Rural 3.222***
(1.223 - 5.221)
0.111
Rural * reform -0.0164
(-0.0779 - 0.0451)
0.930 Number of
Medicare beneficiaries
0.00422*** 0.00442*** 0.00416*** 0.00417***
(0.00311 - 0.00534)
(0.00327 - 0.00556)
(0.00305 - 0.00527)
(0.00308 - 0.00526)
336.2 336.2 336.2 336.2
Number of office visits
-0.00188*** -0.00189*** -0.00191*** -0.00195***
(-0.00226 -
-0.00151) (-0.00230 -
-0.00148) (-0.00228 -
-0.00154) (-0.00231 - -0.00158)
835.7 835.7 835.7 835.7
Number of services -0.000414*** -0.000429*** -0.000414*** -0.000413***
(-0.000531 - -0.000297)
(-0.000548 - -0.000309)
(-0.000532 - -0.000297)
(-0.000529 - -0.000298)
2,278 2,278 2,278 2,278
Percent female
patients 0.0965 -0.00411 0.0996 0.176
(-0.800 - 0.993)
(-0.904 - 0.896)
(-0.799 - 0.998)
(-0.718 - 1.069)
0.587 0.587 0.587 0.587
Average age of patients
-0.0782*** -0.0700*** -0.0798*** -0.0879***
(-0.129 - -0.0271)
(-0.120 - -0.0196)
(-0.131 - -0.0287)
(-0.137 - -0.0385)
70.94 70.94 70.94 70.94
Average clinical
severity of patients 1.842*** 1.786*** 1.896*** 1.968***
(1.122 - 2.562)
(1.063 - 2.509)
(1.185 - 2.606)
(1.250 - 2.685)
1.099 1.099 1.099 1.099
Percentage with HS education
0.0626* 0.0642* 0.0511 0.0603*
(-0.00845 - 0.134)
(-0.00689 - 0.135)
(-0.0179 - 0.120)
(-0.00950 - 0.130)
86.91 86.91 86.91 86.91
Percentage with
college education -0.0840*** -0.0877*** -0.0709*** -0.0616***
(-0.117 -
-0.0510) (-0.121 -
-0.0548) (-0.104 -
-0.0374) (-0.0927 - -0.0305)
30.13 30.13 30.13 30.13
Average dual- eligible months of patients
-0.00635 -0.00776 0.00180 -0.00614
(-0.0865 - 0.0738)
(-0.0879 - 0.0724)
(-0.0786 - 0.0822)
(-0.0852 - 0.0729)
2.594 2.594 2.594 2.594
Constant 6.428* 6.133* 6.701* 6.181*
(-0.714 -
13.57) (-0.997 -
13.26) (-0.332 -
13.73) (-0.801 - 13.16)
Observations 98,884 98,884 98,884 98,884
R-squared 0.030 0.030 0.031 0.031
*** p<0.01, ** p<0.05, * p<0.1
Note. Data are from Medicare claims and include 98,884 physician-years. CI = confidence interval. The coefficient of 2014 reform estimates the effect of a $1,000 increase in facility fee potential on the probability of integrating with a hospital. Outcome is a binary indicator of a physician being integrated with a hospital in any year from 2014-2016. Model 1 includes the reform variable as a continuous measure. Model 2 adds to Model 1 dummy variables for large- loss and large-gain physicians and interactions between those variables and the reform measure.
Model 3 adds to Model 1 a dummy variable for physicians in concentrated hospital markets and an interaction between that variable and the reform measure. Model 4 adds to Model 1 a dummy variable for physicians in rural areas and an interaction between that variable and the reform measure. Covariates include physicians' 2013 values for number of Medicare beneficiaries, number of office visits, number of total services, percentage of female patients, average patient age, average patient severity score, average patient number of months dually enrolled in
Medicaid, and area characteristics of education and income. Standard errors are clustered by hospital referral region.
Appendix Table A4. 2013 characteristics of primary care physicians, stratified by exposure to Medicare office visit facility fee reform
Characteristic
Full sample
Large-loss physicians
Large-gain physicians
Middle of distribution
(neither large-gain nor large-
loss)
Sample Size 98,884 2,568 6,808 89,508
High school education or higher 86.9 86.4 84.4 87.1
Four-year college or higher 30.1 29.1 25.5 30.5
Percent Black 10.7 12.1 7.8 10.9
Percent Hispanic 2.7 2.4 2.5 2.7
Percent White 80.6 80.8 81.4 80.6
Percent Other 6 4.7 8.2 5.8
Percent under 138 percent of
poverty line 14.3 15 16 14.2
Average number of unique patients 336 522 675 305
Average patient age 71 72 73 71
Average patient Hierarchical
Condition Category score 1.10 1.27 1.17 1.09
Average number of office visits 836 1,493 2,357 701
Average facility fee potential, all
2013 services 106,399 260,106 264,758 89,944
Average facility fee potential of 2013 office visits
72,336 155,785 183,499
61,487 Average facility fee potential of
office visits after 2014 reform 77,331 138,108 218,100 64,880 Average change in facility fee
potential due to 2014 reform 4,995 -17,677 34,601 3,393
Notes: Author’s analysis of Medicare claims data. Large-loss (large-gain) physicians are those in the lowest (highest) decile among physicians whose facility fee potential fell (rose) as a result of Medicare’s office visit reform. Physicians in the middle of the distribution are those who are in neither the large-loss nor the large-gain groups.
Appendix Table A5. 2013 characteristics of primary care physicians, stratified by hospital market concentration
Characteristic
Full sample
Concentrate d
Standardize d Mean Difference
Sample Size 98,884 23,114 N/A
High school education or higher 86.9 87 -.00
Four-year college or higher 30.1 26.5 .08
Percent Black 10.7 8.1 .09
Percent Hispanic 2.7 1.5 .08
Percent White 80.6 86.8 -.17
Percent Other 6 3.6 .11
Percent under 138 percent of poverty line
14.3 15 -.02
Average number of unique patients 336 385 -.15
Average patient age 71 71 .00
Average patient Hierarchical Condition Category score
1.10 1.08 .05
Average number of office visits 836 934 -.12
Average facility fee potential, all 2013 services
106,399 116,430 -.06
Average facility fee potential of 2013 office visits
72,336 80,562 -.12
Average facility fee potential of office visits after 2014 reform
77,331 86,402 -.13
Average change in facility fee potential due to 2014 reform
4,995 5,840 -.07
Notes: Author’s analysis of Medicare claims data. Concentrated hospital markets are those with Herfindahl-Hirschman Indices of 2,500 or higher.
Appendix Table A6. 2013 characteristics of primary care physicians, stratified by rurality
Characteristic
Full sample
Physicians in rural areas
Standardize d Mean Difference
Sample Size 98,884 10,938 N/A
High school education or higher 86.9 84.5 .07
Four-year college or higher 30.1 19.5 .25
Percent Black 10.7 6.5 .15
Percent Hispanic 2.7 0.7 .16
Percent White 80.6 89.2 -.24
Percent Other 6 3.5 .12
Percent under 138 percent of poverty line
14.3 17.6 -.09
Average number of unique patients 336 429 -.28
Average patient age 71 71 .00
Average patient Hierarchical Condition Category score
1.10 1.07 .08
Average number of office visits 836 1,104 -.32
Average facility fee potential, all 2013 services
106,399 136,477 -.18
Average facility fee potential of 2013 office visits
72,336 93,738 -.30
Average facility fee potential of office visits after 2014 reform
77,331 102,145 -.32
Average change in facility fee potential due to 2014 reform
4,995 8,407 -.26
Notes: Author’s analysis of Medicare claims data. Rural areas are counties with rural-urban continuum codes greater than or equal to 4.