Description

Katie Leinenbach
1 Summary Statistics

Figure 1: Summary Statistics on Medicare Spending, Claims, and Patients
2 Mean Total Claims for Integrated vs. NonIntegrated Physicians

Figure 2: Mean Total Medicare Claims by Integration Status
3 OLS
Table 1: OLS Results
Integration on Claims
Dependent Variable: log(Total Claims)
Intercept:
Fixed Effects: -0.3223*** (0.0047)
NPI: Yes
Year: Yes
S.E. Clustered: By NPI
Observations: 2,553,058
R2: 0.90194
Within R2: 0.00890

4 Estimated Treatment Effect Bounds
The table below shows the different estimated bounds for the treatment effect based on varying values of ρ and . As ρ and vary, the estimated right bound varies significantly. This indicates that there is a serious problem with the OLS estimation. These results support the need for an instrumental variable.
Table 2: Estimated Treatment Effect Bounds
Rho R Squared Max Left Bound
Right Bound
0 0.5 -0.3465131 -0.3465131
0 0.6 -0.3465131 -0.3465131
0 0.7 -0.3465131 -0.3465131
0 0.8 -0.3465131 -0.3465131
0 0.9 -0.3465131 -0.3465131
0 1 -0.3465131 -0.3465131
0.5 0.5 -0.3465131 1.3293314
0.5 0.6 -0.3465131 0.908344
0.5 0.7 -0.3465131 0.4873565
0.5 0.8 -0.3465131 0.0663691
0.5 0.9 -0.3465131 -0.3546183
0.5 1 -0.3465131 -0.7756058
1 0.5 -0.3465131 3.0051759
1 0.6 -0.3465131 2.163201
1 0.7 -0.3465131 1.3212261
1 0.8 -0.3465131 0.4792513
1 0.9 -0.3465131 -0.3627236
1 1 -0.3465131 -1.2046985
1.5 0.5 -0.3465131 4.6810204
1.5 0.6 -0.3465131 3.4180581
1.5 0.7 -0.3465131 2.1550958
1.5 0.8 -0.3465131 0.8921334
1.5 0.9 -0.3465131 -0.3708289
1.5 1 -0.3465131 -1.6337912
2 0.5 -0.3465131 6.3568649
2 0.6 -0.3465131 4.6729151
2 0.7 -0.3465131 2.9889654
2 0.8 -0.3465131 1.3050156
2 0.9 -0.3465131 -0.3789342
2 1 -0.3465131 -2.0628839
5 2SLS
Table 3: 2SLS Results – First Stage and Reduced Form
Dependent Variable: Integration log(Total Claims)
Practice Revenue Change 1.35e-5*** (6.56e-7)
Intercept:
Fixed Effects: -3.688*** (0.1991)
NPI: Yes Yes
Year: Yes Yes
S.E. Clustered: By NPI By NPI
Observations: 2,252,137 2,252,137
R2: 0.89231 0.81312
Within R2: 0.00247 -1.0330
6 DWH Test
Below are the results from the DWH test when done manually. R also reports the values. I get the same results for both.
Table 4: DWS Results
Dependent Variable: log(Total Claims)
Intercept: -3.688*** (0.1217)
Vhat:
Fixed Effects: 3.363*** (0.1218)
NPI: Yes
Year: Yes
S.E. Clustered: By NPI
Observations: 2,553,058
R2: 0.90926
Within R2: 0.01293
7 Wald Statistic
When using the Anderson-Rubin Wald statistic, I find that the results are same as the traditional t-test. For both, I will reject the null hypothesis.
Table 5: Anderson-Ruben Wald Staistic Test
Dependent Variable: log(Total Claims)
Practice Revenue Change: Fixed Effects: -5e-5*** (1.67e-6)
NPI: Yes
Year: Yes
S.E. Clustered: By NPI
Observations: 2,252,137
R2: 0.90836
Within R2: 0.00312
The For the second part, I will not need to adjust because my F statistic is above 100. The results are the same then.
8 BH Re-Centering
Table 6: Borusyak and Hull
Dependent Variable: log(Total Claims)
Intercept:
Fixed Effects: -3.658*** (0.1984)
NPI: Yes
Year: Yes
S.E. Clustered: By NPI
Observations: 2,252,137
R2: 0.81485
Within R2: -1.0141
9 Discussion
When first estimating the effect of integration on the log of claims, the OLS estimate finds a small, significant, negative result. However, the treatment bounds calculated based on varying values of ρ and Rmax2 show we meed to use an instrumental variable.
Using the 2010 fee schedule update as the instrument, the 2SLS, Durbin-WuHausman test, and Borusyak and Hull all find similar results. There is a large negative relationship between integration status and log of total claims, using the update as an instrument, meaning those who are integrated with the hospital do not claim
10 Reflection
For this assignment, I learned how important separating your code files are. When I went to rerun the code, it recreated the data files, which was completely unnecessary. I need to better organize my files so I do not run into this issue again.
I tried to make my tables more attractive in Latex, but I am still struggling to make effective graphs in R. More practice is required to learn how to edit graphs effectively.
The main takeaway from this assignment is the importance of workflow. Not only do I need to understand how my directory can impact my flow, I also need to learn how to effectively include multiple code files in my workflow.