Medical payment is one of the most controversy topic in our daily. In this project, we are going to use the dataset “Medicare_ Provider_ Util_ Payment_ PUF_ CY2013”, linking it to the referral network dataset and do some analysis about the relaton between medical payment, specialty and referral network. Due the the large amount of data, we only focus on Wisconsin State.

In Wisconsin, we have 83 types of specialties. One of the domain type is the Internal Medician. In our payment data, we have the nearly 20000 pieces of information regarding to 1500 internal medicians in Wisconsin. Our project will be focus on the referral network within internal medicians and the network between internal medicians and other specialities. Also, we will include the geographical factor in our analysis to inspect the referral network in a local community.

First of all, we use SBM to figure out a general picture of referral relations regarding to the internal medicians. We divide the internal medicians into three groups,low, median and high, by their median charge. Here, for each group, we only list the top five referral relations for each group.

We can see there exists really strong referral relations within the internal medicians, and also there are a lot of referral between the internal medicians and diagnostic radiology. And we also notice that there are a lot of referral to the cardiology and family practice. Here, it just indicates some basic routines for the internal medicians. A patient come in, send him or her for diagnostic radiology first. If it is minor aliments, patients will be referred to family doctors for regular treatment and later trackting. While if diagnostic radiology can indicates the signals for cardiology, then a more professional speciality will be referred.

## 3 x 3 Matrix of class "dgeMatrix"
##                            r_1  r_2   r_3
## Internal Medicine low     1.30 0.29 11.77
## Internal Medicine median  4.04 5.23  3.80
## Internal Medicine high   11.05 0.90  3.51
##                  unitype_group freqtype
## 2       Internal Medicine high      243
## 1  Diagnostic Radiology median      193
## 3     Internal Medicine median       91
## 5            Cardiology median       79
## 8    Emergency Medicine median       51
## 12        Family Practice high       45
##                  unitype_group freqtype
## 5     Internal Medicine median      164
## 2  Diagnostic Radiology median       83
## 4    Nurse Practitioner median       77
## 14      Family Practice median       74
## 32           Cardiology median       68
## 10        Family Practice high       55
##                  unitype_group freqtype
## 8  Diagnostic Radiology median       68
## 23    Internal Medicine median       36
## 4        Internal Medicine low       35
## 18      Family Practice median       34
## 7            Cardiology median       27
## 26      Nurse Practitioner low       23

Next question is that when the internal medicians refer their patients, whether they will consider the factors of potential payment or geography? Based on common sense, geography will be a significant factor since physicians usually will should be willing to refer the patients to someone they know and patients usually do not want to travel too far away. Therefore, we apply CASC to combine the information of zipcode and referral network, and divide them into 10 groups. In the plot on the left, we can see the clear pattern of geographical clusters. As a constrast, we also plot those points with the median payment. Here, we can see, for each community on the left, the physicans’ charge varies a lot. In another word, if we assume that a higher-charge physician is the better one, then the reason that a patient do not need to be referred far way is probabily because there exists better enough physicians in each local area so that basic demand can be fulfilled. Our projetc will also focus on each community later and try to figure out the referral patterns associated with the payment.

## Warning: closing unused connection 5
## (http://pages.stat.wisc.edu/~karlrohe/netsci/code/globalTransitivityFunctions.R)
## 
## 10  4  9  1  3  5  7  8  2  6 
## 20 22 22 23 23 25 26 27 28 28

We also consider the relation between referral distance and payment, to see whether they are correlated. Due to the complicate algorithm, here we only plot 150 referrals. Unfortunally, even though the slope is positive, due to the limit size of point and clear cutoff on the plots, we cannot draw any conclusion significantly. Later, we may try to include more points and also try to figure out wehther the cutoff has any practical reasons.