Statistical Methods - Text Mining Techniques for Healthcare Provider Quality Determination - page 105

Table 20. Coefficients of independent variables

Coefficient Normal Gamma Negative Binomial

Age 0.0326 -0.0078 -0.0078

Female 0.1746 0.0074 0.0074

Hemodialysis 1.6820 0.4333 0.4333

Continuous Ventilation > 96 Hours 10.6845 1.3606 1.3606

Insertion of endotracheal tube 0.5237 0.0669 0.0669

Injection of antibiotic 0.5237 -0.0149 -0.0149

Respiratory medication administered by nebulizer 0.7881 -0.0076 -0.0076

Diagnostic ultrasound of heart 1.5878 0.4477 0.4477

Esophagogastroduodenoscopy with closed biopsy 2.9541 0.5803 0.5803

Continuous positive airway pressure 2.3769 0.6475 0.6475

Closed endoscopic biopsy of lung 1.1786 0.3529 0.3529

Computerized axial tomography of head 0.9279 0.3149 0.3149

Transfusion of packed cells 2.9279 0.4823 0.4823

Venous catheterization 4.3954 0.7121 0.7121

Continuous Ventilation < 96 Hours 0.7652 0.6255 0.6255

Closed biopsy of bronchus 3.8546 0.6023 0.6023

Thoracentesis 3.8546 0.6195 0.6195

Where code>0;
Run;

This code not only finds all of the patients with COPD; it identifies the column where COPD is
listed. The diagnosis codes are supposed to be entered in their order of importance, which is not always
the same for all providers.
Given the sample size, every independent variable is statistically significant. Using the procedures
listed in Table 9, we examine the contribution of each procedure to the patient demographic factors.
These are listed in Table 20.
For the normal link function, all procedures indicate that the patient will stay longer compared to
patients who do not have these procedures. However, the gamma and negative binomial link functions
indicate the patients receiving an antibiotic, or medication by a nebulizer will not stay as long compared
to patients who do not receive them. Race and income quartile are also statistically significant (Table
21).
Note that the negative binomial and gamma link functions generate the same coefficient estimates.
For the three link functions, Caucasian reduces the length of stay; Hispanic patients will stay longer.
The normal result indicates that Asians stay less; the gamma and negative binomial indicate that they
stay longer.
There is also a considerable difference for each of the standard errors. The normal link function has
a much larger error for each coefficient (Table 22).
The standard error for the gamma and negative binomial are smaller because these distributions better
fit the shape of the true distribution. The normal distribution assumes a symmetry that does not exist,