Statistical Methods - Text Mining Techniques for Healthcare Provider Quality Determination - page 75

variable for the variable itself in an attempt to have equal variances. Transformations require consider-
able experience to use properly.
Another assumption is that the independent variables are independent of each other. While the model
can tolerate some correlation between these variables, too much correlation will result in a poor model
that cannot be used effectively on fresh data. It will show statistical significance when such significance
does not exist. A similar problem occurs if the independent variables have different range scales. If most
of the variables are 0-1 indicator functions with patient's age on a scale of 0-100, the value of age will
completely dominate the regression equation. The variable scales should be standardized before the
model is developed, or the model will be biased.
Probably the most worrisome requirement is the assumption that the error terms are identically
distributed. In order for this assumption to be valid, we must assume the uniformity of data entry. That
means that all providers must use the ICD9 codes in exactly the same way. Unfortunately, such an as-
sumption cannot possibly be valid. Consider, for example, the condition of “uncontrolled diabetes”.
The term, “uncontrolled” is not defined. Therefore, the designation remains at the discretion of the
provider to define the term. For this reason, different providers will define it differently, and will try to
define it to benefit themselves. For example, one study defined “uncontrolled” to mean that a glycated
haemoglobin (A1C) was greater than 7.5%, but less than or equal to 11%, and their fasting plasma
glucose (FPG) was less than or equal to 15 mmol/l.(Rosenstock et al., 2006) A second study defined
it as plasma glucose over 33 mmol/l and/or venous bicarbonate less than 14 mmol/l.(Gale, Dornan, &
Tattersall, 1981) More recently, the definition of uncontrolled diabetes shifted to an HbA(lc) of more
than 7%. (Barnett et al., 2007)
However, in a web page designed to optimize coding rather than treatment, it states that optimizing
reimbursement requires a focus on the documentation of diabetes as controlled or uncontrolled, identifica-
tion of previously unrecognized diabetes, and accuracy in the specification of diabetes complications. It
suggests still another definition of uncontrolled: blood glucose above 180(-200) mg/dL at admission or
two or more measurements while admitted above 180(-200) mg/dL OR lesser persistent hyperglycemia
outside guidelines for hospital management. For example, a fasting blood glucose above110 mg/dL and
other blood glucose levels above 180 mg/dL of patients in non-critical care units, could also be considered
consistent with uncontrolled diabetes.(Anonymous-uncontrolled, 2008) Note that a blood glucose above
110 will increase the proportion of patients identified as uncontrolled when a patient is not considered
to have diabetes at all unless a fasting glucose exceeds 126 mg/dL, according to the American Diabetes
Association. In other words, the standard for uncontrolled diabetes is now lower than the actual defini-
tion of diabetes. In fact, this provides an example of “upcoding” by taking advantage of the vagueness
of the terminology to increase the number of patients identified as having a specific coded condition.
This same document also gives suggestions for assigning a diagnosis of diabetes while the patient is
in the hospital. The first suggestion is a blood glucose above 200 mg/dL, but the source also suggests
that the value of 126 mg/dL can be used if the attending physician so documents. It also states that
such a diagnosis must be confirmed once the patient has been discharged. This same report discusses
the increased revenue to the hospital if uncontrolled diabetes is better documented. For this reason of
financial benefit, uniformity of data entry just cannot be assumed.
Any statistical test has four parameters: Type I error, Type II error (or power), sample size, and effect
size. Specifying three of the parameters fixes the fourth. In clinical trials, a power analysis computes
the sample size after first fixing Type I error, power, and effect size. The effect size is half the width of
the confidence limit surrounding the hypothesized population measure. For a simple test of the mean,