| |
A variety
of statistical challenges confront researchers who want to use
administrative data to study end-of-life care and symptoms. Many
of these issues are standard for topics related to health care.
Keeping them in mind is, nonetheless, important.
First, much
of the data is quite skewed--meaning that the assumptions of normality
underlying many of our statistical techniques do not always apply.
For instance, length of stay in hospice is not normally distributed;
median is approximately 21 days, while mean is closer to 60 days.
(One feature of a normal distribution is relatively equal means
and medians). This suggests that statistical models that do not
require normal distributions (e.g. non-parametric methods) might
be more correctly employed to study factors influencing length
of stay. A more appropriate approach would be to use survival
analysis techniques such as Kaplan-Meier curves and Cox Proportional
Hazards modeling which both allow for censoring (missing information)
and non-normal distributions. See Virnig BA, Ash A, Kind S, Mesler
DE. Survival Analysis using Medicare Data: Examples and Methods.
Health Services Research, 35(part III):85-101, 2000 for examples
of the non-normal distribution of hospice stays. http://www.hospitalconnect.com/hsr/ArticleAbstracts/Volume35.html
Secondly,
there may be problems related to non-independent observations
because subjects are grouped into hospices, hospitals, and counties.
The extent that this issue biases study results must be assessed
on a case-by-case basis. Intraclass correlation is a statistical
term that describes the problem that persons in the same class
(county, hospital, hospice) will be more similar to each other
than to persons in another class (county, hospital, hospice).
While point estimates (means, regression coefficients) are not
affected by this problem, ignoring clustering may result in an
under-estimation of the variance. In other words, confidence intervals
may be too narrow. Using claims, however, this is less of an issue
because for national or state estimates, power is usually ample.
Finally, despite
large numbers, low power is always an issue. Although approximately
380,000 individuals were enrolled in the Medicare hospice program
in 1999, analysis may still be problematic when focusing on small
geographic areas, race groups, or specific diagnoses. When hospice
users are aggregated by county, age, sex, and race, cell counts
quickly become small.
|