Medical Uses of Statistics by Hoaglin David C. Bailar John C. & David C. Hoaglin

Author:Hoaglin, David C.,Bailar, John C. & David C. Hoaglin
Language: eng
Format: epub
Publisher: John Wiley & Sons, Inc.
Published: 2012-01-10T05:00:00+00:00

INTERPRETATION OF RESULTS

A review of statistical methods in papers published in the New England Journal of Medicine in 1978 and 197995 indicated that a reader who knew 11 categories of methods would have access to the statistical methods reported in 90% of those articles. The results in Chapter 3 show that this would not be true in 2007–2008. This change is due in part to developments in statistical methods, but more to growth in the complexity of the data, attributed to the increase in the number of outcomes and the number of time points and to the need to deal more appropriately with missing data. This section discusses simple interpretations of statistical tests and parameters common in reports of RCTs and other parallel comparisons of treatments.

Summaries of results often combine statistical tests of hypotheses and estimates of parameters. The tests yield numbers, such as a p-value or a test statistic, that do not translate into the scale of the data. For example, for continuous data, a t-statistic for a difference (a ratio of the difference between groups to the standard error of that difference) is less informative than the difference and its confidence interval. In addition, means and variances are generally much easier to combine in meta-analyses than are p-values. As a general rule, estimates of parameters, with confidence bounds, are much more informative than p-values, and statistical software can now calculate confidence bounds for nearly every statistical test. Their use is strongly recommended.

Table 6 lists interpretations of common parameters. With the exception of regression analyses of linear or mixed models that did not present regression coefficients, all of these parameters were commonly cited in the 73 reports of RCTs published in the New England Journal of Medicine in 2003. For categorical data the most prevalent analysis rendered the endpoint binary and carried out some form of survival analysis on the time to reach the endpoint, not just cumulative rates. When cumulative rates were the outcome and the event was rare, logistic regression was appropriately used because the odds ratio approximates the risk ratio, which is more easily interpretable for clinical decision making. (The odds ratio has strong theoretical and computational advantages over the risk ratio, though it is harder to interpret.) When events were common,85, 86, 96 log-linear models (such as Poisson regression) were appropriately applied, yielding estimates of relative risk rather than relative odds.

The number needed to treat (NNT) was rarely presented,97 but it is easily calculated from the crude rates98, 99; it equals 1 divided by the difference in rates. Table 7 shows calculations of NNT for five studies involving persons with serious medical conditions where treatments were effective in improving outcome in at least one subgroup. McConnell et al.97 calculated the NNT for three pharmaceutical combinations and placebo from the difference in life-table estimates of the cumulative incidence probability at four years. To illustrate, the data presented in their Table 2 indicate that the rate of clinical progression of prostate cancer was 17% in the placebo group, but only 5% in the combination doxazosin and finasteride group.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.