The sound practice of medicine requires the ability to use scientific evidence that is based on data and is published in the peer-reviewed medical literature. In this literature, investigators publish their findings using descriptive statistics to summarize data and inferential statistics to test hypotheses.
Judgment is required in the choice of statistical tools and in the interpretation of statistical analyses. Physicians must have a basic understanding of statistics to benefit their patients by being informed and critical users of the medical literature.
[...] The decision on sample size (i.e., how many patients to study) is based on desired power, on estimates of what would be a clinically significant difference (i.e., the minimum difference in mean SBP between the groups that the investigator wishes to be able to detect), and on the known variability in SBP (its SD). If a clinically significant difference between the two groups should be at least 10 mm Hg, and a power of at least 80% is desired patients need to be randomized to each group. [...]
[...] Discriminant analysis accomplishes a similar purpose of modeling a polytomous outcome variable that is determined by several independent variables. Another multivariable analytic tool frequently found in the medical literature is Cox proportional hazards regression, in which the outcome variable is time to occurrence of a certain event. A randomized controlled trial evaluating two different treatments for lung cancer might take survival time as its outcome variable. Cox regression allows one to model the effect of the treatment on survival time, while adjusting for variables such as age, gender, and stage at diagnosis. [...]
[...] Consider the previous example of the antihypertensive trial in which the goal is to compare the mean SBP for subjects given placebo versus subjects given active drug. Assuming that the SBP distributions are bell-shaped, the appropriate test of statistical significance for this situation is the unpaired Student t-test. Expanding the antihypertensive trial to include three (or more) randomization groups, such as diet, drugs, and placebo, means that analysis of variance (ANOVA) is the most appropriate test of statistical significance. Student t-test and ANOVA assign P values to the differences across randomized groups. [...]
[...] If a difference between two population samples is observed (e.g., SBP is lower in the treatment than in the control group), the key statistical question is whether this difference could have been due to random sampling (i.e., chance). The likelihood that an observed difference or an even more extreme difference is due to chance alone is called the P value. If a certain difference between mean SBPs in treatment and control groups were observed, P < .05 would mean: Given that the null hypothesis is true, there is a less than likelihood that this (or a larger) difference is due to chance. [...]
[...] Nominal and ordinal variables usually are summarized by computing the proportions of individuals in each category; means and SDs are not appropriate in these situations. Hypothesis Testing Much clinical research is concerned with detecting associations between an exposure (e.g., tobacco smoking) and a disease (e.g., lung cancer) or between an intervention and a clinical response or outcome. For a typical study, investigators first develop a hypothesis, for example, that the mean SBP in a group given an antihypertensive agent is lower than the mean SBP in a group given a placebo. [...]
Online readingwith our online reader
Content validatedby our reading committee