The Stats Corner and Stats Seminar Series are designed to provide accessible and informative statistical education of a variety of topics that are of interest to faculty and students. The series is presented by Dr. Andres Azuero, Assistant Professor and Center for Nursing Research Statistician.

Dr. Andres Azuero, Assistant Professor and Center for Nursing Research Statistician

Be sure to attend the next Stats Seminar, May 7, 2010 @ 11:00 in the Center for Nursing Research.

A Note on Power Calculations…

The power of a statistical test is the probability of rejecting a null hypothesis (for instance, of no treatment effect) when the alternative is true (e.g.; the treatment actually works). More often than not, the sample size for a study is restricted by a budget. In such cases, power calculations at the design stage of a study provide crucial information regarding the magnitude of the effects that investigators can expect to detect if present. Often times, a calculation under simplified assumptions is sufficient. Standard software is readily available for a variety of simple statistical models. More complex designs may require even simulations. Assuming a fixed sample size constrained by a budget, for quick calculations of detectable differences between two groups (treatment vs. control, or exposed vs. non-exposed, or younger vs. older, etc.) a few items are needed. The first thing needed is to define a primary outcome of interest. Usually continuous or binary outcomes are selected. Second, an attrition rate is needed in order to calculate the effective sample size. The third item needed is the proportion of participants each group or “allocation ratio”. The combination that achieves greatest power is equal numbers on each group, which is standard practice in experimental trials but not necessarily possible in observational studies. For continuous outcomes, an estimate of standard deviation is needed. For binary outcomes, a reference proportion is needed. With this information in hand, power and significance are set at the traditional levels of 0.8 and 0.05 and an estimate of detectable difference can be rapidly calculated. More complex designs such as repeated measures models or group-randomized trials require a little more information.

Remember that as part of its mission the CNR offers statistical consulting services such as power and sample size calculations for grant proposals and manuscripts.