How should I think about sample size assessments when I’m collecting data versus using existing secondary data sources?

To determine how many observations to collect, researchers consider cost, the desired effect size, power, type I error, and the number of statistical comparisons needed. Typically, studies seek 80% power, less than 5% error, minimal costs, and effect sizes based on previous studies (often conducted by other researchers). If a researcher collects his or her own data, while it is ideal to meet all of these criteria, it is often difficult because new ideas generally warrant pilot studies (with fewer funds), which means fewer observations. This means the research question should be well defined and clearly informed by the previous literature as much as possible to get the biggest bang for your buck!
When using existing data, general sample size rules suggest that 10-25 observations per variable to be included in the analysis are required to minimize serious bias and ensure sufficient variability within the data to make valid inferences. The inclusion of too many variables in a single analysis limits the validity of the study and their findings, and should be avoided. This means that the researcher may be required to run various sets of models to examine the sensitivity of the analyses to the variables included in the model- choose carefully and intentionally.

About the Author

Dr. Vicki Lawrence is an academic researcher who studies the epidemiologic nature of social conditions in relation to cardiovascular and other disease outcomes. More specifically, her work focuses on studies of poor health among African Americans and health disparities that may occur my age, race, and gender in cardiovascular and mental health outcomes. Utilizing her background in epidemiology and biostatistics, she has provided statistical support on multiple studies with various investigators commonly focused on physical and mental health data. In addition, she has worked with clinicians, research investigators, and tutored multiple graduate students as well in public health, epidemiology, social work, medicine, education, and nursing to tackle statistics related issues.