SAMPLE SIZE AND POWER
SAMPLE SIZEAND POWER
Sample Size Planning Forces specification of outcome
variables, clinically meaningful effect sizes, and plannedstatistical procedures
Leads to a specific recruitment goal
Encourages development of appropriate timelines and budgets
Discourages the performance of small, inconclusive studies
Sample Size Planning (cont)
To estimate sample size requirements, you need to:
1.Specify your null hypothesis and alternative hypothesis(1-tailed or 2-tailed).
2.Select the appropriate statistical test, based on the types of dependent and independent variables you are using.
Sample Size Planning (cont) Estimating sample size requirements (cont):
3. Determine the minimum effect size (relative or absolute risk difference) that you would like to be able to detect.
4. For continuous outcomes, estimate the standard deviation. For dichotomous outcomes, estimate the baseline risk or incidence/prevalence of the event.
5. Set limits for Type I (α) and Type II (β) error.
Type I and II ErrorsWhen the conclusions of your study, based on your data, differ from the truthin your intended sample (e.g., lack of internal validity):–Type I error (α)—Concluding that a difference
exists (i.e., rejecting the null hypothesis) when there actually is no difference (i.e., the null hypothesis is true). –If you set type I error at .05, then all
results with p<.05 will be labeled statistically significant.
Type I and II Errors (cont)
Type II error (β)—Concluding that a difference does not exist (i.e., failing to reject the null hypothesis) when there actually is a difference.Power (1- β)—The probability of finding a difference (i.e., rejecting the null hypothesis) when a difference exists.You can fix your sample size (post hoc) and estimate power, or fix power (a priori) and estimate your requisite sample size.
Sample Size SoftwareCommonly used programs
EpiInfo for cohort and unmatched case control studies: http://www.cdc.gov/epiinfo/
Southwest Oncology Group for clinical trials: http://www.swogstat.org/statoolsout.html
UCLA options for Poisson analysis, correlation coefficients, Fisher's exact test, and more: http://calculators.stat.ucla.edu/powercalc/
Commercial products offer great flexibility (see also SAS, SPSS, Stata): http://www.power-analysis.com/ http://www.statsol.ie/nquery/nquery.htmhttp://www.ncss.com/pass.html
How to Get Bywith a Smaller Sample
1. Double-check your assumptions:– Type I (α) error rate (0.10 instead of 0.05?)– Type II (β) error rate (0.30 instead of 0.20?)
2. Double-check your hypothesis:–One-tailed instead of two-tailed?
3. Double-check your desired effect size:– Is the intervention likely to have an effect of
this size?–Can I settle for being able to find a larger
effect?
How to Get Bywith a Smaller Sample (cont)
4. Use a more frequent dichotomous outcome variable:–Less serious outcome (don=t
sacrifice clinical face validity)–Composite outcome (watch for
undesired heterogeneity)–Lengthen follow-up period–Select higher-risk study subjects
How to Get By with a Smaller Sample (cont)
5.To decrease variability (increase reliability) of continuous outcome measures, use:
–A continuous variable instead of a dichotomous variable, if possible, but don't give up clinical face validity to gain statistical power (e.g., mean birthweight).–More precise, reliable equipment or data collection techniques; consider repeated/ multiple measurements.–Paired (pre-post or matched) measurements to reduce variability; consider crossover design.
How to Get Bywith a Smaller Sample (cont)
6. Increase your sample size cheaply:–Consider unequal group
sizes (2–3 controls per case)to increase efficiencyin a case-control study
7. Compromise or get more money
Common MistakesRelated to Sample Size
Failure to discuss sample size or power in published papers (especially studies with negative results)
Unrealistic assumptions related to disease incidence or prevalence, or effect size
Failure to explore sample sizerequirements for a range of possible values of key variables (sensitivity analysis)
Common MistakesRelated to Sample Size (cont)
Treating Type I (<0.05) and Type II (<0.2) threshold error rates as if they were divinely inspired
Failure to account for attrition by boosting planned sample size
Thank You !