Content
Summary
Determining a sufficient sample size (N) is essential for statistical power, generalizability, and robustness against outliers, though no universal 'magic number' exists. Researchers should use power calculators to estimate N based on specific methods and effect sizes, then collect at least 10% more data to account for potential losses. This approach ensures the final analyzed dataset retains adequate statistical properties despite artifacts or attrition.
Intuition
Think of sample size as the foundation of your statistical inference. If N is too small, your results may be dominated by noise or outliers, making them unreliable and non-generalizable. However, you cannot simply pick an arbitrary large number; the 'right' N depends on the complexity of your method, the expected effect size, and data quality. For example, in human scalp EEG, N=20 is a common rule of thumb, but this is not a strict standard for all fields. The key is to estimate what you need using power calculators and then buffer that number. Since data collection is rarely perfect, you will likely lose some data points due to technical artifacts or subjects dropping out (attrition). By collecting extra data upfront, you ensure that your final, clean dataset is still large enough to yield valid conclusions.
Derivation
- Statistical power and generalizability require a minimum N to avoid sensitivity to outliers.
- No universal N exists; it is determined by method, paradigm, effect size, and data quality.
- Power calculators (e.g., Faul et al., 2007) are used to estimate the required N.
- Data loss occurs due to artifacts, technical problems, or attrition (subject non-completion/death).
- To compensate for loss, collect at least 10% more data than the estimated N.
- Final analyzed N must still meet the power requirements after exclusions.
Worked example
A researcher estimates they need N=50 subjects for a new EEG study using a power calculator. How many subjects should they recruit to account for potential data loss?
- Identify the estimated required sample size: N_estimated = 50.
- Determine the buffer percentage recommended for data loss: 10%.
- Calculate the additional subjects needed: 50 * 0.10 = 5.
- Add the buffer to the estimated size: 50 + 5 = 55.
- Conclusion: The researcher should recruit at least 55 subjects.
Answer 55 subjects
Common mistakes
- Assuming there is a universal 'magic number' for sample size that applies to all research contexts.
- Failing to account for data loss due to artifacts, technical problems, or subject attrition when planning sample size.
- Treating N=20 as a strict standard for all EEG studies rather than a rule of thumb.