Read on Twitter
Yesterday at the #preregistration session at #LiveMEEG I mentioned that pilot data can be used to get an estimation of an effect size. How can this be done in an unbiased way? Thread 
1/n
1/n
First – a VERY IMPORTANT clarification – in this case, “pilot” doesn’t mean collecting a small number of data points. It actually means collecting a large enough sample. 2/n
The reason this is called a pilot is that you don’t actually run the stats, just use it to get the effect size. So how would that work in practice? 3/n
Suppose you plan 5 experiments, designed to interrogate some effect. Collect an initial pilot dataset with a big enough n (e.g., assuming a medium sample size). Then calculate the effect size based on your pilot. Use this to inform the required n in the subsequent experiments 4/n
Pros: (1) this is similar to estimating effect size based on previous literature, but not subjected to publication bias, and (2) you have more control on the design which can be highly similar (or exactly the same) as the one you’ll actually use. 5/n
Cons: requires additional time and resources, so might not be feasible for expensive (eg imaging) studies. 6/n
More info in these talks: https://bnacredibility.org.uk/recordings @BritishNeuro and thx
@rikhens
for this idea and discussion. n/n
@rikhens
for this idea and discussion. n/n
