Dear all
I am measuring learning of an event using an online Reaction Time (RT) task. The task has got three blocks, lets say B1, B2 and B3. Of course, it is the same stimulus-response routine across all the blocks. A general idea is learning gets better, i.e., RT gets slower as the blocks progress. So, B3 will be significantly faster than B1, if the learning was positive. I know I can use 'Block' as a factor to see if the learning happened. But, we already have a 2x3x3 (in which if we include block it becomes 2x3x3x3), which will make the interpretation even more complex. So, we think we shall get rid of block as a factor, by taking the difference between B1-B3. I am somehow not comfortable with this idea of 'B1-B3' as I feel we are missing on B2. Is there a way I can include B2 in to this and yet receive a single value (the dependent variable). I use 'R' for analysis, so analysis specific suggestions are welcome too.
If you have only 3 time points, and data for all people, (B3-B1)/2 is the empirical slope of B1, B2, B3. However, with just three time points I urge caution with any method estimating the slope. The OLS estimate (with is (B3-B1)/2 without missing data) has a large standard deviation and therefore doesn't correlate that high with the true underlying slope (when looking at simulations). Methods like random slope multilevel models (and actually the first example for the R function lmer is a random slope model for reaction time) tend to have correlations closer to the true slope, but sometimes given odd answers with three time points (e.g., negatively correlating with the true values). So, if you are at the designing stage, I suggest more time points. There is lots on response time modeling (e.g., Luce, Ratcliff, ...). THey would stress the need for a model of the cognitive processes underlying "learning" in your example.
If you have only 3 time points, and data for all people, (B3-B1)/2 is the empirical slope of B1, B2, B3. However, with just three time points I urge caution with any method estimating the slope. The OLS estimate (with is (B3-B1)/2 without missing data) has a large standard deviation and therefore doesn't correlate that high with the true underlying slope (when looking at simulations). Methods like random slope multilevel models (and actually the first example for the R function lmer is a random slope model for reaction time) tend to have correlations closer to the true slope, but sometimes given odd answers with three time points (e.g., negatively correlating with the true values). So, if you are at the designing stage, I suggest more time points. There is lots on response time modeling (e.g., Luce, Ratcliff, ...). THey would stress the need for a model of the cognitive processes underlying "learning" in your example.
Thanks Dan.
I am at the designing stage. The B1 (similarly B2, and B3) I mentioned were actually the median of five time points (because each block (B) had five sets). Yes, therefore, there are more time points, than B1, B2, B3. I am trying to explore Luce and others'. So, there must be a way that can give me a single value (as a representative of RT slope for each participant) out of 15 time points, which will better suite my study. Thanks
- Badges
- Science topic
- Similar topics
- Psychology
- Mental Processes
- Learning