I attached a data here called image_study. The “image_study.csv” dataset contains data from a study examining the speed of classification of two different types of image: artworks, and natural images.
There are four variables in the data as follows:
- participant – a unique identifier for each participant
- image – a unique identifier for each image
- image_type – a factor indicating whether a given image is an Artwork or a Natural image.
- RT – reaction time in milliseconds
I want to model reaction times (RT) as a function of the other variables in the dataset. The purpose here is to examine the effect of Image-type on RT.
1. Which of the fixed effect, random effect, or mixed effect model modes should I choose?
2. Which participant or image variables should be considered as a random effect in the model?
One of the statisticians gave me the answer to this question:
Since you say:
The purpose here is to examine the effect of Image_type on RT.
then we need a regression model with RT and the response and image_type as the main exposure (fixed effect).
Since you have repeated measure within both participant and image then you should fit random intercepts for both of these.
Furthermore, since participant and image appear to be crossed (from the data picture all images were seen by participant 1 and that appears to repeat for participant 2), then the following model should be appropriate:
RT ~ image_type + (1|participant) + (1|image)
which is the syntax you would use with popular R packages such as lme4. eg:
library(lme4) model
Dear Mehdi Azizmohammad Looha
The choice of estimators should be guided by your goals & facts characteristics in preference to simplest by a Hausman check since the test isn't always accurate in all cases. considering that you've got larger among outcomes than inside outcomes, it might no longer be accurate to move for constant effects estimation, simply by way of judging from the Hausman take a look at. The FE version makes an assumption of homogeneity (which does no longer exist on your case) whereas RE version will permit for modelling heterogeneity throughout devices.
A RE model calls for that the group-degree effects & the explanatory variables ought to be uncorrelated; in such cases, RE estimation is unbiased, regular & efficient as it uses each inside-and- between group version while FE makes use of best inside-group variation. In instances wherein institution-level results are correlated with the regressors, RE estimation becomes biased. using RE with organization means then, would be a better alternative. choice of a model is essentially a change-off among precision & bias.
Hope this helps.
Kind Regards
Qamar Ul Islam
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification