In an experimental study with, for eg, 3 conditions, is Little's MCAR test sensitive to variability in non-response (ie mising data) across conditions if condition (eg 1, 2, 3) is included as a variable in Little's test?
Hi Kate,
I interpret your question as asking whether Little's MCAR test can accommodate categorical variables in its test of missing completely at random. The answer is yes. Not sure what software you are using, but in SPSS for example this can be done in the MVA module and by assigning variables as either continuous or categorical. This will work for your experimental design. As an addition check, I would also suggest examining the drop-out rate across the 3 treatment groups and test whether differential dropout is associated with key variables.
What Little's test does is test whether the means of some variables differ between responders and non-responders on other variables. Thus, if you include the categorical variable condition, its difference in means will be tested, which is probably not very informative (e.g., for responders the mean condition is 1.3 and for non-responders 1.8). Also, the difference in means of other variables will be tested between responders and non-responders of the variable condition. This is not what you want to know I think, as it does not take into account the condion itself, but only whether there is a missing on the variable condition. That is, it does not test whether the difference in means between responders and non-responders on some variable (e.g., your outcome variable) differs per condition. Which is more interesting.
Many thanks for your thoughts and suggestions, James and Mark.
To be more clear, what I would like to know is whether missingness on values for variables is related not only to mean scores on variables across the entire sample, but also whether missingness is related to which condition participants were randomly assigned to (ie 1, 2, or 3). From your response, Mark, it appears that the answer is no. Is there a way of testing this?
Btw, James, yes - I am using SPSS and Amos to analyse results.
Thanks for any further thoughts.
Hi Kate,
You can test differences in means on key variables in every group. Or maybe even do anova's using factors condition and a dummy for missingness. Or maybe it is possible to perform Little's test within each condition (by using split file in SPSS). But first I would simply check response rates in each condition. Crosstables with dummy for response and condition is also an option, including the corresponding chi-square test and phi-coëfficiënt.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis