I am currently running a multilevel-analysis, using two levels. My data consist of 6400 intervals of observations of 60 children. The observations were done in rounds, each consisting of 30 intervals . The dependent variable I want to analyse is a variable that for methodological reasons has been aggregated on the round level(N=214). It is an ordinal variable. I would like to add some binomial predictors that have not been aggregated and are on the interval-level (N= 6400)- Is it allowed to do this, or should I first also aggregate these predictors (with the risk of getting less refined outcomes?
Sincerely hope someone can help me since I can't find information about this in the FAQs and courses.
multilevel-analysis, predictor
Re: multilevel-analysis, predictor
Hello to you all,
Earlier on I posted my question regarding a multi-level-analysis I want to execute. Untill now nobody has answered it. Could it be that I have not explained my problem well enough? If so, please let me know if you need more information?
I sincerely hope someone can help me out!!
Thank you very much for all the hints or solutions you can give,
Ine
Earlier on I posted my question regarding a multi-level-analysis I want to execute. Untill now nobody has answered it. Could it be that I have not explained my problem well enough? If so, please let me know if you need more information?
I sincerely hope someone can help me out!!
Thank you very much for all the hints or solutions you can give,
Ine
Re: multilevel-analysis, predictor
Hi Ine,
One always has to model at the level of the response i.e. if you are looking at what factors are affecting something measured at the round level then the predictors have to also be at the round level. If you have predictors that are measured at a finer granularity then you have to aggregate in some way to the round level. There are off course lots of ways to aggregate - mean, variance, min, max, range, percentage-yes etc.
Hope that helps,
Bill.
One always has to model at the level of the response i.e. if you are looking at what factors are affecting something measured at the round level then the predictors have to also be at the round level. If you have predictors that are measured at a finer granularity then you have to aggregate in some way to the round level. There are off course lots of ways to aggregate - mean, variance, min, max, range, percentage-yes etc.
Hope that helps,
Bill.
Re: multilevel-analysis, predictor
Is it not possible to have a household-level outcome variable (say household income) and yet model data with the individual as the lowest-level unit? Therefore simply allocate members of the same household the same value on the outcome variable (same household income), and then account for that interdependence by including a random effect for the household membership? So model the individual's income (though really it is measured as household income) using individual-level and also household-level predictors?
Re: multilevel-analysis, predictor
Hi Ine,
What you are proposing would be a false replication i.e. you will be pretending you have more data than you have (particularly if you didn't put in the level 2 random effect). In fact once you put the random effect in unsurprisingly it will show all the variance is at level 2 and might in some software even cause estimation errors due to the 0 level 1 variance. This is different from a household level predictor for an individual level response which is fine.
As I said in my original reply you should aggregate any predictors to the level of your response and in fact if you try your approach and include the random effect at level 2 then you'll be doing something similar although you will be giving more weight to larger households in your analysis.
Hope this helps,
Bill.
What you are proposing would be a false replication i.e. you will be pretending you have more data than you have (particularly if you didn't put in the level 2 random effect). In fact once you put the random effect in unsurprisingly it will show all the variance is at level 2 and might in some software even cause estimation errors due to the 0 level 1 variance. This is different from a household level predictor for an individual level response which is fine.
As I said in my original reply you should aggregate any predictors to the level of your response and in fact if you try your approach and include the random effect at level 2 then you'll be doing something similar although you will be giving more weight to larger households in your analysis.
Hope this helps,
Bill.