Multilevel Modeling Flashcards
(32 cards)
These are other names for Multilevel models
-Hierarchical linear models
-Mixed Effects Analysis (linear and non-linear)
-Random Coefficient Regression Models
You have to “account or model” the dependencies you have in a within-subjects design… How do you do that?
The Fixed Effects Approach
Why fixed effects?
Fixed because subjects were not randomly sampled. They are considered to be specifically chosen subjects. I generalize to a subject but I can’t generalize to a population
When do you choose a Fixed Effects approach?
When you have small samples. But you can’t generalize
When do you choose a Mixed Effects / multilevel models approach?
When you have a lot of subjects, lots of groups, lots of clusters.
Another name for Multilevel Effects
Mixed Effects, Mike doesn’t like this name but basically “mixed” means you are using both: random and fixed effects here
This approach doesn’t treat clusters or groups as fixed effects… why?
Because it assumes they are random from a bigger population of groups, samples, and stimuli. Pros you can generalize. Cons you lose power
How do you combine multiple modeling with generalized linear modeling
By running the multilevel effects but specifying the data distribution (binomial, gamma, posion)
What is the interpretation of a Random Effect
In essence, they are random deviations from the baseline in a group or a condition. The subject’s intercept or mean level of performance in a particular variable
In a multilevel model we observe two kinds of variance:
Variability within subject and variability between subjects
In a multilevel model what is the population intercept
Fixed effect
Shrinkage
Is what is going on with the group as a whole. The estimates for the individuals will shrink toward the group mean. If you don’t include the fixed effect component you are not allowing shrinkage. And thas no gud
How do you interpret the fixed effects?
They represent the typical effect. It is not considered the average because is fitting at both (group and individual) levels simultaneously
What are some issues when running an overly complex model?
Software struggles more to analyze it. You have more convergence problems, and singularity errors
What is lagged analysis
It is a type of thime service analysis. This allows you to approximate a certain type of time analysis. This when your effect is likely to occur when looking at the prior trial, for instance how much previous behavior is affecting current behavior.
In time series analysis what are the auto regressive models
Models that look compare the actual value to the previous value could be of x or y variables. Is a regretion
Why choose a multilevel model?
Because it allows us to handle:
Repeated measures data
NAs
Imbalances in the levels of your factors
It allows continuous within-subject predictors without doing the wrong thing and having you categorizing them.
Core concept of repeated measures in multilevel modeling
Recognize that the data from one subject is going to be more correlated with one another that it is from another subject
Do this model uses RMSE, MLE?
MLE (maximum likelihood estimation)
Why is it called multilevel modeling?
Because we are estimating parameters at different levels (the average/typical and the individual level)
What are the fixed effects?
Fixed Effects: Are nothing more than the variables you are using as predictors. Can be within-subjects or between-subjects predictors.
What are the random effects?
In essence, they are random deviations from the baseline in a group or condition. When you add a subject, ID or any variable that indicates is an individual, as a predictor, random because technically subjects were randomly chosen.
What is the structure of the simplest multilevel model
Has only intercepts varying across your subjects, no predictors, and no fixed effects at all. Here, what the model is actually doing is estimating the grand mean.
In a multilevel model analysis, this is not happening
Calculating the mean of everything and having a sd and so on (that is aggregation). What it happens is that we specify that some data belongs to one subject, other to another subject, etc. This way we see variability between subjects and within subjects.
