In the video steps of a multilevel analysis, we discuss which steps you need to take to fit a multilevel model to your data. In this video and in the video on centering, we'll talk about some methodological issues that might come up when running a multi level analysis. The first issue has to do with including variables in your model and specifically, what level do we need to include the variables on? In the video my first multilevel model, we saw that multilevel analysis is basically running several regression analysis that were tied together. We also saw that on level two, we basically try to predict the average dependent variable score over level two units using the level two predictors. We discussed this using the examples of students in schools where we try to predict school achievement using IQ and school size. Basically, what happens was that we moved a variable from one level to another. By calculating the average school achievement of school and using that as a variable on level two. We moved the variable achievement, which is clearly a student characteristic to the school level. Where average achievement becomes a characteristic of the particular school. This is called aggregation, moving a variable from a lower level to a higher one. This has some consequences, most notably on sample size. After all, we transform school achievement for which we have a score for each student. Into average school achievement for which we have one score for school. Since there will be fewer schools than pupils, the number of observations on the school achievement variable goes down. Now there's nothing wrong here, on level two the schools are units of interest, and the sample size on level two is just lower than it is on level one. We didn't make a mistake, but it's good to be aware of the fact that the sample size on level two is lower. This means that we have less power on level two than on level one. And therefore that when we move a variable to a higher level, we lose power for defecting an effect of it. Now an example, we obviously also use the raw school achievement scores on level one, so this isn't an issue. The dependent variable is used on both levels, on level one in this raw form and on level two in it's aggregators form. However, predictors usually only exist on one level, and for them it is therefore important on what level they are placed, level one or level two. If we only choose to use an aggregated form of a variable on level two, average pupil, for example. We have a smaller end to estimate the effect of this variable and therefore lower power. So far, we only discussed aggregation moving variables to a higher level. The opposite is also possible, moving variables to a lower level, which is called disaggregation. With disaggregation the opposite happens than with aggregation. We move a variable from a higher level with smaller end to a lower level with higher end. Now Initially, this may seem like a good idea, as a larger end means more power, however disaggregation is rarely a good idea. When we move IQ or school achievement to the school level, we calculated the averages of those variables. Basically, we use a summary measure for them on level two, disaggregating a variable does not transform the variable however. We would just assign the score of the level two unit to every unit that is nested within it. For example, if we disaggregate school size, we give every student from a certain school the same score on the school size variable. Notice the subtle difference, when aggregating we turn to the larger number of school achievement. Or IQ scores into a smaller number of average school achievement, or average IQ scores, there was nothing wrong with that. Average IQ and average school achievements are school characteristics. And a number of schools just happens to be smaller than the number of pupils. However, when disaggregating we assigned the same schools size score to each pupil from a certain school. But school size is not a pupil characteristic. Moreover, this gives the illusion that the size of the school was observed more than once, one time for each pupil in the school. This is obviously nonsense, schools have one size which can be observed only once, just like average school achievement and average IQ. This aggregation then leads to an artificial inflation of sample size and power, while aggregation leads to a natural reduction in sample size. That stems from the fact that we turn the variable into a summary measure. That is a characteristic of the higher level units, of which there are just less than of lower level units. That's it for this video, be sure to also check out the video on centering where I discussed another important methodological issue.
