Hi, everybody. The next topic is studies with multiple levels of clustering. So far we've talked about longitudinal studies which have repeated measures over time, and we've introduced the idea of clustering by taking a look at a single level study. In this lecture, we are going to define multi-level studies, be able to recognize them, and describe their advantages. Just like with single level studies, we will make sure we need to understand the different authors, and different fields use different terms to describe the same study designs. Then, we will finish up by discussing how multilevel design induces correlation. In multilevel study, in which groups or participants are defined along two levels. Again, as we mentioned, there's different terminology used to describe this including; two-level study, two-level group design, two-level cluster design, or two-level hierarchical design. Just like a single level study, members of a level also known as a group or cluster share experiences which leads to their measurements being correlated. Think about student within the same classroom or patients that go to the same doctor, all sharing experiences that lead to correlation. Multilevel studies involve two or more of these types of layers of correlation. Keeping with our school example from the previous lecture, a multilevel study design in education could include; the schools as the independent sampling unit, classrooms within schools as the first level that induces correlation, and the students within these classrooms as the second level than induces correlation. Once more, these correlations are induced by shared experience of the clusters members. So, should having the same teacher or the same administrative leadership within the school. In the healthcare example, a clinic would be an independent sampling unit with each of its physicians within the clinic as the first level of clustering, and the patients of each physician as the second level of clustering in the model. You are probably familiar with randomizing individuals into intervention and control group. In the following example, we're going to talk about just the same type of randomization just with clusters of multiple people instead of just individuals being randomly placed into treatment and control groups. The example study will look at until researchers conducting a multilevel cluster randomized trial to evaluate the effectiveness of a web literacy intervention called ABRACADABRA. The reason is clustered randomized trial is the randomization was of schools, meaning, schools were randomly assigned to different treatments. For clarity of presentation, some of the details we present as for most of examples are quite different from the published study reference. Here's some important details of the study. There are 12 schools in the district in which the study took place, and each school has two classrooms for a total of 24 classrooms. While the schools are all in the same district, they are under the same control, they are considered independent of one another. Schools were randomized in this study to either the intervention group or the control group. The effectiveness of the [inaudible] intervention was measured using a pre-post test taken by all of the student. Now that we have the background information for the study out of the way, here is the flowchart. First, schools were randomized into the intervention group, the ABRACADABRA schools or the control group. The difference between the pre-post literacy measure were examined for each group. The null hypothesis is that there is no difference in literacy between students in the two groups. So, it's just the mean for change and this is a pre-post design. We're going to measure the literacy at the beginning and measured after, so it's longitudinal as well. Now that we can simplify this study by thinking about different schools, however. As we talked about earlier, the multiple levels of hierarchy induced correlation layers, to isolate the impact of the intervention, will need to adjust for the correlations involved in this study. We will talk about the data analysis and how we actually make these adjustments at a later time. For now, we're focused on recognizing correlation so that we understand the need for an adjustment. The levels of correlation come from students within classrooms and classrooms within schools. If we don't make an adjustment, it will inflate the Type 1 error rate. Inflating the Type 1 error rate will affect the ability to get accurate calculations for power and sample size analysis. On a side note, you may be wondering why it is not being treated as a longitudinal study with correlated scores from the pre and post-test. This is because the researchers decided to only look at the different scores between pre and post rather than considering the scores as measurement. Therefore, we only have one measurement per student which is the difference score, and we will need to consider the longitudinal nature of the study in the correlation. So, the independent sampling unit is the score as they are independent of one another and the researchers randomized whole schools into treatment and control groups. The unit of observation is actually two levels below that, which is the difference between pre and post test scores for individual children, and the actual statistical analysis involves taking an average of the kids in a classroom, and then the kids in a school to get scores for the independent sampling unit. The between independent sampling unit was the randomization group, with schools being randomized into intervention and control. The within factor was the cluster member, but notice that there were two levels of clustering within factors, with classrooms in schools and students within classrooms. In this example, we're not interested in the within factor or cluster members from his specific scientific perspective. In other settings you may be interested in this. For example, you might be interested in looking at the differences within a school if you were the school principal, but this is not the scientific objective here. It turns out the intervention was in fact associated with greater improvement in literacy. Researchers were able to land on this conclusion by isolating the intervention by adjusting for correlation structures induced by the multilevel design. Let's do a quick review. Summary. Multilevel studies have multiple organizational levels, like schools that have classrooms within them, then classrooms to have students within them. This induces multiple layers of correlation that we have to account for. This correlation is the result of similar experiences being shared within levels. For example, students all being taught by the same teacher or teachers working all within the same school culture. Accounting for this so correlation, allows us to isolate the factors being studied like treatment, for example, as the primary causes of the outcome we want to find. Well, this wraps up this lecture. Thank you for your time.
