Multi-level modeling?

I've spent the last two days reading and attempting to find an answer to my question but I am unsure if the answers that I am finding are even accurate because to be honest, I am not sure what to call the analysis that I am attempting to use. I am using SPSS for analysis procedures, so any help that is relevant to SPSS would be helpful.

I am examining how several different measurements predict academic achievement. The participants included students from six different classrooms. A reviewer asked if the classrooms were equivalent. After running some analyses, I came to the conclusion that the classrooms were not equivalent. I previously was utilizing a regression procedure for this analysis, but I am unsure of how to proceed. I believe that I want to examine my data from a nested, HLM, MLM perspective? I attempted to include the class into the regression analysis; however, the class is a nominal variable with six levels. From my searching, regression cannot handle more than two levels of a nominal variable without dummy coding. I completed dummy coding and attempted to run the analysis and the results became very confusing to me because I was now comparing the results of five classes against one class and to be honest, I wasn't really sure what I had anymore. I spoke with a statistician who recommended that I use an ANCOVA design with the class as the factor and predictors as covariates and a custom model to examine the main effects of all variables. I did this and have run analyses in two separate ways. (1) I ran this with the class variable entered as the factor and (2) with the 6 dummy variables of the class variable as the factor. Which one is correct? The results are wildly different and I am not sure if I am answering my research question anymore. I hope this made sense, but if not, I will try to clarify. Thanks in advance.


Less is more. Stay pure. Stay poor.
Academic achievement is continuous?

Measurements were not randomized to classrooms, so you have risk of confounding, correct?

Each classroom has a single unique measurement of interest?
academic achievement and the four predictor variables are all continuous. All students completed the same four measures that constitute the predictor variables.


Less is more. Stay pure. Stay poor.
What is it about the classrooms that would effect the outcome or the predictors effects on the outcome?
Some of the classes have different levels of performance on the outcome variable. Other classes display differences in the level of predictor variables. You might argue that the students who perform better on the outcome variable are just smarter or higher achievers, unfortunately, the prior achievement variable that I do have is riddled with missing data.

To be honest, I am unsure if a nested model design is appropriate here. Relatedly, the sample size is also relatively modest (N=100) so I would be underpowered for such analyses.


Less is more. Stay pure. Stay poor.
So on average each class is missing about the same amount of data and there is no known reason?

If you have reason to think certain classes may have different effects and you are not able to control that any other way - run multilevel modeling.
You may have power issues regardless. Do you know about random effect and random slops?