Asking validation of Use of Mixed Model

[#1]

I have tens of thousands of rows of data, where each row is an object and columns are predictors. My columns of interest are sensory measurements on a 10-point scale. From variable screening, I have reduced my predictors to 3 categorical variables.  I am interested in identifying which levels are most influential (I have 33 total levels), and the direction of the influence.  I do not need to make numerical prediction targets; an explainable qualitative model is sufficient. 

I am thinking a mixed model would serve my purpose; the fixed effects parameter estimate p-values would allow me to identify those specific levels to focus on.  Question: Would their fixed parameter estimate be a way to rank impact, and does the sign of the estimate indicate the direction of the effect (i.e., + means increasing the level increases the response, and decreasing the level decreases the response)?

My apologies but I have never used a mixed model before :(.  Thanks in advance for any feedback.

[#2]

Is it possible to show a picture of the top rows of your data table? It would help me imagine what you are working with.

Why a mixed effects model? What is the random effect that you need to account for?

Yes, your interpretation of the parameter estimates is correct. You might also examine the Effect Details for each significant term. The details include the LS Means for each level. Optionally, you can plot the LS Means and you can test differences between specific levels or combinations of levels with Contrasts.

The Prediction Profiler is also a great tool to understand the effect of each level of each categorical factor. Note that it is based on the model, so if you have significant interactions among the predictors, this interactive plot is even more valuable to understand the combined effects.

[#3]

Hi Mark, I sent the headers to your email box.

[#4]

Actually, I meant the top so I can see both the column heading and a few rows of data. I am sure that your description is accurate, but it is still sufficiently vague for me.

A 'linear mixed model' contains a mixture of fixed and random effects that are represented by terms in the linear predictor. So, if you do not have a term for a random term, then you do not need a mixed model. A fixed effect is one for which you want inference and expect it to be reproducible. You are interested in the deliberately selected values of the predictor. A random effect is one for which you want to estimate the variance. The value represent only a sample of a larger population.

[#5]

Thanks. I'll stick to the standard regression suite.