Violation of assumptions mixed ANOVA

[#1]

Hello everybody

In the scope of my master's thesis, I should analyse a mixed ANOVA. I study postural control of children and adults during different surface conditions. The test procedure was the following: all participants, divided into age groups (16 children and 17 adults), stood on 4 balance boards of different difficulty. I went for a mixed ANOVA, with age group and surface condition as fixed effects and participant as random effect.

When checking the assumptions, the following problems appeared:

1) The residuals were not normally distributed. Transforming the data (log) solved this, but then I'm unsure about the quality of the conclusions. I have read about ANOVA's being robust and not to worry too much about this, is that right?
(Note: First, I ran a fit model (personality: mixed model), then I saved the residuals. Afterwards, the residuals were analyzed for distribution.)

2) When checking for homoscedasticity, this also appeared to be violated. When checking this for the data (without transformation), the next graph was found. I concluded this assumption to be violated. 

When checking this for the Log(data) (with transformation), the next graph was found. I'm not sure about this one... This seems to be okay, but I'm still unsure about the transformation.

So right now, I'm not sure about how to handle these assumptions. Should I stick with the transformation, use another correction, or just rely on the robustness of an ANOVA?

Could you help me out with this one? 

Thank you in advance!

P.S.: I'm not very confident using JMP and statistical models, so it would be nice if you replied using rather simple language.

[#2]

@Sam_VH , welcome to the JMP community!  As a rule of thumb, it is good to check the assumptions of your quantitative analysis. You do this to get a better understanding of your model, how well it describes the data and are there times when your model is inadequate and therefore does not describe the data well. Since you did not attach the dataset, I am left with guessing about what may be going on in your dataset.  It is true that there is some degree of robustness of ANOVA to the underlying data distribution as you are performing analysis on sums of squares.  What exactly is the response variable? Is the measurement system consistent?  Have you quantified the measurement error? Cleansing of the data prior to quantitative analysis is, unfortunately, required for almost any data set.  There are a number of things to look for.  For your within treatment variability (the individuals within age group), are there any unusual data points?  Transformations can "clean up" data irregularities, but this also makes the interpretation more challenging as the analysis is on the transformed data.  What you can do is perform analysis on different transforms as well as the raw data and look for how the results compare.  Do you find similar significant effects in all analysis (not worrying about the actual value of the p)?  If so, then perhaps transformation doesn't matter.  If they are different, then you should seek to understand why.

[#3]

Thanks a lot! I will check the data again for any irregularities. Afterwards, I will compare it to the raw data to see if anything went wrong in the data processing. Unfortunately, I don't know what the University policy is considering the sharing of data, so I have to check in on that before I do that.

[#4]

In addition to @statman's help, you might consider an alternative change to your model. Transformations are a well-proven solution, but the change the meaning of your response. Another solution is to model the variance, too. Add an effect to the variance model that might stability the residual plot (eliminate the pattern).