Hello everybody, I'm using cluster-mean standardized repeated measures data (intensive longitudinal data) in multilevel models. However, I only used part of the data in the models (e.g., the first, second, or third responses of each sampling day). The question I’m wondering about is whether what I did is correct – standardizing the whole dataset but only using part of it in the models. Would standardizing the data used in each model separately be a better practice?
Thanks in advance!
Does it make much of a difference whichever way you choose; if so, why? If you report on this , it gives the reader/ referee more confidence that you have been thorough.
From an efficiency point you might want to only standardise what you use. However when there are omissions or outliiers you might want to go back to source and include additional data.
preference is to standardize all source data en a similar manner.
a good approach is ISO 13972 on clinical information models. The core of that is specification of each data element in a standard manner. It includes data types (number, coded value, Boolean, text etc).
datatypes might differ, ridit analysis (Bross, 1958) might be useful to standardize at analysis level.
Thanks, Kelvyn Jones for answering this question. Of course, reporting what had been done is essential. But if I want to compare the standardized betas, would standardizing the datasets used in each model separately make the standardized betas more comparable than standardizing the whole dataset?
Personally I am not a fan of standardized betas ( I have learnt). I prefer working with the raw betas and then doing a range of predictions for plausible values of the predictor variable (one at a time setting other predictors at their mean value) seeing what is predicted response if the value is changed from the lower quartile (say) to the upper quartile. This is then done in terms of the actual measurement of the response value and I can make a (scientific) judgement about what is a large effect or a trivial one. I often make a comparative plot for each predictor variable using the same scale for the vertical axis so as to directly compare..
In our software we built a tool to do this for complex models with interactions and discrete outcomes and multilevel structures with repeated measures; it works by simulation which is needed for the complex models and provides confidence intervals and coverage bands: Article Manual supplement for MLwiN Version 2.14
For the problem for standardised betas see http://www.pmean.com/09/StandardizedBetas.html
I had thought standardizing on the cluster mean was the key question, on which see
Article Understanding and misunderstanding Group mean centering: a c...
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA