I am trying to figure out how to analyze mixed-effect model («lme4» package, lmer function in R). Due to lack of experience, I have some questions that I would be grateful for help with. Currently I am analyzing the dynamic of substance concentration in different patients during the 6 days. As my sample has many missed values, and data are paired, it seems that mixed-effect model is the only suitable analysis here. Questions:
I would really appreciate any help.
library(lme4) library(lmerTest) library(ggplot2) library(MASS) library(multcomp) R4
Alexey Zulkarnaev
You may employ multiple imputation (MI) because you can create multiple plausible replacements of the missing data.
Based on ANOVA, using MI has the additional benefit that it allows taking covariates into account that are relevant for the missing data but not for the analysis. If you run MI, then it consists of three steps.
1.The missing data are imputed multiple times.
2.The imputed data sets are analyzed separately.
3.The parameter estimates and hypothesis tests are pooled to form a final set of estimates and inferences.
Try employing the mice and mitml packages to conduct MI.
Good luck.
Alexey Zulkarnaev
You may employ multiple imputation (MI) because you can create multiple plausible replacements of the missing data.
Based on ANOVA, using MI has the additional benefit that it allows taking covariates into account that are relevant for the missing data but not for the analysis. If you run MI, then it consists of three steps.
1.The missing data are imputed multiple times.
2.The imputed data sets are analyzed separately.
3.The parameter estimates and hypothesis tests are pooled to form a final set of estimates and inferences.
Try employing the mice and mitml packages to conduct MI.
Good luck.
Hi Alexey Zulkarnaev
You most probably you FIRST need to deal with missing data through multiple imputation (MI) to create multiple plausible replacements of the missing data based on three steps: (1). imputation of missing data over multiple times, (2). analysis of the imputed data set separately, and (3). pooling of parameter estimates and hypothesis tests to form a final set of estimates and inferences. Thereafter you can do preliminary analysis to check key properties.
Dear Sir, I intended to implement MI as suggested by Pradeep Paraman. Yet, I found out you - in my opinion - have way too much missing values. In the script in attachment I reshaped your dat from long to wide format. When looking at the data in this format I discovered you have too much missing values, which most brobably are not missing at random. The most conservative test would be replacing all missing values by the mean for that day. When I run the one-way repeated measures ANOVA on these data I found the Time (6 Days) are significantly different. Post Hoc each day from the other.
The script in attachment may take a while. Be patient! It takes about 5 minutes.
Pradeep Paraman , Elvis Munyaradzi Ganyaupfu - thank you for your recommendations.
Dear Jos Feys, thank you very much for your response and your time taken. I have a little bit more questions.
1. Does the analysis implemented by the code you quoted fully fit for the original dataset (imputation of means instead of NA)? Wouldn't it be a source of bias?
2. NA are replaced with means. Considering that the distribution of conc is different from normal on almost all time points (days), wouldn't it be better for me to change NA to medians or geometric mean? Or is imputation of means a common practice in such a case?
3. If I got it correctly, you gave three methods:
- “classic” ANOVA (linear model) – «aov»,
- repeated measures ANOVA - «ezANOVA»,
- nonlinear mixed-effects model – «nlme».
I worked with ANOVA, RM-ANOVA and simple linear mixed effects models before. I don't have enough experience to determine which of these methods is more appropriate in this case. Would you be so kind to give me some advice?
4. Did I get it right that since the assumption of sphericity is not met (ezANOVA, Mauchly's Test for Sphericity p
Yes, you cannot avoid the fact that you have way too much missing values and that these missings are most probably not 'missing at random'. So, in reporting any analysis, you should first admit these failings. For dealing with such problems, see:
https://www.researchgate.net/post/If-missing-data-is-not-MCAR-how-to-proceed-further-using-SPSS
Yes, you can replace the missings with whatever you think might be more appropriate. Maybe you can run several inputation methods. If they all give the same results, you are 'safe'. For many different way to impute, see:
http://www.stefvanbuuren.nl/publications/mice%20in%20r%20-%20draft.pdf
and attachments.
An ANOVA is one of the many ways to test a linear model. They essentially yield the same results. Just use the one you are most familiar with.
nlme is not a non-linear model; nlmer is non-linear. Yes, do correct for Sphericity (heterogeneity of covariances).
When you use nlme, yes the error term is essentially: random = ~1|patient, because you have repeated measures. These have smaller error terms.
Dear Jos Feys , thank you for your help in both theoretical and practical points!
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA