1

I'm trying to compare three groups data. But the data set is about a new drug trial. The data set has these characteristics:

  1. Follow-up set. That is, after administration of the drug, a series of parameters were collected in the following date, day0, day1,day2,...,day28.
  2. Unbalanced set. Because some patients died and some recovered, not all patients were followed to 28 days; for instance, some were followed for only 20 days.
  3. Non normally distributed parameter. Especially, day25-day28, sometimes sd >> mean.

So, I try to use the linear mixed-effect model with group, time and group*time for the comparison. ANOVA is not suited for the unbalanced set.

However, I wonder if it's correct model for non normal data set? Or is there another method suited for this kind of set?

I use SPSS 19. Some of my friends use STATA or JMP.

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
user14866
  • 21
  • 2
  • Can you clarify your concern regarding the non-normality? Are you worried that the raw response data aren't normal, or that the residuals aren't? (NB only the residuals matter.) Is it that the response is categorical (eg, 0 or 1), or a count? What do you mean by the *parameter* being non-normally distributed? – gung - Reinstate Monica Oct 11 '12 at 16:36
  • Independent variable is IL-6 while factors are group (control, drug1, drug2) and time (0,1,2,3,...,28). Raw data of IL-6, at day0-day7, are normal while at day28, mean+/-sd is 47.2 ± 53.1. Obviously, skewness distribution. Like this, Linear Mixed Model is OK for it? Thanks a lot! – user14866 Oct 11 '12 at 18:23
  • That's helpful, @user14866. Do I understand correctly that your concern is that the raw response data are positively skewed towards the end of the study? – gung - Reinstate Monica Oct 11 '12 at 18:54
  • Sorry, accurately, I should say that raw response data become non-normal by normality test. You may see it via mean – user14866 Oct 11 '12 at 19:36
  • I think you should probably just ignore the normality test (see here: [is-normality-testing-essentially-useless](http://stats.stackexchange.com/questions/2492/)). Beyond that, is the situation here that your response variable can only be >0, &, since SD>mean, the raw response data are positively skewed towards the end of the study? Is that what you are worried about? Is that what motivates this question? – gung - Reinstate Monica Oct 11 '12 at 20:04
  • Exactly, responsible variable can only >0, never <0. Now, I know, we needn't care whether the normal distribution. Only the descriptive shouldn't use mean+\-sd instead of Median[Q25,Q75], right? But someone suggest to use NLMIXED model in SAS for this kind of data set. I don't understand it. What's your opinion? – user14866 Oct 11 '12 at 20:38
  • I'm out of practice with SAS, so I'll have to let someone else answer that part. I would suggest you look for a transformation of the response variable (all of it, not just the last few days) than will diminish the skew. I don't know about the nature of your response variable, but commonly w/ vars that are >0, taking the log works well. The result doesn't have to be perfect, just OK. Then proceed as normal. – gung - Reinstate Monica Oct 11 '12 at 21:02
  • I found a literature on analysis of repeated measures designs (British Journal of Mathematical and Statistical Psychology 2001,54:1-20). It's a review. very helpful. – user14866 Oct 12 '12 at 09:04

0 Answers0