Hello all, I need you advice on how to analyze my data. The data comes from the ophthalmic field (eyes). A small study was conducted to test a new treatment. There was no control group, a single arm study. n subjects were chosen by random, and one eye was chosen by random for each one of them - the chosen eye received a treatment. The second eye did not. Before treatment, some measure of the eye (continuous), I'll call it Y, was measures for both eyes. Then the treatment was applied on one eye, and there were several follow up visits: 1 week, 2 weeks and 1 month. The question was to test if the treatment works, to simplify, I could simply take one time period, say 2 weeks, and run a paired t-test (compared to the Y value before treatment). Ideally, I would like to also check how long does the treatment effect works. The problem was, that I plotted the untreated eye (which was measured at every follow-up visit) and found a correlation (illustration is attached - a plot of 1 subject, no real data). Now if I compare the 2 weeks data point with the before data point, using a paired t-test, I do find a statistically significant difference, as can be seen in this representative graph of 1 subject, there is a decrease in Y. However, I see the same pattern in the blue (untreated) line, and I think there must be a better way to model this. My data currently look like this: Subject Time Y treated Y untreated 1 Before 1 1 week 1 2 weeks ... n I wanted your advice on how to model this data (I thought on mixed model, or GEE). Should I transfer the file so each subject will have 4 additional rows (the untreated eye), or keep it like this and add this variable (untreated eye) as a covariate ? Usually in this field only the treated eye is analyzed. However, in his small study the population is different and the untreated eye vary more than in the usual population. I tried this: proc mixed data = A;   class Subject_ID Time;   model Y_Treated = Time Y_untreated;   random Subject_ID;   lsmeans Time / adjust=TUKEY; run; Am I close ? For some reason the diff option did not work (did not become blue). Should I use the contrast or estimate statements if I want to compare a specific time point to the baseline ? I tried and got different results than the multiple comparisons gave me, I must have written it wrong. The most important question I have is if this way (what I did) takes into account the variability of the untreated eye, as it is very correlated to Y. Thank you ! Edit: This is weird, I ran a paired t-test with the Y levels after 2 weeks and before the treatment, and it was statistically significant. Then I ran a similar code to the above just without the untreated_Y in the model, and although I got the same means, it was not statistically significant. I can't figure out why (anything to do with power ?). When I added the untreated_Y, it is significant again, my assumption is that it's due to the fact that I have taken into account the variance with time (represented by the untreated value). Is it valid and correct (my way of thinking) ? ... View more

Hello all, I am trying to analyze a data which was collected in a trial. The design of the trial is bad in my opinion, but it is way too late and the data is already gathered. To make story short, this is the situation: A new kind of surgical stitches was invented, and was compared to 2 existing kinds of stitches, for convenience, I will call these groups from now on "New", "Control 1" and "Control 2". A few patients were enrolled, needing to be stitched up, in 1 or more places. So for every patients, there are 1 or more observations. Each stitch in each person, was done using of of the 3 kinds. For some patients, all stitches were the same kind, however, for some patients, there were different kinds being used. My experimental unit is the patient, and the observational unit is the "gaps" needing stitching up (I hope I ain't mixing experimental and observational). In addition to this whole mess, there are two different stitching techniques, A and B. A look at the data reveals that each patient was treated with either A or B, in all stitching, regardless of the stitch kind (new, control 1 or control 2). It is possible that there are other techniques not being used here, it is even most likely, but is this factor random or fixed ? I am not sure, I'll get back to it later. The response variable is the time it took until the bleeding stopped. It is recorded in minutes, given values 0,0.5,1,1.5,....with all the inaccuracies involved. Here is an illustration of the data (with faked data): Stitch ID Subject ID Treatment Technique Time 1 1 New A 0 2 1 Control 1 A 1 3 2 New B 3 4 3 New A 6 5 3 New A 3.5 6 3 Control 1 A 4.5 7 4 New B 2 8 4 Control 2 B 1.5 I tried fitting a mixed model in several ways, these are my attempts: proc mixed data = Data;   class Treatment SubjectID;   model Time = Treatment;   random SubjectID;   lsmeans Treatment; run; Rational: The subjects are the blocks, ignoring technique as if it ain't there (clinicians claimed it is not important, I wouldn't bet on it !) Treatment is statistically significant AIC=250, BIC=253 ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- proc mixed data = Data;   class Treatment SubjectID Technique;   model Time = Treatment;   random Technique;   random SubjectID(Technique);   lsmeans Treatment / cl pdiff=control('New') adjust=dunnett; run; Rational: The subjects are the blocks, but since each subject had only one kind of technique, I imagined two blocks of technique (two rectangles), in each one patients (smaller rectangles), and in each one of these, stitches, each one could be different kind in other words, subjects are nested within technique Treatment is statistically significant AIC=249, BIC=245 ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ proc mixed data = Data;   class Treatment SubjectID Technique;   model Time = Treatment Technique;   random SubjectID(Technique);   lsmeans Treatment / cl pdiff=control('New') adjust=dunnett; run; Rational: Maybe there are differences due to technique. Maybe the fact that each patients had one and only technique doesn't mean it has to be a blocking factor ? I used it as fixed effect, and surprise surprise, it was statistically significant ! Treatment was still strongly statistically significant. AIC=244 BIC=246 My questions are: 1. Is my syntax correct ? How do I know if to take technique as a random effect or fixed effect ? How do I choose the best model ? Am I correct to choose the Dunnett adjustment ? 2. I tried adding a plot to my lsmeans statement, I wanted a means plot with either standard error or confidence interval, or maybe even a box plot. But for some reason, despite reading in the SAS documentation that this is possible, the keywords plot or plots did not appear blue, and I got an error message (expecting one of the following....). Any ideas why ? example: lsmeans Treatment / cl pdiff=control('New') adjust=dunnett plot=means; 3. Something basic: When we have a single sample, a correlation between observations will change the variance, the standard error and the CI, but it will not affect the means. So why are the adjusted means (lsmeans) differ from the descriptive means I calculate by treatment without specifying subjectID ? (using proc means or univariate) Thank you in advance ! Edit: I want to add and say, that I just tried one more thing and that is to add the interaction between Treatment and Technique, and it is statistically significant. The AIC is now down to 223 and BIC to 225. In case you'll tell me that THIS is now the preferred model, how do I interpret the interaction lsmeans ? By the way, now the lsmeans of the Treatment groups are MUCH closer to the "descriptive ones" (not taking into account the subjectID). ... View more