================================================================== The gateway between this list and the sci.stat.edu newsgroup will be disabled on June 9. This list will be discontinued on June 21. Subscribe to the new list EDSTAT-L at Penn State using the web interface at http://lists.psu.edu/archives/edstat-l.html. ================================================================== . Thank you for having time to look at my request.
For a better understanding I will try to explained in more details the problem. There was tree main effects : two sex, two potential growth rate, and tree slaugter weights. I also want to considere some random effects : sire, damn, and pen. I have weigths for pigs that were raised in the the same environment. Animals started the trial at the same date, and were weighted at different times before being slaughtered. Animals were all weighted at the same time when the group had an averaged weight of 50 and 75 kg. However, there was tree targeted slaughter weights : 107, 115 and 125 kg. Slaughters occured on 6 weeks, once per week, each animal being evaluated to decide if it has to be slaughtered on that week according to its targeted slaughter weight. Following the first slaughter, remaining animals were weighted each week if they were over 100 kg. As we can see, this leads to have animals with unequal number of data, but also with data collection ending at different ages. I am interested in producing a growth curves for each level of sex and group of growth rate to make comparisons betweens curves, and doing this while considering the random effect mentionned earlier. I would like to determine at what moment differences become significants. As I mentionned previously, random regression seemed a good way to make the tests, but leads to computational problems. (Others) Repeated measures analysing lead to artificial inflexion of the mean curves due to the different ending ages. For sure, somebody can say that it is not a proper design for testing what I want to test. Analysing performances per period allows to answer the main objectives of the projects, but I would like to exploit with more depth the data collected. My suggestion is to work with predicted values of individual growth curves and analyse those predicted value with a mixed model to test for differences between sexs and groups of potention growth rate. R2 of individual regressions (third degree polynomial) are high, but prediction error can be moderately high as live weight measurements are subject to some variations like the gut filling. Moreover, the prediction error for a given animal increase greatly if the age is greater then the age at which that animal has been slaughtered, and this should be considered. May be it is a silly problem... Thanks, Jo�l Rivest Once again, apology for my english writing if it as to be. Richard Ulrich <[EMAIL PROTECTED]> wrote in message news:<[EMAIL PROTECTED]>... > posted to sci.stat.edu and e-mailed. > > I haven't seen any answers attempted for this one - > > On 3 Jun 2004 06:16:36 -0700, [EMAIL PROTECTED] (Jo?l Rivest) > wrote: > > > Hi, > > > > I am looking for references or suggestions on how to analyse predicted > > values (e.g. as if they were observed ones) in a mixed model. I have > > individual growth curves providing predicted values of weight for a > > given age and I would like to analyse those predicted values in a > > mixed model, considering their prediction errors. Those prediction > > errors are specific to each individual. > > Are you trying to extract the original model for prediction? or what? > This is unfamiliar to me. It seems to me as if there would not be > much point to it, unless the expected R-squared is very high, that is, > the prediction is going to be very good. > > > > > Some methods such as random regressions and repeated measures > > analysis with different covariance struture were considered, but > > didn't worked properly because either of the data structure or the > > complexity of the mixed model (I am using SAS - I know there are > > packages that would be OK for the job, probably ASREML). > > > > A suggestion is to use weighted analysis, but it seems that the > > weighting variable is more complicated to calculate than just taking > > the inverse of the prediction error. > > > > Are worried that the inverse-of-error is going to disrupt the > statistical tests? Yes, but the only modeling I imagine will > assume and require a good fit. So I would look at R-squared > and sums of squares, not tests. > > What happens with no weighting? If weights are in a narrow > range, that should be a good approximation to the model. > If weights have a large range, then try the model when > deleting the points that matter least - 10%, 20%, 50%. > > I'm curious about what application wants and ANOVA that > uses predicted growth curves, especially if you do know the > basis of the original predictions.
