Hi, Does any one know how we can decide on the correct transformation in (avas and ace) after having drawn the graphs y,g(y) x ,s(x) and g(y) ,s(x) . Is it possible by only looking at patterns the graphs follow for example when y ,g(y) shows a logaritmic pattern can we say that log transform on y is suitable? Thanks for your help. Regards [[alternative HTML version deleted]]
On 01/20/04 17:34, frat zdemir wrote:>Hi, >Does any one know how we can decide on the correct transformation in (avas and ace) >after having drawn the graphs y,g(y) x ,s(x) and g(y) ,s(x) . Is it possible by >only looking at patterns the graphs follow for example when >y ,g(y) shows a logaritmic pattern can we say that log transform on y is suitable?It seems to me that the point of avas() and ace() is to arrive at a transform that meets the criteria of each procedure but is not necessarily based on some simple functional form like the logarithm. The lack of restriction to any particular functional form is what is particularly useful. You might, however, compare a simple function to the results of avas or ace. The output of each function consists of the transformed values ty (the dependent variable) and a matrix of predictors tx. You also get y and x. You can access these as follows: ace1 <- ace(..[all your stuff]..) ace1$ty ace1$y etc. If you think that ty is the log of y, then plot ty as a function of log(y) and see if you get a straight line. Or test for linearity however you like. -- Jonathan Baron, Professor of Psychology, University of Pennsylvania Home page: http://www.sas.upenn.edu/~baron R page: http://finzi.psych.upenn.edu/
On Tue, 20 Jan 2004 17:34:24 +0200 f?rat ?zdemir <firat.ozdemir at deu.edu.tr> wrote:> Hi, > Does any one know how we can decide on the correct transformation in > (avas and ace) after having drawn the graphs y,g(y) x ,s(x) and g(y) > ,s(x) . Is it possible by only looking at patterns the graphs follow for > example when y ,g(y) shows a logaritmic pattern can we say that log > transform on y is suitable? Thanks for your help. > Regards >The strategy I use is to not try to do this. One reason is that you may be tempted to fit a parametric model with such simple transformations, without accounting for the hidden degrees of freedom from the uncertainty in estimating the transformations. The areg.boot function in the Hmisc package will give you bootstrap confidence bands for ace and avas transformations, taking into account almost all sources of uncertainty. --- Frank E Harrell Jr Professor and Chair School of Medicine Department of Biostatistics Vanderbilt University