Meng, This really comes down to what question you are trying to answer. Before worrying about details of default contrasts and issues like that you first need to work out what is really the question of interest. The main difference between declaring a variable ordered or not is the default contrasts. Defaults are provided because there are many cases where which contrasts are used internally does not matter, so why make someone think about it. In cases where the choice of contrasts matter, it is rare that any default coding is the correct/best choice and you should really think through what contrasts answer the question of interest and use those custom contrasts.
For example, to answer the question if Tension has any overall effect it does not matter which contrast encoding you use (as long as it is full rank), the test statistic and p-value for testing the whole effect will be the same. The predictions of the means of groups will also be the same regardless of which contrasts are used (and this is often a clearer way to present/explain the results). A case where the specific contrasts would matter would be if we want to see if we can reduce the number of groups by combining groups together, or interpolate to certain groups. The treatment contrasts will test if low and medium can be combined (which makes sense) and if low and high can be combined (which does not make sense unless the first is true and in fact the overall factor is not significant), what makes more sense would be to compare low to medium and medium to high (it could be that low is different from the other 2, but med and high can be combined). The polynomial contrasts give a different view, the quadratic term in this case tests whether the medium group is the average of the low group and the high group (so we could interpolate medium), this only makes sense if the medium tension is centered (in some sense) between the other 2, i.e. the difference from low to medium is exactly the same as the difference from medium to high, but if that were the case then I would expect a numerical term rather than an ordered factor. So, to summarize, it depends on the question of interest. For some questions the contrasts don't matter, in which case it does not matter, in other cases the correct contrasts to use are determined by the question and you should use the contrasts that answer that question (which are rarely a default). On Tue, May 21, 2013 at 11:09 PM, meng <laomen...@163.com> wrote: > Thanks. > > > As to the data " warpbreaks", if I want to analysis the impact of > tension(L,M,H) on breaks, should I order the tension or not? > > > Many thanks. > > > > > > > > > > > > > At 2013-05-21 20:55:18,"David Winsemius" <dwinsem...@comcast.net> wrote: > > > >On May 20, 2013, at 10:35 PM, meng wrote: > > > >> Hi all: > >> If the explainary variables are ordinal,the result of regression is > different from > >> "unordered variables".But I can't understand the result of regression > from "ordered > >> variable". > >> > >> The data is warpbreaks,which belongs to R. > >> > >> If I use the "unordered variable"(tension):Levels: L M H > >> The result is easy to understand: > >> Estimate Std. Error t value Pr(>|t|) > >> (Intercept) 36.39 2.80 12.995 < 2e-16 *** > >> tensionM -10.00 3.96 -2.525 0.014717 * > >> tensionH -14.72 3.96 -3.718 0.000501 *** > >> > >> If I use the "ordered variable"(tension):Levels: L < M < H > >> I don't know how to explain the result: > >> Estimate Std. Error t value Pr(>|t|) > >> (Intercept) 28.148 1.617 17.410 < 2e-16 *** > >> tension.L -10.410 2.800 -3.718 0.000501 *** > >> tension.Q 2.155 2.800 0.769 0.445182 > >> > >> What's "tension.L" and "tension.Q" stands for?And how to explain the > result then? > > > >Ordered factors are handled by the R regression mechanism with orthogonal > polynomial contrasts: ".L" for linear and ".Q" for quadratic. If the term > had 4 levels there would also have been a ".C" (cubic) term. Treatment > contrasts are used for unordered factors. Generally one would want to do > predictions for explanations of the results. Trying to explain the > individual coefficient values from polynomial contrasts is similar to and > just as unproductive as trying to explain the individual coefficients > involving interaction terms. > > > >-- > > > >David Winsemius > >Alameda, CA, USA > > > > [[alternative HTML version deleted]] > > ______________________________________________ > R-help@r-project.org mailing list > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide > http://www.R-project.org/posting-guide.html > and provide commented, minimal, self-contained, reproducible code. > -- Gregory (Greg) L. Snow Ph.D. 538...@gmail.com [[alternative HTML version deleted]] ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
