Oswald NEWS -- history of user-visible changes -*- text -*- Changes in Oswald version 3.4 * Improved Makefiles so works with more versions of make * ksmooth.rsc has been updated -- it now makes use of the "by" argument, and returns a 2-column matrix suitable for plotting. Changes in Oswald version 3.3 -- BETA TEST VERSION * The latest versions of Carey's GEE and ALR libraries are included. The changes are mostly bugfixes, but there is one change with gee and gee.fit: an offset term must be included in the model formula in the standard S way ( y ~ x + offset(a)), instead of with an offset= argument. * lme and nlme work with Oswald loaded once again. * The ouput of summary() on ldframe objects has changed. * plot.ldframe has a new argument x.expr, and the selector function now ignores NAs. * ksmooth.lda, ksmooth.res and ksmooth.rsc have a new argument x.expr, and the "by" argument now ignores NAs. Changes in Oswald version 3.2 -- BETA TEST VERSION * Oswald now runs under Splus for Windows version 4.0. Also, the help files are now included in the Windows version. (Thanks to the "nr2hlp" script discussed in the Venables and Ripley online complements.) * The Oswald datasets are now available in the new ldframe data type, with names like "milk.ldf" and "lograts.ldf". A transcript of their creation can be found in Data/datasets.St. * New function: na.add, for adding columns to data frames with missing values. * The default plotting options have changed. It is recommended that you remove the object .ldats.options from your working directory if it exists, with remove(".ldats.options"). * "variogram" is now a generic function, with a new method for "ldframe" objects. Note the calling sequence is different for ldframe objects than for ldamat and balanced objects. * The function "reml.fit" is no longer included in Oswald. You can fit the same models with the "mixed" function using the "PD" fixed covariance structure. * It is now possible to assign into singly-indexed ldframe objects. For example, statements like "obj.ldf[2][1,1] <- NA" and "obj.ldf[3:4][1:10,3] <- NA" now work as expected. Due to an S-plus bug, statements like "obj[3][,4] <- NA" do NOT work; you must supply BOTH indices in the second index pair [ , ]. * ksmooth.lda has a new optional argument "by" for smoothing within groups of subjects. (Using this facility with zero bandwidth is a handy way of calculating within-groups-by-times means, and serves as a replacement for `olsfit' for ldframe objects.) * A bug which prevented ksmooth.rsc from working has been fixed. * New function: ksmooth.res, for residuals and fitted values from kernel smooths. * plot.ldframe no longer requires objects to be sorted by Time within Subject. * New function: identify.ldframe. * Chapters 1 and 2 of the manual have been updated, and work has begun on Chapter 3. * There has been some reorganization in the Routines directly; Oswald should now be easier to install on most systems. Changes in Oswald version 3.1 -- BETA TEST VERSION * A bug which prevented plot.ldframe from working has been fixed. * There is a new function anova.pcmid for likelihood-ratio tests between models. (This was included in 3.0, but not previously mentioned in NEWS.) Changes in Oswald version 3.0 -- BETA TEST VERSION * Version 3.0 is a major revision, and will certainly contain bugs. Instead of delaying it any further, is has been released as a beta test. Use with caution! * The Oswald Home Page has changed URL. It is now http://www.maths.lancs.ac.uk/Software/Oswald/ * Oswald now provides an all new data type for longitudinal data - the "ldframe". This is used for balanced and unbalanced longitudinal data, and single time series. The functions "ssv" and "tsv" can be used to add replicated columns. See the manual for more details. Not all functions (notably, olsres and variogram) have yet been rewritten to take full advantage of the new data type. * The plot method for "ldframe" objects is a much improved version of the plot method for "balanced" objects. See the help page for "plot.ldframe" for more information. * There is a new function "nths" which us useful when handling "ldframe" objects. * Attaching the Oswald library no longer causes "lme" and "nlme" to fail. Both these routines are very useful for longitudinal data, especially with the "ldframe" representation. You will usually use "cluster=~Subject" in calls to lme and nlme. * There is a new function for kernel smoothing longitudinal data: "ksmooth". The associated function ksmooth.rsc may be used for Rice-Silverman cross-validation of the bandwidth parameter. See the help files for further information; this is not yet documented in the manual. * There is a new function "mixed" implementing models for longitudinal data with random effects. It extends the functionality of "pcmid" without the dropout modelling, and is much much faster. See the help file for information; it has not yet been documented in the manual. * Various bugs in "pcmid" have been fixed. "pcmid" also has a new argument "trace" which causes evaluation information to be printed. * Oswald 3.0 includes new versions of the Carey routines GEE and ALR. * as.balanced and as.ldamat look for columns named "Subject" and "Time" instead of "id" and "time" in data frame arguments. They use the first column as the data component, instead of a column named "y". * "alr.fit" now returns the raw output from "alr", not the summary output. A bug in "alr" prevents "summary" from working on its output; this will hopefully be fixed. * The manual has been only partially updated. Some of the examples use the new "ldframe" type; in these cases they will refer to a data object like "cows.ldf". You can create this object by applying the "as.ldframe" function to the old "balanced" object, as with cows.ldf <- as.ldframe(cows) Changes in Oswald version 2.6 * The interface to pcmid for dropout models has changed. You can now specify a model with covariates for the constant term of the dropout process; parameters for this part of the model are specified separately from regressors on historic values. * You can fit sub-models of the correlation model with pcmid, by supplying zero as the initial value of any variance or dropout parameter. * pcmid no longer supplies default parameter values (which would probably not have been sensible in any case). An error is generated if you fail to supply initial values for variance or dropout parameters. * Data size limitations in pcmid have been removed. As a consequence of this, Oswald now requires much less memory to run for most problems. Changes in Oswald version 2.5 * Much better inter-conversion between the classes balanced, ldamat and data.frame. Most functions now accept any of these three classes as arguments. See the manual for information on how data.frame objects are treated in Oswald. * There is a new function `ldaplot' for plotting longitudinal data with the trellis graphics routines. * You can now convert balanced objects into data frames with columns `y' `time' `id' and `groups' with `as.data.frame(obj)'. You can similarly convert ldamat objects with as.data.frame. * The ldamat function now only accepts vector, matrix and ldats arguments. You can use as.ldamat to convert balanced and data frame objects to ldamat form. * The as.balanced function can now be used to convert data frames with `y' and `id' columns into balanced objects. * Oswald no longer re-defines the function `sort' to be generic. To sort ldats objects, you must now do `sort.ldats(obj)'. * Informative dropout models can now be fitted using the `pcmid' function. See the manual for further details. Note that unlike reml.fit, pcmid can also be used with ldamat and data.frame objects as well as balanced objects. * reml.fit has been deprecated. The reml.fit algorithm has been reimplemented and incorporated with the `pcmid' routines. That function should now be used to fit models which reml.fit used to fit. General random effects (the `lvmodel' argument to reml.fit) can not be fitted with pcmid, and for this reason reml.fit is still included. It will not be included in future releases. * The latest revision of the GEE code from Vince Carey is included, which should fix the memory leaks that have been reported. Changes in Oswald version 2.4 * Oswald now comes supplied with the GEE and ALR libraries * Oswald for Windows now comes supplied with all Watcom pre-compiled object files for faster execution, and with a live Splus _data directory for easy installation. * reml.fit has a new optional argument drop.missing.covars which, if T, will eliminate any observations within a subject for which any covariates specific to that observation are missing. Changes in Oswald version 2.3 * Some small code changes to allow successful linking with Splus 3.3 * Now uses dyn.load instead of dyn.load2, which is available on more systems. Changes in Oswald version 2.2 * Dynamically loaded C code is used on DOS systems (previously, an external EXE file was used). This is much faster. * Transformed objects, such as log(rats), are now allowed as responses in model formulae * reml.fit now allows ldamat and data.frame objects as responses, and uses a more efficient algorithm to make the estimations. Also, the nu parameter is now reported correctly (in previous versions, the square root of nu was reported). * new function `ml.fit' * Operations (such as "+" and sqrt) are now permitted on ldamat and balanced objects Changes in Oswald version 2.1 * New function: conv.ldamat Changes in Oswald version 2.0 * Bugfix release for DOS systems * Extensive manual update Changes in Oswald version 1.7 * Improved installation procedure for DOS systems * alr.fit has been updated, and a help file added * New function: read.balanced * New dataset: seizure Changes in Oswald version 1.6 * reml.fit now allows an arbitrary matrix to specify the random effects component of the model. Now you can have both a random intercept and a random slope, for example, rather than being constrained to just a random intercept. * reml.fit now uses a more efficient minimisation engine, so that results should be generated quicker. * gee.fit has been added, to allow estimations using the GEE methodology. * alr.fit has been added, for Alternating Logistic Regressions analysis of binary data. The help file has not been written (yet). * There is now an INSTALL file describing the installation procedure, and the README has been updated significantly.