June 14, 2004 SPLIDA (S-PLUS Life Data Analysis) Version 6.5, June 14, 2004 These release notes document changes since the December 11, 2003 release of SPLIDA. The following outlines the changes since the December 2003 version of SPLIDA. More details on the use of these new capabilities are given in the SPLIDA documentation. The Splida documentation has been completely revised to bring it up to date with the current version of the software. The ML estimation procedures now handle all combinations of censoring (not censored, left-censored, right-censored, interval-censored) and truncation (none, left, right, interval). Object refresh buttons have been added to a few dialogs where they were missing. A message is now printed whenever the lock-axes option is on and previous axes are being used to generate the current plot. It is now possible to fit from the GUI the largest extreme value distribution and the corresponding Frachet (reciprocal Weibull) distribution. With our newly written algorithms, it is now possible to estimate the recurrence data MCF with much larger data sets and to use in the data set weights to indicate large groups of similar observations with no events (such groupd arise frequently in warranty da ta where only a small fraction of units have events). It is now possible to fit an NHPP model to recurrence data with multiple units. The recurrence data procedures now allow estimation of the nonparametric MCF and NHPP models with multiple window-observation periods on each unit. It is now possible to use the GUI to plot residuals against x variables not in the model. This is done by initially choosing such x variables for the model, but then leaving them out of the model equation. There is an SPLIDA option to choose alternative methods for coding dummy variables used in categorical explanatory variables (based on the S-PLUS options(contrast=...)). The coding method is printed out in the model output summary. The S-PLUS default is the helmert method in which the intercept of the regression model is the overall mean and coeffieicents are deviations from this value. The the SPLIDA default is the "contrast" method which uses one of the levels as a baseline and coefficients measure deviations from this level. When making a life data event plot in the case where the first character of all failure modes is not unique, numbers are printed instead of characters to indicate the failure mode. In this case, SPLIDA now prints a table of failure modes and the corresponding number. Some improvements were made to the interface for the bootstrap procedures. SPLIDA (S-PLUS Life Data Analysis) Version 6.3, December 11, 2003 The conversion to the new S engine (SV4) has been completed. The current version has been tested with S-PLUS versions 6.0, 6.1, and 6.2. This version has been extensively tested in our University courses as well as many short courses. Most of the changes since the previous version have been to make the necessary conversion to SV4 and should be transparent to the user. A number of minor changes were made in the new methods for the analysis of accelerated destructive degradation data and in the methods for planning accelerated life tests. The documentation has been reworked considerably to reflect continuing development, but the documentation continues to lag the development of new procedures and features. Luis Escobar and I hope to bring the documentation up to date by the end of the first quarter of 2004. A new algorithm for the analysis of recurrence data has been implemented, allowing the analysis of larger data sets. Fitting of nonhomogeneous Poisson process models to such data is also now available from the GUI. A tool have been developed to allow users to lock the axes of a sequence of similar plots (the red button with an L on the tool bar is for Lock and the blue button with a U is for unlock). "Similar pots" are defined as plots that have the same types of x and y axis (e.g., linear, log, or other transformation) and the same titles on the x and y axes. This feature is useful when you want to make direct comparisons between two plots and is much easier to do than the old method of specifying the axis limits directly. New data sets have been added to the example files. Plans for future work in SPLIDA include the development of new and better methods for the analysis of degradation data, recurrence data, warranty data, output options in life data analysis, and test planning. As always we are pleased to receive comments and suggestions. --------------------------------------------------------- This is the first released version for Splus 6.x (with the SV4 engine). It runs only under Splus 6.1.2 for Windows. There have been a number of minor changes to the dialogs and some small bugs have been fixed, but most of the changes needed to convert to the SV4 engine should be transparent to the user. New functionality includes accelerated test planning tools for accelerated tests with more than one accelerating variable. SPLIDA Version 6.1 (beta), December 5, 2002 --------------------------------------------------------- This is the first released version for Splus 6.x (with the SV4 engine). It runs only under Splus 6.1.2 for Windows. There have been a number of minor changes to the dialogs and some small bugs have been fixed, but most of the changes needed to convert to the SV4 engine should be transparent to the user. New functionality includes accelerated test planning tools for accelerated tests with more than one accelerating variable. SPLIDA Version 5.9, October 22, 2001 --------------------------------------------------------- The following outlines the changes since the October 24, 2000 version of SPLIDA. More details on the use of these new capabilities are given in the SPLIDA documentation. Although the documentation for some new features is not yet complete, the updated SPLIDA User's Manual contains screen shots of dialog boxes and output for the examples. More detailed descriptions and explanation of these examples will be contained in the next version of the manual (once again, development of new features has out-paces the writing of detailed documentation).. There have been a large number of minor improvements, but three major additions to the capabilities in SPLIDA. The default regression model continues to be a simple main-effects model (which handles categorical explanatory variables as well as continuous explanatory variables). It is, however, now possible to specify a regression model using the very powerful and general S-PLUS modeling capabilities. Thus it is easy to specify terms for interactions, polynomial factors, and nesting. There are new capabilities for planning accelerated life tests with a single accelerating variable (these capabilities will be generalized to multiple accelerating variables in a future release). In particular, there is now a dialog that allows the user to generate plans according to various criteria, including optimum, traditional, and compromise. There is also a new tool for exploring the results of test plan simulation (explained below). There is now a complete set of tools for planning and analyzing accelerated degradation tests involving _destructive_ measurements (so that only one measurement per unit tested) with a single accelerating variable (this will be generalized to multiple accelerating variables in a future release). There are examples in the draft documentation that is being prepared. Simulation is a powerful tool for exploring the properties of proposed experimental plans, particularly when faced with issues of censoring and other things requiring nonlinear estimation. New tools/dialogs have been developed to allow the user to do detailed evaluation of accelerated life test plan and accelerated destructive degradation simulation results. The dialog will allow the user to display the results of test plan simulations in a number of different useful ways to help assess the strengths and weaknesses of a user-specified test plan. The residual analysis dialogs are now easier and quicker to use. When choosing an object from a list, in many dialogs, a summary or listing of that object is provided in the text/report window to remind the user of the contents without having to go to the object browser. In most dialogs users chooses particular kinds of objects for use for analysis evaluation or other uses. Now there is a "refresh" button that will refresh the list. This is useful when using several dialogs in parallel and dialogs are creating new objects. In the past it was necessary to remember and type the name of the object (which was quite difficult with long automatic object names) or to re-launch the dialog. Data subsetting options have been added to more dialogs (I intend to continue adding these to other dialogs as time permits and needs arise). Following the lead from S-Plus, the term "data frame" in the GUI is being replaced by "Splus data set." When analyzing the results of a designed experiment or an accelerated life test, it is useful to proceed through three levels of model structure, in each case comparing results with nonparametric estimates of the fraction failing over time. These analyses are a) individual analyses at each combination of the explanatory/accelerating variable levels, b) individual analyses in which the log-location-scale distribution shape parameter (location-scale scale parameter) is held constant, and c) a regression model in which the log-location-scale distribution scale parameter (location-scale location parameter) is a (log)-linear function of the (possibly transformed) explanatory/accelerating variables. The procedures that fit these models have been fixed so that the legends come out with combinations listed in a consistent order (depending on the order of choice of the variables into the model) and so that a consistent set of plotting symbols and line colors is used in all cases, even when some individual combinations are not estimable. On many monitors, line types other than solid do not show up well in graphs. With differentiation through different colors, there is little need for stylized line types in graphics. There is now an option to switch between stylized line types and solid lines, with solid lines being the default. Of course it is possible to use the SPLIDA preferences dialog to change this option. SPLIDA automatically generates character strings to label factor level combinations for a data set (e.g., to allow choosing combinations for which special output is desired or for making legends). There is now a long option and a short option for these names. The short-name option pastes together the variable levels (e.g. 20;300;New). The long-name option pastes together both the factor names and levels (e.g. 20Stroke;300Temp;NewMethod). The long-name option is the default. Warnings and errors generated when using the GUI launch a pop-up box with the message. The message also goes to the output window (command or report). When warnings and errors are generated from the command line, there are no longer popup boxes. Examples and data sets have been added to illustrate the analysis of truncated data. If truncation variables are included in a life data object, then the SPLIDA estimation algorithms automatically take account of the truncation. For nonparametric estimation, SPLIDA has a feature (described in Chapter 11 of Meeker and Escobar) that make a parametric adjustment to the nonparametric estimate, to allow distributional assessment by using the standard probability plotting methods. This feature has always been available at the command level, but was not operational from the GUI. Now this correction option is the default. Additional example data sets have been included to illustrate new methods. For regression modeling, new, easier-to-use options are available for specifying "new.data" at which estimates are desired. -------------------------------- Version 5.7 October 28, 2000 The following outlines the changes since the August 24, 2000 version of SLIDA. More details on the use of these new capabilities are given in the SLIDA documentation. Although the documentation for some of the new features is not yet complete, the updated Slida User's Manual contains screen shots of dialog boxes and output for examples for all of the new functionality. More detailed descriptions and explanation of these examples will be contained in the next version of the manual. A more complete User's manual will be available in early 2001. If you delete your _Prefs folder, the SLIDA GUI will be recreated automatically. In some cases, however, you may have to give the CreateSlidaPrefs() command. A number of minor changes have been made in the underlying SLIDA GUI development code to make SLIDA compatible with the August 24, 2000 Release 3 version of SLIDA. Previously, due to a change in Splus diagnostics, the GUI re-create would fail. Also, when remaking the SLIDA GUI (generally necessary when you install a new version of SLIDA to update your _prefs directory) in your S_PROJ (SlidaUser) folders, the messages about what is happening are fewer and more informative. In the analysis of multiple failure mode data, tables of estimates of the series-system combined model failure probabilities or quantiles can now be printed. It is also possible to include confidence intervals for the series-system cdf on the multiple failure mode series-system combined estimate probability plot. When doing multiple failure mode analysis, in some cases, it may be desired to use distributions to different different failure modes. The SLIDA multiple failure mode functions now allow the user to specify different distributions for different failure modes for multiple failure mode analysis. This capability is presently available at the command level (examples in echapter15.q). This capability will be added to the GUI in a subsequent version of SLIDA. Options in the dialog for getting contour, perspective, and profile likelihood plots have changed. Now it is possible to request a profile plot without having to generate the contour plot. It is now possible to generate bootstrap confidence intervals for the standard single-distribution model fits through a special dialog under single distribution analyses. Confidence intervals are available for parameters, quantiles, or failure probabilities. This capability is experimental. There are plans to extend this capability in the future to allow bootstrap intervals for regression models and to allow the user to specify a global option to produce all confidence intervals generated by the standard modeling procedures with either normal approximation (quick, but crude) or bootstrap (slow, but generally more accurate) methods. Simple pedagogical Bayesian analysis tools have been added for standard single distribution analyses. These parallel the Smith and Gelfand "Bayes without tears" approach used in Chapter 14 of Meeker and Escobar (1998). These tools allow graphical comparison of samples from the prior, the likelihood, and samples from the posterior, providing a visual perception of the information-updating process in Bayes theorem. Inferences are provided for parameters, quantiles, or failure probabilities. The acceleration factor dialog has been enhanced to provide graphical output of the acceleration factor versus level of the accelerating variable and more than one acceleration factor can be inputted at a time to get two or more acceleration factor lines to show on the same plot. The Arrhenius model is commonly used for increased-temperature accelerated testing when the underlying mechanism can be approximated by a one-step chemical reaction with a given activation energy. By default, the units of activation energy in SLIDA are eV. Now, however, it is also possible to choose either KJoules/mole or Kcalories/mole instead. It is now possible to extend the range of evaluation in the conditional model plot in a manner that is consistent with other SLIDA dialog boxes. The line width for confidence intervals on probability plots has been increased so that they show up more clearly. The postscript graphic driver will now work on all plot-making actions. Additional example data sets have been included to illustrate the new methods. Ability to fit some of the less commonly used distributions for reliability data, including the three parameter Weibull, three parameter lognormal, generalized gamma, Birnbaum-Saunders, and the inverse Gaussian distributions. These capabilities are experimental and the ML iterations may, in some cases, fail to converge, especially if the data had a large amount of censoring. Remedial measures (such as specifying starting values for the ML iterations) are only available at the command level. The option to allow a log transformation in degradation trellis plots now works correctly (a browser debugging command had been left in the code). The August 24, 2000 version 5.6 of SLIDA. The following outlines the changes since the June 5, 2000 version of SLIDA. More details on the use of these new capabilities are given in the SLIDA Users Manual. Previously, SLIDA allowed one to obtain a contour or a perspective plot of the likelihood for location-scale and log-location-scale distribution with respect to the location and scale parameters. A number of users have asked us to provide capabilities for making additional likelihood contour plots, like those in Sections 8.2 and 8.3 of Meeker and Escobar (1998). It is now possible to make likelihood contour or perspective plots for a specified quantile and sigma (the spread/shape parameter) for a location-scale or a log-location-scale distribution. Such plots are also available for comparing several distributions. When making a likelihood contour or perspective plot, one can also request the corresponding profile plot for sigma (or beta for the Weibull distribution) and the position parameter (mu or the selected quantile). When doing multiple regression analyses, it is now possible to specify "new data" (splus terminology for particular combinations of explanatory variables for at which to evaluate failure probabilities and distribution quantiles) by specifying a data frame object (for simple regression it is simple enough to list the desired values of interest and this option is still available). An event plot can now be made using special markers to indicate different failure modes. The name "stress plot" has been changed to "relationship plot" in SLIDA and its documentation. When using the new recommended SlidaUser folder method of operating Slida (described in detail in the SlidaGUI.doc SLIDA Companion) one only needs to update the Slida folder in a common place without affecting various working SlidaUser working folder(s) that users may have established. For those using this method, the echapterxx.q files containing commands to do the examples in Meeker and Escobar (1998) have been updated so that the commands reading data from the Slida_text_data folder work without modifying the path to the folder. The Slida_text_data folder contains ASCII files for all of the data sets used in Meeker and Escobar (1998), the SLIDA documentation, SLIDA examples, in addition to a number of other previously published and unpublished data sets. Data frames are also provided for some (but not all) of these data sets. The README.txt file in the Slida_text_data folder provides a draft version of an index and cross reference making it easy to identify the data set used in any of the examples in Meeker and Escobar (1998). References providing the source of other previously published data sets is also contained in this file. Slida uses objects to transfer information (e.g., data and results) from one procedure to another. Generally these are, by default, saved from session to session. It is possible to delete these at the command line or by using the object browser. To make the management of such objects easier, a Delete Slida objects dialog box, under "Slida -> Slida tools" has been written allowing such objects to be listed by type and selectively deleted. Also under "Slida -> Slida tools" is a tool allowing one to make blank probability (e.g. Weibull or lognormal paper) and relationship plotting paper (e.g. time vs temperature Arrhenius paper), something that many of us still find useful when we are not hooked up to a computer. Ability to analyze truncated data with both parametric and semi-parametric methods (as described in Chapter 11 of Meeker and Escobar), and including models with explanatory variables (not described in Meeker and Escobar). Additional example data sets have been included to illustrate the new methods. Under "Slida -> Plan accelerated life test" there is a tool that allows one to compute an acceleration factor for a simple acceleration model. For multiple regression modeling, new, easier-to-use options are available for specifying "newdata" at which estimates are desired. It is now possible to get a plot similar to a relationship plot (but without a model linking the plotted densities) for data when individual distributions are fit to several groups. -------------------------------- The June 5, 2000 version 5.5 of SLIDA. 1. The most extensive improvement since the last version of SLIDA has been in the documentation. Luis Escobar (who also has been providing technical advice to this project for many years and who is co-author of our Wiley text book that provided the technical blueprint for SLIDA) has agreed to join me as co-author on the SLIDA User's Manual. The SLIDA User's Manual now documents all of the functionality in the SLIDA GUI (there remains a considerable amount of more esoteric functionality at the command level that is less well documented). 2. Ability to split/subset data objects (e.g., with a factor explanatory variable comparing two or more groups) so that individual detailed analyses can be performed more easily on the individual groups or subsets of one's data. The "Make life data objects for individual groups" menu item now appears in the comparison and regression menu lists. 3. A procedure to compute the probability of successful demonstration for a censored-data life test reliability demonstration. The curves generated by this procedure are a generalization of similar curves provided in Chapter 9 of Hahn and Meeker (1991, Statistical Intervals, John Wiley & Sons) for the normal distribution and complete data. The new procedure will generate similar curves for lognormal, Weibull, and loglogistic distributions for type II censored data. 4. The routines used to convert (in some cases by extrapolation) degradation data into failure data are now more robust, allowing for left censoring (when the degradation path is beyond the failure definition at the first inspection) and to check for degradation paths that seem to be going in the direction opposite from what is expected. 5. The approximate sample size determination figures allow specification of either a censoring time or a fraction failing as input in specifying a proposed test plan. 6. Mean time to failure is now reported when estimation output for a single distribution or a distribution conditional on fixed levels of explanatory variables (in a regression or accelerated test model) is printed. 7. SLIDA was developed largely around the analysis of single distribution data of various types with complicated censoring and experimental regression and accelerated test data with with relatively few test conditions and a considerable amount of replication. In previous versions of SLIDA, methods were not carefully tuned for regression problems with many observations, at many different levels of explanatory variables, with little or no replication. A number of changes have been made to make it easier to analyze such data. Such data sets now show up in the lists of data objects for more procedures (By defaults that can be overridden, SLIDA tries to identify data sets appropriate for a particular analysis and puts them in the list of available data objects. For example, data without explanatory variables will, by default, not show up in lists for procedures doing regression analysis). -------------------------------- The January 21, 2000 version of SLIDA. 1. The SLIDA User's manual is much more complete than in the past. It is still in draft form, but contains much more useful information. 2. A few problems of compatibility with Splus2000 have been fixed. 3. When fitting a log-location-scale distribution (e.g., Weibull or lognormal) it is now possible to fix the distribution's shape parameter when doing ML estimation. Correspondingly, when fitting a location-scale distribution (e.g., smallest extreme value or normal) it is now possible to fix the distribution's scale (spread) parameter. 4. It is now possible to generate trellis plots on repeated-measures (degradation) data objects, making it easy to compare individual degradation paths. 5. There are similar capabilities for making degradation residual plots. 6. Changes have been made so that the control of the axis ranges and range over which the cdf is evaluated when plotted on probability plots are handled consistently over the many dialog boxes that can create such probability plots. 7. Optionally, the time/date or another specified string is placed on the bottom right of plots. Now this is automatically suppressed in the plots with multiple plots on one sheet. 8. Ability to edit an existing data object (e.g., to change titles, units, or other details mapping a data frame to a life data object). 9. Ability to easily and dynamically modify a data object by editing the associated data frame. 10. Additional example data sets have been included to illustrate the new methods. 11. A number of users have requested access to functions used to do bootstrapping, as described in Chapter 9 of Meeker and Escobar. Now, file echapter09.q contains the commands used to do the examples in that chapter. These basic tools can be used to run similar bootstraps for other data sets. Work is being planned to extend these capabilities to regression problems and to better integrate with other functions and, eventually, to have bootstrap intervals as an option from the GUI. We envision giving the user the option to do bootstrap intervals (or likelihood-based intervals) instead of the normal approximation intervals, if the user is willing to use the necessary amount of computer time to compute such intervals (with the newest generation of PCs, this now takes on the order of minutes for small to moderate sized data sets where such improved approximate intervals are important). 12. There is an option to include the relationship into results object names (making it easier to keep track of such results when fitting different models). 13. There is an option to include the response name into a repeated measures data object (especially useful when there is a data frame with multiple degradation readings that need to be analyzed separately). ----------------------------------------------------- The October 3, 1999 version of SLIDA. 1. Life data event plot (similar to the previously available recurrence data event plot). 2. Option to request print out the parameter estimate variance-covariance and correlation matrices from the GUI (it was always available from the command level). 3. Ability to make persistent additions and changes to the observation-type aliases (e.g., users can introduce their own word or words to indicate which observations are censored and which are failures, extending the build-in list). 4. Better control over SLIDA default options (e.g., confidence level, number of digits in tables, and whether the response should be plotted on the x axis or the y axis, list of quantiles used in estimation, names for censoring indicators), including the ability to make global changes to some options and, if desired, to save such changes across sessions. 5. More consistency in how defaults are changed across GUI dialog boxes. 6. Ability to see a perspective or contour plot of the likelihood function for location-scale and log-location-scale distributions. 7. Ability to compare likelihood contours for location-scale and log-location-scale distributions when comparing samples from different populations or processes. 8. Option to mark each plot with time and date and/or name. 9. An alternative color scheme has been chosen and this one works much better for SLIDA graphics. In addition, as explained in the user manual, there is an option to have all graphics come out as pure black and white (no grey scale). These options (requested by several users) avoid the "washed-out" look that some graphics had in the past when multiple colors were used on one plot. 10. SLIDA seems to work fine on the latest version Splus2000. However, because Splus4.5 and Splus2000 have different menu structures, it is necessary for me to post different versions of SLIDA for the two different versions. I have now done this. Also the documentation has been extended and the graphics in this document are now black and white so that they print much better on a laser printer. Additionally, I have figured out how to make the pdf using 3.0 compatibility mode in Adobe Acrobat, but I still recommend downloading the latest free version of Acrobat Reader from www.adobe.com. ----------------------------------------------------- The April 24, 1999 version of SLIDA. 1. Six-distribution probability plots, 2. Ability to present the response on the y-axis in life versus stress plots, 3. Regression diagnostics (especially residual plots) 4. Sensitivity analysis procedures