Option 1 of the CODA Output Analysis Menu produces a graphical summary of the iterates for each monitored variable and chain in the BUGS output. By default, 2 plots are produced for each variable. The first shows the traces from each chain as separate time series on the same graph. The second shows a kernel density estimate calculated by combining the samples of each variable from all chains. Figure 1 shows this output for the line example.
Note that the S-Plus density function used by CODA to produce these plots yields estimates whose support extends the range of the sampled values by 0.75 times the bandwidth (see §5.5.4 for further details on bandwidth specification). This can result in negative estimates for variables such as variance parameters, whose sampled values must be positive but may be close to zero. CODA avoids this problem by using the reflection method described in silverman:86 pp.30-31 to calculate kernel density estimates for variables specified to be positive or restricted to the range (0,1).
Figure 1:
Graphical summary of the output from the line example produced by selecting Plots (Option 1) from the CODA Output Analysis Menu
The the following menu will appear on screen the first time graphical output is requested during a CODA session:
Select graphics device required: 1: openlook (UNIX) 2: motif (UNIX) 3: X11 (UNIX) 4: win.graph (DOS) Selection:
The user should respond by typing the number of the required graphics device after the prompt. UNIX users may choose any of the first 3 devices, whilst DOS users should select the fourth option. (Please contact the authors if other options are needed). Note that CODA occasionally crashes when trying to open a graphics window . This problem only happens when the BUGS input file is very large, and is due to the way S-Plus uses up memory during a session. One solution may be to first read the BUGS files into CODA and then immediately quit and save these as in S-Plus format. Re-start CODA and select option 2 (`Begin a new CODA session using data saved from a previous CODA session') to read the data directly in S-Plus format. This should leave sufficient memory available to then open a graphics window and proceed with the CODA session.
In CODA, kernel density estimation is carried out using the S-Plus density function. A Gaussian kernel is used with default window width given by 0.25 times the range of the sampled values for each variable. This leads to a smooth estimate, but may hide local features of the density. Other bandwidth functions may be used if desired (see §5.5.4).
The user may also change certain other plotting options in order to customize the CODA output produced. Possible alternative options include: