Notes on Rotation vs. Shear by Joe Krahn:

Preface:
I am not an expert in computer graphics, but I have a good understanding of 3D transformations.  These remarks come from personal observations of the 3D display process.

Problems with rotation transforms for stereo views:
There are some significant problems with the rotation method. First, the image produced for each eye is valid for only for a direct perpendicular view, and does not account for perspective distortions that are different for each eye. Second, the equations are only valid if the viewer is looking directly at the center of the display window, and is positioned along the centerline of the display window. The result of these two effects is that rotational views, as described by SIGGRAPH, are actually not physically correct. It also means that perspective effects differ for each eye, and will display corresponding points for each view on different horizontal positions. This produces poor stereo results and significant eye strain, especially when viewing away from the center of the screen. Another problem with rotational transforms is that front and back clipping planes will not clip the two views equally. This can result in some object fragments being displayed for only one eye.

Advantages of Shear transforms:
Shear transforms are less physically accurate, but are appropriate for a viewer looking at any part of the display. Equivalent points will be on the same horizontal line, and clipping planes will work the same for both views. Another result of shearing is that left and right views of objects are more similar. Even though this is an inaccurate representation, it improves stereo perception. In real life, details at the left or right edge of an object will be visible by only one eye. Our brains learn to ignore such artifacts. Even though this situation is realistic, it detracts from stereo perception.

Exact approaches:
It is certainly possible to create an exact transform, accounting for left/right screen parallax differences, and may be fairly easy. Still, you will have problems of unequal Z-clipping, and will need the viewer to remain centered on the screen. Some experimental stereo viewing systems actually track the viewer; this could enable dynamic adjustment of the parallax corrections, allowing for proper implementation of the perfectionist approach. Personally, I'll stick with shear transforms and avoid all the hassle.

Here is a summary of the rotation and shear methods:

Stereo rendering with rotations, according to SigGraph 98/99:

glMatrixMode(GL_MODELVIEW);
glLoadIdentity(); /* the default matrix */
glPushMatrix()
glDrawBuffer(GL_BACK_LEFT)
glTranslatef(-IOD/2.f, 0, 0)
glRotatef(-angle, 0.f, 1.f, 0.f)
<viewing transforms>
<modeling transforms>
draw()
glPopMatrix();
glPushMatrix()
glDrawBuffer(GL_BACK_RIGHT)
glTranslatef(IOD/2, 0., 0.)
glRotatef(angle, 0.f, 1.f, 0.f)
<viewing transforms>
<modeling transforms>
draw()
glPopMatrix()

Where angle is the inverse tangent of the ratio between the fusion distance and half of the interocular distance.  Each viewpoint is rotated towards the centerline halfway between the two viewpoints.

Stereo Rendering, using a shear transform:

 

glMatrixMode(GL_MODELVIEW);
glLoadIdentity(); /* the default matrix */

glPushMatrix()
glDrawBuffer(GL_BACK_LEFT)
m={1,0,sin(angle),-sin(angle)*FD, 0,1,0,0, 0,0,1,0, 0,0,0,1}
glMultMatrix(m)
<viewing transforms>
<modeling transforms>
draw()
glPopMatrix();

glPushMatrix()
glDrawBuffer(GL_BACK_RIGHT)
m={1,0,-sin(angle),sin(angle)*FD, 0,1,0,0, 0,0,1,0, 0,0,0,1}
glMultMatrix(m)
<viewing transforms>
<modeling transforms>
draw()
glPopMatrix()