Most chips have one bias system -

1. Bandgap, creates one refernce voltage
2. Reference voltage is used in a  V to I converter
3. Currents are distributed around the chip, and get locally reconverted back to voltages as needed or used as reference currents.

There are exceptions to the above, but this is the most common methodology. Some devices with RF front ends will use local bias structures for better isolation and noise source control.

Some important items in a mixed signal chip:
Dont distribute ground referenced voltages across a chip. Too much noise.
When creating a current source, put the diode connected transistor that biases it right next to it.
Decouple the bias control with a small capacitiance, enough to suppress the fundamental of the system clock.

There are a million others as well. Here read my white papers on the topic here:

http://www.effectiveelectrons.com/HelpIC.htm

For all mixed signal designers the two part series on noise are a must read.

Faisal,

 Why do you need to have multiple bandgaps? You could have a single bandgap
drive level shifters to generate any additional reference voltages? You did not
mention it, but triming a voltage reference increases costs since it requires test time
and additional die area. Or you could just include the level shift in the LDO.
A single Band Gap may be more susceptable to cross-talk but they maybe
manageable depending on the system.

  My thought would be that there is no absolute answer to this question, the answer
depends on the application: power, cost, performance, off-chip components, package,
and 1000 other factors.

  Have you built a system model: board, package, bondwire parasitics, with estimated
interconnect parasitics and tried to analyze the crosstalk, the effect of Band Gap drift
on overall acccuracy, ...? That is, this maybe a good problem to address with behavioral
modeling.

                                                                  Best Regards,

                                                                    Sheldon