There are a lot of variables. Mains voltage, of course. +/-10% pretty much covers it - a nominal 117V line might give you as little as 103V or as much as 128V. Probably not in most cases. If you know you're only going to use the box in your own house and your own mains are always 121.3VAC, you could design around that number. But it makes more sense to design a circuit that will anticipate a range of variables and not screw you later when your circumstances change.
Next variable: Transformer regulation. The Amveco torroids in the Digikey catalog range in regulation from 15.5% up to 45%. This means that the unloaded output voltage will be 15.5% to 45% higher than the nominal voltage. So, for the worst-case scenarios (the poorly-regulated 5VA torroid in a variety of environments), your 18VAC transformer can put out anywhere between 16.2VAC and 28.71VAC. Go with the 50VA unit and it's only a range of 16.2VAC to 22.9VAC.
Next you have your rectifier topology and the number of diode drops associated with it, whether the ripple is 120Hz or 60Hz, the amount of ripple you can tolerate, and the size of your filter caps.
Finally you come to the regulator itself, which brings in two issues: The regulated voltage plus the dropout differential must be greater than the minimum DC rectifier output minus the ripple; and the regulator must be able to tolerate the maximum possible voltage seen. Regulator in-out differential limits are fairly easy to account for, just like capacitor voltage ratings. The trickier issue is the heat dissipation in the regulator. Watts dissipated is the current through the regulator multiplied by the voltage across it (in-out voltage differential). Once you know the wattage dissipated, you can multiply it by the thermal resistance of the regulator and its heatsinking to determine the temperature increase over the ambient temperature (inside the chassis). Of course, if the circuit generates a lot of heat, then the ambient temperature inside your box will increase. This all has to work out so that your regulator doesn't get hotter than its rated operating limits. But it's tough to tackle, because the only way to limit the heat dissipated is to limit the input voltage. Simply using a lower-voltage transformer is one way to do it, but if you run into drop-outs on the other end then you're just as screwed. You don't have much choice but to choose the transformer to accommodate the voltage minimum (90% mains voltage at full transformer load). Then you simply have to build in the necessary protections against higher input voltages. I tend to use voltage limiters that have no effect below a certain threshold. To keep things simple, you can just use big heatsinks.
As for your specific application, I can give an example. I have a box on the bench right now that uses an 18V+18V torroid transformer to produce 18V bipolar rails along with some other supply needs. I very nearly went with the 15V+15V transformer, but decided it wouldn't work well in Japan, where 100V is nominal. Here in my shop, where the mains are a high 123VAC, I get over +/-30VDC out of the raw supply, unloaded. Once everything is powered it's down just under +/-29VDC. This is a 15VA transformer, which is a bit oversized for the circuit. Going large on the power transformer generally gives you a unit with better regulation, and it means waveform coming off it (both power output and radiated field) tend to be cleaner, without serious harmonic content. I can calculate that my +/-29VDC rails from 123VAC mains would give me +/-27.1VDC rails from nominal 115VAC mains; +/-24.4VDC rails from 90% (103.5VAC) mains; and +/-21.22VDC rails from nominal Japanese mains. That doesn't leave much leeway for low mains in Japan. Everybody else should be fine.
You can do the math yourself and decide if you think a 15V transformer will cover you in all the places and circumstances you'll be using it. The advantages are lower heat dissipation, and higher current capacity in the same size transformer. The disadvantage is the risk of regulator drop-out.
If you're building a power supply for a mike preamp, there's another factor to consider. A good voltage doubler on an 18V transformer can produce a solid 48V phantom supply in most cases. A 15V transformer generally can't. That would mean you'd need a tripler, which will produce a bit TOO much no-load voltage, but will be more difficult to keep stiff if you're pulling a bit of current from it (like if you're powering relays or LEDs off the phantom supply). Maybe not something you need to worry about, I don't know your application.
