If you are getting sick cards, then most of the time it's simply too high memory clock or core clock, or cards running too hot.

Hook card up to a monitor, and run OCCT GPU stress test with error checking mode. Or you could use furmark which has similar features.

Run the card at the clocks you intend to mine at, using the most brutal shader/ram settings in those programs you can. This will reveal if the card is stable at the card's "maximum utilization" aka maximum power and heat that any miner algo could possibly throw at it.

Be sure to let the card temperature ramp up and equilibrate for some time, before judging if it's stable. If it's not, improve fanspeed and/or lower clocks.

Also something to keep in mind, with the new algos that use very little power, is on many cards you can undervolt the cards quite considerably. Because efficiency/clock speed is a non-linear relationship, you can find that for example a decrease in voltage of 10% will give a much greater than 10% improvement in efficiency. For some R9 270 cards for example, a 10% reduction in voltage will give about a 30% improvement in efficiency. This improvement is so dramatic because most of the R9 series is clocked at the factory right up to its stability limit anyway. tl;dr - unless you have free power its better to improve efficiency rather than striving for top-notch clock frequencies.