There's a lot of advice dispensed on the topic. Very little of it is backed by evidence. Google's study on disk failures seems to be the only study that's based on a statistically significant sample of consumer drives.
Keeping disks cool is a common recommendation — “everyone knows” that heat is bad for computer parts. However, one of the results of the Google study is that
Surprisingly, we found that temperature and activity levels were much less correlated with drive failures than previously reported.
Very hot disks apparently fail more than merely warm disks, but cool disks even more so. (Maybe humidity is a contributing factor.) However, you're unlikely to reach the more failure-prone cool temperatures (<30°C) outside of a datacenter.
One piece of the “conventional wisdom” about disk failures that is corroborated by the study is that older disks are more prone to failures. On the other hand, the data does not conclusively show that using a disk more intensely increases the probability of failure.
I think I've heard “I'm not buying any other brand” and “I'll never buy this brand again” about every brand of hard disk. My feeling is that everyone generalizes on one or two anecdotes and any remark about brand reliability should be ignored. There may be good and bad models, or more likely good and bad manufacturing runs; but recommendations based on these would have a useful life measured in weeks, months at most.
<added>
The data recovery firm Storelab has published a breakdown of the disks that have been submitted to them by brand. Non-Russian-speakers may prefer to read the writeup at Tom's Hardware. It shows a marked difference between market shares and failure numbers. I don't know how much observation bias there is (in particular the market share figures seem to be global, whereas the failure numbers are localized).
</added>
I don't see how defragmenting would affect the reliability of the drive. It does have the advantage of reading every data block, revealing any error, so it's indirectly useful as a data consistency verification.