The received wisdom is that users don't hang around long on a web page, so you'd better get it to load quickly and then capture their attention immediately.
Some recent research by Liu et al, and discussed in an article by Jakob Nielsen, offers some more insight (and a big statistical study to back these up). As you'd expect, users make a quick decision - in maybe 10 secs or less - and then move on or stay. The longer they stay, the more likely they are to stay some more: the probability of them leaving drops pretty exponentially.
Statistically, the pattern follows The Weibull Hazard Function, a concept from reliability engineering. The pattern follows a "negative aging distribution". You would get the same results by asking "how likely is a given component in a machine to fail" when the quality of the components is highly variable (so that, the longer this particular component has been reliable, the more likely that it is one of the good ones off the line and will keep functioning). If the manufacturer had better QA in this hardware example, then the componets would begin to follow a "positive aging function" instead - they are all as well made as each other, and so the probability of failure simply rises over time due to wear and tear.