I have not looked at the regressions that I did quite a while back, but the form I remember perfectly. Try a regression of pace on log(distance) (a best fit includes the second-order term as well, which should be small be useful. I found that the WRs (for men) fit extraordinarily well, and all of the WR/bests for "non-championship distances" have "positive errors" except the 3000. (I think that the R Squared statistic was something like 0.99999).
This implies that when you plot them, use the log distance vs pace. It does not work well when you get to shorter distances where the start matters and where entirely different energy systems rule. Similarly, where different energy system constraints come into play there is a drop off (marathon).