When the claim is made that driving distance has become more ‘important’ to performance in professional golf, I think what is meant, intuitively, is that having above-average driving distance now affords a larger advantage than it did in the past. Statistically, we might rephrase the question as, does driving distance ‘account’ for a larger portion of the variance (i.e. the spread or dispersion) of scores in today’s game than it did in the past?

If we agree that this is the claim being made, there are then two different ways it could be true. First, it might be the case that being 10 yards above average in driving distance is now “worth” more (in terms of total strokes-gained) than it was in the past. This could be true due to changing course setups: as courses get longer, players are forced to hit driver off every tee, which makes having 10 extra yards more useful than when players were sometimes choosing shorter clubs off the tee for strategic purposes. (There are of course other reasons for why this claim could be true, or false, too). Second, it might be the case that 10 yards is still worth the same in terms of strokes-gained, but now there are larger differences across golfers in their driving distances. That is, maybe it’s the case that being in the top 10% in driving distance in 2019 means you hit the ball 15 yards further than average, but in 1990 it meant you only hit it 10 yards further than average. If 10 extra yards is still worth the same in 2019 as it was in 1990, this would also result in driving distance being more ‘important’ in 2019, in the sense defined in the opening paragraph.

For a moment let’s focus on the question, “how much is 10 extra yards of driving distance worth?”. One interpretation of this is, “holding all other attributes of a golfer constant, how much of an increase in performance can a golfer expect by increasing their average drive by 10 yards?”. Another interpretation is, “on average, how many strokes better per round are golfers who hit it 10 yards further than the average golfer?” The answer to these two questions will differ because golfers who hit the ball far also tend to do other things well, or poorly. For example, players who hit the ball above-average distances on the PGA Tour are also above-average approach players. Therefore, the simple correlation between a golfer’s average driving distance and their performance in part captures the fact that longer players are, on average, better approach players. This part of the correlation may or may not be of interest. (For more thoughts on this, see [1]). In the analysis that follows, we examine how both the ‘raw’ relationship and the ‘conditional’ (on other attributes) relationship between distance and performance has evolved over time.

Let’s now turn to the data. Below we do two related analyses, which are likely best explained with examples. First, for each PGA Tour season from 1984-2019, we estimate by how much we would expect golfer A to beat golfer B given that A hits it 10 yards further than B; second, we estimate by how much we would expect A to beat B given that A is 1 standard deviation longer than B, where the standard deviation is specific to the PGA Tour season. The latter estimate accounts for the fact that differences across golfers in terms of driving distance have been changing over time; it will provide us with a more complete answer to the question we laid out in the introduction. Also, as previously mentioned, for each analysis we present two versions of the estimate: one that holds constant, or controls for, each golfer’s driving accuracy, strokes-gained approach, strokes-gained around-the-green, and strokes-gained putting (the conditional relationship), and one that does not hold any golfer attributes constant (the raw relationship).

The statistical details are here [2]. Briefly, we are correlating a golfer’s skill in each attribute (driving distance, driving accuracy, etc.) at the time of a tournament, with their subsequent performance (total strokes-gained) in that tournament. The conditional relationships are estimated using regressions with all 5 attributes included, while the raw relationships are just simple regressions with the relevant attribute as the only independent variable.

This first plot shows the raw relationship between driving distance (measured in yards) and total strokes-gained, as well as that between driving accuracy (measured in % fairways hit) and total strokes-gained, from 1984-2019. Stated with examples, the raw relationship tells by how much we would expect golfer A to beat golfer B given that the only thing we know about them is that A hits it 10 yards further than B; the conditional relationship (use toggle to view) between driving distance and strokes-gained performance will tell us by how much we expect golfer A to beat golfer B given that A hits it 10 yards further than B, but they are equally skilled in all other respects.

Focusing first on the raw relationship, we see that the average difference in performance between two players with a 10 yard driving distance difference has bounced around 0.3-0.35 strokes per round since 1984, with a recent uptick since 2015. Toggling to view the conditional estimates, we see an upward trend since 2004 (however, we also see this in the raw correlations). Note that since we are holding constant driving accuracy (percentage of fairways hit per round), the driving distance estimates are likely higher than they “should” be, since driving accuracy should mechanically decrease when distances goes up. Finally, note that, all else equal, you would probably expect the impact of 10 yards to decline as the average driving distance increases: e.g. 10 yards is worth more when the average is 270 yards than when the average is 300 yards. (We will discuss the driving accuracy patterns in the next plot, as they are virtually identical to these.)

For the next plot, we repeat the analysis above except that each attribute is normalized to have a mean of zero and a standard deviation of 1 in each season. (Recall that for normally distributed variables the standard deviation is equal to the difference between the median and the 84th percentile.) Returning to our earlier example with golfers A and B, the interpretation is the same except now the difference we are considering between A and B is equal to 1 standard deviation, and therefore can vary by season. We also include in this plot the estimates for the strokes-gained categories; hover over a data point to see its standard deviation in that season.

Looking at the raw relationships, we see that the distance trend is increasing slightly more than the first plot because the standard deviation in driving distance has increased over time. The driving accuracy estimate has steadily decreased since 1984, with a slowdown in the decline occuring since 2004. It is interesting to note that the highest coefficient by far among the raw correlations is that of strokes-gained approach. This is in large part due to the fact that good approach players tend to be good at all other aspects of the game (see correlation matrix below).

Moving to the conditional relationships, the distances estimates are again trending up; we also see that the driving accuracy estimates have been flat or slightly increasing since 2004. It is also interesting to note the relative sizes of the estimates here: improving driving accuracy by 1 standard deviation (~4.5 percentage points) yields an expected improvement in overall performance that is similar in magnitude to improving SG: approach by 1 standard deviation (~0.35 strokes).

It might be illuminating to consider the interpretation of the strokes-gained category estimates. These estimates tell us by how much we would predict golfer A to beat golfer B given that A is 1 standard deviation better in a given SG category, and equal in all other respects. Given the nature of our skill estimates, we would expect that a 1 stroke difference in skill for an SG category would result in a 1 stroke difference in total strokes-gained going forward. If this were true, the estimate in the plot above would equal the standard deviation in that season. (We see deviations from this due to a combination of noise and the fact that some categories predict others, e.g. ARG predicts APP slightly.) The fact that putting has a smaller coefficient than approach in every season reflects the fact that skill differences across players in SG: putting are smaller than skill differences across players in SG: approach [3]. Therefore, approach play is more important in the sense that improving from the 50th best approach player to the 10th best approach player is worth more than the same improvement in ranking in putting.

The differences between the raw and conditional relationships can be, for the most part, reconciled by examining the correlation matrix below, which displays the pair-wise correlation between our skill estimates in each season.

There are many interesting patterns here, a few of them are:

There is a lot of information to absorb from this analysis. To wrap things up, here are the key takeaways with respect to the question we set out to answer: