Does averaging stats work?
Tutelman -- November 20, 2017
Does the average of the function
equal the function of the average? That's an odd-sounding question, but
it is important to know the answer for some kinds of golf research --
particularly when your data consists of averages and you want to use
to find or test a mathematical relationship.
Lots of times in golf research (or any kind of investigation), you want
to test some mathematical function that lets you calculate how the
world works. For instance, suppose we want to test the common assertion:
distance = 2.5 * clubhead speed
worth noting this
assertion is false, but you see it often enough in popular
articles -- and for this discussion it has the advantage of simplicity.
You want to test the function against real live data, which is exactly how you should
test it! Unfortunately, unless you own a TrackMan or
equivalent, it will be hard to get that data.
wait! TrackMan publishes annually a table of data
showing how tour players hit the ball. (The image is a snapshot
directly from the TrackMan web page.) We should be able to use that
Not so fast! The formula works (or doesn't work) for a drive -- one drive. Measure
the clubhead speed, measure the carry distance, and see if they are
related by a factor of 2.5. But this chart isn't about individual
shots. Each number there is an average of many shots by many different
golfers. The shots had many different clubhead speeds and many
different distances. Before we use the data from such
combined-statistics charts, we need to answer the question: Does the average of the
results come out to be the same number as the result of the average?
If not then it is not valid to check the formula assuming each row is
an individual golf shot.
The words may be confusing, so let's show the question visually. Let's
consider a function to go from clubhead speed to carry distance. For
instance, for the assertion above, the function would be f(x)=2.5x.
Here are two ways we could use the function.
In the diagram, clubhead speeds are black and carry distances are blue.
Data for individual shots are thin arrows and averages are fat arrows.
these two different calculations give the same answer?
That is a very important question if we want to use charts of averages
like the TrackMan chart or similar charts from the PGA Tour.
- The calculation on the left takes each individual
shot's clubhead speed and applies it to the function to get a carry
distance. Then it averages the carry distance for a final answer. Mathematically, it is Avg(f(x)), the average of the function.
- The calculation on the right averages all the
clubhead speeds, and applies the resulting average clubhead speed to
the function to get a final answer. Mathematically, it is f(Avg(x)), the function of the average.
If they do give
the same answer, then we can apply functions to the cells
in the charts and expect them to make sense.
If they don't give
the same answer, it isn't valid to use the data averages
for anything interesting.
Simple numerical examples
Example #1: One that works
Let's start with a very simple function: f(x)=2x+1
We'll compute the answer both ways. Excel makes the
trivially easy, so here is the spreadsheet direct from Excel. We have
take individual data points -- values of x
-- from zero to ten. (That is 0, 1, 2, 3,... 10) We apply the function f(x)
to each data point individually; that is the second column. The blue
cells at the bottom are the averages of the columns. Let's see how the
calculations worked out.
So for this function and this data distribution, it all works. I'll go
further and say (without proving it here) that with this function it
will work for any distribution of input data x.
- The average value of x
- The average value of f(x)
- If we apply the function f(x)
to the average of x,
we get the average of f(x).
= 2*5 + 1 = 11
Example #2: One that doesn't work
So far the results are encouraging. Let's try another simple function: f(x)=x2
Again, we compute the answer both ways. This time:
have a problem. 25 (the function of the average input) is not
the same as 35 (the average of all the fuction outputs).
What this means is that the function f(x)=x2 will
not give valid results when applied to a table of data averages.
- The average value of x
is 5. (Not surprising; we used the same input distribution.)
- The average value of f(x)
- If we apply the function f(x)
to the average of x,
= 52 = 25
Why does it work this way? Can we tell anything about
which functions will work and which will not?
Let's start by looking at a graph of the two functions we just tested.
We see right away that:
That is a critical difference. If the slope of the function does not
change, the function is linear,
and linear functions do not distort averaged data. If the function is
curved -- if the slope changes -- then you can't trust it not to
distort the data. It might give correct answers by coincidence, with
the right data; but you can't depend on it, so it isn't valid to test
- The one that worked (2x+1)
is a straight line. The slope does not change from 2.0 over the range
- The one that didn't work (x2)
has a lot of curvature. The slope is zero (horizontal) at x=0 and all
the way up to 20 at x=10.
How can we tell whether the function we are using is linear? Remember
when you learned to graph a straight line back in high school or
perhaps even middle school? The form I learned for the equation at the
y = mx + b
is the slope of the line and b
is the place it crosses the y-axis. The function is linear if, the only
appears in the function, it is only multiplied by a constant. No
squares, no inverses, no exponents -- just multiplied by a constant. If
we have a
function of several variables (x1, x2,...),
then the form for it to be linear is:
y = k0
+ k1x1 + k2x2
where all the ks
are constants. Functions in this form are linear, and will not distort
(For those who are interested, the proof is in the Appendix.)
This topic came up most recently in a discussion of how to find launch
angle from angle of attack (AoA) and dynamic loft. The golf community
has gradually come to accept a relationship of the form:
L = A + p(D-A)
is launch angle, A
is AoA, and D
is dynamic loft. The percentage p
is generally acknowledged to vary with the loft. (Specifically, it
varies with spin loft, which is equal to D-A.)
The person doing the work in this project had found a table of tour
averages, and was trying to use it to validate the value of p
for various clubs.
is valid to use averages in this case?
Well, the function is certainly linear in D
each of those variables appear only multiplied by p
or 1. But p
varies with loft, so maybe it's a problem. Let's look at the variation
I have curve-fit a formula for p
as a function of spin loft S.
Specifically, it is:
= .96 - .0071 S
If we apply that to the launch angle, and then S=D-A
A + (.96 - .0071S)(D-A)
L = A + .96(D-A) - .0071(D-A)2
Hmmm! Looks like we have a square in the function. How badly will that
mess up research based on averaged data?
Let's do some graphing. We'll set AoA to some harmless value (zero is
pretty harmless). For zero AoA, the dynamic loft equals the spin loft.
Here are graphs of the launch angle and its slope, plotted against spin
loft. Remember that D
is the same as S
and the slope of the curve is the coefficient p.
It is clear from looking at launch angle that the function shows plenty
of curvature. Remember, curvature is what distorts statistical
distributions and invalidates using averages as if they were individual
But there is a silver lining in the cloud of doubt here. The table of
averaged data is grouped by club: driver, 3-wood, and down to pitching
wedge. Each club has its own row of averaged data, and the averaging is
confined to just one type of club. How does that help? If the averages
are only over one type of club, that greatly limits the range of spin
loft we have to consider.
For instance, the driver for a tour player is not likely to be outside
the range of 10°-13° of spin loft. True, the clubhead's loft is lower than that for the
static club, but shaft bend adds a little loft. In any event, it's
about a 3° range. When we look at the slope on the graph on the right,
a 3° range is about a 0.021 range of slope. For a driver, that's about a
2.4% change of slope over the range -- not much curvature at all. And
you can tell that by eye as well, looking at the graph on the left. By
eye, the curve from 10° to 13° is indistinguishable from a straight
So it's probably quite all right to use the data, as long as we confine
our conclusion to one row of data at a time. Over the variation in that
row, the function is linear, for practical purposes if not
Relationship to series expansions
If you took advanced algebra or calculus, here's an interesting way of
looking at it. If you didn't, you won't miss much by not reading this
Do you remember expanding functions as Taylor's series? It looked like:
= k0 + k1(x-a)
You could get a very good representation of any function in the
vicinity of the point x=a
How good? For a given range around the point a
you can be arbitrarily good; you just have to take enough terms in the
series. Example: for a range of 3≤x≤5
you want to use a=4
you might need to use a third-order series to represent your function
accurately within 0.1%.
This is directly applicable to what we have here. For the range of
interest, if we get a good representation of the function from a first-order
Taylor series, then
we can say the function is effectively linear over the range of interest. Why? Because a first-order
series is a linear equation:
f(x) = k0
In golf research (and other types of research as well) we often find
data available in the form of averages over a lot of trials. It is seldom
the raw data for each individual trial, unless we did the experiment
ourselves. Can we use such averaged data to test or even derive
- If the mathematical relationship is linear in the
data variables, then we can draw valid conclusions about individual
trials from the averages.
- To the extent the mathematical relationship contains
significant curvature, then any conclusions we draw about individual
trials cannot be trusted.
Proof that a linear function allows averaged inputs and outputs
The major idea in this article is that a linear function gives the same
answer for the average of the function that it does for the function of
the average, and that a nonlinear function cannot be trusted to give
the same answer. The latter is easy to show; we already did when we
looked at the second simple numerical example, f(x)=x2.
But we need to prove the former. So far we just showed that it
works for one linear function with one distribution of data; that does
not prove the general case. So here's a proof of the general case.
start with the diagram we used at the beginning of the article, and the
function we use will be the perfectly general linear function, y=mx+b. The data distribution will be a perfectly general set of xi.
We work through the calculation for each side. When we have the result,
we can see by inspection that they give the same answer.
|y = Avg ( f(xi) )
|y = f( Avg (xi) )
Both columns end with exactly the same expression, so both methods of calculation give the same result.
Last modified - Nov 22, 2017