www.eamoncaddigan.net

Content and configuration for https://www.eamoncaddigan.net
git clone https://git.eamoncaddigan.net/www.eamoncaddigan.net.git
Log | Files | Refs | Submodules | README

index.md (5651B)


      1 ---
      2 title: "Average percentage change is not the percentage change of the average"
      3 date: 2024-08-11T19:42:59-07:00
      4 draft: false
      5 katex: true
      6 categories:
      7 - Science
      8 - Data Science
      9 tags:
     10 - Statistics
     11 ---
     12 
     13 When I present data to non-technical stakeholders, I sometimes express
     14 differences (typically changes over time) in terms of _percentage change_.
     15 This indicator has the advantage of being unitless and normalized[^norm],
     16 and is familiar to broad audiences (although some aspects are
     17 counter-intuitive---see below).
     18 
     19 The ratio of the difference between two observations (_x₀_ and _x₁_) and the
     20 starting observation provides the _relative change_:
     21 
     22 $$ \frac{x_1 - x_0}{x_0} $$
     23 
     24 Multiplying this proportion by 100 yields the percentage change.
     25 
     26 Usually I have more than two observations separated by time---I’ll have _x₀_
     27 and _x₁_ values for many observational units. Consider the following set of
     28 (made up) test scores:
     29 
     30 | Student     | Test 1 | Test 2 | Percentage Change  |
     31 |------------:|-------:|-------:|-------------------:|
     32 | 1           |  75.0  |  81.0  |      8.00%         |
     33 | 2           |  80.0  |  82.0  |      2.50%         |
     34 | 3           |  80.0  |  86.0  |      7.50%         |
     35 | 4           |  96.0  |  90.0  |     -6.25%         |
     36 | 5           | 100.0  |  90.0  |    -10.00%         |
     37 | **Average** |  86.2  |  85.8  |      0.35%         |
     38 
     39 The average score on Test 2 was lower than that for Test 1, so one might
     40 conclude that the students did worse. However, the average percentage change
     41 of students’ scores is positive; in other words, relative to their starting
     42 scores, students typically did better on the second test. This toy example
     43 illustrates an important point: the average of the percentage change (0.35%
     44 here) is not the same thing as the percentage change of the averages
     45 (approximately -0.46% in this case).
     46 
     47 I find the mathematical definition of these two measures illustrative. I’ll
     48 stick with relative changes to keep the equations tidier.
     49 
     50 The average of the relative change is:
     51 
     52 $$ \mathbb{E}\[ \frac{x_1 - x_0}{x_0} \] = \mathbb{E}\[\frac{x_1}{x_0}\] - 1$$
     53 
     54 On the other hand, the relative change of the averages is:
     55 
     56 $$ \frac{\mathbb{E}[x_1] - \mathbb{E}[x_0]}{\mathbb{E}[x_0]} =
     57 \mathbb{E}[\frac{x_1}{\mathbb{E}[x_0]}] - 1$$
     58 
     59 One way to think about these equations is that the _average relative
     60 change_ scales each _x₁_ by its associated _x₀_ before calculating the mean
     61 of the scaled value. The _relative change of the averages_ scales each _x₁_
     62 by the _mean_ of _x₀_, and then takes the mean of these.
     63 
     64 ## Which one to use
     65 
     66 Since they measure different things, each measure can be appropriate in
     67 different circumstances.
     68 
     69 If _x₀_ and _x₁_ comprise a sample from a larger population, then the sample
     70 mean of _x₀_ and _x₁_ are reliable estimates of the population mean of these
     71 values. In this case, the percentage change of the averages is probably most
     72 appropriate; it will estimate the percentage change in the population.
     73 
     74 On the other hand, if the observational units represent an entire
     75 population, or when using relative change to compare the behavior of
     76 different measures over the same time period, then calculating
     77 statistics---including the mean---of the relative changes is useful.
     78 
     79 ## Why I’m sharing this
     80 
     81 Perhaps this distinction is obvious to most folks[^folks], especially since
     82 there are other transformations that behave similarly[^transform]. As
     83 somebody accustomed to wanting the percentage change of the averages,
     84 I recently saw---in real world data---a fairly large discrepancy between
     85 that and the average percentage change, after deciding that I needed the
     86 latter. Only after checking my code for obvious errors did I confirm that
     87 the math made sense. The experience reminds me of encountering [Simpson’s
     88 Paradox](https://en.wikipedia.org/wiki/Simpson%27s_paradox)[^simpson] in
     89 real data.
     90 
     91 ## Other measures of relative change
     92 
     93 I avoid using percentage change as defined here (when I can) because it can
     94 lead to confusion. Among percentage change’s problems is its lack of
     95 “symmetry”; for example, an increase from 4 to 5 represents a 25% change,
     96 while a decrease from 5 to 4 represents a -20% change. This is addressed by
     97 other measures of relative change, such as the “arithmetic mean change”
     98 (where the difference is scaled by the mean of _x₀_ and _x₁_ rather than
     99 _x₀_ alone) or “logarithmic change” (where relative change is represented by
    100 the natural logarithm of the ratio of _x₁_ to _x₀_). Note that even with
    101 these measures, the average of relative changes are still distinct from the
    102 relative change of the averages; however, a symmetrical measure is probably
    103 a better choice when averaging changes.
    104 
    105 ## Percentage change in practice
    106 
    107 Since percentage change is only a _point estimate_, its presence should
    108 supplement---and never replace---an appropriate model-based approach to
    109 estimating the size of the change and the uncertainty of that estimate.
    110 Still, percentage change is a useful measure to include in reports and slide
    111 decks when it will be familiar to the audience, which is typically the case
    112 in a business setting. I hope this helps somebody else select the right
    113 calculation, and explain why the results are totally different from
    114 a “wrong” one.
    115 
    116 [^norm]: In the colloquial sense, not the statistical one.
    117 
    118 [^folks]:  At least the ones who’d read a post like this.
    119 
    120 [^transform]: For example, the mean of the logarithm of a variable is not
    121     the same as the logarithm of the mean of a variable.
    122 
    123 [^simpson]: I’m not sure if this could be considered a special case of
    124     Simpson’s paradox.