We have heard the phrase statistical significance used quite a bit already when we are discussing a statistical finding that has a p value of less than or equal to 0.05. Often in research and academic journal articles, the phrase statistical significance is shortened to just state that it was a “significant finding.” This is problematic for many who dont understand that the word significant in a statistical perspective may be different than the more straightforward and widely used definition for significance that most people use. For example, consider a study that demonstrates a statistically significant difference between weight loss outcomes in two groups. One that took a weight loss supplement and one that did not. This may be described as a significant difference in outcomes between the groups because the p value was less than 0.05. Think about how the supplement company might use that for marketing purposes. They might say something like lose significantly more weight with our product or any of the other synonyms shown above that comes from thesaurus.com. But that isnt exactly what statistical significance means. If we have a large enough sample, we can find a statistically significant difference even when that difference is a small one (demonstrated at the end of the last chapter). So, we should also be concerned with practical differences.

Another way to describe practical difference is meaningfulness. How meaningful is the statistical difference we found? This information is what is probably most important to the consumer or end user, but it might not be the easiest to spin for marketing. Whenever we perform statistical analyses, whether they are looking for similarity (i.e. correlation) or difference (i.e. t test or ANOVA), we should also include a measure of meaningfulness.

Meaningfulness can be quantified by effect size estimates. We will discuss two main types of effect sizes, one of which you are actually already familiar with.

Effect Sizes

Effect sizes, or more appropriately effect size estimates, help us statistically quantify practical significance. Effect sizes have also been described as a measure of “meaningfulness.”^[1] From a purely practical and applied perspective, the effect size should be the primary outcome of research inquiry.^[2] No matter what we are measuring, we are likely most interested in the size of the difference or relatedness between variables. The reliability of our finding or how confident we are in the finding is important also, but it shouldnt be the only product.

This has become such a problem in many fields of research that more than 800 researchers recently published a paper condemning the usage of p values since many rely too heavily on it and not enough on practical measures of significance that demonstrate realistic effects.^[3] Completely abandoning measures of statistical significance like p values is probably a little too extreme, but their point is well taken. We should be publishing measures of both statistical and practical significance or meaningfulness.

So, lets define effect size. An effect size is a measure that describes the magnitude or size of the difference or relatedness between the variables we are measuring. This means that it is describing the practical/meaningful significance. As we will see in this chapter, it really all boils down to how much the data overlaps between our groups or variables. If we are comparing groups for differences, we would likely not expect to see much overlap in the data. If we were evaluating our data for similarity or association, we are looking for more overlap.

Types of Effect Sizes

As mentioned above, effect sizes can be created for both relatedness and difference tests. This essentially means that there are 2 families of effect sizes; r for relatedness and d for difference. Conversions can also be completed between the two if it was necessary to make comparisons across different study types. That won’t be demonstrated in this chapter, but a quick internet search for “r to d effect size conversion,” or vice versa, will bring up several online calculators and formulas.

Cohen’s d Effect Size Estimates

Looking on the difference side of effect sizes (ES), Cohens d effect size estimates are most commonly seen. They are named after Jacob Cohen. They can be quickly calculated when both means and the pooled standard deviation is known. The means can be from independent or paired groups. The pooled standard deviation is the standard deviation calculated from both groups as one.

[asciimath]"Cohen's d ES" = (M_1 - M_2)/"sd"_"pooled"[/asciimath]

[asciimath]M = mean, sd = "standard deviation"[/asciimath]

With this data, we simply subtract one mean from the other and divide by the pooled standard deviation. If the value is negative, that simply means that the second mean or the one that was subtracted from the other was larger than the original mean. This is dependent on how the dataset was ordered or how they were called into the formula. So, the direction doesnt really matter here since the magnitude is the largest concern. For this reason, the negative sign is usually not included. Effect size estimates can be interpreted with the scale provided. Anything less than 0.2 is considered a trivial difference, 0.2 to 0.49 is a small difference, 0.5 to 0.79 is a medium difference, and anything ? 0.8 is a large difference.

As mentioned previously, effect sizes really all boil down to the amount of overlap in the data. Figure 5.3 demonstrates this with 2 examples. The first one (on top) is a comparison of 2 groups. One distribution has a mean of 50, the other has a mean of 55, and the pooled standard deviation is 10. Solving for the Cohens d effect size, we find that there is a medium difference between these 2 groups with a value of 0.5 (d=(55-50)/10). The second example has a distribution with a mean of 50, a distribution with a mean of 110, and a pooled standard deviation of 60. Here we find a large difference with a Cohens d effect size of 1.0 (d=(110-50)/60).

Take a look at the amount of overlap between the distributions in each example. Can you see why the Cohens d effect size estimate is larger in the 2nd example? It has a lot less overlap between the distributions. What if the data distributions were completely separate with no amount of overlap? What would happen to the effect size then? It would increase quite a bit. What if nearly 100% of the data were overlapping? The effect size would shrink quite a bit and likely be trivial. This is of course if we are still using Cohens d as our measure of effect size, since it is looking for difference. But what if we wanted to look for similarity?

r Effect Size Estimates

When looking for similarity, the PPM r value should be used as an effect size. The fact that r is an effect size is a good thing, since we already know how to quantify it and use it from Chapter 3. Recall that r values can range from -1 to 0 to +1 and we have already seen how to interpret them in Chapter 3.

When interpreting effect sizes for relationships or associations, we are still concerned with the amount of overlap in the data, and weve seen this before in Figure 3.12 (now 5.4 in this chapter). When utilizing a bivariate correlation or simple linear regression with only 2 variables (which are represented by the ovals above), we are mostly concerned with the area where they have shared variance. That is represented as the area of overlap in this figure. Remember the coefficient of determination r² helps us explain the amount of shared variance. Please review Table 3.4 (reproduced below) and go back to Chapter 3 if you need a refresher.

Going back to our previous example with Cohen’s d in Figure 5.3, we see the first example that produced a Cohens d effect size estimate of 0.5 which is interpreted as a medium amount of difference between the variables. You might be thinking that you still see quite a bit of overlap in the data distributions. If there is a medium amount of difference between the variables, how much similarity is there? Intuitively, you might think that there is a medium amount of similarity, but that is incorrect as you can see in the scatterplot prodcued below the distributions (in Figure 5.5 below). In fact, it has a PPM r value of 0.053 which is considered a trivial amount of similarity. This demonstrates that difference and similarity are not necessarily on the same continuum. Just because there isnt a large difference between two variables, they arent necessarily related.

This also demonstrates why we use specific data visualization techniques with each statistical test. Whenever you are testing for difference, you should likely use a box and whisker plot, density plot, or distribution plot. Whenever you are testing for similarity, you should use a scatterplot. If the wrong plot is used, you may confuse or mislead your audience.

Calculating Effect Sizes

Fortunately, our r values are already effect size measurements, so we dont need to do anything else to them. However, whenever testing for difference, we still need to calculate Cohens d effect size estimates. These can be produced relatively easy in most statistical programs. Some are as simple as checking a box, but others may require the formula be typed in. As always, you can download the dataset if you would like to follow along with the tutorials below. This dataset includes performance data from jump testing, 30 yard sprints, and 1RM bench press. The sample was grouped based on bench press strength with those above 200 lbs in the “stronger” group and those below in the “weaker” group. This puts 10 into each group. An independent samples t test on the 30 yd sprint time variable produces a p value of 0.018, so we know there is a statistical difference between the 2 stronger and weaker groups. Calculating a Cohen’s d effect size will help understand how meaningful this difference is.

Other Types of Effect Sizes

There are quite a few other forms of effect sizes that can be used and should be with specific scenarios. Cohen’s d and the PPM r value are the most common and they are the most widely understood, so it is easy to see why many prefer them. No matter what effect size is used, it is important to remember that they are estimates of practical significance and should be reported along sidep values that provide a measure of statistical significance. Relying too much on one of the values may lead to a misinterpretation of the findings.