Linear Relationships – Correlation
So far we have visualized relationships between two quantitative variables using scatterplots, and described the overall pattern of a relationship by considering its direction, form, and strength. We noted that assessing the strength of a relationship just by looking at the scatterplot is quite difficult, and therefore we need to supplement the scatterplot with some kind of numerical measure that will help us assess the strength.
In this part, we will restrict our attention to the special case of relationships that have a linear form, since they are quite common and relatively simple to detect. More importantly, there exists a numerical measure that assesses the strength of the linear relationship between two quantitative variables with which we can supplement the scatterplot. We will introduce this numerical measure here and discuss it in detail.
Even though from this point on we are going to focus only on linear relationships, it is important to remember that not every relationship between two quantitative variables has a linear form. We have actually seen several examples of relationships that are not linear. The statistical tools that will be introduced here are appropriate only for examining linear relationships, and as we will see, when they are used in nonlinear situations, these tools can lead to errors in reasoning.
Let’s start with a motivating example. Consider the following two scatterplots.
We can see that in both cases, the direction of the relationship is positive and the form of the relationship is linear. What about the strength? Recall that the strength of a relationship is the extent to which the data follow its form.
The purpose of this example was to illustrate how assessing the strength of the linear relationship from a scatterplot alone is problematic, since our judgment might be affected by the scale on which the values are plotted. This example, therefore, provides a motivation for the need to supplement the scatterplot with a numerical measure that will measure the strength of the linear relationship between two quantitative variables.
The numerical measure that assesses the strength of a linear relationship is called the correlation coefficient, and is denoted by r. We will:
- give a definition of the correlation r,
- discuss the calculation of r,
- explain how to interpret the value of r, and
- talk about some of the properties of r.
Calculation: r is calculated using the following formula:
However, the calculation of the correlation (r) is not the focus of this course. We will use a statistics package to calculate r for us, and the emphasis of this course will be on the interpretation of its value.
Once we obtain the value of r, its interpretation with respect to the strength of linear relationships is quite simple, as these images illustrate:
In order to get a better sense for how the value of r relates to the strength of the linear relationship, take a look the following applets.
If you will be using correlation often in your research, I highly urge you to read the following more detailed discussion of correlation.
Now that we understand the use of r as a numerical measure for assessing the direction and strength of linear relationships between quantitative variables, we will look at a few examples.
- Note that in both examples we supplemented the scatterplot with the correlation (r). Now that we have the correlation (r), why do we still need to look at a scatterplot when examining the relationship between two quantitative variables?
- The correlation coefficient can only be interpreted as the measure of the strength of a linear relationship, so we need the scatterplot to verify that the relationship indeed looks linear. This point and its importance will be clearer after we examine a few properties of r.
We will now discuss and illustrate several important properties of the correlation coefficient as a numerical measure of the strength of a linear relationship.
- The correlation does not change when the units of measurement of either one of the variables change. In other words, if we change the units of measurement of the explanatory variable and/or the response variable, this has no effect on the correlation (r).
To illustrate this, below are two versions of the scatterplot of the relationship between sign legibility distance and driver’s age:
The top scatterplot displays the original data where the maximum distances are measured in feet. The bottom scatterplot displays the same relationship, but with maximum distances changed to meters. Notice that the Y-values have changed, but the correlations are the same. This is an example of how changing the units of measurement of the response variable has no effect on r, but as we indicated above, the same is true for changing the units of the explanatory variable, or of both variables.
This might be a good place to comment that the correlation (r) is “unitless”. It is just a number.
- The correlation only measures the strength of a linear relationship between two variables. It ignores any other type of relationship, no matter how strong it is. For example, consider the relationship between the average fuel usage of driving a fixed distance in a car, and the speed at which the car drives:
Our data describe a fairly simple non-linear (sometimes called curvilinear) relationship: the amount of fuel consumed decreases rapidly to a minimum for a car driving 60 kilometers per hour, and then increases gradually for speeds exceeding 60 kilometers per hour. The relationship is very strong, as the observations seem to perfectly fit the curve.
Although the relationship is strong, the correlation r = -0.172 indicates a weak linear relationship. This makes sense considering that the data fails to adhere closely to a linear form:
- The correlation by itself is not enough to determine whether or not a relationship is linear. To see this, let’s consider the study that examined the effect of monetary incentives on the return rate of questionnaires. Below is the scatterplot relating the percentage of participants who completed a survey to the monetary incentive that researchers promised to participants, in which we find a strong non-linear (sometimes called curvilinear) relationship:
The relationship is non-linear (sometimes called curvilinear), yet the correlation r = 0.876 is quite close to 1.
In the last two examples we have seen two very strong non-linear (sometimes called curvilinear) relationships, one with a correlation close to 0, and one with a correlation close to 1. Therefore, the correlation alone does not indicate whether a relationship is linear or not. The important principle here is:
Always look at the data!
- The correlation is heavily influenced by outliers. As you will learn in the next two activities, the way in which the outlier influences the correlation depends upon whether or not the outlier is consistent with the pattern of the linear relationship.
Hopefully, you’ve noticed the correlation decreasing when you created this kind of outlier, which is not consistent with the pattern of the relationship.
The next activity will show you how an outlier that is consistent with the direction of the linear relationship actually strengthens it.
In the previous activity, we saw an example where there was a positive linear relationship between the two variables, and including the outlier just “strengthened” it. Consider the hypothetical data displayed by the following scatterplot:
In this case, the low outlier gives an “illusion” of a positive linear relationship, whereas in reality, there is no linear relationship between X and Y.