# What does correlation measure

## Correlation

Here is a scatterplot of heights and weights for a sample population:

Looking at this graph, you should get the sense that there is some relationship between a person's height and their weight. In particular, if you know that a person's height is greater than average, you might expect that their weight is also greater than average. The strength of this association between two variables is called the correlation. We will use the symbol ** r** to represent the measure of correlation.

What should a measure of correlation ** r** depend on? Obviously it should reflect how strongly the two variables are related. If knowing the value for the variable on the x-axis of a graph implies perfect knowledge of the value for the variable on the y-axis, then this point should fall on the regression line. If knowing the value for the variable on the x-axis gives a strong ability to predict the value for the variable on the y-axis, then this point should fall near the regression line, depending on the accuracy of the prediction. If knowing the x-value for the variable gives only a weak ability to predict its value on the y-axis, then the y-value for the point will tend to fall further from the regression line. Thus, it is reasonable to expect that the strength of a relationship would be visible on a scatterplot of the two variables, such as shown in the graph above. When the points on the scatterplot are clustered close to the regression line,

**should be high. For example, if knowledge of an individuals' height (x-axis) is helpful in predicting weight (y-axis), then the points on the scatter plot should cluster around the regression line, and**

*r***should be positive. The next graph shows the plot of weights vs. heights with the regression line added.**

*r***Pearson's correlation coefficient** ** r** is the most commonly used measure of association in the social sciences.

**summarizes the linear relationship between two variables having ranked categories. That is, the variables may be continuous. ordinal. interval. or ratio. Its formula is as follows:**

*r*In this equation, ** n** is the sample size,

**. is the observed sample mean for variable**

*x***.**

*y***is the standard deviation for variable**

*S*_{x}**and**

*x***is the standard deviation for variable**

*S*_{y}**.**

*y***and**

*X*_{i}**represent the values of variables**

*Y*_{i}**and**

*x***for the i**

*y*_{th}individual in the sample. The values for

**generated by this equation will range from -1 to 1. A value of -1 indicates a perfect inverse correlation between the variables (that is, when the value for one variable is high, the value for the other variable is low). A value of 0 indicates no relationship between the variables. A value of 1 indicates a perfect correlation between the variables.**

*r*In the example data set above, the scatterplot and regression line lead us to believe there is a correlation between height and weight. The value of ** r** for this data set is 0.76. As we expect, this is much higher than a value of zero, thus indicating a correlation between the variables. But the value is less than 1, indicating that the relationship between the variables is not perfect.

Regression and correlation are intertwined. In fact, the slope of the regression line is directly related to the value of ** r** . as will be explained below. For now, we only need to recognize that, as the value of

**moves away from 0, the data points will cluster closer to the regression line. This concept is illustrated in the following activity.**

*r*Activity 1

Below is a scatterplot for a hypothetical set of points. You can grab the blue line (initially lying on the x-axis) and move it up or down to adjust the value of ** r** through a range of -1 to 0 to 1. As you move the line, the points move as well to yield the desired level of correlation.

Try moving the blue line and observe the corresponding movement in the data points. You will notice that the data points align closer to the blue line (which is also the regression line for the data points) as you approach a 45 degree angle from the x-axis. The value for ** r** (shown in the upper right corner of the graph) changes with movement of the regression line.

## Units of Analysis in Frequency Distributions

Consider annual incomes for a large sample of families that have been converted from U.S. Dollars to the currency of another country, such as British Pounds or Japanese yen. Below is a graph that can show any of three separate distributions (one for Dollars, one for Pounds, and one for Yen). Click on one of the types of currency to observe the frequency distribution.

Activity 2

Now, let's reinforce the ideas of Activity 2. Suppose that an international corporation headquartered in the U.S. sends a large group of employees to its main offices in each of three countries -- Bolivia, Denmark, and Hong Kong. Each team consists of comparable numbers of executives, supervisors, technicians, secretaries, etc. who are about to leave the U.S. to take up residence in one of the three countries. All of the workers will be paid in an amount equivalent to their American salaries. However, these salaries will be converted into local currencies.

If each frequency distribution of income scores is superimposed on a grid where each square of the grid represents 10,000 units of currency, then the overall width of the x-axis will change depending on which currency is used. However, the distribution of incomes does not change. Somebody making a relatively high salary (at least, relative to the other team members) under one unit of currency will still be making a relatively high salary if the currency is converted properly.

Below, you may choose between three histograms. Each histogram represents an income frequency distribution for one of three countries). The sample size and frequency is identical for each histogram (thus, ** n _{Bolivia} = n _{Denmark} = n _{Hong Kong} ** ).

## Correlation and Standardized Distributions

In the previous section, you saw that the size of the units used to measure a variable may affect the appearance of a frequency distribution even though the actual frequencies remain unchanged. Now let us see if changing the units of analysis affects the correlation between two variables.

Activity 3

Assuming that the side of an individual square on the grid stands for a constant number of units, the ratio of the sides of a scatterplot will vary with a change in the units of either of those variables. Let us consider again the example of comparing peoples' heights to their weights. If we change the measure of height from inches to centimeters, the value on the x-axis (weight) will stay constant, while the y-axis (height) will change when the unit of measure for income is changed. Of course, changing the unit of measure for the height does not affect the measure for the strength of the relationship.

We can eliminate the effects of the unit of measurement by transforming the values to Z-scores. Just as the strength of the correlation remains unchanged by changes in units, the strength of the correlation remains unchanged when we transfrom the values to their Z-scores.

To transform scores on ** X** and

**to**

*Y***and**

*Z*_{x}**. use the following equations:**

*Z*_{y}where ** X** is a score on an independent variable,

**. and**

*X***is the standard deviation of the scores on**

*s*_{x}**. and similarly for**

*X***. the standardized scores for**

*Z*_{y}**. the dependent variable.**

*Y*Recall that the raw regression coefficient ** b_{yx} ** indicates the average change in the

predicted value for scores on ** Y** that occurs when there is a one-unit increase in the value for

**.**

*X***is the**

*B*_{yx}**standardized**regression coefficient. It indicates the amount of change in the predicted value for

**that occurs when there is a one-unit increase in the value of**

*Z*_{y}**. As you saw above, the changing size of the empirical units of**

*Z*_{x}**and**

*X***may affect the**

*Y***appearance**of a frequency distribution, but not the

**relative**distance of scores from each other on either variable. Similarly, a change in empirical units of

**and**

*X***may affect the appearance of the relationship when presented in a scatterplot. This change also affects the size of**

*Y***. the raw regression coefficient. But, changing the units of measure does not affect the size of**

*b*_{yx}**. the standardized regression coefficient. By converting all scores on**

*B*_{yx}**and**

*X***to standardized scores, you standardize the measure for the correlation coefficient.**

*Y*Changing from raw scores to Z-scores affects the range of values possible for the regression coefficient. Since raw scores can exagerate the height of the value on the y-axis, the slope of the regression line can be arbitrarily high. Thus, the raw regression coefficient can be anywhere in the range -

to +. When the scores are standardized, the regression coefficient must fall between -1.00 and +1.00, inclusive.The next graphic represents the standardized regression slope (** B_{yx} ** ). The colored (green) area shows the possible range for the standardized regression slope when

**depends on**

*Y***.**

*X*The correlation coefficient can be defined in terms of the regression coefficient as follows:

Thus, knowing the values of the raw regression coefficient (** b_{yx} ** ),

**(standard deviation of variable x) and**

*s*_{x}**(standard deviation of variable y) allows us to compute**

*s*_{y}**. the correlation coefficient.**

*r*_{yx}When we convert to Z-scores, the standard deviation of variables x and y becomes 1. Thus, we see from the above equation that the value of the standardized regression coefficient is the same as the value of the correlation coefficient.

If height is related to weight, then obviously weight must be related to height. In other words, if knowing the height helps us to predict the weight, then knowing the weight should be equally helpful in predicting the height.

The first graph in this tutorial showed height on the** X** axis (the independent variable) and weight on the

**axis (the dependent variable). This is an arbitrary selection, and we can flip the axes as follows:**

*Y*If we flip the axes, then we also change the slope of the regression line. The next graphic shows the standardized regression slope (** B_{xy} ** ) -- that is, the indication of how

**relates to**

*Z*_{x}**. rather than the more typical indication of how**

*Z*_{y}**relates to**

*Z*_{y}**. The colored (red) area represents the possible range within which**

*Z*_{x}**must fall.**

*B*_{xy}## Correlation versus Causation

If two variables are correlated, then they are related in the sense that knowing the value for one variable can help to predict the value for the other. It also means that high values of one variable tends to occur along with high (or low) values of the other variable.

Correlation is quite different from a cause-and-effect relationship. Often we would like to know what causes something to happen. For example, what causes a person's height to be a given value? We have already seen that height and weight are correlated. Does this mean that a person's height is "caused" by their weight? This seems to be a silly proposition, and it is. One way that we can clearly see that this proposition is silly is to recognize that a person can increase or reduce their weight by dieting, and this will have no affect on their height.

If we know that a person's height and weight are correlated, and we know that the weight does not cause the height, does this necessarily mean that the height "causes" the weight? In other words, must there always be a cause-and-effect relationship between two correlated variables?

The answer is no. One possibility is that the correlation we see is merely coincidental. That is, we may simply have gathered a peculiar sample that has two correlated variables, but the general population does not have this same correlation.

Even when two variables are statistically correlated, it is not necessarily the case that one variable causes the other. It is possible that both variables are caused by a third variable. In the case of our example of height and weight, the individuals sampled are children in the age range 3 to 13. Consider the following two graphs. One shows the relationship between age and height, while the other shows the relationship between age and weight.

In this example, height and weight are correlated to each other, and they are also each correlated to age. It is reasonable to hypothesize that a person's age "causes" their height and weight to some degree. Clearly, neither height nor weight "causes" age.

## Coefficient of Determination

Recall that we have described correlation as an ability to make predictions. Let's examine this idea more closely. Consider your ability to predict the height for an individual when all you know is the mean height of the population.

Activity 4

Imagine that you are trying to predict the score that a random individual will have for some variable ** Y** . What value should you guess if you know nothing about the individual? The best guess that you could make is the mean for the variable,

**.**

*s*_{y}Now, consider what happens when you try to predict the score for some individual on variable ** Y** . but you already happen to know the value for that individual on variable

**. If**

*X***and**

*Y***are correlated, then you can make a better prediction. If the variables are perfectly correlated (that is, the value for**

*X***is 1 or -1), then your prediction for**

*r***will be completely correct if you know**

*Y***. The data points, when shown on a scatterplot, all fall on the regression line. In this case, given a value for**

*X***. the value that you predict for**

*X***should be taken directly from the regression line.**

*Y*If the variables are not correlated at all (that is, the value for ** r** is 0), then the best prediction you can make for

**will just be the mean for**

*Y***. and the average error in your prediction will be**

*Y***.**

*s*_{y}When the correlation between** X** and

**has an**

*Y***value somewhere between 0 and 1, then the average error in the prediction will be somewhere between 0 and**

*r***. The value of the correlation**

*s*_{y}**explains**some of the variance in

**. The amount of predictability is called the**

*Y**coefficient of determination*. In visual terms, as the value of

**increases, the data points appear closer and closer to the regression line. The "standard deviation" from the regression line drops as the correlation increases. Thus, the ability to predict improves as the correlation increases. We can see that this is so from the following graphs. The graph on the left again shows height vs. weight. The graph on the right shows the "residual" weights after we have removed the variance that is "explained" by the value of the height for a given data point. This is equivalent to rotating the graph so that the regression line lies on the**

*r***axis.**

*X*Source: simon.cs.vt.edu

Category: Forex