Correlation Analysis - Full Course in 30 min

Correlation analysis measures the relationship between two variables, such as a person's salary and age. It aims to determine the strength and direction of this relationship, represented by a correlation coefficient ranging from -1 to 1. A positive coefficient indicates that variables move in the same direction, while a negative coefficient indicates they move in opposite directions. The strength is categorized from no correlation (0-0.1) to very strong (0.7-1).

The Pearson correlation coefficient (R) quantifies the linear relationship between two metric variables. It indicates the strength and direction of this linear relationship. The video explains the formula, showing how individual values, their means, and sums influence the coefficient. It also discusses hypothesis testing for Pearson correlation, where the null hypothesis states no linear relationship exists between variables in the population, and the alternative hypothesis states there is one. Assumptions for testing include normal distribution of both variables.

The Spearman rank correlation is the non-parametric counterpart to Pearson correlation. Unlike Pearson, it uses the ranks of the data rather than raw data. The video illustrates this with an example of reaction time and age, showing how ranks are assigned and then used to calculate a Pearson correlation on these ranks. If there are no rank ties, a simpler formula can be used. Similar to Pearson, Spearman's coefficient also ranges from -1 to 1.

Kendall's Tau is another non-parametric correlation coefficient, suitable when data is not normally distributed or variables are at ordinal scale levels. It is preferred over Spearman's when there are few data points with many rank ties. The calculation involves identifying concordant and discordant pairs (pairs that move in the same or opposite direction) and using a formula based on these counts. Kendall's Tau also ranges from -1 to 1, and its significance can be tested.

The Point-biserial correlation is a special case of Pearson correlation, used to examine the relationship between a dichotomous variable (e.g., male/female, passed/failed) and a metric variable. To calculate it, the dichotomous variable is converted into numerical values (e.g., 0 and 1). The video demonstrates that calculating Pearson correlation with this encoded dichotomous variable yields the same result as using the dedicated Point-biserial formula. Interestingly, the p-value for Point-biserial correlation is identical to that obtained from an independent samples t-test. Assumptions for hypothesis testing include the metric variable being normally distributed.

The video emphasizes the critical difference between correlation and causation. Correlation indicates a relationship between variables, but not necessarily that one causes the other. An example of ice cream sales and sunburns shows how a third variable (sunny weather) can cause both. Causality implies a clear cause-and-effect relationship, where one variable directly influences another. Conditions for establishing causality include a significant correlation, chronological sequence, evidence from controlled experiments, or a well-founded theory. Without these, only correlation, not causation, can be inferred. A humorous example of head lice and body temperature illustrates a common misinterpretation of correlation as causation.