How to Compute Correlations

From Q
Jump to navigation Jump to search

Q automatically computes correlations between Numeric Questions. If you wish to compute correlations involving categorical data, this is done as follows:

  1. Ensure that the two variables have a Question Type that is either Number or Number - Multi. This is done on the Variables and Questions tab, by changing the Question Type drop-down menu (one of the right-most columns). Often it is a good idea to first take an Exact Copy, so that any tables that have already been created do not change when the Question Type is changed.
  2. Check that the values are appropriate. This is done by pressing the ValuesButton.png on the Variables and Questions tab next to the relevant variables, ensuring that appropriate numbers appear in the Value column, and pressing OK. For example:
    • With a scale measuring 'once a week or more often', '4 to 6 times a week', down to 'less often than once a year', a value of 10 could be assigned for 'once a week or more often' (i.e., as perhaps once a week or more corresponds to about 10 times a week), a value of 5 for '4 to 6 times a week', down to 0 for 'less often than once a year'.
    • A scale measuring 'trust a great deal', 'reasonably trustworthy', 'trust a bit' to 'don't trust at all' could be given values of 4, 3, 2 and 1 respectively (i.e., where higher numbers correspond to higher levels of trust).
  3. On the Outputs Tab, select one question in the Blue drop-down menu and the other question in the Brown drop-down menu.
  4. Optionally, change the Correlations - Comparing Two Numeric Variables setting to Spearman or Kendall tau-b. By default Q uses Pearson's correlation, which is the main method of correlation in use in survey analysis. However, at a technical level Pearson's correlation assumes that the underlying data is numeric, whereas the other two methods instead make the milder assumption that the values simply reflect relative ordering. Stated differently: if the values that are shown in the Value column represent arbitrary assumptions that you have made, using Kendall tau-b or perhaps Spearman is, technically, more appropriate (however, it is rare that this makes much of a difference with survey data).

Options for computing and displaying different kinds of correlations are also available in Create > Correlation.

See also