I want to run a multiple regression analysis for a given dataset in SPSS.

However, the dataset violates the assumption of normality of residuals, as depicted in the picture. The values for the dependent variable are given in percent. To transform the data and get a normal distribution for the residuals, I have applied various transformation recommended by IBM:


Is there anything else I could do? Thanks in advance!

Histogram:Normality of residuals

  • 3
    $\begingroup$ If your dependent variable is in units of percent, a standard linear multiple regression might not be appropriate. Please say more about the details of your data (for example, what is the dependent variable a percentage of, whether there are any dependent-variable values exactly 0 or 100 percent, how many cases and how many predictors you have) to get a more helpful answer. $\endgroup$ – EdM May 11 '19 at 14:08
  • $\begingroup$ I have a dataset containing information regarding sponsors of events. The DV indicates how much percent of sponsors for an event stems from a specific industry (I want to see if a high creative reputation of an event attracts more sponsors that stem from a creative industry). E.g. 0.35 of all sponsors work in a creative industry. I have about 115 observations, one independent variable, one moderator variable and eight control variables, which are all dummy variables. Thank your very much! $\endgroup$ – nisch May 12 '19 at 8:43
  • $\begingroup$ Also, there are no cases with exactly 100 percent or 0 percent. $\endgroup$ – nisch May 12 '19 at 9:04

You are to be commended for examining whether the assumptions underlying your initial linear regression model were satisfied. There are better ways to model data in which the outcome variable is a fraction between 0 an 1.

Two standard ways to model such outcomes with regression are beta regression and logistic regression. This answer describes the differences between them and provides links that show how to implement them. Beta regression is most appropriate when the fraction is continuous. In your case, where the outcome variable is discrete (the ratio of the number of sponsors that are in a creative industry to the total number of sponsors, for each sponsored event), logistic regression is probably a better choice.

Although logistic regression is typically used with a binary outcome variable, this answer shows how to proceed with situations like yours where for each case you have counts of "successes" (creative-industry sponsors) and "failures" (other sponsors) for each sponsored event. Logistic regression will also take into account the number of sponsors in each case; one would expect an event with more sponsors to provide more information about the relationship between creative-industry sponsorship and the nature of the event (as modeled in your other predictor variables).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.