I have a dataset that looks at the default and non-default of companies. Thus it contains one variable 'company default' that is 1 or 0 and some extra descriptive variables. However my dataset is highly unbalanced. 97 percent of the samples are non-default, i.e. 'company default' equal to 0. Only 3 percent of my dataset has 'company default' equal to 1. Now I am not intereset in a predictive study, but I am interested in finding which variables have what impact on my independent variable default. I will perform a logistic regression on the dataset and use the coefficients of the variables that are produced to perform my expanatory analysis.

I know that if I would do a predictive analysis I should make sure my dataset is balanced by, for example oversampling my minority class or undersampling my majority class. However I cannot find any literature online that says how to adress this when I am only interested in expalantory analysis. Can I just use my unbalanced dataset to perform logistic regression on the data and interpret the coefficients that result from the model for the expanatory analysis?

Anyone more familiar in this area or with econometrics to approach this?

  • 1
    $\begingroup$ Welcome to Cross Validated! Misconceptions abound when it comes to class imbalance. For the most part, class imbalance presents no problems for proper statistical methods. $\endgroup$
    – Dave
    Aug 4, 2022 at 19:32
  • $\begingroup$ So can I just perform logistic regression with this unbalanced data and interpret the coefficients in an explanatory way? I mean, will these coefficients be reliable to explain some variables in my data? I am not looking for prediction, I am interested in knowing how my independent variables impact the dependent variable. Thank you in advance! $\endgroup$
    – cmrn
    Aug 4, 2022 at 20:08
  • $\begingroup$ What would you mean about knowing how the independent variables impact the dependent variable if you have no ability to predict? $\endgroup$
    – Dave
    Aug 4, 2022 at 20:13
  • $\begingroup$ I don't want to predict with the data. I just want to see what variables have an impact on the probability of default. So just by using the data I have available, I want to examine what independent variables influence the probability of default in the dataset. Is logistic regression an appropriate method for this? Or is it better to use another method. Because I thought, based on the coefficients of the logistic regression, I can derive their impact on the probability of default (by calculiting their increase in odds if a one unit increase in the independent variable). Is this a valid method? $\endgroup$
    – cmrn
    Aug 5, 2022 at 8:02