Statistical Inference in Logistic Regression

This post discusses statistical inference in logistic regression, the assumptions related to logical regression compared to simple regression, and the types of variables used in each.

Logistic regression does require many of the assumptions that regular regression requires (Anonymous, n.d.).  Logistic regression does not assume there is a linear relationship between independent and dependent variables.  It does not assume the independent variables are multivariate normal.  The residuals (regression error terms) need not be multivariate normally distributed.

Logistic regression does have several assumptions that must be met (Schumacker, 2014).  The model should be fitted correctly.  To avoid overfitting, only meaningful variables should be included.  The variables being utilized should match the type of regression being performed.  For example, if binary logistic regression is being performed the variables should be binary.  Coercing a variable can result in loss of specificity.  Observations should be independent; avoiding relationships such as before-after.  Logistic regression requires larger sample sizes than simple regression.  In standard regression, the response variable is a continuous variable while in logistic regression it is binary (Ledolter, 2013). 

References

Anonymous. (n.d.). Assumptions of logistic regression.   Retrieved from http://www.statisticssolutions.com/assumptions-of-logistic-regression/

Ledolter, J. (2013). Data mining and business analytics with R: John Wiley & Sons.

Schumacker, R. E. (2014). Learning Statistics Using R: SAGE Publications, Inc.