No

Eshetu Andarge, No. Instead, you should find the appropriate methods/procedures that can remedy the encountered issues.

Kelvyn Jones , thank you for your reaction. However, why no? Can you give me some more explanations and a reference to read?

By dichotomizing a continuous dependent variable you are getting rid of valuable information and losing statistical power.

Can you tell us about the nature of the dependent variable, how was it measured , and its distribution? Remember that the normality assumption concerns the residuals conditional on the predictor variables. Unless there are outliers, it is not a particularly important assumption in comparison to non-independence and heteroscedasticity. Moreover, the generalized model is designed to deal with different distributions. Thus Gamma regression, is useful for highly positively skewed data while Inverse-Gaussian regression is useful when the response is strictly positive and skewed to the right.

https://bookdown.org/ndphillips/YaRrr/regression-on-non-normal-data-with-glm.html

Thanks Kelvyn Jones, the variable under consideration is physical activity among school adolescents. The questionnaire used is the physical activity questionaire for adolescents(PAQ_A). The developers recommended mean score from 8 items in a likert scale. The finding among adolescents in my setting inclined consistently to the lower scores with a significant number of outliers breaking the assumptions of linear regression. What if I categorize it as " higher physical activity" for those who scored above the mean and " lower..." for the reverse.

Ordinal regress is designed for Likert like responses

see

Ordinal responses: multiple categories ordered at

http://www.bristol.ac.uk/cmm/software/mlwin/mlwin-resources.html#discrete