Assa Gory Multiple regression assumes that the residuals of the DV are normally distributed, not the DV data itself. If the residuals are normally distributed, run the regression!

If your residuals are nonnormal, you'll probably want to use a Generalized Linear Model instead.

You can use ranks instead of the actual observations. But there are alternative transformations: https://xdplot.blogspot.com/p/data-transformations-guidelines.html

Blaine Tomkins Andrey Davydenko Thank you!! Will definitely look into these suggestions.

For continuous outcomes, the most commonly used statistical method is multiple regression analysis. The least squares procedure yields minimum-variance unbiased estimators of the regression coefficients when the variance of the error term is assumed to be the same regardless of the values of covariates (homoscedasticity). Since these assumptions are rarely fulfilled in practice, certain adjustments should be made to improve the precision and validity of results. Such adjustments include retransforming outcome variables using methods which include multiple imputation and attrition weights, smear factor retransformation, and improving statistical inference with bootstrap methods.

If the outcome variable has some missing observations/values, multiple imputation can be undertaken to account for missing case. The multiple imputation process involves two steps: (1) Simulation of the posterior distribution of the regression coefficients; and (2) using the simulated parameter values to estimate each missing value from its regression on the observed variables for the same subject, and add to this estimate a random error selected from the distribution of residuals obtained in the data set. The process can be repeated several times, resulting in a number of datasets, each with a different set of imputed values.

For positive right-skewed data, a natural log transformation often results in a more symmetrical distribution. As such, taking the log makes the distribution closer to being symmetric, thus better approximating a normal distribution. Therefore, retransforming the outcome variable using the Smear Factor can be done to compute a quantity called the smear factor, which is based on the distribution of the residuals. In addition, you can consider conducting bootstrapping - a normal approximation method which uses the percentiles of the normal distribution to compute the confidence limits based on the estimated mean and variance of the sampling distribution.