What *is* your dependent variable? This is much more important to know when one needs to find an appropriate regression model.

Generally, the simple regression has no restriction on the dependent variable. Only the residuals should be roughly normal distributed.

Thanks Prof. Jochen for answering my question. My  research model dependent variable is the correlation between capital markets and independent variables are macroeconomic indicators . I am wondering unrestricted specification might affect my results.   

There is no problem with negative values for a dependent variable in Multiple Regression. It would be misleading to change your negative and positive values to absolute values.

You explain that your DV is a correlation coefficient. That is, values potentially range from -1 to +1. In linear models such as regression, variables are assumed to be able to continue beyond those limits. You should consider transforming your correlations into Z-scores before applying regression. This is called a Fisher Transformation.

Thanks Prof Hume for answer. Is it appropriate to do fisher transformation when variables violate normality assumption? My bivariates of correlation are not normally distributed.   

Yes it is quite okay to use a Fisher transformation on correlations where your sample of correlations is not normally distributed.

It is also quite okay to run regression with data that are not normally distributed - your results are still BLUE (Best Linear Unbiased Estimates). The standard errors of your coefficients may be biased in such situations, so be careful with your interpretation of significance levels in marginally significant regressors.

It is not necessary to use absolute values of your DV. I believe you can go ahead with your regression analysis.

It does not make a difference. You can change negative values by adding the same constant to the variable keeping in mind that adding or subtracting a constant affeects the mean but does not affect variation (SD and Variance).

It does not make a difference. You can change negative values by adding the same constant to the variable keeping in mind that adding or subtracting a constant affeects the mean but does not affect variation (SD and Variance).

It does not make a difference. You can change negative values by adding the same constant to the variable keeping in mind that adding or subtracting a constant affects the mean but does not affect variation (SD and Variance).

Hey Muhammad,

I do not understand what you are trying to model, exactly.  What are you trying to model in the capital markets. What is your endogenous variable exactly?

Hi,

I think Abdulrazak Charbaji is correct when you add any constant to your variable, the correlation coefficient as well as regression coefficients remain unchanged. If you are using t-test for the significance of regression coefficient then normality is required otherwise for z-test normality is not required.

Adding an arbitrary constant to your independent variables indeed would not introduce any difficulty. You should not take the absolute value of the variable unless the absolute value of the DV is what you are trying to predict/understand. In either case, neither of these 'solutions' are necessary or even helpful. 

Hume suggests that 'It is also quite okay to run regression with data that are not normally distributed.' Indeed, this is true. But in fact we would generally not expect the data to be normally distributed. We may, however, expect the residuals to be normally distributed. Although, as Hume suggests, for a linear model, even if the residuals are not normally distributed, the parameter estimates from ordinary least squares regression are still the best linear unbiased estimator (BLUE). Best is a little bit subjective, what we really mean to say is minimum variance unbiased estimator (MVUE). In this case, we still require that the residuals are heteroskedastic for the estimator to remain BLUE. However, they are no longer the maximum likelihood estimator and the standard errors calculated in the usual manner are no longer meaningful.

For further information of basic regression methods, I would recommend Frank Harrel's book, 'Regression Modelling Strategies.'

Regressions run fine with negative values. There is no need to add a constant.

 I agree Richard simple opinion. Today, I finish to write my paper.

I may be free after it. If you offer the sample data by Excel, I can analyse your .

Variable names are first row. After  second row, data are imputed.

I have a similar question, with running OLS regression when the dependent variables has negative postive and zeros.

How does one interpret the results, if the dependent variables has -1,0,1 and -4 through 4.

Lets says B1=gender (0=male 1=female) and the dependent variables is -1, 0. or 1, who would you interpret B1= -.002??