Non normality of data

[Fadi Ansar]

Dear All,
I am having dataset consist of T=14 and N=59 and having observation around 1167.for running regression ,I had run normality test for my variable of interest. Here my variable of interest is NPL_asset and which is having missing values (total observation 970 and out of that 196 is missing) I was getting below histogram and understood that my data is not normal(shapiro wilk test also showing the same).Since I am new learner in the econometrics can any one suggest me is it okay for me to perform a quantile regression for the data .I am doubtful about the way how my normality look like. Do I need to worry about that .your suggestion are highly appreciated .
Can any please guide me on this




Thanks in advance
Fadi ansarGraph.gph

[Carlo Lazzaro]

Fadi:
normality is a (weak) requirement for residuals distribution only.

[Fadi Ansar]

Thank you @Carlo Lazzaro.my only concern is that the way my graph look like -is it okay to have a histogram I got while checking for normality .

[Jackson Monroe]

You might consider a log transformation because your data is restricted to be positive, but really what must be normal (sometimes) are the residuals, not the dependent variable itself.

[Fadi Ansar]

Thankyou so much @ Jackson Monroe.I was getting a similar kind of graph while considering log transformation and residual plot. Can you just tell me is it okay if a normality plot look like what i got

[Fadi Ansar]

Fadi:
normality is a (weak) requirement for residuals distribution only.

Dear Carlo ,
I know that normality is a weak requirement .but my concern is that does my histogram look anything unusual or it it is okay to have an histogram like what i got.

[Carlo Lazzaro]

Fadi:
your variable seems to follow a gamma distribution (ie, positively skewed due to a long right tail), as it is often the case with such variables.
I do not think this in an issue.
Log-transformation can bean option, provided that a log-linear model makes sense.

[Nick Cox]

Log transformation of such an outcome is certainly a possibility -- so long as all data are positive, which isn't clear from any post so far..

I would tend to go further and choose a logarithmic link function in a generalized linear model (GLM). An advantage of GLMs is that you can choose a link function and that doesn't commit you to a particular error family: there can be a choice. Also, even with a logarithmic link function it is not an absolute assumption that all outcomes are positive: the main idea is just that the mean function is exp(Xb) in predictors X (so positive always)

Depending on where you were brought up, you may think of that as Poisson regression, which is much more general than just a flavour of regression that is for counted data (and even then whether any marginal or conditional distribution is, or is approximately, Poisson distributed is a secondary issue).

For Stata users https://blog.stata.com/2011/08/22/us...tell-a-friend/ is an easy way into this territory.

[Fadi Ansar]

Thankyou so much for your suggestions @ Carlo Nick Cox I just want to add that all my data is positive and from literature some are employed quantile regression by considering the non normality of variable. What will be your opinion going for a Quantile regression model.