centiles for income data, skewed

[CatM]

Hello, 

I am trying to do a multilinear regression of income data as the outcome with disability as the main predictor and controlling for confounding of age education and gender. the income data is from two different years where i am trying to see if there have been any improvements or socioeconomic gains for people with disabilities. In order to look at this from a more relative angle, it was suggested that i convert the income data into centiles (100 by 1)  and use linear regression. However, i am running into issues because the income data includes negative and 0 income values which we do not wish to take out but analyze all together. These are being grouped into one centile which makes the data skewed when running the regression analysis. Is there any ideas on how this data should be analyzed instead. there is a weight for the data however this is not helping meet the assumption of normality. 

proc rank data=incdis groups=100 out=cent_incdis ;
var atinc42;
ranks atinc42_centile;
by year;
run;

proc surveyreg data=cent_incdis;
class ecsex99_n (ref='0') education (ref='0');
model atinc42_centile = disabs26_n ecsex99_n education ecage26 / solution;
weight icswt26;
run;

[Reeza]

The assumption is that residuals are normally distributed, not necessarily the variables. 

I think centiles is great idea, not sure how it can be skewed when by definition, 1% is in each group. Do you mean the residuals become skewed?

[Reeza · Accepted solution]

Doh! You have a lot of tied data...look at the options for handling tied data, it’s under the PROC RANK statement documentation I believe.

[CatM]

Ok great thank you! i just checked the residuals and they look fine. i'll look at tied data documents

[Reeza]

I’m not sure that I’d be that concerned anyways, the idea behind centiles is to see if the groups improvement is occurring at same rate as general population, so if that group has A larger number that’s reflective of reality. When this happens though youre more likely to see something like Simpsons Paradox occur due to an imbalance so be careful if you do subgroup analysis.