Factor analysis using stata "predict" command and get negative value for non-negative variable? I am running factor analysis on stata to reduce a few variables to a single explanatory variable which means "experience" of a manager (should be non-negative value), however, after using "predict" command I check the range of the new variable and found that there are many negative values, how do I avoid this?
The original variables are 1) number of projects done by a manager 2) number of people interacted for each project 3) number of unfinished projects. Since these 3 variables are highly correlated so I want to use factor analysis to reduce them to one single variable.
If there is any other method that can also tackle this problem, please let me know! Thank you guys!!
 A: This is expecting much more of factor analysis than it will give, at least by default. My own view is that this is somewhere between non-standard and downright weird as an application of factor analysis, but there is considerable variation among statistically-minded people on the merits of factor analysis and how it might be well used, so conflicting advice is highly likely. 
Factor analysis works on either the correlation or the covariance matrix. Necessarily information on means is discarded and new factors have zero means, so negative values are inevitable. You have missed that in your reading! I suppose that using zero means could be regarded as a convention, not an absolute requirement, but it seems to be a universal convention. 
You could, I suppose, rescale a factor using what you think are an appropriate mean and standard deviation, but there is no real gain in that, especially as you want a predictor, not a proxy for your response variable. That rescaling  wouldn't guarantee all values being positive. What would guarantee all positive values would be to factor analyse the logarithms of the variables, predict a factor and then exponentiate, but whether that would work well depends on many things. There would be still a units of measurements issue: in effect units are discarded and you would have to work to put them back. I would expect all these variables you name to be skewed, so transformation might be a good idea on other grounds; however, if there are zeros then logarithmic transformation is not possible, unless you add a constant to make all values positive, in turn in my view not a good idea, but views differ on that. 
All that said, I think your whole strategy would strike most statistical people as dubious at best, both statistically and scientifically. You have three predictors. It is easiest and probably most illuminating to keep them separate and to use them all in some regression-type model, possibly a generalised linear model with logarithmic link given that your unnamed (!) response variable is also likely to be non-negative. The correlation between predictors is clearly something you have to think about, but throwing away the detail on those predictors is more likely to make your model difficult to interpret and discuss. 
I see no Stata-specific issues here. 
A: The three variables you want to combine sound to me like resources that result in a manager having more experience rather than characteristics of managers that are influenced by his/her experience. The difference between the two is that I think that it is the number of projects done by a manager, number of people interacted for each project, and number of unfinished projects should influence experience rather than experience influencing those three observed variables. If you also think that this is the way that the three observed variables relate to the latent experience variable then you have a problem in that factor analysis assumes that the latent variable influences the observed variables rather than the other way around. (Bollen 1984; Bollen and Lennox 1991) In that case you will probably want to look at something like a sheaf coefficient (Heise 1972) rather than factor analysis. In Stata this is implemented in the sheafcoef package.
Bollen, Kenneth A. 1984. "Multiple Indicators: Internal Consistency or No Necessary Relationship" Quality and Quantity 18(4): 377-385.
Bollen, Kenneth A. and Richard Lennox. 1991. "Conventional Wisdom on Measurement: A Structural Equation Perspective" Psychological Bulletin 110(2): 305-314.
Heise, David R. 1972. "Employing nominal variables, induced variables, and block variables in path analysis." Sociological Methods & Research 1(2): 147-173.
A: FA is for unconstrained variables.  First transform the data to make it this way.  Use any inverse CDF (I use the quantile function for the normal distribution, qnorm in R) for variables that start out between 0 and 1; use any log function for positive variables (if zero can occur, add 1 first.)  Then do factor analysis.  The results will be for the transformed variables, use their inverses to recover the original ones, if needed.
