When to Use Wide Format versus Long Format

[Jeff Thompson]

I'm working on longitudinal data involving the variables "employment" and "payroll" for each year of a company's life. Currently it's in wide format with rows being each companies' name and many columns of emp(years) and pay(years). The data unbalanced because companies appear and disappear over the time span of the data set. I'm looking to observe that as the number of employees increases or decreases how the payroll increases or decreases for each company. It's important that the employment change and payroll change are compared on a company to company basis. I want to show that employment increases at a different rate than payroll for each company- not just show that higher employment means higher payroll. I hope that makes sense.

Now, will the wide format work in this case? Everywhere I've looked insists that longitudinal data needs to be in long form, and I anticipate you saying the same. But will this allow me to compare two correlated variables as they change overtime for each firm? Can I make firm specific dummy variables?

Also, something that is a little harder to explain is that I'd like to account for both the year (from economic activity) and the age of the company (older and younger). Could this also be done in the long format?

Any and all advice are welcome!

[Clyde Schechter]

Well, you don't spell out the specific commands you plan to use for your analyses, but frankly I cannot envision an appropriate analysis for your goals that would be workable in wide layout. In long layout, it should be fairly straightforward.

If you want to calculate company age, this is simplicity itself in long layout:

by company (year), sort: gen age = year - year[1]

assuming that their first year in the data set is the year in which the company was "born." (Also assumes that observations with missing data are -drop-ped after you -reshape- your data to long layout.

Once you have your data in long layout, you will be able to -xtset- it, and then you can handle lags, leads, and differences with Stata's time series operators. Run -help tsvarlist- to see how those work.