Data manipulation details

• How to move between one-case-per-person and one case per person-wave panel data:
• Stata's reshape command is very powerful
  1. Let's say we have data in one-case-per-person format:

     PID jbstat1 jbsemp1 jbstat2 jbsemp2 jbstat3 jbsemp3
       1       2       3       4       5       6       7
     

     and we would like to have

     PID wave jbstat jbsemp
       1    1      2      3
       1    2      4      5
       1    3      6      7
     

  2. In Stata:

     . reshape long jbstat jbsemp, i(pid) j(wave)
     

     will move from the first to the second from almost instantaneously, while

     . reshape wide jbstat jbsemp, i(pid) j(wave)
     

     will do the reverse.

• In SPSS (pre version 11) it is somewhat slower (given variables in the right order!)

  vector jb=jbstat1 to jbstat3
        /js=jbsemp1 to jbsemp3.
  loop #i = 1 to 3.
  . compute jbstat=jb(#i).
  . compute jbsemp=js(#i).
  . compute wave = #i.
  . xsave out = "longform.sav"
     /keep=pid, wave, jbstat, jbsemp.
  end loop.
  

• From SPSS 11 on:

  varstocases 
    /id=pid /index=wave
    /make jbstat from jbstat1 to jbstat3
    /make jbsemp from jbsemp1 to jbsemp3.
  

• The reverse is more tedious (pre SPSS 11):

  get file "longform.sav".
  vector jbstat(3) /jbsemp(3).
  compute jbstat(wave)=jbstat.
  compute jbsemp(wave)=jbsemp.
  agg out = */break = pid
    /jbstat1 to jbsemp3 = first(jbstat1 to jbsemp3).
  

• But easy in SPSS 11:

  casestovars /id=pid /index=wave.
  

• This is a transformation we often need to make: both data structures suit different sorts of analysis.

• Another important transformation is turning spell-level data into state calendars.
• This is easy enough: let's assume we have PID, state, start and end (both dates in integer format, e.g., months since 1900)

  vector month(1200).
  loop #i=start to end.
  . compute month(#i)=state.
  end loop.
  

• The vector has to be at least as long as the range of dates, thus potentially very big.
• The researcher must generate integer dates from the year/month variables.
• The reverse transformation (calendar to spell) is difficult.

• For many purposes the spell format is superior
• For instance, duration is end minus start (plus 1, depending on how dates are recorded)
• We can easily cumulate duration in different states using aggregate (assuming 5 states):

  vector dur(5).
  compute dur(state) = duration.
  agg out = * /break = pid
    /dur1 to dur5 = sum(dur1 to dur5).
  

• The vector alternative is easy but very slow.