More data modification

Earlier we created a simple binary variable using the command

generate binalc=alcohol>2

There are two problems with this method. The obvious one is that it only works for creating binary variables. Less obviously, it gives the wrong answer when there are missing observations. In Stata a missing observation behaves like a very large number and so will be recoded to 1 along with the higher levels of alcohol consumption.

There are a number of other useful ways to recode variables

• The recode statement. This modifies the variable rather than creating a new one, so you will want to create a copy first. For example

  generate agegp=age
  recode agegp 1=0 2/4=2 5/max=4
  

  After this command 1 is recoded to 0, value from 2 to 4 are recoded to 2 and values from 5 to the highest value (6 in this case) are recoded to 4. Any missing values will remain missing. You can also recode a numerical value to missing (indicated by .) or vice versa.

  Don't confuse this recode statement with the similar recode function that is also available in Stata. The one given here is the one that the Help system tells you about easily.

• The xi: command. If you put xi: before a Stata command, Stata will create dummy variables and two-way interactions. Interactions are specified by an asterisk ( *) and dummy variables are requested by preceding the variable name with i., for example:

  xi: logistic case i.age i.alcohol*i.tobacco
  

  This creates a logistic model with 5 dummy variables for age, 3 each for tobacco and alcohol and 9 for the interaction between them. The xi: command is most useful with model-fitting commands but can be used in other contexts.

• The tabulate command, which produces one-way or two-way frequency tables, has an option that creates dummy variables. This is useful if you only want some of the possible dummy variables in your model. For example

  tabulate alchohol,generate(alcgp)
  

  creates 4 dummary variables called alcgp1,alcgp2, alcgp3, and alcgp4.

• The egen command does a number of complicated things. The most useful to us is a way of creating a single variable indicating all the different covariate patterns for a set of variables. Uses for this include creating stratum indicators for conditional logistic regression and performing the Tsiatis test of goodness of fit. For example, consider creating a variable with a different value for each level of alcohol and tobacco consumption

  egen alctob = group(alcohol tobacco)
  

  The new variable alctob will have values from 1 to 16, indicating the 16 possible combinations of alcohol and tobacco consumption.