Alternatives to margins command (too slow)

[Dominique Chaufree]

Hi,

I have a simple program that estimates that regresses an outcome on (a) event-time dummies (i.e. dummies for time relative to year in which a particular event happens), (b) age dummies and (c) year dummies. I then estimate margins at the unique values of the event-time variable (e.g., 1 period before event, 2 periods before event,...) that abstract from the effect of the event (i.e., I set the event-dummies to zero; put differently, I consider the effect at the value of the constant). Here is the program:

    cap program drop event_study
    program define event_study, rclass    

        *** Syntax
            syntax [varlist] [if], [                 ///
                outcome(string)                  ///
                time(string)                          ///
                baselevel(string)                ///                
                age_F(string)                     ///
                years(string)]                                    
                
        *** Regression
            reg     `outcome'                ///
                    ib`baselevel'.`time'    ///
                    i.`age_F'                    ///
                    i.`years', r
                    
        *** Margins
             margins, over(`time') at(`time' = `baselevel') nose
    end

The code works well. However, the margins command is incredibly slow. Is this because I use factor variables? Is there an alternative way to estimate the margins faster? Would you do it by hand? I am not sure how to do it in a general form (the program allows for a flexible set of covariates).

Thank you so much for your help!

[William Lisowski]

You tell us nothing about the size of your dataset - this wouldn't perhaps be the 36,553,056 observations you mentioned in your previous topic?

[Dominique Chaufree]

Sorry about that: it varies quite a bit. The lower bound is 1 million obs. The upper bound is around 10 million obs.

[Dominique Chaufree]

If someone is interested in the solution. I did the following to calculate the margins:

                    local coefs: colnames e(b)                                 
                    local coefs = regexr("`coefs'","..+\\`time'","")           
                    tokenize "`coefs'"                                       
   
                    mat Betas = e(b)                                          
                    mat b = Betas[1, "`1'".. "_cons"]                       
                    local coefs = regexr("`coefs'","_cons","")                
                    mata: b     = st_matrix("b")
                    mata: X     = st_data(. , "`coefs'")
                    mata: rows  = rows(X)
                    mata: cons  = J(rows, 1, 1)
                    mata: X     = X, cons
                    mata: y_hat = X*b'
                    cap drop y_hat
                    mata: st_addvar("float", ("y_hat"))
                    mata: st_store(., "y_hat", y_hat)
                   
                   levelsof `time', local(levels)                         
                   local size: list sizeof local(levels)                 
                   mat Fcounterf_`outcome' = J(`size', 2, .)            
                   local i = 1
                   foreach l of local levels {                                
                      if `sweight' == 1 {                                 
                                qui sum y_hat [weight=weight] if `time' == `l'       
                      }
                      else {
                               qui sum y_hat if `time' == `l'                  
                       }
                    
                        mat Fcounterf_`outcome'[`i', 1] = `l'
                        mat Fcounterf_`outcome'[`i', 2] = r(mean)
                        local i = `i'+1        
                        }        
                        
                       mat list Fcounterf_`outcome'