The LaLonde Problem

The National Supported Work program randomly assigned applicants to a job-training treatment. We know the true effect on 1978 earnings from the experiment. The question is whether regression could have found it without the experiment, using a comparison group drawn from the Current Population Survey instead.

Work through these

Your answers are saved in this browser only. Copy them out if you need to hand them in.

    Which control group?
    Which predictors go into lm()?

    185 NSW treated units are in every regression. Outcome is re78, real earnings in 1978 (dollars).

    Estimated effect of treat—
    95% CI width—
    Experimental benchmark (difference in means)—

    Treatment effect estimate and 95% confidence interval

    What R ran

    
  
    
  
    

    


    
  
    Specifications you have tried
    
  
    
  
    
    
    #Control groupPredictors (besides treat)Estimate95% CIn
    
    
  
    
  
    
  
    Change the control group or a predictor to start the log.
    
  
    

    


    
  
    
    What this page assumes and leaves out
    
      
      
    • Data are the Dehejia–Wahba (1999) subset of LaLonde's male sample: 185 treated and 260 experimental controls with 1974 earnings observed (LaLonde's original had 297 and 425), plus the CPS-1 comparison group of 15,992 men. Earnings are in 1978 dollars.
      • 
      
    • Every model is ordinary least squares on re78 with treat and the chosen covariates, estimated in R 4.3.3 with lm(); the output shown is R's own, computed in advance for every one of the 512 predictor subsets for each control group. Standard errors are the conventional homoskedastic ones summary() reports.
      • 
      
    • The benchmark is the raw difference in mean re78 between experimental treated and control units, with a Welch t-interval. It is what the experiment delivers without any modelling.
      • 
      
    • Left out on purpose: propensity-score matching, trimming, subclassification and the PSID comparison groups. Those are the next chapter; this page is only about what regression does and does not do.
      • 
    
    
  
    

    


    Sources: LaLonde, R. (1986), "Evaluating the Econometric Evaluations of Training Programs with Experimental Data," American Economic Review 76(4). Dehejia, R. and Wahba, S. (1999), "Causal Effects in Nonexperimental Studies: Reevaluating the Evaluation of Training Programs," JASA 94(448). Data files as distributed by Dehejia.