Diff-in-Diff using Proc ttest (or anything)?

[braam]

Hi All,

I would like to implement a univariate diff-in-diff test presumably using proc ttest. Any procedure would be fine with me as long as I can do so.

The below dataset includes four groups by two class variables: TYPE (SUV and Sedan) and ORIGIN (Asia and Europe).

Let me put it this way arithmetically:

Mean of mpg for Asia and Sedan: a

Mean of mpg for Asia and SUV: b

Mean of mpg for Europe and Sedan: c

Mean of mpg for Europe and SUV: d

The t-stat I want for is (a-b) - (c-d).

This way, I can see whether the mean difference in mpg between SUV and Sedan is different between Asia and Europe (i.e., diff-in-diff).

Ideally, I would like to implement this using the dataset generated in the below SAS code (i.e., the one with three variables, mpg_city, origin, type). Hope that somebody can help me.

data temp;
	set sashelp.cars;
	keep mpg_city origin type; 
	where origin in ("Asia", "Europe") and Type in ("Sedan" "SUV");
	run;

proc sort data= temp;
	by origin type mpg_city; run;

proc print data= temp; run; *172;
proc tabulate data= temp;
	class origin type;
	var mpg_city;
	table origin, type*mpg_city*( n mean)  ; run;

[FreelanceReinh · Accepted solution]

Hi @braam,

According to Usage Note 61830: Estimating the difference in differences of means you can use PROC GENMOD:

proc genmod data=temp;
class origin(ref='Europe') type(ref='SUV');
model mpg_city = origin type origin*type;
run;

Result (excerpt) with point estimate, standard error, confidence interval and significance test of the requested difference in differences:

                               Analysis Of Maximum Likelihood Parameter Estimates

                                                    Standard     Wald 95% Confidence          Wald
Parameter                         DF    Estimate       Error           Limits           Chi-Square    Pr > ChiSq

Origin*Type    Asia      Sedan     1      0.5076      1.6221     -2.6716      3.6868          0.10        0.7543

[braam]

Thanks! it looks like running a regression without control variables.
Let me have a look at it and give it a try. Just one question. Would this method be applied to both continuous/dummy variables?

[FreelanceReinh]

I think in principle this method could also be applied to continuous independent variables (or one continuous and one discrete variable), but the interpretation would be different: You would consider the change in the dependent variable per unit increase of the continuous variable rather than compare means of subgroups.

[braam]

I was asking about the dependent variable, not independent variables. Sorry that I brought up some confusion.

[FreelanceReinh]

Ah, I see. Yes, the method can also be applied to models with a discrete dependent variable such as logistic regression models: see section "Generalized linear models with non-identity link" in Usage Note 61830.