Indicator variable vs Dummy variables and output change

[Peregrine Dalziel]

I am using Stata 12.

I am doing a regression and have been using Dummy variables in order to use eststo: prefix.

Simply out of idle curiosity I did the regression instead using the i. prefx for the variable used to generate the Dummy (because I think the output looks nicer.

But it generates different levels of significance for the cateogries that if I used the Dummy variables.

When using
reg logmBeginToProv i.FCReg Period if DayShift ==1 & Outlier==0

Output

logmBeginT~v	Coef.	Std. Err.	t	P>t	[95% Conf.	Interval]
FCReg
3	0.239057	0.065421	3.65	0	0.110774	0.36734
4	0.222279	0.063113	3.52	0	0.098521	0.346037
5	0.067246	0.060392	1.11	0.266	-0.05118	0.185668
6	-0.15131	0.079708	-1.9	0.058	-0.30761	0.004992
7	0.444792	0.07781	5.72	0	0.292215	0.597369
8	0.769023	0.061541	12.5	0	0.648348	0.889698
15	0.84092	0.051907	16.2	0	0.739137	0.942703
17	0.134585	0.122669	1.1	0.273	-0.10596	0.375126
19	0.00783	0.055735	0.14	0.888	-0.10146	0.117121
21	0.395971	0.083419	4.75	0	0.232395	0.559547
25	-0.11284	0.061729	-1.83	0.068	-0.23388	0.008207
27	0.283143	0.054459	5.2	0	0.176354	0.389931
28	0.298593	0.06775	4.41	0	0.165742	0.431443
29	0.511729	0.055338	9.25	0	0.403217	0.62024
30	0.575785	0.065874	8.74	0	0.446613	0.704958
Period	-0.32571	0.027863	-11.69	0	-0.38034	-0.27107
_cons	4.563697	0.035962	126.91	0	4.49318	4.634214

Where as

. reg logmBeginToProv FCReg_Dummy* Period if DayShift ==1 & Outlier==0

Yields

logmBeginTo~v	Coef.	Std. Err.	t	P>t	[95% Conf.	Interval]
FCReg_Dummy1	-0.39597	0.083419	-4.75	0	-0.55955	-0.2324
FCReg_Dummy2	-0.15691	0.088609	-1.77	0.077	-0.33067	0.01684
FCReg_Dummy3	-0.17369	0.088775	-1.96	0.051	-0.34777	0.000387
FCReg_Dummy4	-0.32872	0.089405	-3.68	0	-0.50404	-0.15341
FCReg_Dummy5	-0.54728	0.101946	-5.37	0	-0.74718	-0.34737
FCReg_Dummy6	0.048821	0.103077	0.47	0.636	-0.1533	0.250943
FCReg_Dummy7	0.373052	0.087218	4.28	0	0.202026	0.544077
FCReg_Dummy8	0.444949	0.080113	5.55	0	0.287856	0.602041
FCReg_Dummy9	-0.26139	0.140065	-1.87	0.062	-0.53604	0.013266
FCReg_Dummy10	-0.38814	0.08357	-4.64	0	-0.55201	-0.22427
FCReg_Dummy11	0	(omitted)
FCReg_Dummy12	-0.50881	0.089372	-5.69	0	-0.68406	-0.33356
FCReg_Dummy13	-0.11283	0.081903	-1.38	0.168	-0.27343	0.047774
FCReg_Dummy14	-0.09738	0.091479	-1.06	0.287	-0.27676	0.082002
FCReg_Dummy15	0.115758	0.08035	1.44	0.15	-0.0418	0.273314
FCReg_Dummy16	0.179814	0.091002	1.98	0.048	0.00137	0.358259
Period	-0.32571	0.027863	-11.69	0	-0.38034	-0.27107
_cons	4.959668	0.076574	64.77	0	4.809514	5.109822

With different p values for the same categories (I have cross checked)

The R2 etc is the same and other independent variable values are not affected.

Is this normal behavior? I thought these different approaches would yield the same result? Is it to do with the ommited variable in each?

It doesn't really matter as I am just controlling for these to isolate another variable effect. But I thought it was curious.

Thanks,

Peregrine

[Phil Bromiley]

All of your dummy parameters are essentially changes versus the omitted category. You've switched omitted categories so all of the parameters change. This is normal. You can control the omitted category in factor variable notation if you want to - look at the documentation.

[Peregrine Dalziel]

Thanks Phil!

Yeah. I figured it was something elementary like that. Figuring out which one was dropped in the factor variable analysis and using the ib# prefix sorted the problem out.


Watch this space for questions about other very basic issues.

Peregrine