Function like SAS: Proc syslin

[Lucas Long]

Hello everyone,

I would like to ask whether Stata has the function like SAS syslin command,

Here is what I trying to do. my model: y= x1+x2+x1*x2, and I would like to test whether x1+x1*x2 >0.

In Sas I could run as below:

Proc syslin data =one;
Model y = x1 x2 x1*x2;
test x1 + x1*x2 ne 0;
run;

Is there anyway we can do in Stata?

Thanks in advance

[Leonardo Guizzetti]

I'm not familiar with RPOC SYSLIN, but this looks to be estimating a linear model (y=x1 + x2 + x1*x2) and then performing a hypothesis test about a linear combination of coefficients for x1 and x2. If this is what you want, then something like the following will work.

Since I didn't see a CLASS statement in your SAS code, I assume that x1 and x2 are both continuous. If they are are not, then this code needs to be changed slightly.

reg y c.x1##c.x2
lincom _b[x1] + _b[x1#x2]

This will test the linear combination of x1 + x1*x2 against the null hypothesis of zero.

You may find it helpful to read the help for -help lincom-, -help test- and -help factor variables-.

[Lucas Long]

I'm not familiar with RPOC SYSLIN, but this looks to be estimating a linear model (y=x1 + x2 + x1*x2) and then performing a hypothesis test about a linear combination of coefficients for x1 and x2. If this is what you want, then something like the following will work.

Since I didn't see a CLASS statement in your SAS code, I assume that x1 and x2 are both continuous. If they are are not, then this code needs to be changed slightly.

reg y c.x1##c.x2
lincom _b[x1] + _b[x1#x2]

This will test the linear combination of x1 + x1*x2 against the null hypothesis of zero.

You may find it helpful to read the help for -help lincom-, -help test- and -help factor variables-.

Hi Leonardo,

Thanks for your reply.

I want to test if the coefficients x1 + x1*x2 > 0. and here both x1 and x2 are dummy variables. in that case, should I change the code?

[Leonardo Guizzetti]

In that case, you could show us a data example. If x1 and x2 only have two levels, and are coded as 0/1, then this would work:

reg y i.x1##i.x2
lincom _b[1.x1] + _b[1.x1#1.x2]

Lincom and -test- both give two-sided Wald-type tests, so you can modify the output in the usual way if you really do want a one-sided test. (Note that testing != 0 implies a two-sided test in your initial post.) For example, see this FAQ.

[Lucas Long]

In that case, you could show us a data example. If x1 and x2 only have two levels, and are coded as 0/1, then this would work:

reg y i.x1##i.x2
lincom _b[1.x1] + _b[1.x1#1.x2]

Lincom and -test- both give two-sided Wald-type tests, so you can modify the output in the usual way if you really do want a one-sided test. (Note that testing != 0 implies a two-sided test in your initial post.) For example, see this FAQ.

Thank you Leonardo, I will check the FAQ, and thanks for your code and time
.

[Lucas Long]

In that case, you could show us a data example. If x1 and x2 only have two levels, and are coded as 0/1, then this would work:

reg y i.x1##i.x2
lincom _b[1.x1] + _b[1.x1#1.x2]

Lincom and -test- both give two-sided Wald-type tests, so you can modify the output in the usual way if you really do want a one-sided test. (Note that testing != 0 implies a two-sided test in your initial post.) For example, see this FAQ.

Hi Leonardo,

I try to use lincom and test, is there anyway that I could directly do a one-side test? the code I used showed as below:

reg y x1 x2 x1*x2
test _b[1.x1]+_b[1.x1#1.x2]=0
local signbeta=sign(_b[1.x1]+_b[1.x1#1.x2])
display "Ho: coef >= 0  p-value = " 1-ttail(r(df_r),`signbeta'*sqrt(r(F)))
display "Ho: coef <= 0  p-value = " ttail(r(df_r),`signbeta'*sqrt(r(F)))

Somehow I am not sure where went wrong the the p value I got for coef>=0 and coef<=0 are exactly the same. Mind if you could tell me where I did wrong?

[Leonardo Guizzetti]

Sorry, it seems the output of -test- and -lincom- are little different, but still usable with some tweaking. -test- outputs an F-test with 1 numerator degree of freedom and k denomintor degrees of freedom, whereas -lincom- gives a t-test with k degrees of freedom.

As an example, let's start with a dataset that you can follow along with on your own copy of Stata. I'm setting up an interaction between two indicator variables, each coded as 0 or 1.

sysuse auto
recode rep78 (1/3=0) (4/5=1) , gen(repair)

reg price i.foreign##i.repair

To use -test- as in the FAQ, you can do this:

[/code]
. test _b[1.foreign] + _b[1.foreign#1.repair] = 0

( 1) 1.foreign + 1.foreign#1.repair = 0

F( 1, 65) = 0.38
Prob > F = 0.5373

. local sign_combo = sign(_b[1.foreign] + _b[1.foreign#1.repair])
. display "Ho: coef >= 0 p-value = " 1 - ttail(r(df_r),`sign_combo'*sqrt(r(F)))

Ho: coef >= 0 p-value = .73134831
[/code]

To use -lincom-, some slight adjustments are needed to account for the lincom. We use the t-statistic instead of the F-statistic, and we can dispense with the sign, since that is included with the t-statistic.

. lincom _b[1.foreign] + _b[1.foreign#1.repair]

 ( 1)  1.foreign + 1.foreign#1.repair = 0

------------------------------------------------------------------------------
       price |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         (1) |   700.4192   1129.369     0.62   0.537    -1555.086    2955.924
------------------------------------------------------------------------------

. display "Ho: coef >=  p-value = " 1 - ttail(r(df), r(t))
Ho: coef >=  p-value = .73134831

[Lucas Long]

Sorry, it seems the output of -test- and -lincom- are little different, but still usable with some tweaking. -test- outputs an F-test with 1 numerator degree of freedom and k denomintor degrees of freedom, whereas -lincom- gives a t-test with k degrees of freedom.

As an example, let's start with a dataset that you can follow along with on your own copy of Stata. I'm setting up an interaction between two indicator variables, each coded as 0 or 1.

sysuse auto
recode rep78 (1/3=0) (4/5=1) , gen(repair)

reg price i.foreign##i.repair

To use -test- as in the FAQ, you can do this:

[/code]
. test _b[1.foreign] + _b[1.foreign#1.repair] = 0

( 1) 1.foreign + 1.foreign#1.repair = 0

F( 1, 65) = 0.38
Prob > F = 0.5373

. local sign_combo = sign(_b[1.foreign] + _b[1.foreign#1.repair])
. display "Ho: coef >= 0 p-value = " 1 - ttail(r(df_r),`sign_combo'*sqrt(r(F)))

Ho: coef >= 0 p-value = .73134831
[/code]

To use -lincom-, some slight adjustments are needed to account for the lincom. We use the t-statistic instead of the F-statistic, and we can dispense with the sign, since that is included with the t-statistic.

. lincom _b[1.foreign] + _b[1.foreign#1.repair]

( 1) 1.foreign + 1.foreign#1.repair = 0

------------------------------------------------------------------------------
price | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(1) | 700.4192 1129.369 0.62 0.537 -1555.086 2955.924
------------------------------------------------------------------------------

. display "Ho: coef >= p-value = " 1 - ttail(r(df), r(t))
Ho: coef >= p-value = .73134831

Hi Leonardo,

Thank you so much for this, this works fine. I am very grateful for your help.

Very best regards,

Lucas