[Paul Leeds]

Thanks everyone for the replies.

Richard, which code are you referring to to generate the squared term? is it

gen SQNAF = sqrt(NAF)

or

gen SQNAF=NAF^2

?

[Paul Leeds]

I would also like to add that the main continuous variable (NAF) is in log so I have created the squared term out of the raw dataset.

[Richard Williams]

The latter. c.NAF#c.NAF means multiply NAF by itself, e.g. square it. Besides saving you the trouble of computing the variable beforehand, this approach has advantages when using post-estimation commands like margins. But in any event, the squared term seems to be a distraction at this point, so get rid of it.

[Paul Leeds]

After rerunning the model once without the interaction term and then with the interaction term the coefficients stands almost the same with the model that had the squared term.
So, as you said earlier, the conclusion is that there is strong effect of NAF on READ before the crisis but not during the after the crisis. This may be theoretically provocative but can be justified by saying that the crisis casts different rules of play on the relationship between fees and text quality.

Many Thanks for your careful comments. This was really helpful.

Regards

[Paul Leeds]

Another scenario where the first model regress READ (text quality, low is difficult) on fees and the coefficient of the main effect (fees) is significant.
the second model includes an interaction term of fees*performance (dummy 1 good 0 bad) and tern out that the term is significant.
Can I say that the coefficient of the main effect in the second model represent the effect of fees in poor performance (dummy code 0) and the sum of coefficients of the main effect and the interaction term represent the effect of fees in good performers (dummy code 1)? and can then compare the level of effect?

[Richard Williams]

is the main effect of performance in the model as well? If not it should be.

It would help to see the output from the command. Your description may be fine but my eyes sometimes glaze over when I just hear things described rather than actually see them.

[Paul Leeds]

Hi Richard

These are the model

Model 1

              |               Robust
      FLESCH |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
          AF |      -1.3      .39        -3.30   0.001      -2.06  -.52
      RETURN |       .74      .74         1.01   0.313      -.70   2.20     
       _cons |      -680      300        -2.5    0.020      -1300   -109

Model 2

            |               Robust
      FLESCH |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
          AF |     -2.3      .68         -3.28   0.001    -3.7    -.91
      RETURN |     -5.9      3.1         -1.93   0.054    -11.8    .092
RETURN#c.AF |
          1  |     1.1       .5            2.3   0.025     .135   2.1

[Richard Williams]

The effect of AF is -2.3 when performance is poor and -1.2 (-2.3 + 1.1) when performance is good, i.e. the negative effect is twice as large for poor performers. I think that is what you were saying before.

[Paul Leeds]

That's alright..

Thanks Richard