Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Interpreting linear model with square term (inverted U) using percentage change notion

    Hello,
    This is sort of a continuation of the discussion here: https://www.statalist.org/forums/for...ally-very-high

    Prof. Wooldridge kindly got me on the right track on interpreting margins dydx for the linear model with square terms. Now, I have a more conceptual followup question. I thought that this better be in a different thread as the topic is quite different from the original thread above.

    Say have my model as :
    ln(y) = b0 + b1 x + b2 x^2 + b3 z1 + b4 z2 + e


    I estimated and have coefficients for it, it does exhibit the inverted U behavior. Please note that x is a variable between 0 and 1, and the inflection point is well within the range.

    Now, if I evaluate ln(y1) = ln(y) evaluated at x=x1 and ln(y2) = ln(y) evaluated at x=x2, then

    ln(y2) - ln(y1) = -b1 x1 - b2 x1^2 + b1 x2 + b2 x2^2

    Taking exponents on both sides and simplifying:

    y2/y1 = Exp(-b1 x1 - b2 x1^2 + b1 x2 + b2 x2^2)

    So, if I now substitute b1, b2 coefficients and select two points x1 and x2 (x2>x1, and both are smaller than the inflection point) in the RHS, can I then interpret that as the ratio between y2/y1?

    Say, that the ratio is evaluated as 1.4, then can I state that moving x from x1 to x2, yields a 40% increase in y?

    From my experiments with predicted values, this does work out. But I am concerned that I am just playing around with point estimates and making some grave errors on not considering standard errors.


  • #2
    For relatively small changes, just use ln(y2) - ln(y1) is the estimated propotionate increase. So, if you compute the difference and it's something like .25, then that's 25% to first order. But what you've done is fine and gets a better estimate for large changes -- which you seem to have. So, yes, 40% is the correct interpretation.Generally, it's 100*[exp(c) - 1] where c is the estimated change in the logs. I would use the nlcom command to get a standard error.

    Code:
    nlcom 100*(exp(_b[x]*(x2 - x1) + _b[c.x#c.x](x2^2 - x1^2)) - 1)

    Comment


    • #3
      Thank you so much, Prof. Wooldridge!

      Comment


      • #4
        ns sn Please note our request to use full real names on Statalist. https://www.statalist.org/forums/help#realnames and #3 at https://www.statalist.org/forums/help#adviceextras explain why we ask that and how to fix your name.

        Comment

        Working...
        X