Announcement

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

  • sivqr now on SSC (for IV quantile regression)

    The command sivqr (smoothed instrumental variables quantile regression) is now available on SSC, thanks to Kit Baum. (Very briefly: it's like a combination of ivregress and qreg.) Details are at
    https://kaplandm.github.io/#sivqr
    including a paper describing the command (revision under review at Stata Journal).

    If you have questions, please email me or mention me (@David Kaplan) in your Statalist post (so I get notified); I'm happy to try to help. =)

    Dave
    David M. Kaplan
    Associate Professor & co-DGS, Economics, University of Missouri
    https://kaplandm.github.io/

  • #2
    Hi David Kaplan, I've just used sivqr and would appreciate if you could comment on my interpretation of the results.

    Context & Question: I'm looking at the impact of propensity to plan (P) on a measure of wealth (W), and have run a sivqr regression for deciles from 0.1 to 0.9. For each decile, I took the value of the wealth measure. Would I be right in thinking that I could calculate the implied % impact of P on W by dividing the sivqr coefficient of P by the value at each decile?

    The conclusion I am drawing is that for lower levels of wealth, a one-point increase in planning is much more impactful than for higher levels of wealth.

    I'd be grateful for any feedback or other ways of thinking about this question.

    Best,


    Joe
    Decile Coefficient of P and significance Wealth value at decile (arbitrary units) Implied % impact on Wealth of a one-point increase in P
    0.1 0.127** 0.484 +26.2% (i.e., 0.127/0.484)
    0.2 0.109*** 0.667 +16.3%
    0.3 0.134** 0.873 +15.3%
    0.4 0.143*** 1.13 +12.7%
    0.5 0.128*** 1.50 (median) +8.5%
    0.6 0.125*** 1.89 +6.6%
    0.7 0.148*** 2.24 +6.6%
    0.8 0.152*** 2.54 +6.0%
    0.9 0.141 (not sig.) 2.85 n/a

    Comment


    • #3
      Hi Joe Rifaat,
      Thanks for your question. I think that would give you an approximately correct value, with some caveats.
      1. If you think of the tau-QTE of changing from planning level P=a to P=b, then ideally I think you'd want to express the tau-QTE as a percentage of the initial wealth level, i.e., the tau-quantile of the potential outcome (wealth) distribution for P=a. Generally, this does not equal the unconditional tau-quantile of observed wealth, although the latter would give at least an approximate idea of scale. Maybe instead of claiming an exact % effect, you could report the coefficient along with the unconditional wealth deciles (and could say what the corresponding % is but without claiming it's an exact value).
      2. If you have control variables, then the IVQR model (Chernozhukov and Hansen, 2005) is for conditional-on-X quantile treatment effects, which further complicates things; but again I think the unconditional wealth deciles would help give a sense of scale even if you can't exactly compute a % effect.
      3. Do you have a lot of non-positive wealth values? If there aren't "too many" then you could use log wealth as your outcome, to directly get a % effect interpretation. Note that unlike for means/ATE, the most negative values have no effect on quantiles, so you can just set something like logW=-99999 if W<=0 (and otherwise logW=ln(W)). For example, the median of the sample {-9, 1, 3, 5, 8} is 3, and the median of {-99999, 1, 3, 5, 8} is also 3, because the value of the minimum has no influence over the median. Similarly, it doesn't affect the tau-quantile for any tau>=0.5, or even any tau>0.2, but it *does* affect the very lowest tau<=0.2. So probably your tau=0.9 results would be fine using log wealth, and probably tau=0.8, ..., but depending how many non-positive W values you have, this might not be possible for the lowest decile(s). (To check, you can run it with logW=-99999 if W<=0, and then run again with logW=-99, and see if the estimates change.)
      Sorry that's a bit more complicated than a "yes" or "no," but I hope you find it helpful!
      Kind regards,
      Dave
      David M. Kaplan
      Associate Professor & co-DGS, Economics, University of Missouri
      https://kaplandm.github.io/

      Comment

      Working...
      X