Announcement

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

  • Running a random effect panel data model incl. interactions with control functions (incl. Gaussian Copula

    Dear Statalists,

    My question in short: Am I calculating my control function terms and the following regression correctly?

    I’m running a random effect panel model to estimate effects on weekly sales of product new releases (first 12 weeks after release, during 6 observation years) with exponentially declining sales on three sales channels. One (endog.) independent has a nonlinear effect that is measured via its squared values. For the other (endog.) independent there is no instrument, thus I rely on a gaussian copula term. Since many controls are time invariant a fixed effect models cannot return useful results, most information is omitted. Due to several interactions that shall be calculated in the following steps, I employ control function terms.

    Model

    Dependent: sales (= cumul. sales of product i in sales week t, t = 1…12. Accrued on three different sales channels simultaneously)
    Independents: w (invariant with t), w^2, and p (variant with t); all likely endogenous.
    Controls: c (time invariant continuous variables), d (time invariant dummies), i.sweek (fixed effect for sales week), i.year (fixed effect for year of product release = 1-6), i.channel (fixed effect for sales channels 1-3)
    Instruments for w, w^2: z1 z2
    Control Functions: w_cf, w2_cf, p_gc (= gaussian copula)
    Panel Setup: Panel Variable = ProductID+Channel, Time Variable = Sales Week.

    Stata commands
    1. Control Function Terms:
    Code:
    cumul p, g(p_cdf)
    gen p_gc = invnormal(p_cdf)
    reg w z1 z2 p p_gc c d i.sweek i.year i.channel, vce(bootstrap, reps(1000))
    predict w_cf, r
    reg w2 z1 z2 p p_gc c d i.sweek i.year i.channel, vce(bootstrap, reps(1000))
    predict w2_cf, r
    1. Regression:
    Code:
    xtreg ln_sales w w2 p w_cf w2_cf p_gc  c d i.sweek i.year i.channel, re vce(bootstrap, reps(200))
    Rob. Check: xtivreg ln_sales (w w2 = z1 z2) p p_gc c d i.sweek i.year i.channel, re vce(bootstrap, reps(200))
    Note/problem: sometimes stata omits the constant term (dependent on included fixed effects), the estimated coefficient for w, w^2 of xtreg vs. xtivreg are not the same.


    Thank you very much in advance, it would help me a great deal

    Kind regards
    Nicolas

  • #2
    I'm curious why you include so many instruments in the first-stage of the w_cf and w2_cf equations. The idea is to use exogenous instruments, and I doubt that all the instruments included meet the criteria. This may lead to multicolinearity issue and make the model less stable, perhaps?

    Comment

    Working...
    X