f**n
发帖数: 401 | 1 I have trouble with fitting data to panel models. The balanced panel
covers 50 subjects and 4 time periods. The data includes retail sales
d_i and price information p_i on six(6) products (i=1,...,6)in the
same product category.
I am trying to use fixed and/or random effect model to estimate a
price-demand relationship as follows:
d_1 = f(p_1, p_2, ..., p_6) where f() is a linear function.
I also tried to transform p and d variables into log forms.
This is what I found:
a) overall the data fits poorly (fixed or random effect); r-square is
typically in the range of 2%~5%.
b) for fixed effect models I almost ALWAYS get the wrong sign, that
is, if p_1 is higher, d_1 is higher.
c) cross elasticity is always insignificant
d) for a couple of random effect models with transformed variables,
the sign is right and statistically significant.
I am trying to understand what happened here. Any comment is
sincerely appreciated. In particular I have questions myself:
q0) shall I go back and collect more data, if possible?
q1) does it make sense at all to inspect the meaning of the model if
overall goodness of fit metrics is poor?
q2) what conclusion do people draw if fixed/random effect model does
NOT yield meaningful results? I can think of many reasons: business
scenario, instrument variables, etc.
q3) for some reason choice models do not apply to this case. What
else shall I try?
Many thanks! | C******n
发帖数: 9204 | 2 内生性。。。联立方程，dynamic panel。不了解这个literature。 |
|
