Implementation
The procedures has been implemented on the Indian data set using the Stata 8 “levpet”
command (Levinshon, Petrin and Poi, 2003).
One important issue for this estimation procedure is the choice of proxies. In fact any
intermediate material can be potentially used as a valid proxy. In our case we could use
both intermediate material inputs and energy consumption.
Often energy (and in particular electricity consumption) is considered the best proxy. In
fact, since it cannot be stored, its use should be highly correlated with the year to year
productivity term.
However, one of the basic estimation assumption is that the input demand function is such
that for any capital level and productivity shock, the firm is really able to obtain mt(wt,kt).
In the case of energy, we have that many of the firms in the sample have reported
pessimistic evaluation of the supply reliability. For example the mean number of power
outages or surges declared per month is about 9,3 and when asked to rate the quality of
power on a 1 to 10 scale, 40 percent of the firms in the sample judged it less than 5.
therefore unreliability of supply might lead, in our case, to the observed energy usage that
is different from the true demand. For this reason, we choose to rely on material inputs as
proxy variable. Furthermore, on the basis of some information reported in the data, the
“number of days of inventory kept for the most important product”48 is on average 30,
therefore we have grounds to consider intermediates as not heavily stored.
The estimation were then performed on each macro sector identified so no assumption of
common production technologies and common return to factors among sectors had to be
made. Outliers were identified by means of the Hadi method. In particular, we dropped the
one percent tails at both ends of the joint distribution of all the variables used in the
production function estimation.
Table A1 displays the coefficient from the production function estimations with the
Levinshon-Petrin methodology and with ordinary least squares. The variation between the
two estimates is close enough to expectations. In fact, in case of simultaneity bias OLS
tends to overestimate the labour coefficient and underestimates the capital coefficient.
Therefore the Levinshon-Petrin procedure, solving the bias, should give lower labour
coefficient and higher capital coefficient.
48 One of the question of the IC-survey.
37