of the ‘x axis’ and of the ‘y axis’ as a device for handling measurement errors. He
did not, however, consider this method in a panel data context.
For materials, the between period (BP) estimates on levels for the original
and the reverse regression imply virtually the same input elasticity, 1 /μ, in the
range 1.00 - 1.09 for the four sectors considered. They are also very close to
the estimates obtained from seven-period differences. The BP estimates based
on differences, βb BPDC, show somewhat larger discrepancies. For capital, there are
substantial deviations between the level BP, the difference BP, and the seven-period
difference estimates. For the BP estimators on levels, the reverse regression gives
systematically higher estimates of the input elasticity of capital (lower estimates
of μ) than the original regressions. This may indicate that the measurement errors
in capital have period specific, or strongly serially correlated, components, which
make both the between period and all period difference estimators inconsistent.
For capital, unlike materials, the results also suggest the presence of period specific
heterogeneity in the relationship.
OLS estimates calculated from levels and from one period differences,
b
βOLS
∑i ∑t(xit - x)(yit - y)
Pi Pt(Xit — x)2
b
βOLSDC =
D∑t(∆Xit - ∆X)(∆yit - ∆y)
Pi Pt(∆Xit - ∆X)2
b
βOLSD =
∑l∑t (∆ Xit )(∆ yit )
Pi Pt (∆Xit)2
are also reported (columns 1, 4, and 8), βOLSDC removing the possible effect of
linear trends. Columns 3 and 6 contain within firm (WF) estimates calculated from
levels and differences,
b
βWF
∑i ∑t (Xit — Xi∙ )(yit - yi∙ )
Pi Pt (Xit - Xi∙ )2
b
βWFDC =
∑i'∑t(∆Xit - ∆Xi∙)(∆yit - ∆yi∙)
Pi Pt(∆Xit - ∆Xi∙ )2
βb WFDC removing the possible effect of linear trends. These three OLS and the
two WF estimates, all of which are inconsistent in the presence of measurement