Quality and quantity in the innovation process of firms: a statistical approach

<<It is not strictly necessary from a technical point of view that highly collinear variables be excluded. For instance, if two perfectly collinear variables were included in the composite, with weights w₁ and w₂, then the particular dimension of performance which they measure will be included in the composite with the weight (w₁ + w₂). This is not problematic if the weights have been chosen correctly>> (Jacobs et al. 2004, pp. 34–35).

The choice of the proper number of principal components takes place on the basis of three criteria which take into account their explanatory power. First we consider a number of principal components which take into account at least 95 % of the variance of each of the k initial variables, which imposes a minimal threshold; second, we keep all the principal components whose eigenvalue is larger than 1; third, we observe the screen plot of the eigen values as a function of the number of principal components; as eigenvalues are obtained in decreasing order, the graph will show a decreasing curve, with a kink in correspondence to the proper number of principal components. In particular, on the basis of the results of the analysis, we choose two principal components.

See Moncada Paternò Castello et al. (2009) for more details.

See Maraut et al. (2008) for the methodology used for the construction of REGPAT.

Please contact Helene.DERNIS@oecd.org to download the REGPAT database.

Sources for exchange rates and deflators are EUROSTAT.

Low [0, 1412.581], medium (1412.581, 2825.162] and high (2825.162, 25651].

Low [0, 742.0396], medium (742.0396, 1484.0792] and high (1484.0792, 59243].