4.2 Heterogeneity in unit values

Using firm-level data minimises the extent to which ERPT estimates are affected by quality and compositional changes. While each product category may contain a range of varieties selling for different prices, restricting to goods supplied by a single New Zealand firm captures a large proportion of within-good variation, as the range of a single firm is more restricted than that of exporters in the country as a whole.

Goods are defined at the 10-digit HS level. Relationships (firm-good-destination combinations) with only a single unit value observation are excluded. Column 2 restricts to relationships with at least 10 unit value observations. NEC stands for not elsewhere classified.

Table 3 demonstrates this by reporting R² from regressions where the log of the price level is regressed on fixed effects at increasingly detailed levels of resolution. Focussing first on column 1, which includes all export relationships, we see that good and good-destination fixed effects capture a substantial proportion of the overall variation in unit values across observations (R² of 0.731 and 0.795 respectively). This reflects the unit of observation available in most aggregate studies, which pool all exporters trading a certain good to the same destination.

Similarily, since firms trading multiple products to multiple destinations represent a large proportion of trades, firm-level controls alone (row 3) account for relatively little of the observed variation in unit values (R² of 0.622). In contrast, relationship (firm-good-destination) level controls add substantial explanatory power, soaking up a total of 92 percent of total variation. Residual variation is likely to include some degree of composition change, alongside exchange rate-induced price changes, other price shocks, and random variation due to, eg, measurement error. However, the use of detailed firm-level data on export relationships minimises the role of composition change as much as possible for comprehensive administrative data.

To check that the high R² at the firm-good-destination level is not driven by the prevalence of relationships which have only a few observations, column 2 repeats these regressions excluding all relationships with less than ten observations. This adjustment makes little difference to the final share of variation explained in the fully disaggregated specification, but leads to reasonably substantial increases in R² at other levels.

Columns 3 and 4 consider the same disaggregation strategy across two subsets of goods - differentiated goods as defined by Rauch (1999) (column 3) and those allocated to residual product groupings, as identified by product descriptions including the term "not elsewhere classified" or "n.e.c." (column 4).[15]

These breakdowns give further assurance that the disaggregated approach provides a good control for composition. In the case of both differentiated and NEC products, the good and good-destination controls provide substantially less explanatory power, as we would expect (because the products themselves are more highly differentiated, or because a range of products have been grouped under a single heading). Adding firm controls effectively closes this "explanatory power" gap, consistent with firms having limited product ranges within these categories, or market segmentation of their multiple varieties.[16]