Measuring Productivity
Once you have a good idea of how many people work for a firm, a natural next step is to think about what those people are able to produce. Measuring productivity in the LBD is complicated by the need to pull together multiple data sources on firm inputs (capital, labour and intermediates) and outputs (gross output or revenue) which have been collected for different purposes, and to address changes in the underlying data collections over time.
The Fabling and Maré (2015, 2019) Productivity Tables
Happily for the rest of the research community, the Fabling and Maré productivity tables provide a cleaned, harmonised, and carefully documented dataset of the necessary components for measuring productivity.
Annual information on gross output, intermediate consumption and capital components comes from two sources: Statistics New Zealand’s Annual Enterprise Survey (AES) and the Financial statements summary - IR10 submissions to Inland Revenue. The former is a collection designed to support the construction of National Accounts (e.g. GDP). The latter is used by Inland Revenue to evaluate tax compliance risk, as well as to form policy and for strategic research. It is also used by Stats NZ to supplement the AES in construct the national accounts.
The labour component is described in the https://nzproductivity.atlassian.net/wiki/spaces/PH/pages/37322812 page.
When brought together, and after excluding observations that do not meet a reasonable quality threshold, harmonised productivity data from AES and IR10s are available for around 65 percent of active in-scope firms and 80 percent of (in-scope) employment.
Updates to the Productivity tables since 2019
Fabling and Maré (2015 and 2019) provide a comprehensive discussion of how AES and IR10 data have been processed to create the Productivity tables. Since the publication of the 2019 paper, the authors have continued to improve the code, resulting in changes to the productivity dataset for the 201903 financial year.
There are some substantial methodology and table structure changes that one should be aware of when using the new tables.
Additional IR10 quality check
Following work by Richard Fabling on firm finance costs, it became clear that there are problems with the IR10 balance sheet reporting under the old (pre-2013) form. Specifically, some firms likely report all assets in a single “other” category to minimise reporting effort (ie, non-itemisation). The effect of this in the old dataset is to include firms in the productivity dataset that have underreported capital. The productivity code now removes these observations, which results in a loss of 1-1.5% of former productivity observations in years prior to 2013, and substantially lower data loss in more recent years.
Changes to MFP estimates
The MFP estimates have been refined and incorporated into the productivity dataset suite. Because these estimates are restricted to a subset of firms, they have also been moved to a separate table (pent_prod_mfp_incumb_emp). Productivity components (y,m,k,l) remain in the main productivity table allowing users to continue to estimate alternative MFP measures as they wish. Key changes are:
Production functions are only estimated on incumbent (ie, non-entering/exiting) employer (ie, FTE>0) firms. The rationale for this change is that this sub-population are likely to have higher quality data (even after the data cleaning steps), with at least some of the observed dispersion in productivity in the WP-only subpopulation being due to measurement error (see http://motu-www.motu.org.nz/wpapers/14_16.pdf). While these restrictions have a substantial impact on sample size, the loss of aggregate output/inputs is much lower ( http://motu-www.motu.org.nz/wpapers/21_01.pdf) implying that: the new estimates cover the bulk of the economy; and that the excluded firm have the “weight of numbers” to substantially affect the estimated shape of the production function
MFP is now calculated simply as output less the estimated contribution of observed inputs. Compared to the last instance of the data, this change adds the industry time effect into MFP. MFP is then normalised so that it is mean zero within industry in 2001 (the first year of data)
OLS estimates of Cobb-Douglas and translog production functions have been dropped
Proxy variable (labelled _pv) estimates of the Cobb-Douglas production function have been added, following the one-step method proposed by Wooldridge ( https://ideas.repec.org/a/eee/ecolet/v104y2009i3p112-114.html ). Users may wish to exercise some caution with these estimates (eg, there is one industry where the estimated coefficient on capital is negative). Please also note that these estimates use the preferred measure of K - including rental, leasing and rates costs - meaning that the required timing assumptions around the relative stickiness of K and L used to apply the proxy variable approach may be less likely to hold in our data, compared to other datasets
Production function estimates with quality-adjusted labour have been added, where the quality adjustment comes from the estimated two-way wage fixed effects included in the labour-productivity suite (ie, ql=ln(L)+avg(WFE+xB))
Variable names associated with estimated translog coefficients have been changed to make it clearer that these production functions have been estimated using demeaned inputs (to aid interpretation of coefficients)
2022 Update
In the 2020 financial year, changes in international accounting standards came into effect, leading to a substantial increase in measured depreciation. The change relates to reporting for depreciation on "right of use" (long term leased) assets. These changes will be explored as part of the 2022 update to the productivity tables.
