Natural resources and income inequality in developed countries: synthetic control method evidence

Similarly, fixed-effects model allows for the individual-specific unobserved heterogeneity but restricts the effect of this heterogeneity to be constant in time.

The exclusion restrictions do not necessarily have to be valid given than commodity price changes are correlated with inequality.

This effect may be mitigated depending upon the ethnic make-up of the country in question, as Fum and Hodler (2010) find that ethnically homogeneous countries have lower income inequality in the face of resource abundance. However, the ramifications of this finding are that countries which are highly fractionalized tend to have investment projects with negative social surplus, using overinvestment as a way to credibly redistribute funding but in a less efficient manner (van der Ploeg 2011). Lessmann and Steinkraus (2019) using global data also show that inequality in the spatial location of mineral resources induces ethnic income inequality and strong institutions cannot mitigate this relationship.

However, Stijns (2006) points out that the relationship between resource abundance and human capital accumulation strongly depends on specific indicators used to measure the natural resources and education. He finds that mineral export shares are negatively correlated with the average years of schooling, while resource rents per capita are positively correlated with average years of schooling, and this effect is especially strong for developing economies. Also, Emery et al. (2012) employ data from the oil-rich Province of Alberta in Canada and suggest that oil booms may lead to postponement of education, but do not reduce human capital accumulation.

To reach these results, Goderis and Malone (2011) identify the impact of natural resource booms through export prices variation and a country’s dependence on the export of particular goods, constructing a country-specific commodity export price index and analyzing its impact on income inequality. The index reflects a country’s export of particular commodities and the same country’s dependence on world prices for those commodities. Thus, the index is larger for commodity-dependent countries.

Smith (2015) and Cavallo et al. (2013) use the synthetic control method to examine (respectively) the impact of natural disasters and resource discoveries on economic growth.

A unit can be, for example, country, city, school, etc. In our analysis, a unit is a country.

A weighted average of countries without natural discovery, which characteristics match the income inequality in the country with natural discovery before the actual natural discovery as closely as possible.

A predictor can be a pre-intervention value of any variable, including the outcome variable, or the linear combination of them. For example, years of schooling 1 year before the intervention can be one predictor, years of schooling averaged over 5–10 years before the intervention can be another predictor, and outcome variable 1, 3, and 5 years before the intervention can also appear among predictors. Predictors help to determine countries with similar pre-intervention characteristics as the treatment country.

To achieve the uniqueness of the solution, the Euclidean norm of \(\mathbf {V}^{\star }\) is normalized to one.

See Saez and Zucman (2016) on the measurement of wealth inequality and its evolution in the USA over time.

An alternative data source on income inequality is internationally comparable Gini coefficients, for instance, from Solt (2016). However, the Solt data exist only starting from the 1960s for selected countries (but, for example, for Norway only from 1970), a period which coincides with the timing of natural resource discoveries. To identify the impact of natural resource discoveries on income inequality, data for a longer time period are required. In addition, the data are also available for the UK but the natural resource discovery in the 1970s has been followed by another major shock—financial market deregulations in the 1980s. This creates difficulties to identify the effect of natural resource shock on income inequality within the synthetic control method. We refer the reader to Tanndal and Waldenstrom (2018), who examine the effect of financial deregulation on top incomes in the UK using the synthetic control method. See Atkinson and Sogaard (2016) on the long-term evolution of income inequality in Denmark.

For resource production data, Smith (2015) uses UN Industrial Commodities Statistics; for oil discovery dates, he utilizes the 2007 and 1994 editions of the Oil and Gas Journal Data Book.

For gas-producing countries, Smith converted natural gas production to its oil barrel energy equivalent using the conversion rate of 0.00586152 oil barrels per terajoule.

The event year can be endogenous to economic growth if countries invest in exploitation only in periods of high oil prices (Smith 2015). However, for our purposes, high oil prices are less of a concern in case of studying income inequality.

We also checked that these event years broadly correspond to the evolution of natural resource rents to GDP from the World Bank Dataset. We also experimented with a different event year, shifting the event year 3 years before to match it closer to the discovery year, which did not have a systematic effect on our results. Smith (2015); Roine et al. (2017) argue that it typically takes about 4–6 years from drilling to production.

Another issue in applying SCM is the data availability requirement. Essentially, data on the outcome variable (in our case, income inequality data) cannot have any missing values during the sample period for the treatment country and for all countries in the donor pool. Predictors that are not lagged outcome variable can have missing values, but they cannot be missing for the entire pre-treatment period for any country, because they are averaged over this period. Those predictors can also be averaged over a subset of a pre-treatment period; accordingly, they cannot be missing for the entire period they are averaged over. Income inequality data were missing for some years for the majority of the possible sample countries, so we decided to fill those missing values using linear interpolation. We opted for linear interpolation in favor of cubic spline interpolation because, even though the latter produced qualitatively similar results, we obtained a lower mean square prediction error (MSPE) with linearly interpolated values. Another solution would be to discard all years with missing outcome variable values for at least one (treatment or control) country, but that would leave us with only a few pre-treatment years available for each analysis.

As a robustness check, we also used more restricted set of predictors but the results remained largely the same.

Only control countries with nonzero weights are included.

According to the OECD tax database, and the changes in tax policies in Denmark, Netherlands, and Norway have been the opposite of any increased redistribution over the past 20 years. For example, Denmark’s top statutory corporate tax rate decreased from 32% in 2000 to 22% in 2018, while in the Netherlands it went from 35% in 2000 to 25% in 2018 and in Norway it went from 28% in 2000 to 23% in 2018. For the income tax rates, the top marginal rate for Denmark, Netherlands, and Norway (in that order) in 2018 was 27.16%, 51.75%, and 24.05%, while in 2000 they were 28%, 60%, and 29.85% (top statutory rates in 2000 for Denmark, the Netherlands, and Norway were 59%, 60%, and 47.5%, while in 2019 they were 55.9%, 51.8%, and 38.2%). In each instance, top tax rates were reduced, suggesting less of a social pot of funding for redistribution—and suggesting that inequality should be increasing. This is the case with social security taxes as well, which have also undergone consolidation in each country and, in each case, a lowering of rates.

Smith (2015) also included Belgium to this region but there is not any data on either the top 1% or top 10% share for Belgium at the World Wealth & Income Database.

See Table 6 for predictors in each analysis and Table 1 for sample periods.

Therefore, we assume the natural resource discovery in 1955 and 1960, respectively, for the Netherlands and 1960 and 1965, respectively, for Norway. We cannot conduct in-time placebo test for Denmark because our data for Denmark start from 1976 and the actual treatment for Denmark is 1981.

We choose the significance parameter in line with Firpo and Possebom (2018), who develop confidence intervals for the synthetic control method. They suggest choosing the significance as the number from 1/(1+J), 2/(1+J), ...., (1+J)/(1+J), where J+1 is the number of countries in the donor pool. However, in our case, the number of countries in the donor pool precludes us to get exactly at the significance parameter of 0.05. We would need at least 20 countries in the donor pool to achieve the ratio 1/20 = 0.05, while our donor pool is between 8 and 15. We set the significance parameter as low as possible: Denmark—top 1% income share: 2/15, Denmark—top income 10% share: 1/11, Netherlands—top 1% income share: 1/13, Netherlands—top 10% income share: 1/8, Norway—top 1% income share: 1/11, Norway—top 10% income share: 1/8. The significance parameter is set to attain the highest significance level possible while ensuring the convergence of the estimator (for this reason, the number in the numerator is not set to 1 in all the cases).