Rethinking Regional Attractiveness in the New Global Environment

Annex A. Technical analysis

Foreign direct investment (FDI)

Table A A.1.  Regression results: Foreign Direct Investment

By dependent variable (column) and independent variables (row)

	Number of greenfield projects (coefficient)	Total capital expenditure of greenfield investments (coefficient)
Model (1): Education– OECD + EU TL2 regions (n = 369)
EDU_TOP500_UNIVERSITY	120.96***	3.781***
Model (2): Education and digital infrastructure – OECD TL2 regions (n = 300; 266 for capex)
EDU_TOP500_UNIVERSITY	124.7***	3.064***
DIG_DOWNSPEED_NAT100	1.33*	56.88*
GDP_PC_REAL_PPP	0.004***	0.02
Model (3): Education and transportation infrastructures – EU + UK (n = 179)
EDU_TOP500_UNIVERSITY	89.51***	2 657.21***
ACCESS_RAILWAY_EC	17.05***	235.0323
ACCESS_FLIGHT_EC	0.177***	-1.95
GDP_PC_REAL_PPP	0.001	0.0370

Note: While the traditional measure of FDI is total capital expenditure, the number of FDI projects is also considered as another way of measuring FDI, which can shed light on how lagging regions can attract relatively smaller-sized projects.

The significance level is expressed as the asterisks added to each coefficient. The smaller p-value signifies higher statistical significance level: *p<0.1; **p<0.05; ***p<0.01.

In the above table, regression analysis is conducted using different regional samples due to the lack of data available regarding train and flight accessibility when including non-European Union (EU) OECD regions. Hence, the analysis is split into three parts: i) a first study covering OECD and EU Territorial Level 2 (TL2) regions; ii) a second study on OECD regions, plus Argentina (due to their participation in the Rethinking Regional Attractiveness community of practice); and iii) a final analysis on EU countries for which the European Commission has produced an indicator of train and flight accessibility and proximity on Eurostat.

In the first instance, correlation coefficients using the dependent variable and the attractiveness database are conducted, along with a literature review, to get a first idea of which variables should be the subject of further exploration. All of the selected variables are significantly correlated to both the number of FDI projects and the amount of foreign-owned capital expenditures; however, the magnitude of correlation varies across indicators. In the selected variables, two are significantly correlated (>0.4) with the number of FDI projects and the sum of capital expenditures: universities and the flight and railway performance indicators, namely edu_top500_university, access_flight_ec and access_railway_ec.

For the three subsamples presented in Table A A.1. , the analysis has been done with the two variables of interest: i) the number of FDI projects; and ii) the amount of capital expenditures to check the consistency of the results.

Depending on the model, all things being equal and on average, one additional university ranked in the world’s top 500 ranking would have brought to a region between 90 and 125 more new FDI projects and between USD 2 657 and USD 3.781 million of additional capital expenditure in greenfield investments.

Then, for the OECD (plus Argentina) subsample, digital connectedness appears to attract foreign investment but the power of the relationship slightly edges down. A 10% increase in the download time from fixed devices (expressed as a percentage of the national average time) would have led, on average and all things being equal, to 13 additional new projects and USD 569 million of foreign capital expenditures. Once the control variables are added, the coefficients are still statistically significant showing little variance.

Regarding the EU countries’ subsample, after universities and research and development (R&D), the main drivers of FDI are train and flight accessibility. Utilising the EC’s passenger rail performance indicator, a measure of accessibility and proximity (calculated as the population within 90-minute travel over the population within a 120-km radius x 100), a 10pp increase in rail performance would have led to 171 additional foreign investment projects. Similarly, 100 more daily passenger flights accessible within 90-minute drive would have added 17 new projects. However, the statistical evidence is not clear when FDI is represented through the amount of foreign capital expenditures.

Talent

The first indicator entered into the stepwise model with time and region fixed effects is the share of international students in the student population in higher education (eter_stu_for_sh), which produces a statistically significant and positive coefficient (1). This shows that the higher share of foreign student population has a positive impact on regional talent attraction (Table A A.2)

An indicator of digitalisation is then added to the panel regression with the first independent variable (2). The share of households with broadband Internet access (bb_acc) exhibits a positive and statistically significant coefficient. This shows the importance of digitalisation for regions to attract talent.

In the third model (3), gross domestic product (GDP) per capita is added as a control variable and the behaviour of the first two variables with its addition is observed. While bb_acc stays relatively stable and continues to have a statistically significant coefficient, the coefficient associated with the international student variable (eter_stu_for_sh) is no longer significant. This could be attributed to an omitted variable bias that underestimates the coefficient.

Moving on from the stepwise panel regression with two-way fixed effects to the time-lag model, 2017 and 2018 data are used for the aforementioned three independent variables. The time-lag model considers the time it takes for the independent variables to affect the dependent variable and a one-year lag is considered. Similar results are shown in the time-lag model as in the panel regression model. The coefficients of bb_acc are statistically significant and the p-value of bb_acc is still less than 0.05, just as in the panel regression.

Finally, the housing affordability variable (afford_h) is added to the time-lag model (5). The data for this variable were only available as the average over 2016-20 and the average over the 2017-21 period (individual yearly data are not published). Thus afford_h could not be inserted into the panel regression but only into the time-lag model. Indeed, afford_h exhibits a statistically significant and positive coefficient, which conveys the importance of housing affordability in attracting and retaining talent. On average and all other factors being equal, a 10% increase in the share of the population satisfied with housing affordability translates into a 1.8% increase in the foreign-born share of foreign-born employed people in the total working-age population (15-64). The coefficient of GDP per capita was positive in the two-way fixed effect panel regression (3) but negative in the time-lag regression (5), perhaps because GDP per capita is positively correlated with many other variables, thus potentially resulting in omitted variable bias. One limitation of afford_h is that it is a subjective indicator of the share of people satisfied with housing affordability in their region. Nevertheless, the link between housing prices and talent attraction is discovered from this analysis, which can open a new series of relevant discussions.

Table A A.2.  Regression results: Talent

By dependent variable (column) and independent variables (row)

	Dependent Variable: the share of foreign-born employed people in the total working-age population (15-64)
	Panel regression model (1)	Panel regression model (2)	Panel regression model (3)	Time-lag model (4)	Time-lag model (5)
ETER_STU_FOR_SH	0.0173** (0.0088)	0.0165* (0.0097)	0.0117 (0.0105)	-0.3211 (0.333)	-0.4300 (0.353)
BB_ACC		0.0221** (0.0105)	0.0200** (0.0101)	0.1725** (0.079)	0.1368* (0.080)
GDP_PC_PPP			0.0001** (4.31e-05)	-0.0006 (0.000)	-0.0010** (0.000)
AFFORD_H					0.1786** (0.086)
Fixed effects	Region, Year	Region, Year	Region, Year	None	None
Number of observations	1 546	1 455	1 455	186	186
Covered regions and features	389 EU and Turkey regions	389 EU and Turkey regions	389 EU and Turkey regions	186 EU and Turkey regions	186 EU and Turkey regions
Covered time periods	2011-19	2011-19	2011-19	2017-18	2017-18 (see note)

Note: The significance level is expressed as the asterisks added to each coefficient. The smaller p-value signifies higher statistical significance level: *p<0.1; **p<0.05; ***p<0.01

The dependent variable was constructed by multiplying the TL2 regional foreign-born employment rate and the foreign-born participation rate indicators from the OECD Regional Attractiveness Database

Regarding the covered time periods for the last time-lag model, 2017-18 data were used for all variables in model (5) except afford_h, whose data were available as two time periods: 2016-20 average and 2017-21 average.

Visitors

A stepwise panel regression is applied with time and region fixed effects to control for the potential variation in tourism drivers across the large sample of regions and the ebbs and flows over time (Table A A.3). As a first step, the firm birth rate (empent_b_ra) is regressed. empent_b_ra is found to have a positive coefficient with strong statistical significance in its univariate model (1) with the dependent variable, the ratio of the number of overnight tourist stays over the number of accommodation beds in the TL2 region.

The second variable to be added to the model is the air pollution variable (air_pol_t), defined as the average level of pm2.5 in micrograms per cubic metre (µg/m³) experienced by the regional population. Both empent_b_ra and air_pol_t exhibit significant coefficients in the second model (2). The firm birth rate indicator remains positive, while the air pollution indicator exhibits a negative relationship with the tourism-dependent variable. This means that the higher firm birth rate and lower air pollution will attract more visitors per accommodation bed.

The share of international students in the student population in higher education (eter_stu_for_sh) is then added to the panel regression. The coefficients of air pollution and firm birth rate indicators both remain stable and statistically significant, and the newly added foreign student share variable shows a positive and significant coefficient.

Table A A.3.  Regression results: Visitors

By dependent variable (column) and independent variables (row)

	Ratio of the number of overnight stays in tourist accommodation to the number of tourist accommodation beds
	Panel regression model (1)	Panel regression model (2)	Panel regression model (3)	Panel regression model (4)
EMPENT_B_RA	0.8147** (0.0334)	0.9140** (0.4035)	1.7872** (0.0109)	1.6283** (0.6804)
AIR_POL_T		-1.8199*** (0.6203)	-1.8247** (0.8391)	-1.1090 (0.7573)
ETER_STU_FOR_SH			0.1335* (0.0806)	0.1131* (0.0671)
GVA_IND_TOTAL				0.0012*** (0.0004)
Fixed effects	Region, Year	Region, Year	Region, Year	Region, Year
Number of observations	1 276	1 276	1 065	1 065
Covered regions and features	168 EU regions	168 EU regions	160 EU regions	160 EU regions
Covered time periods	2008-20	2008-20	2011-20	2011-20

Note: The significance level is expressed as the stars added to each coefficient. The smaller p-value signifies higher statistical significance level: *p<0.1; **p<0.05; ***p<0.01

Finally, the regional gross value-added (gva_ind_total) is added as a control variable. With the new addition of gva_ind_total, eter_stu_for_sh maintains a similar level of statistical significance and a positive coefficient positive. While the coefficient of air_pol_t stays negative as in every previous model, it loses its statistical significance with the inclusion of gva_ind_total. Although this could have been caused due to an underestimation of the coefficient, for now, there is not enough evidence to conclude that air_pol_t exhibits a causal link with the tourism-dependent variable. In the case of empent_b_ra, the coefficient remains positive and with strong statistical significance as before, a result that leads to the conclusion that the firm birth rate is the strongest driver of tourism among the three aforementioned independent variables. On average and all other variables being equal, a 1% increase in the firm birth rate leads to a 1.6% increase in the ratio of the number of overnight stays in tourist accommodation to the number of tourist accommodation beds.