Fig 1.
Sample characteristics: Countries, cities and city innovation.
Panel a indicates the countries included in the sample, and the number of cities in each country. The final sample contains 96 countries and 164 cities. Inclusion criteria are explained in the Methods section. Panel b highlights the cities in the sample, the GDP per capita of the sample countries, and the aggregate innovation score of each city. Our regression analysis is at the firm level, and we do not use city aggregates; in this figure they are used for illustration. The innovation index is derived from the WBES data as explained in the Methods section. For cities that appear twice in the WBES data, the average NTL and/or innovation values for those two years were employed. Maps of city boundaries come from the Database of Global Administrative Areas (https://gadm.org/download_country_v3.html). We use ArcGIS to match GPS coordinates of firms with city locations. Source: own calculations.
Table 1.
Summary statistics of key variables.
Table 2.
Innovation and night light intensity for cities in developing countries.
Fig 2.
Innovation, nightlights and a continuous measure of city size.
Figs a-d report the relationship between innovation and nightlights for the continuous spectrum of city population sizes. The graphs plot the predicted outcomes of innovation (z-axis) based on the interaction between nightlight intensity (x-axis) and city population (y-axis) using an ordered logit model. The outcome of innovation in a and b is innovation intention, calculated as the cumulative (predicted) probability of having low, medium and high levels of innovation. The outcome of innovation in c and d is innovation intensity, calculated as the predicted probability of having a high level of innovation. Figs a and c present the 3D graphs from the front angle and b and c present the same graphs from the back angle. Data on city population size is obtained from the most recent census data for each country, and only counts the population in the city proper or urban areas, whichever is consistent with the city boundary defined in our analysis. Darker color indicates higher night light luminosity. S3 Table reports the regressions upon which these 3D graphs are based.
Fig 3.
Innovation and night light in large and small cities.
This Figure summarizes predicted outcomes of innovation from ordered logit models with 95% significance levels. The ordered logit models regress two innovation outcomes on a quadratic equation of logarithm of nightlight density: the outcome in a refers to innovation intensity, calculated as the cumulative (predicted) probability of having low, medium, and high levels of innovation; the outcome in b refers to innovation intensity, calculated as the predicted probability of having a high level of innovation. Confidence intervals in b are derived directly from the ordered logit regression. For a, as it combines three out of the four outcomes Pr(Innovation Index = 1, 2 and 3) in ordered logit regressions, we calculate the confidence intervals for Pr(Innovation Index = 0) from ordered logit regressions first. We then use one minus the above intervals to approximate the confidence intervals for Pr(Innovation Index = 1, 2 and 3). The results from this construction are very similar to the ones obtained directly from logit models where the dependent variable is a dummy variable based on whether the innovation index is greater than 0. In each graph, we estimate these ordered logit models for large and small cities separately. Large cities are defined as those with >5 million population and small cities as those with <5 million population. Nightlights are measured in natural logarithms and lagged. All equations include lagged GDP per capita, regional dummy variables and year fixed effects. Reported coefficients are all statistically significant at 95% levels. Full results of regressions on the respective sub-samples upon which these figures are based are reported in S4 Table. Red lines refer to large cities and yellow lines indicate small cities. The grey shaded areas are 95% confidence intervals. The point estimates of the linear and quadratic forms of the logarithm of night light density and the corresponding two-tail p-values of the estimated parameters are written next to the curves. The results for each individual component of the index are similar. In a robustness check, we also use 10 million population as the threshold for large cities because 10 million population is the threshold for “megacities” in the UN Report on World Urbanization Prospects 2018. We do not use the 10 million threshold in our main analysis because it results in a much smaller sample of giant cities: only 12 cities are identified as megacities in our sample using that measure.
Fig 4.
Geographical distribution of large cities and small cities.
This Figure plots the distribution of large and small cities and the average innovation index among all firms in these cities. As before, large cities are defined as those with >5 million population. Panel a plots 36 cities in our sample that are defined as “large” cities and b plots 128 “small” cities with <5 million population. The Figure also highlights some representative cities in each category with their corresponding nightlight density and innovation levels. Representative cities are selected to ensure maximal coverage of continents. The Figure reveals two patterns: 1) the maximum level of innovation is higher in large cities (2.53) than in small cities (2.02); 2) among large cities in a, innovation levels in general decrease with nightlight density. On the contrary, among small cities in b, innovation levels in general increase with nightlight density. National/regional specificity and the level of country development both have weak explanatory powers in predicting which cities have positive agglomeration effects and which have negative ones. Nightlight data is based on the harmonised version derived by Li et al. (2020) because different satellites were used in 1992–2013 (Defense Meteorological Satellite Program (DMSP)) and 2012–2018 (Visible Infrared Imaging Radiometer Suite (VIIRS) from the Suomi satellite). City location comes from the Database of Global Administrative Areas (https://gadm.org/download_country_v3.html).