Climate change and urban total factor productivity: evidence from capital cities and municipalities in China

Abstract

This paper studies climate change impacts on total factor productivity (TFP) in China using economic and climatic data for provincial capital cities and municipalities from 1998 to 2017. We employ a novel nonparametric quantile method to decompose historical temperature data into multiple temperature quantiles, which are then used in our regression analysis to avoid estimation bias caused by seasonal heterogeneity of temperatures across China. Specifically, we create three temperature quantiles for each city to represent their extremely high temperatures in summer, extremely low temperatures in winter, and mild temperatures in spring and fall. In general, we find that a warming climate has a significant negative impact on TFP in the long-run, while in the short term, only increases in extreme temperatures exert significant negative effects on TFP growth. However, the temperature effects on TFP vary substantially across coastal capital cities, inland capital cities, and municipalities due to their differences in geography, development levels, and political positions. Finally, our results are robust when spatial spillover, temporal lagging, and labor intensity effects are taken into account.

Appendices

Appendices

The TerraClimate data source

TerraClimate (Abatzoglou et al. 2018) is a high-resolution global climate dataset that is publicly available and can be accessed as a collection of images using the API (application programming interface) to the Google Earth Engine Platform.⁸ It is developed with a focus on temporal analysis, and contains monthly climate and climatic water balance for global terrestrial surfaces with a relatively long temporal coverage from 1958 to 2018. Moreover, TerraClimate is derived using three global gridded climate datasets (that are mostly weather station data collections), the WorldClim climatology version 2 (Fick and Hijmans 2017), Climate Research Unit (CRU) time series data version 4.0 (Harris et al. 2014)) and the Japanese 55-year Reanalysis (Kobayashi et al. (2015)). In particular, it employs climatically aided interpolation (e.g., Willmott and Robeson (1995) and Mosier et al. (2014)) to combine higher-spatial (< 5-km) but lower-temporal (1960–2000) resolution climatological normals from the WorldClim dataset with coarser-spatial but higher-temporal (1958–2018) resolution data from CRU Ts4.0 and the Japanese 55-year Reanalysis (JRA55) for a high-spatial-resolution data with a broader temporal coverage. The advantage of using such data is that one can extract climate data for temporal analysis in any particular geographical region across the Earth given its high spatial resolution of 2.5 arc-minute (smaller than 5 km).

Fig. 1

The monthly land surface temperature series from the two sources, the China City Statistical Yearbook and TerraClimate, for capital and municipality cities in China, 1998–2018

Fig. 2

Comparing the monthly land surface temperature between the China City Statistical Yearbook and TerraClimate. The vertical axis denotes the temperature sourced from the China City Statistical Yearbook for each city, and the horizontal axis denotes the temperature sourced from the TerraClimate for each city. The blue dashed line is the 45-degree line

Specifically, to obtain the land surface mean temperature (denoted by $T_{avg}$) for a particular provincial capital or municipality city in China in a certain month, we first leverage the shapefiles which can identify the boundaries of these cities in the global map obtained from GADM⁹ and compute the medians of the maximum and minimum temperatures denoted by $T_{max}$ and $T_{min}$ respectively over all the pixels within the city of interest, where it is assumed that the regions within the same city should have similar surface temperatures and the median is used to avoid the effect of extreme temperature values. Then, we simply derive the mean temperature as $T_{avg} = (T_{max} + T_{min})/2$ as this is the formula employed in the construction of CRU time series dataset (e.g., Harris et al. (2014)). The temperature data is coded based on the Celsius degrees ($^{\circ }$C) at 0.1 scales.

It is worth noting that in the TerraClimate data the JRA55 data is mainly utilized for those regions not covered by the CRU Ts4.0 (including all of Antarctica, and parts of Africa, South America, and scattered islands) and has been used exclusively for solar radiation and wind speeds, so the temporal information of TerraClimate is inherited from CRU Ts4.0 for most global land surfaces for temperature, precipitation, and vapor pressure. Therefore, the information of the panel data set for the land surface mean temperature we have constructed in this paper is essentially from the CRU Ts4.0. Moreover, the CRU climate data is also the primary source for the annual and monthly temperature data provided by the World Bank’s Climate Change Knowledge Portal (https://climateknowledgeportal.worldbank.org).

Temperature data validation

To confirm the accuracy and validity of the use of TerraClimate data to present the changing pattern in the land surface temperature of China, we compare the derived temperature data with those published in the China City Statistical Yearbook across the overlapping period from January 1998 to December 2018. In particular, Fig. 1 demonstrates the monthly land surface temperature series sourced respectively from the China City Statistical Yearbook and TerraClimate data in the same graph. The general observation is that temperature series from the two sources are aligned with each other for most cities. However, the TerraClimate temperature is much lower than the temperature in the China City Statistical Yearbook for Lhasa and Xining. Meanwhile, Chongqing, Fuzhou and Urumqi exhibit minor differences in the temperature series between the two data sources. To further investigate the difference found in these cities, we undertake a comparison in Fig. 2 by placing the TerraClimate temperature and China City Statistical Yearbook temperature as the x-axis and y-axis respectively in one graph with the 45-degree reference line. The correlation coefficient between the two temperature series for each city is also shown in this graph to examine their alignment (e.g., Harris et al. 2014; Fick and Hijmans 2017). Particularly, a point on the 45-degree line implies that the temperature values from the two data sources are exactly the same in the corresponding month. A point above this reference line indicates the higher China City Statistical Yearbook temperature compared to the Terra-Climate temperature, whilst a point below it means that the Terra-Climate temperature is greater than the China City Statistical Yearbook temperature for the relevant month. As can be seen from this graph, the temperature series for Chongqing, Fuzhou and Urumqi present the same patterns between the two sources with extremely large correlation coefficients over $99.50\%$, noting that all the data points in the corresponding charts are located rather closely around the 45-degree line, which further endorses the very strong linear relationship between the series from the two sources. Moreover, the correlation coefficient between the two temperature series is respectively $98.18\%$ and $99.56\%$ for Lhasa and Xining, which are both high. Although the straight lines formulated by the points are slightly distant from the 45-degree line for these two cities. They are in parallel, indicating that for Lhasa and Xining the temperature series from the two sources generally differ from each other by just a constant. This means that the temperature series from the two sources show the same pattern of movements, and have identical temperature change (i.e., the first difference of the temperature series) which is used in our empirical study. Overall, our validation implies that the constructed temperature data is sufficiently accurate for studying the impact of climate change on the total factor productivity in this paper considering the China City Statistical Yearbook temperature data as a benchmark.

TFP and temperature quantiles

See Figs. 3, 4 and 5.

Fig. 3

The total factor productivity versus the estimated $5\%$ quantile for the land surface mean temperatures

Fig. 4

The total factor productivity versus the estimated $50\%$ quantile for the land surface mean temperature

Fig. 5

The total factor productivity versus the estimated $95\%$ quantile for the land surface mean temperatures

Fig. 6

The cumulative change heat maps over time since 1978 for the cities of Beijing, Tianjin, Shijiazhuang, Taiyuan, Hohhot, Shenyang, Changchun, Harbin, Xian, Lanzhou, Xining, and Yinchuan in China across the estimated temperature quantiles from 0.05 till 0.95 at a 0.05 increment

Fig. 7

The cumulative change heat maps over time since 1978 for the cities of Urumqi, Shanghai, Nanjing, Hangzhou, Hefei, Fuzhou, Nanchang, Jinan, Zhengzhou, Wuhan, Changsha, and Guangzhou in China across the estimated temperature quantiles from 0.05 till 0.95 at a 0.05 increment

Climate change patterns across China

This section briefly discusses what climate change patterns the temperature data yields for different regions across China. In particular, we focus on the specific period from the year 1978 when the Chinese economic reform started till 2018. In particular, The Figs. 6, 7, 8 below demonstrate the heat maps of the estimated cumulative change (specified by Equation (6)) since 1978 over time with respect to different temperature quantiles beginning from 0.05 till 0.95 at a 0.05 increment for each capital city. Loosely speaking, we can consider the quantiles lower than the $25\%$ quantile reflecting the temperatures for the three months of summer in these figures, while the quantiles above the $75\%$ quantile are regarded as measuring the temperatures in the winter. The main observations of the Figs. 6, 7, 8 are summarized in what follows.

• All the cities included in this analysis seem to start showing a persistent warming pattern in common since the 1990s, although the degree to which a certain city becomes warm for each year is both time-varying and individual-specific.

• Compared to the rest of the seasons within a year, the warming trend in winters appears more severe over time, where in the extreme cases the increase in the temperature quantile is estimated to be in the range between 2$^{\circ }$C and 2.5$^{\circ }$C since the year of 1978.

• In general, the region of North, Northeast, and Northwest China (including all the cities shown in Fig. 6 and Urumqi in Fig. 7) have exhibited the most significant temperature increase over the past 40 years, whilst the region of South and Southwest China (with Guangzhou in Fig. 7 and the cities in Fig. 8) are affected the least by the global warming. The severity of the warming effect on East and Central China (with the remaining cities present in Fig. 7) is in the middle compared to the two regions just mentioned.

Fig. 8

The cumulative change heat maps over time since 1978 for the cities of Nanning, Haikou, Chongqing, Chengdu, Guiyang, Kunming, and Lhasa in China across the estimated temperature quantiles from 0.05 till 0.95 at a 0.05 increment

Moran I test results

See Table 11.

Table 11 Moran I test results

Results on panel unit root tests

To understand the degree of integration and stationarity properties of the variables used in our empirical study, three panel data unit root tests are applied: the unit root test proposed by Levin et al. (2002) (LLC), and the Fisher-ADF and Fisher-PP tests following Maddala and Wu (1999). In addition to the above first-generation panel unit root tests, we also employ the Pesaran-CIPS test following Pesaran (2007), which takes into account the cross-section dependence. For these tests, the null hypothesis is the presence of a unit root and the alternative hypothesis is no statistical evidence for a unit root. The results of the panel unit root tests are shown in Table 12 below.

Table 12 Panel unit root test statistics