The Impact of Climate Change on Agricultural Total Factor Productivity: A Cross-Country Panel Data Analysis, 1961–2013

Abstract

The literature has mixed findings regarding the impact of climate change on agricultural production, probably due to different choices of dependent variables. Based on panel data analysis of 140 countries from an extended period, i.e., 1961 to 2013, this study applies the explicit method of using TFP as the dependent variable, but also delivers estimations with the implicit approach of using agricultural output as the dependent variable, so as to be able to directly compare the results with the mainstream trends in the existing literature. We found that using agricultural TFP as the dependent variable generates more consistent and robust results. We also found a strong negative impact of climate change on agricultural productivity. A one-unit downward deviation of precipitation (i.e., drought) and one unit of upward deviation of temperature (i.e., global warming) decrease the TFP by 0.530 and 0.494, respectively. As we have one of the largest sample sizes when it comes to studying the effect of climate change, we carried out an exploration by dividing the sample into different categories of developed vs. developing countries and cold vs. warm climates, as well as by splitting the time series into two distinct periods. We found that climate change has had a heterogeneous impact on countries with different development levels, with negative impacts on developing countries and positive effects on developed countries, which leads to the rising concern of the impact of climate change on inequality. This heterogeneity and the uneven impact found in this study strongly suggest the need for countries to work together to coordinate and mitigate these adverse effects rather than to adapt to the situation by focusing on the input variations, as the latter will increase the inequality effect of climate change and add to an already unstable global environment.

1. Introduction

The world climate shows a clear and present threat because of increased rates of global warming. The IPCC Fifth Assessment Report stated that between 1880 and 2012, the global mean surface temperature increased by about 0.85 ℃. It further reported that, although the global average precipitation shows no significant difference, the inter-regional differences are significant, and in extreme cases, they lead to extreme drought and frequent floods in disperse regions [1]. The combined value of damage increases quadratically in temperature, costing 1.2% of GDP per +1 °C on average [2]. The agriculture sector has become the focal point, both globally and regionally, regarding the adverse effect of climate change. The consensus is that such climate change will have a significant impact on agricultural production and its related processes.

However, empirical studies have lagged behind the rapid change in the climate and its potential calamity. Their findings are often contradictory, both in terms of regional differences and impact mechanisms. To compound the matter, there has been debate regarding what the dependent variable should be, among other issues. This study attempts to take a broader perspective to this important question by rigorously selecting the appropriate dependent variable and developing sound theoretical methods in the hope of extending our understanding of the effect of climate change and reconciling some of the differences in the existing literature.

Based on the choice of the dependent variable, we contribute to the research in the field in several ways. First, we compare the effects of weather shocks on agricultural TFP and output, which establishes the robustness of the empirical results using TFP as the dependent variable. Second, this study, to our knowledge, is based on one of the most comprehensive datasets (encompassing 140 nations), with a long time series (1961–2013), similar to some recent papers such as that by Ortiz-Bobea et al. [3] and Lesk et al. [4], which allows us to compare various sub-samples to present a wider perspective. Third, we explore regional differences by performing a robustness analysis between different sub-samples. Therefore, the current design and testing the TFP data make a positive contribution in this direction.

The remainder of the paper is structured as follows. Section 2 surveys the literature, followed by the theoretical framework and empirical specification in Section 3. Section 4 presents the data and variable descriptions, with the empirical results depicted in Section 5. A summary and conclusion section concludes this study.

2. Literature Review

The adverse impact of climate change on agriculture threatens the sustainability of the sector. As policymakers grapple with ways to alleviate the adverse effects of climate change, academics try to understand the true nature of such effects on agriculture and to estimate the magnitude of the effect. A sizable empirical literature has accumulated on the relationship between climate change and agriculture. Academics have used a variety of methods towards this end, so that we need to carry out a comprehensive literature review to explore the research gaps [5]. Some research has relied on experiments; other studies have relied on crop growth simulation; and the majority rely on econometric methods [6]. Among the three designs, the third approach is most common for obvious reasons. Time and space limit the experimental method, and strict conditions limit the crop growth simulation method [7]; therefore, the econometric method has become a heavily relied upon research strategy for estimating the effects of climate change and agricultural output [8].

In principle, econometric methods estimate the impact of climate change on agricultural production by establishing a statistical model in the agricultural production system. Three econometric models have dominated the literature and can be distinguished by the selection of the dependent variable. The Ricardian model uses land rent as the dependent variable (e.g., [9,10,11,12]). The profit model selects, as its name dictates, the profit as the dependent variable [13,14,15]. A third model uses total agricultural output as the dependent variable and has become the model of choice in the analysis of the effects of climate change [16,17,18], as the aggregate agricultural output can be roughly linked to food availability in a country.

The Ricardian approach in Mendelsohn et al. (1994) [9] was a reaction to earlier studies based on biophysical approaches that did not take into account farmer adaptation to climate change properly [3]. This method reflects the broad and long-term impacts of climate change, thus avoiding some disadvantage of the production function analysis. However, the Ricardian model suffers from three shortcomings. First, the Ricardian model does not account for adjustments costs [11]. Second, it has a temporal dilemma. In the long term, land value is directly correlated with agricultural output, but in the short-term, it is difficult to rationalize that land value reflects the impact of climate change. Third, the land value used in the Ricardian model is only a partial productivity indicator and cannot reflect the total factor productivity. As a result, most of the Ricardian model studies show slight positive effects. [9] argued that due to the aforementioned shortcoming, the Ricardian model underestimates the negative effects of climate change on agricultural output.

In response to these issues, researchers have turned to using agricultural profit per unit of land as the dependent variable [13,14,15]. This method has two advantages over the Ricardian model. First, it reduces the influence of other omitted variables that plagued the Ricardian model. Second, it captures the short-term (e.g., one year) effect of climate change on agriculture [10]. Due to these fundamental merits, the agricultural profit per unit of land has it a large presence in the literature, but its estimation process is complicated and suffers from data errors [19]. Further, the approach could not account for inventories properly so that the profits were precisely calculated [3]. Moreover, like the Ricardian model, the land-based measure of profit is also only a partial reflection of productivity.

The downside of the previous two methods has led to the choice of the total output as a dependent variable, but this has its own methodological problems, and it raises theoretical issues in response to thew control variables used in the model [20]. To avoid these biases, some researchers have turned to using ‘crop yield’ as the dependent variable (e.g., [21,22,23,24]). As the yield analysis is based on the specific association between crop output and land input, it measures the partial factor productivity of the agriculture sector, differing from total factor productivity. Despite the improvements, the yield method has also suffered from a number of disadvantages that weaken its general appeal. First and foremost, the yield analysis offers an inadequate explanation of the impact of climate change on agriculture because, contrary to the assumption of a single-crop yield, farmers can adjust crop varieties to at least partially offset the negative effects of climate change in the real world [25]. The second disadvantage is a direct consequence of the first problem: the yield analysis exhibits a high level of sensitivity to crop adaptability and land-use change [26]. In other words, whenever farmers change crop variety or land usage, the yield curve shows discrete and disproportionate deviations. Finally, due to the previous two disadvantages, the coefficients of this type of analysis tend to have a downward bias and show a smaller negative effect due to climate change [9].

Due to the limitations of the four methods and different choices of research regions, the findings are often contradictory and lack consensus. Research suggests that poorer countries or regions are more negatively affected by a changing climate, which may lead to the problem of increased global inequality [27,28]. Ref. [29] found that a 1 °C increase in temperature decreases macro TFP growth rates by 1.1–1.8 percentage points for poor countries, whereas the impact is insignificant in wealthy countries. As for the agricultural sector, [18] used cross-country panel data on total agricultural output and found that the impact of climate change on sub-Saharan African countries was more pronounced than other developing countries. Ref. [13], using county-level panel data on farm profits, showed that climate change had a minimal but positive impact in the United States. These conflicting studies suggest that the impact of climate change on agriculture may be partially due to the selection of the dependent variables.

To address this issue, we argue that any reasonable dependent variables need to meet two conditions to answer the research question of how climate change affects agriculture productivity. First, the dependent variables need to consider not just agricultural output but also the input, as the variation in input captures the adaption of the farmers [30]. Second, the dependent variables need to have a minimal data requirement, so as to allow the researcher to compile consistent cross-country data with less risk of serious errors and thus be able to generate accurate estimates across nations. Only then can we address the primary question of whether climate change has a negative impact on the agriculture sector. Conditioned on the findings related to the primary research question, we also need to answer a second question: Does the impact of climate on agriculture differ across countries and periods? These questions assess the real effect of the impact of climate change on agriculture and will serve as the foundation for developing viable policies to alleviate the negativity of global warming.

In this study, we chose to use TFP as the dependent variable, as the estimation of the TFP includes both agricultural inputs and outputs, and as such, it reflects a more comprehensive impact of climate change on agriculture (as discussed in the criteria in selecting the dependent variable). TFP data can even account for changes in inventories [31]. By doing this, we can also avoid the over-controlling problem, which is caused by the introduction of variables of agricultural inputs into the explanatory variables and leads to a restriction of necessary impact pathways [20].

Though most of the literature uses land rents, profits, or total agricultural outputs as the measured dependent variables, there is a rising trend in exploring the impact of climate change on TFP (e.g., [3,30,32,33,34,35,36]) as shown in

Table 1. Representative papers with various dependent variables.

Dependent Variable	Representative Papers
Land rent	Mendelsohn et al., 1994 [9]; Kelly et al., 2005 [10]; Schlenker et al., 2005, 2006 [11,12]
Profit	Deschenes and Greenstone, 2007 [13]; Seo and Mendelsohn, 2008 [14]; Wang et al., 2009 [15]
Total agricultural output	Nelson et al., 2005 [16]; Ackerman and Munitz, 2012 [17]; Barrios et al., 2008 [18]
Yield	Feng et al., 2010 [21]; D’Agostino and Schlenker, 2016 [22]; Ortiz-Bobea and Tack, 2018 [23]; Shew et al., 2020 [24]
TFP	Liang et al., 2017 [32]; Njuki et al., 2018 [33]; Ortiz-Bobea et al., 2018 [34]; Chambers and Pieralli, 2020 [35]; Plastina et al., 2021 [36]; Chen and Gong, 2021 [30]; Ortiz-Bobea et al., 2021 [3]

. Ref. [32] use TFP as the dependent variable to examine the impact on US agriculture. Ref. [30] present some empirical results about the impact of climate change on agricultural TFP and inputs and output in China. A very good recent example is [3], who use a similar dataset as our paper and show that climate change has reduced agricultural TFP by 21% since 1961.

The choice of agricultural TFP as the dependent variable has several advantages, as outlined above. However, we also estimate the effects of climate change on total output, and by doing so, we contribute to the literature in several ways. First, by comparing the effects of climate change on both agricultural TFP and output, we create a bridge between the two main strands of literature and can reconcile and establish the inter- and intra-robustness of the empirical results. Second, this study comprehensively examines the relationship between climate change and TFP by using a representative sample of global economies. In doing so, we claim a third important contribution, as we can explore regional differences by performing s robustness analysis between different sub-samples. Only by doing this can we explore the heterogeneous effects of climate change, including North–South, warm–cool areas, etc.

In summary, research on the link between climate change and agriculture has demonstrated advancements in recent years. However, the literature lacks consensus on some critical issues. First, there is disagreement on what dependent variable truly represents the impact of climate change on agriculture. As discussed earlier, each researcher has reasons for their choice of dependent variable, and this choice may have a significant impact on the results.

In addition, despite the seemingly agreed-upon finding that climate change seems to have a more negative impact on developing countries and a less negative or probably positive impact on developed countries, the literature has failed to answer why climate change hurts different countries differently. As such, we believe the choice of TFP as the dependent variable offers a more comprehensive measurement that eliminates the disadvantages of the other choices and can better establish a clear and strong link between climate change and agricultural output. We now turn to our proposed analytical model and empirical specification.

3. Theoretical Framework and Empirical Specifications

The TFP-based approach differs from output- or profit-based approaches in several ways. Using output as the dependent variable integrates two sources of effects of climate change: the productivity effect and the input effect. In our TFP-based model, we focus on the productivity effect. The existing literature tries to capture productivity effects while controlling the input factors and thus obtains the productivity impact of climate change implicitly. For instance, we introduce a production function that includes climate change $q(A(w),X(w))$, where $w$ stands for the weather variables, $A(w)$ stands for total factor productivity, and $X(w)$ stands for factor input. This inclusive approach leads to the total derivative of the production function in Equation (1):

$$\frac{dq(A(w),X(w))}{dw}=\underset{\mathrm{Productivity} \mathrm{effect}}{\underbrace{\frac{\partial q}{\partial A}·\frac{\partial A}{dw}}}+\underset{\mathrm{Factor} \mathrm{effect}}{\underbrace{\frac{\partial q}{\partial X}·\frac{\partial X}{dw}}}$$

(1)

Equation (1) shows that climate change affects agricultural output through the total factor productivity ($\frac{\partial q}{\partial A}·\frac{\partial A}{dw}$) and factor inputs ($\frac{\partial q}{\partial X}·\frac{\partial X}{dw}$). The “true” impact of climate change on agriculture should be the effect of TFP, but by using agricultural output as the measured dependent variable, the impact of climate change is distorted by the mitigating effects of changing inputs. In Equation (1), the model measures the overall impact of climate change on the agricultural output and necessitates the inclusion of both types of predictors in the regression to assess the impact of climate change. By controlling for the impact of factor inputs, the model will indirectly estimate the productivity impact of climate change ($\frac{\partial q}{\partial A}·\frac{\partial A}{dw}$). In an ideal world, the implicit method can capture the impact of climate change on productivity just as well as the explicit method of using TFP as the dependent variable. However, in practice, that is not the case. First, both TFP and output measures may have measurement errors. However, TFP poses less of a challenge econometrically, as the measurement error is linked with the dependent variable [37]. In comparison, the measurement issue is more serious when the output is used as the dependent variable. This is due to the necessity of controlling for input variables in the regression to capture their impact on the output. By doing so, the measurement error of inputs interacts with the variables of interest and creates an endogeneity problem [38]. Finally, if there are any omitted control factors, their impacts will also be captured by the productivity effect. Hence, the output approach may be biased by including the input factors as the control variables.

As mentioned earlier, there are some similarities between the TFP-based method and the profit-based method—both consider agricultural inputs and outputs to account for the adaptive behaviors of farmers—and there are significant differences. The profit-based method mainly examines the impact on unit land profit, which is only a partial factor of productivity, with the land as its denominator. Compared to the TFP method, this partial factor productivity approach overlooks other factors such as capital and labor, which constrains its feasibility analytically, as well as its usefulness in forecasting and policy formation.

Taking these issues into consideration, our econometric specifications include three models, one Equation (2) on TFP to show the explicit impact on TFP, and two Equations (3) and (4) on output to show the implicit impact on TFP and to be used as a comparison.

The effects of climate change on TFP can be captured by using the two-way fixed-effects model in Equation (2), where the subscripts k and t denote the country and year, respectively. $c_k$ and $y_t$ are the country and year dummies, and ${\epsilon }_{k,t}$ is the error term. $TF{P}_{k,t}$ is the Total Factor Productivity, which is the dependent variable and the main focus of this study. $w_{k,t}$ is a vector of climate variables, including linear, squared, and interactive terms of precipitation and temperature, i.e., pre, tmp, pre², tmp², and pre × tmp, where the variables pre and tmp are defined in

Table 2. Descriptive statistics.

Variable	Abb.	Definition (Unit)	Mean	Standard Deviation	Minimum	Maximum
Total factor productivity	TFP	Agricultural TFP index (base year 2005 = 100)	85.765	21.958	22.639	179.155
Agricultural output	value	Gross value added (USD billion)	10.532	33.118	0.020	587.675
Agricultural capital	cap	Total farm machinery in “40-CV tractor equivalents” (thousand)	202.617	675.677	0.001	11,239.590
Livestock input	live	Total livestock capital on farms in “cattle equivalents” (USD billion)	12.225	33.292	0.024	304.104
Labor	labor	Number of economically active adults in agriculture (millions)	6.808	31.823	0.002	390.980
Land	land	Total agricultural land in hectares of “rainfed cropland equivalents.” (million hectares)	14.493	40.789	0.011	325.350
Fertilizer	fert	(thousand tons)	832.520	3179.002	0.015	49,887.500
Precipitation deviation	pre	Deviation from the mean (standard deviation)	0.036	1.243	−7.097	7.249
Temperature deviation	tmp	Deviation from the mean (standard deviation)	0.931	1.588	−6.294	12.032
Precipitation	prev	(thousand mm)	1.111	0.773	0.022	3.687
Temperature	tmpv	(°C)	18.071	8.391	−7.400	29.800

Note: Number of countries and observations are 140 and 7153, respectively.

and represent precipitation and temperature, respectively. Crops normally have an optimal temperature and precipitation span for their best yield, and this may not be a simple linear relationship; hence, adding a quadratic term of the climate variable to approximate this nonlinear effect of climatic variables in the estimation is relevant [18,25,27]. Fixed-effects models can largely avoid the problems of heterogeneous samples by presenting the within-country estimation results [37]. As climate variables can be regarded as exogenous, there is no need to include other control variables [39]. (The methodology in Diffenbaugh and Burke (2019a) [27] was questioned by Rosen (2019) [40], who notes: “Importantly, this regression equation for changes in economic growth is a function only of temperature, precipitation, and time. Thus, this regression equation has no economically based independent variables that would explain most of the time dependence of economic growth over 49 years. That is the key theoretical flaw in the research, because for any regression equation relevant to explaining the changes to gross domestic product (GDP) growth over time all theoretically important economic variables that explain economic growth need to be included”. However, Diffenbaugh and Burke (2019b) [39] replied to this challenge vividly: “Rosen (2019) argues that because our statistical model relating temperature to economic growth does not explain all of the variation in economic growth over the last half-century it cannot uncover the relationship between temperature and growth. This is a little like saying a medicine cannot possibly be effective at reducing headaches if it does not also address all other ailments. Just as there are many determinants of overall human health, there are many causes of variation in economic output. The point of our analysis is not to explain all of these sources of variation, but rather to plausibly isolate the role of temperature from these other sources.” They further argued that their methodology is the “standard econometric approach for causal inference in panel data for at least 2 decades (see Wooldridge (2010) for the textbook treatment)”.

$$TF{P}_{k,t}={\beta }_1{w}_{k,t}+c_k+y_t+{\epsilon }_{k,t}$$

(2)

Equations (3) and (4) show the possible impact mechanism of climate change on output, where the dependent variable $Outpu{t}_{k,t}$ represents agricultural output. $X_{k,t}$ is a vector of five types of inputs (agricultural capital, livestock input, labor, land, and fertilizer).

$$Outpu{t}_{k,t}={\beta }_2{w}_{k,t}+c_k+y_t+{\epsilon }_{k,t}$$

(3)

$$Outpu{t}_{k,t}={\beta }_3{w}_{k,t}+{\beta }_4{X}_{k,t}+c_k+y_t+{\epsilon }_{k,t}$$

(4)

In Equation (2), ${\beta }_1$ represents the impact on TFP, i.e., $\frac{dA}{dw}$ in Equation (1). Equations (3) and (4) differ in whether inputs (X_k,t) are included as control variables. If the inputs are not included, as shown in Equation (3), ${\beta }_2$ will represent the total impact of climate change on agricultural output, i.e., $\frac{dq}{dw}$. If the inputs are included, as shown in Equation (4), ${\beta }_3$ will represent the impact via the total factor productivity, i.e., $\frac{\partial q}{\partial A}·\frac{\partial A}{dw}$, and the impact via inputs are excluded. Hence, we can estimate the impact on TFP in Equation (2) and show the impact mechanism on TFP by comparing coefficients in Equations (3) and (4).

In order for the linear coefficients to reflect the marginal effects at the sample mean, the climate variables are centered to the mean by the country in the regression [37]. To make the coefficients easy to explain and to reduce the excessive influence of abnormal observations, both agricultural output and input variables are logarithmic in the equations.

In line with the literature, this article compares the results for developing countries with developed countries. To do so, we run the same sets of regressions for different types of countries. In addition, the regression results for different temperature levels and time periods are compared.

4. Data and Variable Description

The data used in this study came from the databases of the USDA (United States Department of Agriculture) (https://www.ers.usda.gov/data-products/international-agricultural-productivity/documentation-and-methods/ (accessed on 4 December 2022)) and the Climatic Research Unit (CRU) (https://crudata.uea.ac.uk/cru/data/hrg/cru_ts_4.04/crucy.2004161557.v4.04/countries/ (accessed on 4 December 2022)). The latter, arguably, maintains the most comprehensive database on climate change. Upon merging the two datasets and excluding the outliers, such as Brunei Darussalam and Singapore, which have a limited scale of agriculture, we obtained panel data of 140 countries with a long period from 1961 to 2013, resulting in a total of 7153 observations. As of now, this study is one of the most comprehensive in terms of the number of countries covered and the size of the data.

Dependent variables: The primary dependent variable is the TFP index (base year 2005 = 100, by country) from the USDA dataset, and it is based on the growth accounting method (for details, refer to USDA, 2018 [41]). Even though the TFP index is normalized to the level in 2005 for each country, it does not resolve the comparability problem across countries; our econometrical specification with fixed effects is aimed at dealing with this issue [3]. The secondary dependent variable is the agricultural output in billion constants of 2004–2006 international dollars, and it is drawn from the agricultural gross output value in the USDA dataset.

Explanatory variables: Precipitation and temperature are the main exogenous variables measuring the change in climate, and got from the “High-resolution Gridded Datasets (and Derived Products)” gathered by the Climatic Research Unit (CRU CY v. 3.22).

In the climate data, the observed countries belong to various grids, at 0.5 longitudinal and 0.5 latitudinal spacing. Annual country averages are calculated using area-weighted means [42]. To reflect climate change and to be comparable across nations, we drew on the popular approach (e.g., [18,43]) to treat the time series climate data for each country, which has also been supported by Merel and Gammans (2021) [25]. This method measures the deviation from the average between 1901 and 1960 (the base period) and is derived by dividing the recorded value by the standard deviation of the base period in said country. (This is important, as a sizeable increase in precipitation in an arid area may be moderate in a monsoon region, whereas for the arid area in which precipitation occurs, it could have a dramatic effect.) This relative measure of climate change, rather than permanent cross-country climate differences in levels (the absolute change of climate), is more meaningful, as a one-degree change in temperature may have quite different implications depending on the latitudinal location at which it occurs [44].

Control variables: The control variables include the main agricultural inputs available in the USDA dataset. The agricultural capital is mainly comprised of machinery, represented by the total metric horsepower of the major farm equipment in use. Likewise, the livestock input is the aggregate value of animals. To approximate livestock capital, “cattle equivalents” are used to measure total inventories of animals. This variable measures units of the constant 2004–2006 international dollar. Agricultural labor is the total number of adults that are economically active in agriculture. The agricultural land includes the area in permanent crops, annual crops, and permanent pasture. Cropland is further divided into rain-fed and irrigated areas. The areas of various types of land are aggregated into a quality-adjusted measure to account for relative land productivity. Similarly, fertilizer as a component of the agricultural input is the amount of major inorganic nutrients, measured as metric tons of N, P₂O₅, and K₂O nutrients.

When output is used as the dependent variable, the measurement errors in these input variables may lead to an endogeneity problem when these variables are used as controls. In comparison, the measurement errors in these input variables have minimal impact on the TFP measure, and when TFP is used as the dependent variable, these measurement errors will be less problematic. The variable definitions and descriptive statistics are shown in Table 2. It should be noted that the average livestock capital was USD 12.225 billion, which is greater than the gross value added to agriculture (USD 10.532 billion). This is indicative of the aforementioned serious measurement error issue when agricultural output or profit is used as the dependent variable.

Compared to the average of the base period (1901 to 1960), global precipitation increased slightly (+0.036 standard deviations), and the increase in temperature was somewhat more significant (+0.931 standard deviations). This, however, hides the much more pronounced inter-regional differences in the impact of climate change. The average precipitation of high-latitude countries increased significantly, e.g., Canada and Russia increased by 1.933 and 1.116 standard deviations, respectively. In contrast, the average precipitation of most African countries decreased, e.g., Guinea and Senegal decreased by 1.257 and 1.151 standard deviations, respectively. The country with the largest annual precipitation deviation was Mauritius (in the year of 1982), with a deviation of +7.249 standard deviations. During that year, the precipitation in Mauritius was 3.333 thousand millimeters, whereas the country’s historical average rainfall was 2.057 thousand millimeters, with a standard deviation of 0.176 thousand millimeters. The relatively driest country was Saudi Arabia in 1970, with a deviation of −7.097 standard deviations.

We see comparable regional differences in terms of temperature. The average temperature of Sub-Saharan African countries (such as Uganda and Rwanda) rose by a more significant margin. The country with the largest annual temperature deviation was Uganda in 2002, with +12.032 standard deviations. During that year, the average temperature was 24.9 degrees Celsius, whereas the country’s historical average temperature is 22.851 degrees, with a standard deviation of 0.170 degrees. Within our sample, only four countries (i.e., Madagascar, Ecuador, Chile, and Paraguay) exhibited a slightly downward deviation in temperature during this period.

The global average TFPs were 74.986 and 113.165 in 1961 and 2013, respectively, showing a per annum increase of 0.734.

Table 3. Top and bottom 10 countries by TFP for 2006–2013.

Rank	Country	Average TFP	Average Temperature for 2006–2013
1	Suriname	142.863	26.000
2	Malawi	138.410	22.525
3	Guyana	134.297	26.100
4	Angola	129.575	21.738
5	Bulgaria	127.386	11.750
6	Iceland	127.130	2.838
7	Sierra Leone	123.715	26.538
8	Cambodia	122.744	27.438
9	Morocco	122.702	17.975
10	Bosnia and Herzegovina	122.216	11.238
/
131	Kyrgyzstan	93.533	2.650
132	Liberia	93.032	25.663
133	Kenya	92.590	25.525
134	Fiji	90.812	24.513
135	Afghanistan	90.471	13.863
136	Bhutan	88.897	8.575
137	Congo	87.335	24.825
138	Yemen	87.060	24.475
139	Belize	84.675	25.988
140	Uganda	83.544	24.138

lists the top ten and bottom ten countries in terms of TFP. As the TFP indices are normalized at 100 in the base year of 2005 for all the countries, the average TFP indices essentially reflect TFP growth between 2006 and 2013. With the exception of Iceland, all the countries on either the high or low end of TFP changes were all developing countries, clearly indicating that climate change has more pronounced effects on developing countries, with the effects being diverse and polarized. (During this period, Iceland experienced a 1.63 SD increase in precipitation and a 1.26SD increase in temperature. Given its high latitude, it should not be surprising that this weather shock from global climate change was favorable in raising the productivity of agriculture.) Certain developing countries (e.g., Suriname) experienced rapid TFP growth, whereas others (e.g., Uganda, Belize, and Yemen) suffered significant declines. One reason for the rapid decline in productivity in some developing countries is the deterioration of climatic conditions. During this period, the temperature in Uganda and Yemen rose sharply, deviating +7.553 and +3.923 SD from the average temperature, respectively. In addition, precipitation in Uganda increased sharply and deviated +1.565 SD from the average precipitation. We show the average temperature in Table 3 to see if there is an obvious correlation between temperature and TFP growth, but no such relationship seems to exist.

In a separate set of computations, we estimated the average annual change from the trend of global climate variables and average TFPs in agriculture, so as to make an observational connection between changes in the time trend of climate change and TFP. A slightly upward deviation was observed for precipitation, as was a more measurable upward trend in temperature, clearly indicating global warming. We identified a productivity decrease associated with the increase in precipitation or temperature deviations; this is a sneak peek of our formal empirical analysis.

5. Empirical Analysis

5.1. Impact of Climate Change on TFP

Table 4. Impact of climate change on agriculture.

Dependent Variable	TFP	lnvalue
Independent Variable	(1)	(2)	(3)
pre	0.530 ***	0.009 ***	0.009 ***
	(0.140)	(0.003)	(0.002)
pre2	−0.064	−0.007 ***	−0.002 **
	(0.066)	(0.001)	(0.001)
tmp	−0.494 ***	0.008 ***	−0.006 ***
	(0.142)	(0.003)	(0.002)
tmp2	−0.038	0.002 **	−0.000
	(0.042)	(0.001)	(0.000)
pre × tmp	0.116	0.003 **	0.002 *
	(0.090)	(0.002)	(0.001)
lncap			0.049 ***
			(0.003)
lnlive			0.366 ***
			(0.008)
lnlabor			0.077 ***
			(0.007)
lnland			0.259 ***
			(0.013)
lnfert			0.100 ***
			(0.003)
Year dummy	controlled	controlled	controlled
Constant	109.368 ***	15.108 ***	7.844 ***
	(1.191)	(0.022)	(0.091)
Observations	7153	7153	7153
Countries	140	140	140
R2	0.460	0.605	0.850

Note: Standard errors in parentheses; *, **, *** denote statistical significance at the 10%, 5%, and 1% level, respectively.

shows the fixed-effect estimates of TFP and output under different specifications (estimations implemented with Stata 15.0). Models 1 and 3 (Columns 1 and 3) correspond to Equations (2) and (4), respectively, and show the estimates of the impact of climate change on productivity via the explicit (Equation (2)) and implicit (Equation (4)) approaches. Model 2 (Column 2) shows the overall impact of climate change on output through both input factors and productivity channels.

The coefficients of climate variables in Models 1 and 3 exhibit the same sign, which shows that the impact of climate variables in the two models is consistent. Model 1 identifies the direct impact of climate variables on productivity, and Model 3 identifies the impact of climate variables on agricultural output through the implicit approach.

According to Model 1, both precipitation and temperature have a significant linear impact on TFP. The results show that at the sample mean (the independent variable normalized to zero by country), one unit of downward deviation of precipitation (in the case of drought) and one unit of upward deviation of temperature (in the case of warming) decrease the TFP by 0.530 and 0.494, respectively. Even though the overall impact of temperature on the output (shown in the second column) is positive, this output increase is mainly due to increases in factor inputs. As shown in

Table A2. Climate change impacts on agricultural inputs across countries.

	Independent\Dependent Variable	lncap	lnlive	lnlabor	lnland	lnfert
All	pre	0.000	−0.000	−0.005	0.004 **	−0.006
		(0.006)	(0.003)	(0.003)	(0.002)	(0.008)
	pre2	−0.001	−0.006 ***	−0.003 *	−0.004 ***	−0.014 ***
		(0.003)	(0.001)	(0.002)	(0.001)	(0.004)
	tmp	0.023 ***	0.021 ***	0.029 ***	0.011 ***	−0.000
		(0.006)	(0.003)	(0.003)	(0.002)	(0.008)
	tmp2	0.002	0.002 ***	0.002 **	0.002 ***	0.004
		(0.002)	(0.001)	(0.001)	(0.001)	(0.002)
	pre × tmp	0.007 *	−0.000	0.008 ***	0.003 **	−0.000
		(0.004)	(0.002)	(0.002)	(0.001)	(0.005)
Developed countries	pre	−0.038 ***	−0.008 *	0.004	0.001	−0.013 *
		(0.015)	(0.005)	(0.006)	(0.002)	(0.007)
	pre2	0.021 **	0.003	−0.002	−0.004 ***	−0.002
		(0.009)	(0.003)	(0.004)	(0.002)	(0.005)
	tmp	−0.024	0.016 ***	−0.019 ***	0.006 *	−0.004
		(0.018)	(0.006)	(0.007)	(0.003)	(0.009)
	tmp2	0.006	0.002	0.003	−0.000	0.004
		(0.009)	(0.003)	(0.004)	(0.002)	(0.005)
	pre × tmp	0.014	−0.000	−0.005	−0.001	−0.011 *
		(0.012)	(0.004)	(0.005)	(0.002)	(0.006)
Developing countries	pre	0.011	0.004	0.000	0.006 ***	0.004
		(0.007)	(0.003)	(0.003)	(0.002)	(0.009)
	pre2	−0.003	−0.007 ***	−0.001	−0.003 ***	−0.015 ***
		(0.003)	(0.001)	(0.001)	(0.001)	(0.004)
	tmp	0.026 ***	0.018 ***	0.023 ***	0.008 ***	−0.014
		(0.007)	(0.003)	(0.003)	(0.002)	(0.009)
	tmp2	0.002	0.002 **	0.002 *	0.002 ***	0.004 *
		(0.002)	(0.001)	(0.001)	(0.001)	(0.003)
	pre × tmp	0.006	−0.001	0.008 ***	0.003 **	−0.002
		(0.004)	(0.002)	(0.002)	(0.001)	(0.005)

Note: Standard errors in parentheses. *, ** and *** represent the 10%, 5% and 1% significance level respectively. Year dummy variables and constant items are controlled. Table 4 shows the respective numbers of observations and countries for diverse groups.

, with the exception of fertilizers, all input factors are increased in response to a larger temperature deviation. The combination of less of an output increase and a significant increase in input factor usage implies a decrease in TFP. Taking into account the annual growth of TFP 0.734, these negative impacts are non-trivial. It is worth emphasizing that the negative coefficient of the temperature variable does not mean the temperature is too high and is damaging global crop production; it means that the deviation from the average local temperature is harmful to agricultural production.

We now turn our attention to assessing the impact of input factors as control variables. This is central to our argument that TFP is the appropriate dependent variable, and this can be seen vividly as we compare Models 2 and 3. To reiterate, both Models 2 and 3 use agricultural output as the dependent variable, and the difference is whether input factors are explicitly controlled for in the regression. As such, the coefficients in Model 2 are really overall effects, as Model 2 is designed to capture the combined effects of the impacts of climate change via productivity and via factor inputs (which, in turn, affects agricultural output). Refer to Equation (1) for a more mathematical and explicit representation of the discussion above.

As can be clearly seen, the impact of climate change on output is significantly different between Models 2 and 3. As we have emphasized, the impact of climate change on output in Model 2 is an overall effect, and hence it can be misleading if one takes it at face value and interprets it as the effect of climate change on agriculture productivity. Take temperature, for example (Model 2). Even if one believes a rise in temperature “may” increase agriculture output, the positive and significant quadratic term is hard to comprehend. If one puts more thought into the matter, even the linear term of the temperature variable being positive is a bit troubling, especially if one considers that the coefficient of the same variable in Model 3 is negative and just as significant. In conjunction with the fact that all the coefficients of the input factors are positive in Model 3, this is strong evidence for the fact that excluding them in the regression confounds the climate change effects in Model 2, which includes both the productivity effect and the factor input effect, with the latter as a dominant effect. This is further reinforced by the fact that most of the climate change coefficients show significant differences between Models 2 and 3, further suggesting that climate change affects the output through input factor adjustment. Therefore, the coefficients of the estimated climate change variables in Model 3 are superior, as they reflectf the net effect after excluding the impact of these intermediary variables by controlling for the input factors as the mediating variables [45].

From the discussion above, the literature based on Model 3 is more relevant when it comes to gauging the effects of climate change on the agriculture sector, and this has been the main trend of the existing literature (e.g., [18]). However, as we argued when we presented Equation (1), in an ideal world, Model 3 should generate an unbiased estimate of the climate change on agricultural productivity, but there are mitigating circumstances that prevent this from being realized. In addition to the measurement error issue discussed earlier, we would like to focus here on the effect of the “adaptive” measures the industrious farmers may take in changing their input usage. (In addition to changes in factor inputs, farmers can change crop rotation, crop selection, and experiment in hybrid crop species. These are beyond the scope of this study, but they are certainly worthy research areas in dealing with ways to alleviate negative climate change.) These may lessen the negative impact on crop yield, but they will most likely come at a higher cost per unit of output, much like a firm switching to a second-best option in the face of government regulations. As Model 3 measures the effect of climate change on crop production, this cost effect is not accounted for and may lead to the problem of endogeneity due to variable omission.

5.2. Effects of Climate Change on Countries with Different Development Levels

As stated earlier, this study is one of the most comprehensive examinations of climate change in the sense that it includes a large number of countries. This affords us the ability to decompose the sample and make direct comparisons between different groups of interest.

Table 5. Impact of climate change on countries with different development levels.

Dependent Variable	TFP		lnvalue
Countries	Developed	Developing	Developed		Developing
Independent Variables	(4)	(5)	(6)	(7)	(8)	(9)
pre	−0.591 **	0.606 ***	−0.009 **	−0.003	0.014 ***	0.010 ***
	(0.231)	(0.161)	(0.004)	(0.002)	(0.003)	(0.002)
pre2	−0.123	−0.078	−0.001	−0.002	−0.007 ***	−0.002 **
	(0.149)	(0.072)	(0.002)	(0.001)	(0.001)	(0.001)
tmp	0.559 **	−0.437 ***	−0.001	−0.006 **	0.005 *	−0.005 ***
	(0.282)	(0.163)	(0.004)	(0.003)	(0.003)	(0.002)
tmp2	0.033	−0.044	0.004	0.002	0.001	−0.001
	(0.148)	(0.046)	(0.002)	(0.001)	(0.001)	(0.001)
pre × tmp	0.168	0.122	0.001	0.003 *	0.003	0.002
	(0.195)	(0.099)	(0.003)	(0.002)	(0.002)	(0.001)
lncap				0.018 ***		0.050 ***
				(0.004)		(0.004)
lnlive				0.458 ***		0.345 ***
				(0.014)		(0.010)
lnlabor				0.100 ***		0.088 ***
				(0.011)		(0.010)
lnland				0.121 ***		0.275 ***
				(0.025)		(0.015)
lnfert				0.170 ***		0.103 ***
				(0.008)		(0.004)
Year dummies	controlled	controlled	controlled	controlled	controlled	controlled
Constant	61.298 ***	79.754 ***	15.185 ***	7.415 ***	14.277 ***	7.654 ***
	(1.778)	(1.344)	(0.028)	(0.229)	(0.025)	(0.106)
Observations	1452	5701	1452	1452	5701	5701
Countries	30	110	30	30	110	110
R2	0.789	0.396	0.574	0.857	0.649	0.856

Note: Standard errors in parentheses; *, **, *** denote statistical significance at the 10%, 5%, and 1% level, respectively.

divides the sample countries into two groups: developed and developing countries (

Table A1. Countries included in the empirical analysis.

Developing Countries (110)						Developed Countries (30)
Africa (42)		Asia (32)		Latin America (24)	Others (12)
Algeria	Mauritania	Afghanistan	Sri Lanka	Argentina	Albania	Australia	Slovenia
Angola	Mauritius	Armenia	Syria	Belize	Belarus	Austria	Spain
Benin	Morocco	Azerbaijan	Tajikistan	Bolivia	Bosnia & Herzegovina	Canada	Sweden
Botswana	Mozambique	Bangladesh	Thailand	Brazil	Bulgaria	Cyprus	Switzerland
Burkina Faso	Namibia	Bhutan	Turkey	Chile	Croatia	Czech	UK
Burundi	Niger	Cambodia	Uzbekistan	Colombia	Estonia	Denmark	USA
Cameroon	Nigeria	China	Viet Nam	Costa Rica	Hungary	Finland
Central Africa	Rwanda	Fiji	Yemen	Cuba	Latvia	France
Chad	Senegal	India		Dominican Rep.	Lithuania	Germany
Congo	Sierra Leone	Indonesia		Ecuador	Romania	Greece
Egypt	Somalia	Iran		Salvador	Russia	Iceland
Eritrea	South Africa	Iraq		Guatemala	Ukraine	Ireland
Ethiopia	Togo	Jordan		Guyana		Israel
Gabon	Tunisia	Kazakhstan		Honduras		Italy
Gambia	Uganda	Kyrgyzstan		Jamaica		Japan
Ghana	Tanzania	Lebanon		Mexico		Luxembourg
Guinea	Zambia	Malaysia		Nicaragua		Malta
Guinea-Bissau	Zimbabwe	Mongolia		Panama		Netherlands
Kenya		Myanmar		Paraguay		New Zealand
Lesotho		Nepal		Peru		Norway
Liberia		Pakistan		Suriname		Poland
Madagascar		Papua New Guinea		Trinidad and Tobago		Portugal
Malawi		Philippines		Uruguay		Korea
Mali		Saudi Arabia		Venezuela		Slovakia

lists the countries in each group). At the risk of stating the obvious, columns (4) and (5) refer to Model 1 for developed and developing nations, respectively; columns (6) and (8) represent Model 2, and Columns (7) and (9) are that of Model 3 for the two groups, respectively. The direct impact of climate change on TFP is different, but reasonable. The linear coefficient of temperature for developing countries is negative, but positive for developed countries. Astounding as this result may seem, this is still fundamentally consistent with the literature. For example, [46] find that both developing and developed countries were harmed by climate change, with the negative effect on developing countries being more severe.

For developed countries, the overall impacts of temperature on output are positive, though statistically insignificant, as shown in the third and fourth rows of Column (6). This, combined with the decrease in labor input (as shown in Table A2), leads to an increase in TFP in developed countries. For the developing nations, there are at least two reasons why the negative impact is more severe. First, we have the more obvious technology advantage of developed countries. This is obviously the factor that comes to mind almost immediately. What is less apparent is the climate differences. This leads us to the second reason: the temperature difference between the developing and developed countries. The average temperatures in developing and developed countries for the period 1961–2013 are 20.129 and 9.992 degrees Celsius, respectively. As such, a rise in temperature may be more detrimental to agriculture in developing countries than in developed countries.

This global perspective is only possible when the data is comprehensive enough, and that is one of the important contributions of this study. If we delve deeper in terms of inputs, as shown by Figure 1, it becomes even more evident that the increase in agricultural output in the developing nations comes with a hefty cost: a continued steady rise in inputs, be it land, capital, livestock, fertilizer, or even labor. In contrast, though agricultural output has continued to rise also for the developed nations, this is achieved with a sharp decline in almost every input. The amount of labor devoted to agricultural production has decreased (for the developed nations) throughout the entire sample period, and all the other inputs showed a significant decline at various points in time. This, combined with the steady increase in output, unequivocally implies strong technical innovation and/or improved production efficiency in agricultural production. This is perhaps the strongest argument for using TFP, as it not only captures the effect of measurable inputs (as the implicit method does), but also the productivity gains through innovations and efficiency.

5.3. Impact of Climate Change on Countries with Different Temperature Levels

The observation above naturally leads us to the examination of whether different levels of temperatures cause a variation in the effect of climate change on agricultural productivity. Intuitively, global warming raises the potential for agricultural production in colder regions, which has a positive impact on agricultural output, yet empirical evidence for this trend is conspicuously lacking in the literature. Taking advantage of this gap, we divide the countries into two groups to assess the effect of temperature differences. The regression results in

Table 6. Impact of climate change on countries with different temperature levels.

Dependent Variable	TFP		lnvalue
Countries	Cold	Warm	Cold		Warm
Independent Variables	(10)	(11)	(12)	(13)	(14)	(15)
pre	−0.010	0.930 ***	−0.008 **	0.002	0.018 ***	0.012 ***
	(0.185)	(0.208)	(0.004)	(0.002)	(0.003)	(0.002)
pre2	−0.056	−0.145 *	−0.004 *	−0.002 *	−0.005 ***	−0.003 ***
	(0.117)	(0.084)	(0.002)	(0.001)	(0.001)	(0.001)
tmp	−0.520 ***	−0.145	0.009 **	−0.008 ***	−0.005	0.001
	(0.198)	(0.219)	(0.004)	(0.002)	(0.003)	(0.002)
tmp2	−0.063	−0.068	0.005 ***	0.000	0.001	−0.001 *
	(0.061)	(0.060)	(0.001)	(0.001)	(0.001)	(0.001)
pre × tmp	0.066	0.147	0.004	0.001	0.004 *	0.002
	(0.137)	(0.121)	(0.003)	(0.002)	(0.002)	(0.001)
lncap				0.033 ***		0.047 ***
				(0.004)		(0.005)
lnlive				0.506 ***		0.256 ***
				(0.011)		(0.013)
lnlabor				0.116 ***		0.085 ***
				(0.008)		(0.015)
lnland				0.041 **		0.423 ***
				(0.020)		(0.017)
lnfert				0.114 ***		0.090 ***
				(0.004)		(0.005)
Year dummies	controlled	controlled	controlled	controlled	controlled	controlled
Constant	83.237 ***	82.354 ***	15.221 ***	8.140 ***	13.898 ***	7.355 ***
	(1.460)	(1.798)	(0.030)	(0.157)	(0.028)	(0.156)
Observations	3561	3592	3561	3561	3592	3592
Countries	71	69	71	71	69	69
R2	0.600	0.347	0.467	0.828	0.767	0.883

Note: Standard errors in parentheses; *, **, *** denote statistical significance at the 10%, 5%, and 1% level, respectively.

show that the impact of climate change on agricultural TFP is consistent for both groups, but it differs in the significance level. In other words, regardless of the warm or cold groups, climate change shows a negative impact on the countries’ TFPs. However, the negative effect is only significant for cold countries.

The intuition that warming temperature increases the agricultural potential of cold regions is substantiated by the increase in inputs (

Table A3. Climate change impacts on agricultural inputs across countries with different temperature levels.

	Independent\Dependent Variable	lncap	lnlive	lnlabor	lnland	lnfert
Cold countries	pre	−0.014	−0.011 ***	−0.012 **	−0.001	−0.020 *
		(0.009)	(0.004)	(0.005)	(0.002)	(0.010)
	pre2	0.002	−0.002	−0.004	−0.002	−0.001
		(0.006)	(0.002)	(0.003)	(0.001)	(0.007)
	tmp	0.033 ***	0.019 ***	0.028 ***	0.017 ***	0.024 **
		(0.010)	(0.004)	(0.006)	(0.002)	(0.011)
	tmp2	0.004	0.004 ***	0.009 ***	0.004 ***	0.008 **
		(0.003)	(0.001)	(0.002)	(0.001)	(0.003)
	pre × tmp	0.006	0.002	0.006	0.003 *	0.000
		(0.007)	(0.003)	(0.004)	(0.002)	(0.008)
Warm countries	pre	0.009	0.008 ***	0.002	0.007 ***	0.006
		(0.009)	(0.003)	(0.003)	(0.003)	(0.010)
	pre2	0.001	−0.005 ***	0.002	−0.002	−0.011 ***
		(0.003)	(0.001)	(0.001)	(0.001)	(0.004)
	tmp	0.004	0.006 *	0.007 **	−0.006 **	−0.065 ***
		(0.009)	(0.003)	(0.003)	(0.003)	(0.010)
	tmp2	0.001	0.003 ***	0.000	0.002 **	0.005
		(0.002)	(0.001)	(0.001)	(0.001)	(0.003)
	pre × tmp	0.007	−0.002	0.008 ***	0.003 *	−0.001
		(0.005)	(0.002)	(0.002)	(0.002)	(0.006)

Note: Standard errors in parentheses. *, **, and *** represent the 10%, 5%, and 1% significance level, respectively. Year dummy variables and constant items are controlled. Table 4 shows the respective numbers of observations and countries for diverse groups.

). However, the empirical evidence suggests that the increase in input factors used in cold countries exceeds the increase in their output levels. In other words, even though cold countries may potentially increase their output due to global warming, they must invest more input factors, and their adaptive behavior may create even higher costs. Another explanation may come from the higher adaptive ability in hotter regions [44].

Another important difference between the warm and cold countries is the threat of drought for the countries with a higher average temperature, though drought has no significant effect on TFP for cold weather countries. One downward standard deviation (i.e., drought) from the mean precipitation level decreases TFP by 1.220 for warm countries.

5.4. Impact of Climate Change in Different Periods

Given the length of our sample period, we wanted to see whether the effects of climate change have worsened over time. We divided the sample into two periods: 1961–1987 and 1988–2013. As shown in

Table 7. Impact of climate change in different periods.

Dependent Variable	TFP		lnvalue
Periods	1961–1987	1988–2013	1961–1987		1988–2013
Independent Variables	(16)	(17)	(18)	(19)	(20)	(21)
pre	0.733 ***	0.334 *	0.009 ***	0.010 ***	0.008 **	0.005 **
	(0.152)	(0.195)	(0.002)	(0.002)	(0.003)	(0.002)
pre2	−0.102 *	−0.183 *	−0.003 ***	−0.002 ***	−0.004 **	−0.001
	(0.057)	(0.095)	(0.001)	(0.001)	(0.002)	(0.001)
tmp	−0.483 **	−0.543 **	−0.006 **	−0.006 ***	0.008 **	−0.004 *
	(0.195)	(0.211)	(0.003)	(0.002)	(0.004)	(0.002)
tmp2	−0.094	−0.029	−0.001	−0.002 *	0.001	−0.000
	(0.068)	(0.051)	(0.001)	(0.001)	(0.001)	(0.001)
pre × tmp	0.130	0.155	0.001	0.002	0.006 ***	0.002
	(0.100)	(0.122)	(0.002)	(0.001)	(0.002)	(0.001)
lncap				0.032 ***		0.053 ***
				(0.005)		(0.007)
lnlive				0.292 ***		0.333 ***
				(0.014)		(0.012)
lnlabor				−0.072 ***		0.172 ***
				(0.015)		(0.012)
lnland				0.394 ***		0.365 ***
				(0.020)		(0.018)
lnfert				0.052 ***		0.074 ***
				(0.004)		(0.004)
Year dummies	controlled	controlled	controlled	controlled	controlled	controlled
Constant	77.280 ***	109.772 ***	14.517 ***	8.790 ***	15.085 ***	6.847 ***
	(0.802)	(0.998)	(0.012)	(0.172)	(0.017)	(0.130)
Observations	3563	3590	3563	3563	3590	3590
Countries	132	140	132	132	140	140
R2	0.041 †	0.499	0.633	0.749	0.436	0.779

Note: Standard errors in parentheses; *,**,*** denote statistical significance at the 10%, 5%, and 1% level, respectively. † The low value of R² shows the poor overall explanatory power of the model in period 1961–1987. The reason lies in the less significant trend of TFP in that period compared with the period 1988–2013, which leads to lower explanatory power of the year dummies. However, the overall explanatory power cannot endanger the explanatory power of the climate variables, as the relevant coefficients are significant.

, both periods have showed negative impacts of climate change. However, counterintuitively, the negative impact was slightly more pronounced in the earlier period, if one takes into account the negative quadratic coefficients of the temperature variable in Columns 16 and 17. The implicit approach shows a similar pattern (Columns 19 and 21). Given the news reports of the growing severity of global warming, one would expect just the opposite. However, if we take into account the adaptive measures that the agricultural sector has undertaken, this result that seems surprising may not in fact be.

6. Summary and Conclusions

This paper tries to fill three interrelated gaps in the literature on impact of climate change on agriculture. The first stems from the choice of the dependent variable: different preferences have led to contradicting results in the literature in assessing the impact of climate change on the agricultural sector. The second gap comes from the lack of analysis of impact mechanisms, as most analysts have misunderstood the role of factor inputs in their research designs. The third gap is a lack of comprehensive coverage that requires the analysis of subsamples of countries, and the extant literature has overlooked this aspect. To address these three interrelated issues related to the effects of climate change on the agriculture sector, we took a comprehensive approach to the study design and implement rigorous analysis.

With panel data of 140 countries, covering a long period (1961–2013), we estimated the effects of climate change on agricultural TFP (the explicit method) and agricultural output in the implicit method. We compared the effects of climate change on agriculture through the above two methods. The results show that at the sample mean (the independent variable normalized to zero by country), one unit of downward deviation of precipitation (in the case of drought) and one unit of upward deviation of temperature (in the case of warming) decrease the TFP by 0.530 and 0.494 respectively. This finding remains robust when we examine countries with different development stages and temperature levels.

Our findings reveal a robust negative effect on agricultural productivity, especially for developing countries, and this adverse effect persists over time. From a global perspective, this raises another layer of concern for climate change, as it may—indeed, it will—increase North–South inequality and may even pose the possibility of famine in less developed nations.

Given that climate change negatively impacts developing countries, but it may benefit developed nations, combined with the fact that both developed and developing nations have shown a significant growth in agricultural output (especially developing nations), another issue is raised. This growth in output is achieved in very different ways. For developing nations, the growth is brought about by significantly increasing all factor inputs, but this is not the case for developed nations, as all their factor inputs have shown a decreasing trend, suggesting that the increased output is achieved via technical innovation and production efficiency. This is perhaps the strongest case in favor of the explicit TFP method we are advocating compared to the mainstream approaches that dominate the literature.

Even though climate change is likely to improve the agricultural production potential of cold countries, the adaptive approach to climate change in these countries will require new investment, resulting in a lower level of TFP. The ultimate result is a robust negative impact on productivity, regardless of whether the countries are in a cold or warm region. In the same vein, climate change has negative impacts during both the earlier periods of our sample (1961–1987) and the latter period (1988–2013). Though this may seem to contradict the worsening trend of global warming, it is really reflective of the adaptive measures that have already been taken in agricultural production.

We hope our study draws attention to and attracts further research towards mitigation and adaptation vis-à-vis global warming [47,48]. With global warming being potentially beneficial to agricultural production in developed nations (be it at a higher cost and hence a lower TFP), we hope a more North–South cooperative approach can be reached to mitigate the impact of climate change. This will not only benefit developing nations, but with a more stable agricultural production and a more peaceful world, developed nations will certainly also have much to gain.

With our sample size, we hope to conduct more detailed study by grouping developing nations into various categories, with several classifications, with the hope of identifying how the impact of climate change varies in different developing regions. By doing so, we hope to identify countries that have been more successful in their adaptation efforts. This can further enhance a global effort to spread such practices and move towards an integrated climate of smart agriculture [49]. To this end, we welcome any interested reader to join us with their ideas and jointly conduct further research using our comprehensive dataset, which we are currently and will continue to expand and update.