Evaluating Impact of Climate Change to Fishing Productivity of Vietnam: An Application of Autoregressive Distributed Lag (ARDL) Regression Model45226

Evaluating Impact of Climate Change to Fishing Productivity of Vietnam: An Application of Autoregressive Distributed Lag ARDL Regression Model Nguyen Thi Vinh Ha 1* 1 VNU University of

Evaluating Impact of Climate Change to Fishing

Productivity of Vietnam: An Application of

Autoregressive Distributed Lag (ARDL) Regression Model

Nguyen Thi Vinh Ha (1)*

(1) VNU University of Economics and Business, Vietnam National University, Hanoi, Vietnam

* Correspondence: vinhha78@gmail.com

Abstract: Fisheries are most affected by climate change Yet studies on impact of climate change to

fisheries in Vietnam are still limited This paper uses production function with Autoregressive Distributed Lag (ARDL) regression and finds out that fishing productivity in Vietnam is negatively impacted by climate change In the long run, if sea surface temperature increases by 1 oC, fishing productivity, as measured by catch per unit effort (CPUE), will decrease by 0.25 ton/CV However, CPUE does not statistically significantly alter if there are changes in precipitation and number of storms

Keywords: fishing productivity; climate change; impact assessment; Autoregressive Distributed Lag

(ARDL) regression

1 INTRODUCTION

Climate change is having profound effects on environment, natural resources and economic, political and social life of economies around the world Fisheries are most affected

by climate change (Williams and Rota 2010) The world's fish stock has been significantly affected by various impacts such as overfishing, pollution, loss of habitats, declining biodiversity, epidemics, etc (Brander 2010) Climate change exacerbates and has faster direct and indirect effects on aquatic species The impacts of climate change to the oceans include increasing sea temperature, reducing dissolved oxygen, changing salinity, falling

pH, and shifting ocean currents, etc (Brander 2010; Sumaila et al 2011) These changes positively or negatively affect growth, reproductive productivity, behaviors and distribution of marine species (Brander 2010; Pinnegar et al 2013; Sumaila et al 2011)

Aquatic invertebrates and fishes are thermoplastic, each species has different tolerance to temperature (Williams and Rota 2010) Rising temperature reduces the ability

of dissolving oxygen in water, restricting respiration and affecting the health of aquatic species Therefore, when the water is warmer, fish will move to cooler areas for their favorite condition

Aquatic species also prefer different salinity levels The alteration of seawater salinity due to climate change is not significant so far However, in the future, the salinity of the oceans is likely to upsurge due to the increasing amount of groundwater containing salt running into the sea In the polar regions, salinity may fall due to augmented rainfall and low-salinity water flow from rivers (Roessig et al 2004)

The indirect impacts of climate change via ecosystems include foods, competitors, predators and pathogens of aquatic species (Brander 2010) The ocean acidification makes marine creatures like mollusks, zooplankton, etc difficult to create shells This disturbs the food webs, thereby shifting species distribution, growth and structure, leading to influences

on organisms in the oceans, estuaries, coral reefs, mangroves, and seagrass beds which are habitats of fishes Under-optimal environmental conditions can lessen feed intake, foster competition for food and shelter These changes deteriorate rate of growth, reproduction, behavior, distribution, species structure, stocks, migration behavior, and survival of aquatic species The likelihood of fish disease also worse due to rise of temperature, sea level, storms and cyclones (Brander 2010; Roessig et al 2004)

According to Roessig et al (2004) and Cheung et al (2009), climate change negatively affect fisheries in low latitude waters while it increases fishing benefits in high latitude waters However, fisheries in both sea areas can be negatively affected by the deterioration

of water quality and the increasing likelihood of disease (Williams and Rota 2010)

Vietnam locates in tropical region, its fishing industry is therefore more likely to be negatively affected by climate change This paper uses production function model with time series data to assess impact of climate change on fishing productivity in Vietnam

The production function approach has been widely used in research on assessment

of effects of changes in environmental quality

The Economic Commission for Latin America and the Caribbean (ECLAC) (Kirton 2011) assessed the relationship between fishery production (both capture and aquaculture) and seafood export price, sea surface temperature and annual average rainfall The results showed that the sea surface temperature and the average rainfall were inversely proportional to the fishery production in Guyana Damage to the fishery sector to year 2050, under the A2 (high emission) scenario, was estimated at USD 15 million (4% social discount per annum) For the B2 (medium emission) scenario, the anticipated damage by 2050 was

12 million USD at 4% discount

Caviedes and Fik (1992) showed that during the El Nino period 1997-1998, the pelagic fish catch yield in Peru and Chile decreased by 50% and 52% respectively, leading

to declines in export values and negative economic impacts (losses of job and income) in both countries Catch per unit effort was modeled to depend on annual average sea surface temperature and El Nino events MARMA regression was applied to correct time series problems such as autocorrelation and non-stationarity

Aaheim and Sygna (2000) used time series from 1980 to 1998 to examine the impact

of El Nino and La Nina on tuna catches in Fiji and Kiribati The study showed that the Southern Oscillation did not significantly affect tuna catch in Fiji while catch increased with

El Nino in Kiribati However, Aaheim and Sygna (2000) acknowledged that the regression model was so simple that it might not produce good estimation results

In Vietnam, Pham et al (2012) studied impacts of climate change on shrimp production in seven ecological regions of Vietnam Research results showed that there was

no correlation between shrimp productivity and temperature between 1990 and 2009 and

seasonal rainfall from 1995 to 2009 in Phu Tho and Hoa Binh provinces In the North Central Coastal region, temperature had an impact while rainfall had no influence on shrimp production in Nghe An and Thua Thien Hue provinces Cao et al (2013) quantified the variation in shrimp production due to changes in temperature and rainfalls in Thanh Hoa and Ha Tinh provinces They found out that there was inverse correlation between shrimp production with temperature and precipitation, in addition to capital, labour and acreage of shrimp ponds Nguyen et al (2015) forecasted the impact of climate change on fisheries production in northern region of Vietnam The study showed that the total damage of fishing industry in 2050 would be 584 billion VND at social discount rate of 3% in the medium (RCP4.5) climate change scenario

2 PRODUCTION FUNCTION MODEL AND DATA

Production function decribes relationship between inputs and outputs of a production process In theory, there are two major inputs of production which are capital and labor In sectors such as agriculture or fisheries, climate could be considered as an additional input Impacts of climate change are measured as differences of outputs as results

of variations in climate factors Let Y denote production output, K capital, L labour, and CC

climate factor, the production function is expressed as formula (1)

If climate change has impact on output, then δY/δCC is different to zero

Adjusted Cobb-Douglas function (Zellner et al 1966) is chosen for model specification and the production function has formula (2)

In which α, β and γ are elasticities of output to capital, labour and climate factors

respectively A is the impact of other factors The logarithm of the two sides are applied to have the formula (3)

LnY = LnA + αLnK + βLnL + γLnCC (3)

In this study, output of fisheries sector is measured in terms of fishing productivity, represented by catch per unit effort (CPUE) Fishing productivity, as measured by catch per unit effort (CPUE) might be used as proxy to fish stock Stable CPUE shows sustainable catch yield while decreasing CPUE means that fish is over-exploited (Quirijns et al 2008)

Since there is no available data on investment in fisheries sector, the study uses the

variable of total fishing vessel capacity (measured in horsepower, or cheval vapeur - CV) as a

proxy to capital

In Vietnam, climate change manifests in increasing temperature and precipitation, which may have influence on fish growth and migration behavior, affecting the fish stocks and then catches Storms (with windspeed of level 8 or higher) cause damages to the fisheries activities and fishing vessels, losses of life and property, and deteriorating livelihoods of fishing communities According to Ngo et al (2013), wind speed from level 0

to level 6 is convenient for fishing activities at sea El Nino and La Nina, which perform

similarly to climate change in short run, also have some bearing on fishing So, variables on temperature, precipitation, storms and El Nino are included in the production function

In 1997, the Vietnamese Government encouraged fishermen to invest in offshore fishing via a preferential finance project (Decision No 393-TTg dated 09 June 1997) In 2003, the National Assembly promulgated the Law of Fisheries After these two milestones, there were major policy changes related to fishing activities in Vietnam and supposed to have positive effects on the catch yields Therefore, we add two dummies to assess the impact of these policies in the production model

The production function has the following form:

CPUE t = β 0 + β 1 LnCapacity t + β 2 LnLabour t + β 3 SST t

+ β 4 LnRainfall t + β 5 Typhoon t + β 6 SOI t + β 7 D 1 + β 8 D 2 + ε t (4)

Data and sources of data for the regression models are described in Table 1

Table 1: Data description

Variables Description Sources Notes

T Time in year From 1976 to 2014

Catch t Catch yield in year

t

1976 - 2010: Ngo et al

(2013)

2011 - 2014: General Statistical Office (GSO)

(2016)

Including marine and inland catch

Capacity t Catch effort in year

t (CV)

1976 - 2010: Ngo et al

(2013)

2011 - 2014: GSO (2016)

CPUE t Catch per unit

effort in year t

(tons)

CPUE t = Catch t /Capacity t Fishing productivity

Labour t Total fishing labor

in year t (persons)

1976 - 2010: Ngo et al

(2013) Missing values (in 1978,

2011 - 2014) are filled by interpolation

SST t Average sea

surface temperature in

year t (°C)

The National Oceanic and Atmospheric Administration (NOAA),

USA

Measured at Halong Bay

Rainfall t Total precipitation

in year t (mm)

Climate Change Knowledge Portal, the World Bank

SOI Southern

Oscillation Index

NOAA Difference in air pressure

between Tahiti (Southern

Pacific) and Darwin (North to Australia)

Typhoon t Number of

typhoons in year t

National Centre for Hydro-Meteorological Forecasting (Vietnam) Dinh (2010)

Number of storms in the Eastern Sea

D 1 Proxied to offshore

fishing finance project in 1993

Value 0 for years before

1997, value 1 for years

1997 and later

D 2 Proxied to the

availability of Law

of Fisheries

Value 0 for years before

2003, value 1 for years

2003 and later

LnX Logarithm of X

β i Coefficients

ε t White noise

Autoregressive Distributed Lag regression

In addition to the fishing inputs and other impacting factors of the same year, annual fishing productivity tends to rely on productivity and factors of previous years due to lagging impacts Therefore, the Autoregressive Distributed Lag (ARDL) model (Pesaran and Shin 1998) is chosen to demonstrate these dependencies ARDL is ordinary least square (OLS) regression, which includes the lagged variables of dependent and independent variables The ARDL model is appropriate when time series have different degrees of integration (e.g I (0), I (1) or a combination of both) and especially when there is a single long-run relationship among variables It is also suitable with small sample size (n ≤ 30) (Nkoro and Uko 2016)

ARDL model has the following form:

Y t = c + α 1 Y t-1 + α 2 Y t-2 +… + α p Y t-p + β 0 X t + β 1 X t-1 + … + β q X t-q + u t (5)

In which Y is dependent variable, X are explanatory variables, α and β are coefficients, p and q are number of lags of dependent and explanatory variables respectively,

c is intercept, t denotes time and u t is white noise

Several tests should be performed to confirm the appropriateness of the ARDL model, including selection of number of lags, tests for stationarity, long-run relationship among variables, model specification, autocorrelation, heteroscedasticity, multicollinearity, white noise (i.e residual series are normal distribution and stationary), stability of the coefficients and convergence of long-run coefficients

Test for stationarity of time series

Normally, time series regressions require all series to be stationary, i.e mean, variance, and covariance at different lags have constant values over time (Gujarati and Porter 2009) Non-stationary series can lead to spurious regression However, according to Nkoro and Uko (2016), ARDL regression is suitable with integrated time series of order zero

or one Augmented Dickey-Fuller (ADF) test and Schwarz information criteria (SIC) are applied to perform unit root tests of the time series The test results (Table 2) show that

CPUE and LnLabour are integrated of order 1 They are difference stationary LnCapacity is

integrated of order zero, which is trend stationary, or its mean trends are deterministic The other time series are integrated of order zero Therefore, while traditional OLS regression is not applicable to this data set, ARDL regression can work

Table 2: Unit root test results

Variable ADF statistics p-value p-value of difference Critical value

1% 5% 10%

SST -3.862 0.005 - -3.616 -2.941 -2.609

SOI -4.358 0.001 - -3.616 -2.941 -2.609

In which D(x) denotes the first-order difference of x, i.e D(CPUE t ) = CPUE t - CPUE t-1

Selecting number of lags for regression models

Vector autoregression (VAR) test and Akaike information criterion (AIC) are applied

to select number of lags in ARDL model The VAR results (Table 3) show that the ARDL should have 3 lags for all variables

Table 3: Selection of optimal number of lags using VAR

0 0 -98.51 - 8.29e-07 5.86 6.17

1 1 68.02 259.04* 1.28e-09 -0.67 1.80*

2 2 116.54 56.61 1.81e-09 -0.64 3.98

3 3 191.46 58.27 1.19e-09* -2.08* 4.69

*Number of lags selected by criteria

Test for long-run relationship among variables

To seek for the existence of long-run relationship among variables, the bound test is performed, using bound F-statistics and t-statistics to determine the cointegration among variables (Nkoro and Uko 2016) Bound test has the following form:

∆Yt = δ0 + + + δ1∆Y t-1 + δ2∆X t-1 + v t (6)

In which ∆ denotes the difference values, for example ∆Y t = Y t - Y t-1 ; p is the maximum lag chosen by author; (δ1 - δ2) depicts the long-run relationship, while (α1 - α2) depicts

short-run one Wald test is applied with null hypothesis is that all coefficients of lagged variables are zero

F-statistics in Wald test does not follow normal distribution It depends on: (1) integration orders of variables (I(0) or I(1)); (2) number of independent variables; (3) the existence of constant and trend variables in the model; and (4) sample size (Narayan 2005) Narayan (2005) provided critical values for ARDL model with small sample size (from 30 to

80 observations) If the F-statistics is larger than the upper bound of the critical value, the null hypothesis is rejected

Results of bound test are described in Table 4 F-statistics in model with dummies D1 and D2 is 7.4652, larger than the critical value 5.797 at significance level of 1% F-statistics in model without dummies is 5.1777, larger than the critical value 4.324 at significance level of 5% The t-statistics in model with dummies is -5.3166, less than the critical value -4.99 at significance level of 1% The t-statistics in model without dummies is -4.5429, less than the critical value -4.38 at significance level of 5% Therefore, we reject the null hypothesis and accept that there is long-run relationship among variables in the models The application of ARDL regression is acceptable with the data set

Table 4: Bound tests

Test statistics Value Significance

level

I(0) I(1)

Sample size n = 40

With dummies 7.4652 1% 3.800 5.643 Without dummies 5.1777 Sample size n = 35

Number of degrees k 6 5% 2.864 4.324 Real sample size 36 1% 4.016 5.797 t-statistics

With dummies -5.3166 5% -2.86 -4.38 Without dummies -4.5420 1% -3.43 -4.99

Test for autocorrelation

Table 5: Breusch-Godfrey test

With dummies Without dummies

Breusch-Godfrey test is applied to seek for autocorrelation It gives p-values of Chi

square at 0.000 for both models (Table 5) So, we reject the null hypothesis and accept that the models have autocorrelation For OLS regression, when there is autocorrelation, the estimates are not biased but ineffective (non-smallest variance), leading to unreliable F and

t tests Newey-West estimates are used as a remedy The correlograms after the Newey-West estimation show that the models are no longer autocorrelated

Autocorrelation Autocorrelation

Figure 1: Correlograms

Test for suitability of model specification

Table 6: Ramsey Reset test for the suitability of model specification

With dummies Without dummies

Degree of freedom df (3.3) (1.7)

p-value of the models are larger than α=0.05, so the model specification is suitable

Test for heteroskedasticity

Table 7: Breusch-Pagan-Godfrey test for heteroskedasticity

With dummies Without dummies

p-value of Chi square is larger than α=0.05, we accept the hypothesis that there is no

heteroskedasticity in models Harvey and Glejser tests provide the same results

Test for multicollinearity

Most of explanatory variables of the models are statistically different from zero, Durbin-Watson d statistics is close to 2, so we can ignore the multicollinearity in the models

Test for stability of the coefficients

Cumulative sum (CUSUM) and cumulative sum square (CUSUMSQ) control charts (Figure 3) show that CUSUM and CUSUMSQ curves lie between the critical curves at significance level of 5% This result confirms that there is long-run relationship among variables and the coefficients are stable

-6

-4

-2

0

2

4

6

8

2009 2010 2011 2012 2013 2014

CUSUM 5% S ignificance

-0.4 0.0 0.4 0.8 1.2 1.6

2009 2010 2011 2012 2013 2014

CUS UM of Squares 5% Significance

With dummies

-10.0

-7.5

-5.0

-2.5

0.0

2.5

5.0

7.5

10.0

2007 2008 2009 2010 2011 2012 2013 2014

CUSUM 5% Significance

-0.4 0.0 0.4 0.8 1.2 1.6

2007 2008 2009 2010 2011 2012 2013 2014

CUSUM of Squares 5% Significance

Without dummies Figure 1: CUSUM and CUSUMSQ control charts

Test for normal distribution of residuals

Using the Jarque-Bera test on the residual series of the models gives the p-values of all models greater than α = 0.05 We accept the hypothesis that the residual series follow normal distribution (Figure 2)

0

1

2

3

4

5

6

7

8

-0.02 -0.01 0.00 0.01 0.02 0.03

Series: Residuals Sample 1979 2014 Observations 36

Mean 8.88e-16 Median 0.001051 Maximum 0.025051 Minimum -0.021440 Std Dev 0.012478 Skewness 0.160345 Kurtosis 2.220148

Jarque-Bera 1.066517 Probability 0.586690 0

1 2 3 4 5 6 7

-0.03 -0.02 -0.01 0.00 0.01 0.02 0.03

Series: Residuals Sample 1979 2014 Observations 36

Mean -3.16e-15 Median 0.000663 Maximum 0.031153 Minimum -0.026961 Std Dev 0.016919 Skewness 0.226024 Kurtosis 2.095594

Jarque-Bera 1.533448 Probability 0.464532

Figure 2: Histograms and Jarque-Bera tests Test for stationarity of residuals

Table 8: Unit root test for residual series With dummies Without dummies Critical value at significance level of 1%

ADF statistics are negative and less than the critical value at significance level of 1%,

p-value is 0.000 Therefore, the residual series are stationary and normal distributed, i.e they

are white noise

The coefficient of the first lag of Error Correction EC(-1) of the Error Correction Model (ECM) is -1.0859 (i.e negative and larger than -2), and statistically significant It confirms that it is dynamically stable (Loayza and Ranciere 2005)

The results show that there is long-run relationship among the variables; the model has no autocorrelation, no heteroskedasticity, no multicollinearity; the residuals are white noise, the coefficients of the models are stable and model specifications are suitable Therefore, it can be said that regression models are appropriate and reliable

3 RESULTS

The ARDL model and its Conditional Error Correction model (ECM) can show the short-run and long-run relationship among variables In case the number of observations of the model is small, the ECM model gives more reliable results (Nkoro and Uko 2016) However, the coefficient of the ARDL model can be explained more easily and visually