Cyclical variation of the fiscal multiplier in Turkey

ABSTRACT

This paper investigates the cyclical variation in the government spending multiplier for Turkey from 1990:q1 to 2015:q4. We use a time series model, namely the local projection method, to estimate the variation in the fiscal multiplier under two different regimes: low and high growth regimes with respect to long-term economic growth. Our results confirm that fiscal policy is more effective in times of low growth than high growth. Turning to the components of government spending, we find that the government investment multiplier is larger than the government consumption multiplier in both regimes. This evidence supports the view that an expansionary fiscal policy via public investment has a more profound effect on output than via public consumption. However, we also find evidence that the influence of government consumption on GDP increases substantially in times of low growth. We, therefore, suggest that policymakers use public investment rather than public consumption to stimulate the economy during economic expansion and increase public consumption during economic slowdowns.

KEY WORDS:

• Fiscal policy
• local projection
• state-dependent fiscal multipliers

JEL:

• E62
• H30

Notes

1. See Blanchard and Leigh (2013) about this issue.

2. One may find detailed information about these models and comparison of their advantages and in advantages in Whalen and Reichling (2015).

3. In the VAR context, there are several papers which solve the identification problem of fiscal policy shocks in a different manner. These are SVAR approach (Blanchard and Perotti (2002)), sign restriction approach (Mountford and Uhlig (2009)), recursive approach (Fatas and Mihov (2001) and Favero (2003)) and narrative approach (Ramey and Shapiro (1998), Edelberg, Eichenbaum, and Fisher (1999) and Burnside et al. (2004)). Within the DSGE framework, see Coenen et al. (2012) for the model-based estimations of fiscal multipliers.

4. While Baum and Koester (2011), Batini, Callegari, and Melina (2012) and Baum, Poplawski-Ribeiro, and Weber (2012) use threshold VAR, Auerbach and Gorodnichenko (2012), Hernández de Cos and Moral-Benito (2013) and Herbert (2014) use STVAR technique for estimating fiscal multiplier. On the other hand, Ko and Morita (2013) and Arin, Koray, and Spagnolo (2015) can be given as examples of the MSVAR technique.

5. See Batini, Eyraud, and Weber (2014) for more information about the determinants of the fiscal multiplier.

6. We used government spending data obtained from the National Accounts (as a component of real GDP) instead of central government budget expenditure. Since the coverage of the budget and budgetary system of Turkey changes frequently, it would be difficult to obtain consistent quarterly data for government spending for a long period. Therefore, we chose the former definition of government spending. According to the economic classification of budget expenditure, total government spending consists of three main components, namely government consumption, investment, and transfer spending. In line with other studies, we defined government spending as the sum of government consumption and investment, which directly affects output. On the other hand, transfer expenditure affects household disposable income and so indirectly affects output.

7. The real GDP series based on 1998 = 100 was extrapolated backward using the growth rate of the GDP series based on 1987 = 100.

8. The Ministry of Finance of Turkey has published monthly data on central government budget since January 2006 while also extending central government budget data back to 2000 on an annual basis. For the earlier periods (1990–1999), tax rebates were subtracted from the consolidated budget tax revenues while the amount of local administrations and fund shares were added to the consolidated budget tax revenues to obtain a proxy for central government budget tax revenues. To convert annual data to quarterly data, we calculated the shares of quarterly tax revenues in total tax revenues for each year using the consolidated budget figures. We then applied these quarterly ratios to the corresponding yearly central government budget tax revenues to get quarterly data for 1990–2005.

9. All series were seasonally adjusted using the Tramo/Seats method.

10. It is quite common in the literature to use four lags for the quarterly data set. However, in section 6, we will display the sensitivity of fiscal multiplier to the selected lag length by estimating the model with different lag lengths as three and five. Although the size of fiscal multiplier change according to different lag lengths, the results support the view that the effectiveness of fiscal policy increases at the times of low growth.

11. See Newey and West (1987).

12. Considering only a quarter to decide the state of the economy may give wrong messages due to the highly volatile nature of quarterly growth rate. Hence, we select 7-quarter centered moving average growth rate as a threshold variable to obtain more stable parameter to decide low and high growth states. As shown in Figure 2, low growth periods, calculated using this methodology, capture very well the shrinks in real GDP growth as in 1991, 1994, 1999, 2001, and 2009.

13. We assume that the annual potential GDP growth rate is 4 for Turkey. Thus, we choose 1 as a threshold value for growth rate in a period, which corresponds to 4 for the annual growth rate. We had 100 observations. Half of all observations (50 observations) were recorded as low and rest as high growth state. We also defined two different threshold values for the two sub-periods marked by the banking crisis in 2001. We then distinguished low/high growth regimes by taking into account corresponding threshold values for the two sub-periods. We calculated the quarterly average growth rates as 0.7% for the fragile period (1990–2001) and 1.2% for the sound period (2002–2015). Based on these threshold values, we determined the state of the economy to calculate the size of the fiscal multiplier. According to our findings, the magnitude of the fiscal multipliers does not change much compared to the baseline model.

14. Ramey (2011) constructed two news series of government spending shocks. One of them is called the `defense news variable’ (news about future defense spending), which is computed by taking into account the expected present discounted value of changes in government spending. The other is based on the forecast errors of professional forecasters. While Auerbach and Gorodnichenko (2012, 2013) used forecast error for the growth rate of government spending, Ramey and Zubairy (2014) used news about future defense spending to identify fiscal policy shocks. Alloza (2014) is an example of both a VAR-based and narrative approach (news about future defense spending).

15. Other identification strategies cannot be used in Turkey. The narrative approach uses news about future defense spending related to wars as a measure of exogenous fiscal shocks. However, Turkey had not experienced war during the period under consideration. Moreover, we are not interested in the multiplier effect of military expenditure. Rather, this study focused on the output responses of government consumption and investment shocks. On the other hand, we could not calculate the forecast error for government spending because there are no official announcements of the quarterly forecast of fiscal variables in Turkey.

16. The cumulative fiscal multiplier for the different horizon H can be shown as follows: $\left(\frac{{\sum }_{j=0}^H\mathrm{\Delta }{Y}_{t+j}}{{\sum }_{j=0}^H\mathrm{\Delta }{G}_{t+j}}\right)$.

17. STATA and MATLAB were used for the estimation. To convert a percentage change to a unit change, we used the sample average of government spending to GDP ratio. When we calculated the state-dependent fiscal multipliers, we used the respective values of government spending to GDP ratio for each regime.

18. The impact multiplier can be shown as $\left(\frac{\mathrm{\Delta }{Y}_t}{\mathrm{\Delta }{G}_t}\right)$ and the maximum of the normal multiplier over any horizon H is defined as $\left(max\frac{\mathrm{\Delta }{Y}_{t+H}}{\mathrm{\Delta }{G}_t}\right).$ The maximum of the cumulative multiplier was calculated as $\left(max\frac{{\sum }_{j=0}^H\mathrm{\Delta }{Y}_{t+j}}{{\sum }_{j=0}^H\mathrm{\Delta }{G}_{t+j}}\right)$.

19. Estimating the magnitude of the fiscal multiplier using linear models (one-regime models) may send inaccurate messages to policymakers because linear models rule out state-dependent multipliers and calculate fiscal multipliers by taking into account the averages of expansion and contraction periods. This implies that the average fiscal multiplier obtained from a linear model underestimates the size of the fiscal multiplier during a recession but overestimates it during an expansion.

20. When the fiscal multiplier is equal to 2, it means that a unit increase in government spending increases output by two units. Note that this argument holds not only for an expansionary fiscal policy but also for a contractionary fiscal policy that has the opposite sign. In this section, we assume that the multiplier effects of an increase in government spending and a decrease in government spending are the same. In other words, there are no asymmetric effects of positive and negative fiscal shocks on output. However, in section 6, we will show that the size of the government spending multiplier depends crucially on the sign of the fiscal shock (i.e. an expansionary or contractionary fiscal shock).

21. It is quite common in the literature to find high values for fiscal multipliers that substantially exceed 1, such as Hernández de Cos and Moral-Benito (2013) for Spain, Herbert (2014) for France, Germany and the USA, and Auerbach and Gorodnichenko (2014) for Japan. For example, Hernández de Cos and Moral-Benito (2013) find that the impact, peak, one-year cumulative and two-year cumulative multipliers in Spain are 0.65, 1.96, 1.26 and 1.25, respectively, during a recession. Similarly, Herbert (2014) finds maximum multipliers in a recession of 1.38 for France, 1.31 for Germany and 2.07 for the USA. Auerbach and Gorodnichenko (2014) report maximum and three-year cumulative multipliers in recession for Japan of 2.51 and 2.73, respectively.

22. To compare the multiplier effects of public consumption and investment, we extended the three-variable model to a four-variable model by replacing government spending with government consumption and investment. Similarly, fiscal shocks were obtained from a four-variable VAR model, where government consumption was ordered first and government investment second.

23. Çebi (2016) estimates a linear VAR model for 2002q1-2014q4 to calculate fiscal multipliers. Breaking down government spending into consumption and investment he finds high values for fiscal multipliers as in our study. He calculates impact and one-year cumulative multipliers as 1.4 and 1.7 for government consumption and 2.1 and 1.7 for government investment, respectively. However, different from our study, the fiscal multipliers estimated in Çebi (2016) represent the whole period without taking into account the state of the business cycle. In this aspect, the added value of the current paper is to calculate the fiscal multipliers for different growth regimes.

24. In contrast to our findings, Özlale and Yüksel (2016) find that the public consumption multiplier is larger than the public investment multiplier for Turkey. On the other hand, Çebi (2016) reports similar results for Turkey, with larger multiplier effects for public investment than public consumption. Similarly, Hernández de Cos and Moral-Benito (2013) find some evidence in favor of the public investment multiplier in recession for Spain while Auerbach and Gorodnichenko (2012) suggest that the public investment multiplier has a greater effect than the public consumption multiplier for the USA. According to Herbert (2014), the long-term output effect of public investment is higher than that of public consumption in recession for France and the USA. However, she reports the opposite result for Germany, in that the government consumption multiplier is larger than the investment multiplier in recession.

25. Although the model used in this study does not explicitly reveal the mechanism behind the high value for the public investment multiplier, one explanation might be as follows: In contrast to public consumption, public investment is a fiscal tool that affects aggregate demand in the short run and may affect aggregate supply in the long run. If government investment is allocated to productive fields, it may help to increase a country’s potential output. Thus, economic agents may perceive public investment as productive expenditure, which offsets the short run cost of government investment in the long run.

26. Real interest rates were defined as $r_t=(1+i_t)/(1+{\pi }_t)-1$, where $r_t$ and $i_t$ represent real and nominal interest rates, respectively, while ${\pi }_t$ denotes the inflation rate using the GDP deflator. Nominal interest rates were obtained from the Turkish Treasury.

27. Note that $\frac{Y_{t+h}-Y_{t-1}}{Y_{t-1}}\approx \left(ln{Y}_{t+h}-ln{Y}_{t-1}\right)$ and $\left(\frac{G_{t+h}-G_{t-1}}{G_{t-1}}\right)\ast \left(\frac{G_{t-1}}{Y_{t-1}}\right)=\frac{G_{t+h}-G_{t-1}}{Y_{t-1}}\mathrm{\%}\approx \left(ln{G}_{t+h}-ln{G}_{t-1}\right)\ast \left(\frac{G_{t-1}}{\mathrm{\%}{Y}_{t-1}}\right).$

28. We re-estimated the model with a 3 lag. However, estimating the model with 3 or 4 lags made no difference to the size of the fiscal multipliers (Table 2). On the other hand, estimating the model with a 5 lag considerably reduced the magnitude of the fiscal multiplier, although this specification still produces qualitatively similar results.

29. We represented fiscal shocks using the non-linear model as follows: First, we created a dummy variable, which takes the value of 1 for a low growth regime and 0 for a high growth regime. Then, we simply interacted the dummy variable with the other variables. Finally, we estimated the government spending equation of the VAR model by adding the interaction terms to obtain non-linear fiscal shocks.

30. The estimation period for this specification starts from 1994:q1 due to data availability.

31. To do this, we applied an HP filter to quarterly growth rates and extracted the trend growth rate. We then compared the 7-quarter centered moving average growth rate to the corresponding value of the trend growth rate that we obtained from the HP filter. If the 7-quarter moving average growth rate is higher (lower) than that of trend growth, then it is accepted as a high (low) growth regime. In contrast to the previous model, which had a fixed value of 1 as the threshold value for the quarterly growth rate over the sample period, this methodology allowed the trend growth rate to vary across the sample period.

32. We separate the data into four groups, namely, low growth/positive shocks, low growth/negative shocks, high growth/positive shocks, high growth/negative shocks. In terms of the low growth period, 26 of 50 observations fall into the first group (positive shocks) and 24 of 50 observations remain the second group (negative shocks). Similarly, high growth period can be divided into two groups with 23 of 50 observations fall into third group (positive shocks) and rest of the observations (27 observations) fall into fourth group (negative shocks).