4. Empirical Results
4.2. ECDP VAR Models
4.2.3. ECDP Model 3—Industrial Production, Iron and Steel Prices, and Diesel Prices
國
立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
4.2.3. ECDP Model 3—Industrial Production, Iron and Steel Prices, and Diesel Prices The next VAR model uses ECDPs differenced by one-year, industrial production of business equipment differenced by one-month, iron and steel prices differenced by one month, and diesel prices differenced by one year. Since all of the variables have undergone a Dickey Fuller unit root test, this step will be skipped. Therefore, the next step is lag order selection criteria tests.
Table 4-28 Lag Order Selection Criteria for ECDPs, Iron and Steel Prices One-Month Differenced, Industrial Production One-One-Month Differenced, and Diesel Prices One-Year Differenced
A total of 13 lags were used in the above selection-order criteria results. As an be seen by Table 4-26, 13 lags were statistically significant with the LR, FPE, and AIC statistics. Continuing to use the AIC statistic is a main foundation of the results, the VAR then uses 13 lags given the results of these tests.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Table 4-29 VAR Equation Test Results for ECDPs, Iron and Steel Prices One-Month Differenced, Industrial Production One-One-Month Differenced, and Diesel Prices One-Year Differenced
The VAR equations were then conducted finding all variables to be statistically significant (See Table 4-29). However, industrial production of business equipment suffered from a very low R-squared statistic suggesting that its variance only accounted for around 34 percent of the variance of the population. Iron and steel prices also suffered from a relatively low R-squared statistic compared to other models, but it was still high compared to that of industrial production.
After these statistics were examined, the OIRFs were executed.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4-31 OIRF for Diesel Prices Impulse and ECDPs Response
The first OIRF examined diesel prices against equipment prices (see Figure 4-31). It was the only OIRF which found 90 percent confidence intervals to be entirely above or below the zero-reference point. As can be seen, diesel prices increase the price of these particular measurement of equipment prices, but it does not do so until after the 9th step and only until around the 18th step.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4-32 OIRF for Industrial Production Impulse and ECDs Response
As can be seen by the above OIRF, industrial production did not enjoy a meaningful relationship with equipment prices. It could be suggested that the impact is almost certainly positive in the first two steps, but beyond this it cannot be determined if the impact is positive or negative for the entirety of the steps after the initial ‘shock’.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4-33 OIRF for Iron and Steel Prices Impulse and ECDs Response
So too is this the case with iron and steel prices and equipment prices. In general, the model suggests that the impact could be negative with the moving average negative for the first ten steps (see Figure 4-33). However, it could also be positive as the 90 percent confidence interval encompasses both positive and negative territory. After the OIRFs were conducted, the Granger causality tests were executed.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Table 4-30 Lagrange-Multiplier Test for ECDPs Model 3
A Lagrange-multiplier test was then conducted after the OIRFs. It showed that only five of the thirteen lags suffered from autocorrelation. Again, since the higher order lag was not statistically significant, it can then be surmised that the model overall does not suffer from a large degree of serial correlation.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Table 4-31 Granger Causality Test Results for ECDPs, Iron and Steel Prices One-Month Differenced, Industrial Production One-Month Differenced, and Diesel Prices One-Year Differenced
The results show that none of the variables Granger cause excavator, cranes, and dragline prices (see Table 4-31). Also, ECDPs do not Granger cause any other variables. Moreover, there are other variables in the model which Granger cause one another. For instance, industrial production and diesel prices Granger cause iron and steel prices while iron and steel prices also Granger cause industrial production. Also, iron and steel prices Granger cause diesel prices along with industrial production.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4-34 Dynamic Forecasts for ECDPs, Iron and Steel Prices One-Month Differenced, Industrial Production One-Month Differenced, and Diesel Prices One-Year Differenced
Again, the above forecasts of the variables demonstrated that the primary variable of analysis, equipment prices, enjoyed the ability to be forecasted throughout the duration of the forecast period. It reinforced the notion that prices would decrease at the end of 2019. Moreover, the forecast suggested that throughout the duration of 2020, prices would be suppressed in negative territory. This would be the case until around mid-2020 when prices would become positive again followed by further increases in price into 2021. However, forecasts were more uncertain for the other variable sin the model. Even so, forecasting these variables is only peripheral to the primary purpose of this thesis and is therefore negligible.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
4.3. Added Variable Plots
As was mentioned in the method section, the added variable plots can determine the degree to which variables are a well fit for the model based on the relationship between the overall model’s correlative components, average values, and the residual values. If the coefficient is 0, then the variable is not a good fit for the model. As can be seen by the results, none of the coefficients in the added variable plots is zero. While some of the variables appear to have a forced linear line through the scatterplot which are horizontal, the coefficient value underneath each scatterplot asserts that none of the variables suffer from this condition. One particular variable, NRCS, appears to have a very small coefficient. However, this is not the case as the slope of the line asserts that the value of the coefficient is likely impacted by the coding of the values in the data extraction process. Moreover, the distribution of all the added variable plots appear to be homoscedastic for the most part.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4-35 Added Variable Plots Matrix for One-Year Differenced Variables
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
4.4. Political Party Impacts
A multiple equation OLS regression implementing four different dependent variables was also used. It included CEPs, GEPs, ECDPs, and PCS. Smaller models were used for these regressions so as to focus on the impact of the dummy variables of the impact of political congress in the US House of Representatives, US Senate, and the presidency. For all of the dummy variables, 0 was denoted as Democrats in control whereas 1 denoted Republicans. As can be seen, The impact of political power on CEPs, GMPs, and ECDPs was somewhat different and suffered from somewhat different levels of strength.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Table 4-32 OLS Regression Impacts of Political Control
As for PCS, only the presidency and Senate were statistically significant. Statistical significance. Moreover, the impact is substantively strong as can be seen in the coefficient value indicating a clear positive impact between these variables. Therefore, it can be deduced that Republicans controlling the Senate and the presidency is statistically likely to increase PCS in conjunction with CEPs, GMPs, and ECDPs. However, the different equipment prices are impacted in different ways from one another. CEPs and ECDPs are similar in their statistical significance
‧
and coefficient directionality. GMPs on the other hand also enjoy statistical significance with Republican control of the presidency. This statistical divergence in significance is likely due to the fact that GMPs are more reflective of the economy overall as opposed to equipment prices since GMPs encompass many more types of machines than what CEPs and ECDPs do.
Moreover, lag order selection tests are conducted. From these results, it can be deduced that only one lag is appropriate for this model. This is due to the fact that the FPE, AIC, HQIC, and SBIC statistics suggested that 1 lag was the appropriate model specification. Because of this, one lag is used for the VAR model which uses the tri-variate political variables and CEPs. While this model may suffer from under-specification, this is likely due to the variable structure of a majority of the variables that are embedded within the VAR vector equations. Under other models, these statistics would perhaps change with the different number of maximum lags selection in the selection-order criteria test specification process. However, with this particular model given these particular variables the maximum lag order for the selection-order criteria tests did not change statistical significance which furthermore buttresses the notion that 1 lag is the most appropriate lag order for this model.
Table 4-33 Lag Order Selection Criteria Tests for the House, Senate, and CEPs
After these the lag order selection criteria tests were conducted, the VAR itself was
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
conducted. As can be seen by the results, all variables are statistically significant. All variables enjoy a large R-square statistic ranging from .83 to .9. The model therefore appears to be a good fit as a representation of the population.
Table 4-34 VAR Equation Results for the House, Senate, and CEPs
After the VAR equations were conducted, a series of OIRFs were conducted. First, an impulse of the House of Representatives was executed. To be clear, because of the Dickey Fuller unit root requirement, this shock does not represent merely a Republican or Democrat control of the House of Representatives and the Senate for that OIRF. Rather, a shock in this model represents a change in control from one party to the next where a value of 1 in the dataset represents a change from Democrat to Republican, a value of -1 represents a change from Republican to Democrat, and a value of 0 represents no change in control of these chambers of congress.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4-36 IRF for House Impulse and CEPs Response
As can be seen, the shock of a change in governance control in the House of Representatives from Democrats to Republicans provides CEPs to increase for more than 25 steps after the initial shock. Additionally, this can be the asserted with a 90 percent confidence until the around the 30th step where the 90 percent confidence bands dip below the zero-reference point.
However, the moving average of the OIRF is positive until the 50th step indicating that the impact could still be positive.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4-37 IRF for the Senate Impulse and CEPs Response
As for a change from Democrats to Republicans in the Senate, the impact is virtually zero as can be seen by the entirely flat moving average. This is not merely an assertion that the Senate has no impact on CEPs. Rather, it indicates that a change in governance in the Senate does not impact CEPs. The OLS models underscore this point which suggests that Republican control of the Senate does in fact elevate CEPs on average.
Clearly, there is an impact from political control and policy on equipment prices. However, the story of this impact likely begins in spending on construction. The dynamics of PCS and NRCS is murky; it is not clear which causes the other. Moreover, what are the political impacts on these
‧
dynamics? A final VAR model is used to come to this conclusion. First, the VAR equations are conducted.
One particular line of argument could be that a vector autoregression with exogenous variables (VARX) model should be used instead of a VAR since political control may be exogenous to the model. The argument asserts that CEPs, NRCS, and PCS cannot determine political control. However, there is much literature on the relationship between economic conditions can dramatically impact elections. Good economic performance then may be responsible for one party being elected over another. First, it needs to be recognized that Republicans are more likely to be in office during times of higher GDP growth. This is evident with the correlation coefficients between GDP and the political indicators which suggests the political variables tend to lean towards Republican control and higher GDP (see Table 4-35).
Table 4-35 Correlation Coefficient Matrix of Macroeconomic Indicators and Political Control Dummy Variables
Since its clear CEPs lag GDP, GDP is negatively correlated with CEPs, and CEPs enjoy a negative relationship with political control, then it appears to be the case that a negative relationship between political control and CEPs could be the result of the positive relationship between GDP and political control. Since this is the case, it could be that a positive shock from CEPs would result in Democratic control since it tends to be the case that Democrats are in power
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
during times of lower GDP and lower prices. The OIRFs show this to be the case at least with the US House of Representatives (see Figure 4-38).
Figure 4-38 CEPs Impulse House Control Response OIRF
A final model was then created including political control and different construction spending indicators along with CEPs. To be brief, it is evident that a change in power from the House has a positive impact on NRCS. This is clear in the OIRF between the two variables..
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Table 4-36 Political Control, CEPs, NRCS, and PCS VAR Equation Results
As can be seen, all variables are statistically significant. All variables enjoy a high R-square statistic. Since NRCS was included in the model, it is important then to keep in mind that the number of observations is quite small compared to other VAR models. This means that only data from 2000 until 2019 was included.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4-39 House Political Control and NRCS
Moreover, there is a positive shock from NRCS on PCS. This is evident in the IRF between the two of them. The impact is also quite strong and long-lasting as can be seen by the duration in which the 90 percent confidence intervals are both above the zero-reference line. The moving average is moreover above the zero-reference line for all 50 steps of the OIRF.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4-40 NRCS Impulse PCS Response
After these two figures were created, the Granger causality equations were conducted to examine which variables in the model Granger cause one another if any. The impact is less strong than that of the House on NRCS, but the 90 percent confidence intervals are still above the zero-reference line for a number of steps. However, it should also be noted that the moving average is above the zero-reference point for all 50 steps indicating that the impact is on average more likely to be positive than negative for all 50 steps.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
As can be seen, the Granger causality table shows that a change in power in the US House of Representatives Granger causes an increase in NRCS. Furthermore, NRCS then Granger causes PCS. Moreover, these are the only Granger causal relationships that are present in this model making the impact much more clear than prior VAR models.