.

Data and Methodology

We use the publicly available quarterly data for Egypt ranging from March 2001 to March 2021.

The data source for all the variables except economic policy uncertainty is the Central Bank of

Egypt (CBE); we use monthly statistical bulletins, quarterly economic reviews, annual reports, and

quarterly time-series data – all retrieved from the official CBE webpage. Data for economic policy

uncertainty (EPU) index has been retrieved from Federal Reserve Bank of St. Louis Economic

Data (FRED), World Economic Uncertainty Index for Egypt. The summary of the variables used

in the research is given in the Table 1 below.

Table 1. Descriptive statistics of the variables

Available Data Qtrs. Mean Median Min. Max. S.D.

Δ Real GDP 2003:Q3-2021:Q1 71 1.272 1.210 -7.580 9.330 3.751

Domestic Debt 2001:Q2-2020:Q2 77 78.813 79.800 66.300 96.700 6.808

Total Debt 2001:Q2-2020:Q2 77 103.840 105.100 81.400 131.200 13.260

Fiscal Balance 2003:Q3-2020:Q3 69 -2.194 -2.200 -4.900 0.4000 0.964

Δ Money Supply 2003:Q3-2021:Q1 71 4.169 4.030 -1.790 13.330 2.900

Consumption 2001:Q3-2021:Q1 79 78.529 78.080 62.870 91.080 6.848

Investment 2001:Q3-2021:Q1 79 16.874 16.420 8.200 27.780 3.941

Export 2001:Q3-2021:Q1 79 20.401 19.680 9.440 35.760 7.094

Inflation Rate 2003:Q3-2021:Q1 71 11.134 10.470 3.200 32.150 5.993

Unemployment Rate 2003:Q1-2021:Q1 73 10.656 10.600 7.200 13.400 1.853

EPU 2001:Q1-2021:Q1 81 0.1446 0.100 0.000 1.010 0.180

5

As can be seen from the Table 1, the overlapping period for all the variables is September 2003 –

July 2020, 68 quarters in total. Since Egypt uses a fiscal year (FY) that starts in July and ends in

June, the time frame of our research matches FY 2003/04:Q1 – FY 2019/20:Q4.

The raw quarterly data for real GDP exhibits a clear pattern of seasonality: there is a significant

increase in real GDP between the last quarter of a previous fiscal year and the first quarter of a

current fiscal year; we observe this pattern for the entire span of the data series. We apply a simple

deseasonalizing method based on the centered moving average and used the deseasonalized data

to calculate the change in real GDP between quarterly periods.

We consider two debt variables as a potential threshold variable in our paper: domestic debt and

total debt, the latter refers to the sum of domestic debt and external debt. Both variables are

normalized to a GDP level (expressed as debt-to-GDP ratios) and are measured in percentage

points of Egypt’s nominal GDP. It should be noted that the public debt reported by CBE includes

the government’s debt as well as the debt by public economic authorities and the debt accrued to

National Investment Bank of Egypt; however, the share of the government’s debt in public debt is

estimated to fluctuate around 80-90% for the period analyzed in the current research.

The fiscal balance is calculated as the difference between total government revenues and total

government expenditures over nominal GDP, the positive (negative) value for the fiscal balance

implies that the government is running a fiscal surplus (deficit) in the current period. As it can be

seen in the Figure 1 below, the government of Egypt was running a fiscal deficit in all but three

time periods during the time span analyzed in the paper, and the median value for the fiscal balance

to nominal GDP is negative 2.2 percentage points.

Figure 1. Egypt's Fiscal Balance in 2003 September - 2020 September

6

We chose the standard control variables used in the economic literature: change in money supply

(M1), consumption, investment, export, inflation rate, unemployment rate, and EP uncertainty.

The inflation rate has been calculated as the change in consumer price index (CPI) relative to

corresponding month of previous year, the weights from January 2010 CPI were used for the entire

data span to preserve the consistency for the calculation exercise. Consumption, investment, and

export are normalized to a GDP level and are measured in GDP percentage points.

We resort to Hansen (2000) sample splitting threshold regression model as a methodological base

of our exercise. Consider a following simple regression equation:

𝑦𝑦𝑡𝑡 = (𝛽𝛽10 + 𝛽𝛽11𝑥𝑥𝑡𝑡 + 𝛽𝛽12𝑥𝑥𝑡𝑡−1 + +𝛽𝛽13𝑥𝑥𝑡𝑡−2 ) 𝐼𝐼[𝑞𝑞𝑡𝑡 ≤ 𝛾𝛾]

+ (𝛽𝛽20 + 𝛽𝛽21𝑥𝑥𝑡𝑡 + 𝛽𝛽22𝑥𝑥𝑡𝑡−1 + +𝛽𝛽23𝑥𝑥𝑡𝑡−2) 𝐼𝐼[𝑞𝑞𝑡𝑡 > 𝛾𝛾] + 𝜀𝜀𝑡𝑡 (1)

where 𝑦𝑦𝑡𝑡 is the dependent variable;

𝑥𝑥𝑡𝑡−𝑗𝑗 is a vector of predictor variables, lagged j period(s);

𝑞𝑞𝑡𝑡 is a threshold variable;

𝛾𝛾 is a threshold value;

𝐼𝐼[𝑞𝑞𝑡𝑡 ≤ 𝛾𝛾]] is an indicator function that is equal to 1 when 𝑞𝑞𝑡𝑡 ≤ 𝛾𝛾 and equals 0 otherwise;

𝐼𝐼[𝑞𝑞𝑡𝑡 > 𝛾𝛾]] is an indicator function that is equal to 1 when 𝑞𝑞𝑡𝑡 > 𝛾𝛾 and equals 0 otherwise.

We are testing the null hypothesis 𝐻𝐻0 : = 𝛽𝛽1𝑖𝑖 = 𝛽𝛽2𝑖𝑖 𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖 = 0, 1, 2,3. If the null hypothesis has

been rejected, then the threshold effect has been established. The threshold value 𝛾𝛾 can be found

by estimating equation (1) though finding the minimum one of the sums of squared errors in a

threshold variable. Under the null hypothesis, the distribution of the p-value statistic is uniform,

and this transformation can be calculated through bootstrap.