However, many have noted that tax amnesty also has some risks. It may instigate public opinion of similar forgiveness in the future (Leonard and Zeckhauser, 1987). It may also reduce government credibility (Stella, 1991) as people saw the weakness and lack of authority from the government to properly collect taxes (Uchitelle, 1989). While amnesty can generate short term revenue, the aforementioned risks make the impact of amnesty to future tax revenues remain uncertain (Marchese, 2014), or outright detrimental (Das Gupta and Mookherje, 1995).
Gauging the Post-Amnesty Effect
As is known, tax amnesty granted several facilities to participating taxpayers. Article 11 para. (5) of Tax Amnesty Law stipulates the annulment of tax arrears, cancellation of administrative penalty, as well as termination of audit for participating taxpayers. While the number of terminated audits were quite numerous (21,177 audit notice were terminated, 12,365 instruction/assignment were also terminated), the amount of settlement of prior notice of tax assessment reached around 74.3% (19.4 trillion rupiah was paid to settle 26.1 trillion rupiah in unpaid taxes including the administrative penalties) (DGT, 2017). Granted, those audits may result in bigger amounts in tax collected were there no tax amnesty. Hence, ex-post testing of tax amnesty's probable counterfactual effect remains difficult.*
On the other hand, the filling compliance ratio in 2017 reached 70.98%, which is an increase from 61% in 2016 (DGT, 2017). Those who report via electronic filing (e-filing) also increased from 59% in 2016 to 70% in 2017 (DGT, 2018). Among tax amnesty participants, 870,000 tax returns were filed in subsequent fiscal year, which correspond to 89.4% of filling compliance ratio –the highest compliance rate of overall taxpayer (DGT, 2017). In terms of money, tax revenues realization grew 4.07% in 2017 compared to previous year (13.75% if non-routine revenues from amnesty is excluded). It further grew to 14.33% in 2018 (15.53% if amnesty is excluded), which is the best tax revenue realization in the last 5 years and the first double-digit growth in the last 7 years (Kemenkeu, 2019). Given this cursory view of the data, it may seem that Indonesian tax amnesty does not significantly impact post-amnesty compliance.
Alm and Beck (1993) had once tried to measure the impact of amnesty to post-amnesty tax revenues. They tested whether the 1985 Colorado tax amnesty affected subsequent years. Similar methodology is employed here.
The Models
Three models are used based on Alm and Beck (1993). First, a simple time trend analysis testing based on OLS in the form of
Where, Yt represents monthly net tax revenue collected by DGT, T is the numeric representation of the month, e represents the error term, and b0 and b are parameters. Additional specifications for a separate intercept or slope change from the amnesty based on periods within and post-amnesty is coded A and its interaction with time is coded A*T. Tests are separated in three ways: full period from January 2014 to March 2019, pre-amnesty periods from January 2014 to June 2016, and post-amnesty periods from July 2016 to March 2019.
Second model employs Autoregressive Integrated Moving Average (ARIMA) by Box and Jenkins (1976) which tests whether the process that generated the tax revenues before amnesty was the same as the process that generated the tax revenues after the amnesty. Additionally, the identified ARIMA processes from before and after the amnesty maybe used to “fit” and “forecast” the tax revenues.
I employ Augmented Dickey Fuller and DFGLS test to find the integration of tax revenues in full period, pre-amnesty period and post-amnesty periods. All three of them are not stationary at level and stationary at first-difference. So first differenced data will be used. Further ACF-PACF and Godfrey tests suggest the existence of autoregression of order 1 and no moving average. So the ARIMA specification will be (1,1,0) instead of (2,1,0) as used in Alm and Beck (1993).
Thirdly, Multivariate ARIMA (MARIMA) intervention analysis is applied. In this case, we introduce three type of intervention assuming that tax amnesty is a discrete “intervention” thus can be represented as an additive effect of the amnesty on revenues (Box and Tiao, 1975). This requires the specification both of a starting point for the intervention and of the shape of the intervention impact. Similarly, ARIMA specification of (1,1,0) will be used here.
The starting point for the intervention (or the amnesty) is simply the time at which the amnesty occurs, which in our case is June 2016. The shape may be modeled by a “step” function with zero values up to the point of the intervention and one for all periods following the intervention, by a “pulse” function where the intervention occurs at one period and the intervention variable has just one nonzero value, or by a “ramp” function in which the step is spread over some period as a ramp response. To sum up, the shapes of intervention may be crudely depicted here:
Results
Turns out amnesty does not significantly affect post-amnesty revenue. A and/or A*T are not statistically significant, as well as partial “chopping” of time in January 2014 – June 2016 and July 2016 – March 2019.
In our robustness check using ARIMA below,
it could be construed that the autoregressive process that generate tax revenues pre- and post-amnesty are quite similar, in terms of sign and coefficient. Indeed, confirming this assumption using Chow breakpoint test and forecast test at July 2016 (when amnesty started) both result in no difference of pre- and post-amnesty revenue (p-values of 0.6776 and 0.2546, respectively).
Lastly, using MARIMA:
either pulse, step, or ramp intervention does not result in statistically significant outcome.
If I were to predict post-amnesty tax revenues from pre-amnesty data, and compare the predicted outcome and actual outcome, this is the result:
You may see that the actual value is actually higher than the predicted value, which confirms the remarkable growth of 2017 and 2018 tax revenues.
Conclusion
It seems that the 2016-2017 tax amnesty in Indonesia does not lower the tax revenues of subsequent years. Yay!
Reference:
Directorate General of Taxes (DGT). (2017). Tax Amnesty Mosaic
DGT (2018). Kepatuhan dan Penerimaan Pajak 2017 Tumbuh Pesat, DJP Optimis Hadapi 2018. http://www.pajak.go.id/kepatuhan-dan-penerimaan-pajak-2017-tumbuh-pesat-djp-optimis-hadapi-2018 (Accessed March 20, 2019)
Kementerian Keuangan. (2019). APBN Kita, January 2019 edition
Alm, J. Beck, W. (1993). Tax Amnesties and Compliance in the Long Run: A Time Series Analysis. National Tax Journal, Vol. 46. Pp. 53-60
Box. G. E. and Jenkins, G. M. (1976). Time Series Analysis, Forecasting, and Control. San Francisco Holden-Day
Box. G. E. and Tiao, G. C. (1975). “Intervention Analysis with Application to Economics and Environmental Problems”. Journal of the American Statistical Association 70 (March 1975) pp. 70-79
Das-Gupta, A. and Mookherjee, D. (1995). Tax Amnesties in India: An Empirical Evaluation. IED Discussion Paper Series, 53. Boston University
Leonhard, D. Zeckhauser, R. J. (1987) “Amnesty, Enforcement, and Tax Policy”. NBER Working Paper No. 2096
Luitel, H. S. dan Sobel, R. (2007). The Revenue Impact of Repeated Tax Amnesties”. Public Budgeting and Finance 27 (Fall 2007): 19-38.
Marchesse, C. (2014). Tax Amnesties. Papers in Comparative Analysis of Institutions, Economics, and Law No. 17.
Stella, P. (1991). An Economic Analysis of Tax Amnesties. Journal of Public Economics 46 (3): 383–400.
Uchitelle, E. (1989). Amnesty Programs in Selected Countries. FRBNY Quarterly Review
Monthly tax revenues data is obtained from internal data. Net total tax revenue is used.
------
*It is possible, given microdata of corporate and self-employed taxpayers which either participated or did not participate in tax amnesty, then conducting a simple ANOVA, ANCOVA, or difference-in-differences analysis in a quasi-experimental research setting.
