 Welcome everyone to this session on improving tax collection, which is a very important topic as we all know, and I'm pleased to say I see so many people here. Improving tax collection has become even more important, I would say, due to the COVID-19 pandemic, and we need to increase tax collection to reach the sustainable development goals. So that's pretty much the point of departure for this session. My name is Simon Björneru, I work with the Norwegian Agency for Development Cooperation. We support the Univider Domestic Research Mobilization Program. The aim of this program is to help improve developing countries' tax system and to strengthen developing countries' domestic capacities to increase revenue collection. Please go to the Univider website to check it out. Improving tax collection is both about policy and administration. It's about changing the rules and better enforce the ones that already exist. Both are equally important. Today we will mostly focus on enforcement, including how this could be improved by better utilizing automation and technology. We have three distinguished presenters with us to shed some light on this topic. They will present the papers 10 minutes each before we open for questions towards the end. You can ask your questions in the Q&A, or you can also take the mic to ask the questions yourself. First, I will ask Ronald Weiswe to come on stage to present the highlights of his paper. Ronald will talk about the potential that lies in simple tax regimes, tax administration innovation and education programs. Ronald Weiswe is a supervisor for research and policy analysis at Uganda Revenue Authority. He collaborated with ICTD on several research projects in Uganda and Ethiopia on issues including taxing and wealthy individuals, public sector agencies, and improving tax compliance. He has also collaborated with Univider on research projects including tax compliance of presenter taxpayers and tax benefit micro-simulation modelling. So please, Ronald. Thank you, Simeon. Because of time, I will go straight into the presentation. It's been a collaborative study between myself, Maria, from Renew Wider, and my friend here, Mille, who is the ACRPD, and here we try to assess the effectiveness of two administrative interventions that are targeted at improving tax compliance of small businesses in Uganda. So in terms of motivation, in developing countries largely, there is a large informal sector. And because of weak institutions and low social norms of paying taxes, there is a challenge to mobilize domestic revenue, domestic revenue researches, which are needed to support development and be less dependent on age. So for most of the African countries that have established presumptive regimes for small businesses, which are commonly known as presumptive tax, now presumptive tax is a tax on turnover. So there are no allowable deductions. So you just declare the turnover and you take a percentage of the tax, the percentage on the turnover. Now we have administrative interventions that you have implemented over time that we assess in this project. And in this study, we assess how effective they have been in terms of increasing the tax base and in terms of increasing revenue corrections. We have three questions we answer here. So one is how has the new simpler e-finding system been effective in increasing the number of taxpayers, and how effective has been the taxpayer register expansion project called the name the trip in increasing the number of taxpayers. So we established the effect of these two administrative interventions. The new e-simpler firing and then the taxpayer register on the number of taxpayers are registered. And then two, we're trying to assess how these reforms have impacted the revenue and for trip we assess how effective, cost effective it has been. Now the trend of taxpayers, presumptive taxpayers over time, this is how it has taken. You, for instance, see that from 2011 it is increasing marginally and then we see a sharp increase from 2015 to 2016. So we try to assess what are the implications of trip and what are the implications of e-finding on this jump. So briefly about trip, trip is a collaborative project between Uganda Reven Authority, local government agencies and the registration below for companies. The work in collaboration reach out to the bigger business community, mainly the small traders. We just start companies in a collaborative environment. Then the new e-firing system is, is, is, is web based. Now originally they were fighting the no more traditional income tax return which is complicated and based on Excel templates. Now in July 2015 at the URA, Uganda Reven Authority rolled out any simpler return form which is web based for presumptive taxpayers. So what the presumptive taxpayers are required to do is to just specify the tax identification number to specify the allocation and turnover and then the system automatically computes the tax for them. Then in terms of the data and methods, we use tax administrative data from here at Uganda Reven Authority returns for presumptive from 2012 to, 2012 to 13 to 2017, 18. And we also use returns for corporate income tax, especially returns for, for tax payers and corporate income tax in the range of 150 and 400 millions, equally small tax payers for comparison purposes. And then we use simpler impact evaluation methods to analyze both the impact of TREP and new fighting separately using the difference and difference approach. Then in TREP, we consider both treatment and control are presumptive tax payers. So because TREP was implemented in phases, so for instance, it started in Kampara, which is the capital city of Uganda. So the control for that are tax payers that were outside Kampara. So both the treatment and control groups are presumptive tax payers. But now when we come to e-firing, the treatment group is the presumptive tax payers and the control group are corporate income tax payers that are equally small just above the presumptive threshold of 150 million to 400 million. So in terms of our main findings, we see here the impact of TREP on the left. And we see where we see TREP 1, TREP 1, TREP 2 and TREP 3, TREP 1 is because TREP was first rolled out in Kampara, which is the capital city of Uganda. And then after the first year, it was rolled out to other bigger towns in Uganda. And then TREP 3 is when it was rolled out to other smaller, smaller municipalities across the country. Now in terms of the impact of TREP on the number of tax payers, as you can see TREP 1 in Kampara, it had a significant impact on the number of tax payers that were newly registered. But then when we split the after TREP period, because in terms of methodology, we had TREP implemented, but there were different methods that the project was using. We had the one-stop border posts. So the one-stop border posts is where services could be built from one point. So you are A, under collaborating distribution, set up an office where all the services could be picked from one point in time. So we see that where you see TREP 1 after, in column 3, for instance, we see that the one-stop border posts, when the collaborating distribution set up an office where all the services were provided in one point, it had a significant impact, a significant impact on the number of tax payers. So it was a better method to implement, to increase the tax payers size. And that is also observed in TREP 2 when it was rolled out to other municipalities. It did have a positive impact on the number of tax payers. And again, we see that when the one-stop border posts in column 6, when the one-stop border posts were implemented in these municipalities, the impact was still much stronger. Now, in TREP 3, the one-stop border posts were implemented at the same time when TREP was rolled out to those municipalities. And that's why we have one column here, which is 7. And we still found a significant impact of TREP on the number of tax payers. So in terms of expanding the tax payers base, TREP had a significant impact, and especially the one-stop border office, one-stop border post. Then in terms of new e-finding, you remember the control group here is cooperating income tax payers that were fighting the traditional income tax return. So here we find it having a positive impact also, significant impact on the number of tax payers. And we see this even after. So that is it in terms of number of tax payers. Then when we check out the two interventions, we find that the effects of the interventions are complementary. So TREP and the new e-finding system came into effect around the same time. So possibly complementary effect because reforms impacted the same group of tax payers. On TREP's objective was to educate. One of TREP's objective was to educate tax payers, not only to formalize them. So maybe it also had implications on e-finding. So we then try to understand who are these tax payers that we actually built on board. So we see that while most of them, most of the new, we got a big number of new registered tax payers, most of them were very small. And you can see here, we're branching at almost the first threshold. It's the threshold here runs from 10 million to 50 million. That is the first threshold. And we see most of them were branching. So we did not register really significant or high value tax payers, unfortunately. Then in terms of revenue, the results are here. TREP 1, TREP 2 and TREP 3, like I discussed before, we still see an impact on tax payables of TREP. And again, we do see the same impact when we put in the one-stop center office. One-stop center office is where I said all the collaborating institutions set up one office and where all the services were implemented in the same place. You could go in one office, you'd just have a business with the Organic Registration Service Bureau, you get a trading license from the local government, and then you pay your taxes in one office. So we still see it having a significant impact here, even in these municipalities where there's TREP 2 and then where there's TREP 3. So same applies to e-finding. We find it having a significant impact on the tax payable. Then in terms of analyzing the cost-benefit analysis of TREP, here we use simplified back of the envelope calculations. We calculate the average additional tax revenue per industry, by area, per year, from TREP 1 and TREP 3. And then we compare this with the URAS budget for TREP and calculate the average expenditure per industry, per area, per year. So we see that on average the additional revenue was approximately 14 million and the average expenditure on TREP was roughly 1.5 million. So therefore we find that TREP was actually cost-effective. So in terms of conclusion, both TREP and the new e-finding system increase the number of tax payers and the tax revenues. Two is that the interventions however complemented each other and the largest effect was after the establishment of one stock shops in TREP. We find that the benefits outweigh costs in terms of costs. Thank you so much. Thank you very much, Ronald. Perfect timing, 10 minutes short. So that was very interesting. And next I will ask Amina Ebrahim to take the virtual stage. Her study discusses how administrative interventions help lower the extent of tax avoidance and evasion. At the same time, it says something about whether technical assistance financed by donor countries has an impact and evaluations like that are rare to come back. Amina is an economist holding a PhD from University of Cape Town. She is a research associate at the university where she is a core researcher in the domestic resource mobilization program. Her research includes labor and public economics. Her work focuses on making large administrative tax data available for research and the research collaboration with African revenue authorities. So Amina, please. Thank you, Simon. And thank you, Ronald, as well for your presentation. I'll actually make some references to it here as well. To just double check, you can see my slides. Yes. Great. Okay. So today I'll be sharing some highlights from a study and maybe one of the first studies of more to come in Tanzania. And this one specifically is on tax examinations. The study was conducted in collaboration with colleagues at the Tanzanian Revenue Authority and the University of Dar es Salaam. But in addition to the collaborators mentioned here, we are grateful to colleagues at the TRA in assisting us with the data extraction and some of the data work and for those at the Finnish tax authority. So in terms of the context for this presentation and the study is what we think is the core focus of any tax administration is broadly twofold, expanding the tax base and sort of what Ronald's not been talking to us about and then enhancing tax compliance while at the same time minimizing tax evasion. I think Ludwig is going to talk about that next. But also to maximize the revenue collection. Now to further this, tax authorities have increasingly been using technology as a powerful tool. And in this case, we're going to talk about technology as a tool for risk assessment by revenue authorities. So this pilot introduces a risk-based non-ordered approach for identifying which tax payers to examine. So the pilot study was a collaboration between the Finnish tax authority, Vero, with the Tanzanian revenue authority. And this is part of a broader technical assistance program. So we're presenting to you some evaluation of the pilot study, but there are other efforts of that collaboration. How does the pilot work? The pilot introduced an AXL sheet to be used at tax officers in the Dar es Salaam region. That flags tax payers for an examination based on a risk assessment. So you have a tax officer who enters information based on the tax form, enter this AXL sheet and at the end of it it says whether or not the tax payer needs to be flagged for an assessment. The pilot was introduced in July 2019 and it was only introduced in the Dar es Salaam region. And it was designed to test the effectiveness of risk-based non-ordered control measures. The goal of this research study is to examine the impact of the treatment or the pilot on incremental tax revenues from firms subject to this examination. And what we'll show today is that there has been an increase in adjusted taxable income collected. So interestingly, how do we calculate this effect? It's actually very similar to what Ronald's just shared with us. It's also a difference in difference approach and quite similar in the sense of some of the areas that we've implemented, how the pilot was implemented to take you through this. So we use a difference in difference model to estimate the impact of tax examinations on revenue. And what this method does, it compares the change in adjusted revenue between the before and after implementation of the project in the pilot area, which is our Dar es Salaam region, and areas that were not run in the pilot, so outside of the Dar es Salaam region. And so this method is commonly used and provides reliable evidence about the impacts of certain, if certain key assumptions hold. The idea, the simple idea behind it is that in the absence of the policy, the revenues in both areas would evolve in the same way. And so for those who are interested in the equation, the Dar, let's see, you can see that. So the Dar region is, that's our treatment variable, post is our time variable, and the interaction between our treatment and our time is what we're interested in. So we're interested in this coefficient Delta. We use data administrative firm level panel data that's collected by the Tanzanian Revenue Authority. So this is data that contains about 25,000 observations each year. And this is for the period 1st of July 2015 to the end of June 2020. And the data includes information both in the region we're interested in, in Dar es Salaam, but also includes areas around that. And I guess the interesting thing to note here, and I'll bring it up a bit later again, is that we have data now on this one year period of implementation of the pilot study. So what we do in the first instance is inspect the trends in the just-attacksable income between the two groups. So what we check is the treatment group year is firms in the Dar es Salaam region, and that's represented by the solid line, and then in the control region. So that's outside of Dar es Salaam, and that's your dotted line. And then this period before the implementation of the pilot, we understand that this is mostly parallel. And so what we're looking for is if there's any difference in the period after, and it looks like there is some divergence there. So we have some indication of what we think might happen. So what do we find? We don't find any increase in the number of examinations. So the pilot hasn't actually flagged a lot more firms than they previously would have, but we do find that the pilot led to a statistically significant increase in additional reported income. And we look at various ways as we look at the taxable income, we look at adjusted income, and then we look at some extra income. And then depending on the model, we see basically between a 10 and 15% percentage increase depending on the modeling strategy. When we try and control for the type of firm or the firm in the industry, and then even if we cluster our standard areas at the firm level, our results remain the same, quite similar. And we also find that while this was implemented both for corporate income tax and in personal income tax, the effects mainly driven by adjusted corporate income tax. And the impact is arising predominantly from the server sector. So where does this leave us? Well, we've been able to show that the pilot led to an increase in reported income. The study is only conducted in the first year that the pilot was available. So we don't know how long this effect will last and how implementation continues. Are tax offices continuing with us? How do they find using it? And I think that would be an important question going forward. However, at the moment, the TRA are investigating implementation in other regions. But the implementation might change because they've also been a change in the way in which corporate income taxpayers file their taxes. And so they've started moving towards e-filing. And this may again bring about this push towards using ICT for these risk-based examinations. So you'll have something that's somewhat more automated instead of an Excel sheet where you have to manually input all the information. But the main message here is that there is potential for cost-effective enforcement improvements that can be devised using tax data in a novel way. And this is even in a low-income setting. And potentially very important as part of a plan to improve tax collection and move towards this post-pandemic period. So this research is not yet available online, but it should be within the next month or so. And we're hoping to publish this alongside a research brief. Thank you. Thank you very much, Amina. It's very, very interesting. So let's go to the last one. Ludwig Wir. He has written several papers on tax avoidance and evasion. He has among several things shown the potential that lies in digital tax enforcement. So Ludwig is a post-doctoral researcher from Berkeley, and holds a PhD from the University of Copenhagen. His research and teaching focus on international taxation, inequality and development. In that context, he is also working as a consultant to the UNI wider IMF and the National Treasury in South Africa. He has previously worked as a public sector government management consultant in Boston Consulting Group and as an assistant economist at the Danish Ministry of Finance. So please, Ludwig. Hi there. My name is Ludwig Wir, and today I'm going to be speaking about my practical insights from IMF and UN missions on how to use advanced analytics in tax enforcement. Now the basis of my talk today will be the experiences from my missions in five different African countries with the IMF and UN, where we promoted the use of advanced analytics to improve enforcement both in domestic taxes and in customs. The overall lessons from all of these missions in my view is that advanced analytics can be easily implemented, and that big data is already there, it's already available to the tax authorities in all of the countries I've seen, and it just needs to be used. Now we went into all of these missions with the target of improving compliance, taxpayer compliance, while keeping the amount of resources constant, and the circle of course that we're trying to square here is that there are lots of taxpayers out there, but there's only limited resources in a tax authority, so we want to make sure that the audits we do, they really count. And how are audits then currently being done in 99.9% of all tax authorities? Well right now the way it works is that each tax authority sits down and designs a set of taxpayer risk rules manually. So a common risk rule can for example be that you compute the taxes paid as a ratio to gross sales for each individual business. Now if that ratio looks far off, there are much less taxes being paid as a ratio of gross sales than in comparable businesses, well then you assign a certain risk score to that firm, and you can do the same using other ratios or looking at the ratios of taxes paid this year compared to last year, etc. The key point here is that all of these rules are created by human beings manually based on intuition, and at the end of the day you then assign these risk scores to all of the taxpayers and you audit X% of the returns based on your results. Now this is the status quo, where we want to go is to a situation where all of these taxpayer risk rules and risk scores, they are not manually recorded by human beings, they are instead designed automatically using all available data and one way to do that is for example to use a machine learning algorithm or other fully data-driven algorithms where you simply tell the computer we want to maximize the return on our limited amounts of audits and examinations or other interventions, and please computer go and find a way for us to do that we don't need exactly to understand why we just want this equation to be maximized. That is the utopia where we... Now the benefits of automated data-driven case selections are fairly obvious, but let me still go through them. So first of all we're sure that all available data is used not only what we as human beings can grasp and equally important we ensure that the variables of interest from this enormous set of data that is available to a tax authority, the variables of interest are chosen by an automated model, not a human being, meaning that we allow the algorithm to see patterns that human beings cannot grasp. So I gave the example before with the ratio of taxes paid to gross sales, that's a meaningful statistic to me, but of course we want to ensure that all of the most important relationships and all of the most important variables they are weeded through a computer looking for all possible patterns. Now another benefit of moving to automated data-driven case selection is that you can continuously and in real time update your risk assessment. You are not restricted to waiting for a tax auditor to have time to sit and update the risk rules, the computer can do it all the time. And finally I think a benefit of having automated data-driven case selection is that firms and citizens know that all of their transactions are constantly being watched and evaluated not only in a case where in order to have the time to audit you which is very rarely in the real world. Now I've sketched out here how the process of moving to data-driven audit selection looks like and it's fairly simple. So on the left we start by importing and merging all the data sets that we can find. So we start with the tax returns, so in all the countries I've seen that's going to be in a digitized format, so we have that already, there's going to be tons of useful information on that, but we can merge that with the customs data which is almost always also digitized. The bill of lading information is extremely useful, so we merge that with the tax return data using the taxpayer identification numbers. Then we go and look for more fringe data sets, so that can for example be a registry of land titles or water usage, etc. Whatever we can find we take that and merge it into the other data. And once we've done that which is the biggest assignment we go ahead and transform the data, so that means doing tabulation, summing over years or whatever we think is useful and it could also mean computing networks amongst taxpayers which previous research has found to be a meaningful statistic when looking for tax cheaters. And once we've done that, imported the data, merged the data, transformed the data, well then we've really done 99% of the work. So actually going from that to building a predictive model that predicts the likelihood and the consequence of auditing a taxpayer, that's going to be a matter of minutes and that's a situation where when we go on these missions we've actually built the code for doing the predictive modeling, we've built it from home, so this literally takes a matter of hours to set up. And once we have the predictive model done and running we're going to have a list of taxpayers that have a high risk and high consequence, so taxpayers that we want to go and audit and we put the model into use and we keep an open eye towards whether the model was right or wrong and whenever the answer is we go back and tell the model such that it can improve the algorithm and we get this continuous loop of deployment and learning. In 2016 the OECD did a report on looking into the implementation of advanced analytics in rich countries and they found that the number one reason for advanced analytics not being utilized in revenue authorities in rich countries was first and foremost a lack of trust in the algorithms producing correct results. Now the obvious solution to this is of course to go ahead and do rigorous impact measurement. The OECD also found that the algorithms would often produce non-actionable intel, so the solution to this is that you make sure that not only does your algorithm tell you that you're going to go ahead and audit this taxpayer, but it also tells you why the algorithm thinks you should go and check this taxpayer, meaning that the auditor actually knows what to look for and finally what the OECD found is that there was a broad resistance to change from auditors who felt that their skills become redundant when an algorithm goes and replicates their job and for that to change you really need to go and work with the softer values in the form of change management and I'm not going to go into that today. Now let me just briefly address how you can approach the issue of rigorous impact measurement. So there's the first stage where do simple statistical testing, so you can do all of these well-known methods where you build a model using a training dataset and you test it on a separate dataset, but that's still going to be statistical testing which is hard for non-economists to trust. So what you really want to fairly quickly move to is infield real-life policy evaluation where you do an actual RCT, a horse race between the old manual risk modeling versus the new computer algorithm and one way to do this that we have advised countries to do is to for example select 300 cases, the best 300 cases that your algorithm tells you to go ahead and audit and then randomly assign these 300 cases based on the new models to auditors and inspectors. Now you ensure that the auditor or inspector does not know whether it was the new risk model that was used to select who was audited or whether it was one of their colleagues and this allows you in an unbiased way to have a horse race between the new and the old case selection. Now if the model is successful it's easy for you to upscale it to 500 cases than 1000. Now a lesson that we also try to convey on this mission is that in the long run we're all dead right, so we want to aim for rapid pilot implementation then learn from that and repeat the exercise. The aim here is not perfection but improvement and what we try to stress is that each day we make wrong choices in a tax authority and these choices are partly informed by poor or missing modeling. Now if we can build a model that improves risk targeting compared to current procedures then we are wasting taxpayer money but by not doing it right now. So we want to deploy our model in small scale, benchmark it with the current alternative and then if successful in a larger scale. Now I'm very briefly going to speak to a model that we built for container examinations. Now what you see here is a completely standard picture of how case selection looks in both rich and poor countries. So you see that for on the y-axis here we have the yield of the container examination, the size of the upliftment and on the x-axis you have the ranked number of examinations and we see that very few, only 7% of the container examinations result in a large upliftment while the bulk 73% of examinations deliver no. Now with predictive modeling we were able to change that ratio a lot. So we proved in a pilot model that we could do 60% of current upliftment with only 10% of the resources or we could do 100% gain 100% of current upliftment with 33% of examinations. So in essence the model could either be used to facilitate trade by examining way less containers or increase the total amount of upliftments that is to do the same amount of inspections but with a higher return and that was all from me. Thank you so much for listening. Well thank you very much all of you. We don't have that much time but I hope that we all learn something about the potential that lies in automation and technology. I don't think there is any silver bullet in the improving tax collection but I think to be sure that this is an ICT avenue to pursue. I see that, I'm not sure if I get this right but I think this one question in the chat to Ronald. Ronald have you seen the question? Is it a technical question? Okay so yeah I saw it and what I can say is yes we did some robustness checks but mainly they did with checking if our results are consistent. So what we did we used the different control groups especially when we're trying to investigate the impacts of ifriding where we investigated that using different control groups especially people that were in the corporate income tax category but and within the same threshold of presumptive 0 to 150 then we also used another control group of 1 million to 150 then 0 to 400 and the results were not changing but also for trip we used the number of small corporate income tax payers and the results were small and also positive in terms of the other square value I think we need to check it again that we didn't but we won't check again. Thank you. Thank you. Kauti have you seen any other questions where I can't There's now one question in the Q&A section from Jukka Pirtle. Can you see that Simon? Yeah so Jukka is asking in love income countries a small number of firms are responsible for a large share of revenues should all these firms be selected to examination irrespective of what prediction model suggests? So anyone feel like they want to answer this? So I think it's a really good point Jukka and I think you'll get to this conclusion maybe quickly if you just do a simple probability times consequence right that if as you as you note there are some countries where one firm you know even if they they just evade a percentage of their their total tax bill then that's going to be more than the entire tail so in these cases it makes a lot of sense to order them no matter what might be. I was asking thinking there the technology you have been talking about today is it a kind of technology that can be employed by like all different countries or is it I mean I guess it has to be adapted to local conditions and local capabilities but I'm not sure if it's is it advanced or is it how is it? So what we usually do is that we we start by automating what the tax authorities already do and we've been to also very poor countries and they are usually you'll have a setup as as Emina was also describing. So the first thing we do is that we help them automate this process of assigning risk scores based on manual rules and I've yet not seen a place where we couldn't assist in doing that and when you've done that then it's actually a fairly easy step to then use more advanced modeling right so whether you do the risk scores manually or you have a computer helping you do them that's actually the smallest part of the equation. So the short answer is yes I think using advanced risk modeling can be useful for all the countries I've at least been to. Thank you very much everyone and I really enjoyed the presentations and I people should also read the papers and I think they're out maybe except Amina's not yet but we should share them and then you can sit in quiet and read them. So thank you everyone for taking the time to join this this session and have a nice day.