 So I guess this paper is titled Authority and Delegation in Online Communities. And what I try to address in this paper is one important question that many platforms face, which is how platforms can incentivize user-generated content. So as you may know, many platforms relies on online communities of users that participate voluntarily. So this is the case of Wikipedia, for example, but many also question and answer websites, product forum support. In all these platforms, there is a community of users that participate voluntarily and provide the content. And this content is useful for the company to survive and to be successful. So there's a lot of literature studying what type of motivations bring these people in and how the platform can incentivize participation. So this paper I'm gonna study one specific channel that could be used by platforms to incentivize participation, which is the delegation of authority over specific tasks that users can make within the platform. And more specifically, I'm gonna look at editing tasks and I'm gonna be specific in a minute. And I'm gonna use for this purpose data from Stack Exchange, which is a question and answering website, well known for computer science questions, but is very much used in many other topics of Q&A. So let me go more specific. So in online communities, mainly people can do two type of things. Either they can contribute with some content. So it can be articles, for example, in Wikipedia, but in the case of Stack Exchange, these are answers that people can provide to other fellow users that have asked some questions, right? And then the second main action that they can take is to edit the existing content. So this editing task is much more diverse across platforms and in particular is diverse the way people can do this task. So in many websites, editing is just a suggestion. So basically people provide a flag, a comment about some content and then other users or managers of the website will be in charge of deciding whether the suggestion is valid or not and then implementing the suggestion. On the contrary, other websites like Wikipedia allow user to directly modify the content, which means that if I'm a user in Wikipedia, I just find that some content is wrong, I directly modify it and publish my modification. So in this case, I said that people have full control and authority over the action of editing because they don't need to wait for a third party to actually confirm that modification and that edit. Okay? And then there is a third type of platforms like Stack Exchange, where people can gain full control and authority over this editing task, but this is conditional on some performance measure. So basically the platform set an observable measure that the pattern can control and once users achieve a certain level of performance, then they are able to fully implement their edits without waiting for third party to approve. So clearly this decision, the pattern has to decide how to allow user to make edits and the objective of the pattern is to find the ways to incentivize users to produce content and to edit content. So they may want to incentivize both content production and editing. So the question is how delegation over this action, so more control over this editing task may indeed help the platform to achieve this purpose and to incentivize participation. So let me explain what type of incentives delegation may create. So consider a situation in which people participating in a community can gain some reputation points. So in Stack Exchange, people answer some questions and then other community members will upvote these answers, for example, if they like it or if they find it useful and then these upvotes will provide points to the author of the answers. So the author will accumulate reputation points and then once a rich at threshold T, then he gains this authority I was mentioning. Okay, so just to recall, before they reach this threshold, people are still able to edit content on Stack Exchange, but these edits are just suggested. Other users or editors have to approve the modification that the user is suggesting. Once they reach this threshold T, then they are able to directly implement the edits. Okay, this doesn't mean that this content can be edited again by others, but they have full control over the actual implementation of their edit. So this creates two possible incentives mechanisms. One is what I call dynamic incentive effect. If users value to gain authority, so if users value to reach this higher level of authority over editing, then they may be willing to put more effort before the threshold to reach this threshold faster because arriving to that threshold will give them specific higher utility. Okay, so and I call it dynamic incentive effect because if people discount time, this incentive is stronger when people approach the threshold. At the same time, there is another type of incentive effect that I call static incentive effect. So if people value to have authority, like if people specifically have a higher utility by making action when they have more authority on those actions, then to provide more authority over editing would relax the participation constraints on some ways affect that participation constraints and in some ways affect the level of participation of the users. Clearly these two incentive effects are maybe conflicting because if I move this threshold up and down, I may shrink or increase the size of these effects. So in this paper then, what I will do is first of all to identify what type of users are acting in the website. And this is mainly because we know that in platforms, there is a lot of heterogeneity of what type of user participate and incentive systems may be differently affecting different type of users. So my first point of the paper is to identify what type of users are out there. And then I will use a dynamic discreet choice model to estimate those preferences for gaining and having authority. In other words, I will try to use a dynamic model to infer how much people value to gain and to have and then how much they are sensitive to the dynamic incentive effect and to the static incentive effect. Okay, so this is only achievable with a dynamic discreet choice, with a dynamic model and not with a reduced form evidence because of the clearly the time at which users achieve this threshold is endogenous to their decision process. So we really need a forward-looking model that incorporates the decision of the users. That's why I'm using a structural model. I will estimate these preference parameters for each type of user that I identified in the first step so that to understand really who is sensitive to what and potentially which type of users the platform is targeting once implementing one incentive or the other. And finally, once with this preference parameter estimated, I will simulate counterfactual analysis. To, I really understand for the platform from the platform perspective, what is the trade-off that the platform is facing when pushing on the dynamic incentive rather than the static incentive by shifting the threshold level, right? So the idea here is that the platform can move from a Wikipedia extreme case to a Quora extreme case and then we can decide how much challenging side to this threshold by and in this way affecting the size of the static and the dynamic effect. So this is gonna be the purpose of the counterfactual analysis. So the literature that I touch on this paper is three-fold. So first of all, I touch the intrinsic value of authority to the territory that studies experimentally how people may value to help control over decision-making. And in this case, I will kind of confirm these are experimental results with using non-experimental data. I then contribute to the organizational economics with territory by bringing to the empirics the theory that studies how delegation can be used as incentive device. And finally, I contributes to the information system literature that studies non-monetary motives and participation of users in online communities proposing this like starting this novel channel of incentivizing participation. So in this paper, I will first, so for the rest of the talk, I will first show you what type of data I use and how I identify the type of users. I will explain you delegation system in place in Stack Exchange, how I estimated the preference parameter through the structural model and the counter-factual analysis. So if there is any question, I can maybe answer now or no. There is no question in the chat, yeah. Perfect, okay. Okay, so let me go ahead. So Stack Exchange, as many of you know, is a platform that hosts many questions and answering website. And each of these website focuses on a specific topic of Q&A. So the most well-known website is Stack Overflow that is focusing on computer science and programming languages, the questions. But then on this specific paper, I'm gonna focus on the one called the English Language Learners, which is a question and answering website that focuses on question related to the use of English, specifically for foreign learners. The reason why I'm focusing on this website is that the text of this question and answers is mainly textual, so there is no code or other type of features that, and then it's more easy to measure quality of these answers with text measures. So, yeah, go ahead now. Sorry, there's now a question from Oren. Oren, would you like to unmute yourself and ask it directly? Yeah, sure. So, just a question about the courses of the model. So you talked about this fact that there's two tier and you talk about incentives of the users. I was just kind of wondering about the incentive of the platform to have this, this two tier system. So it seems to me like it would be maybe by identifying who are the high-qualityers and then letting them do the edit. So that means that the users that have editing rights are going to be not just in different senses, but are going to be at the type of users. So just wondering if the model kind of speaks to that as well. Yes, so you're mentioning about the idea of quality of users in the editing task, for example, or? Yeah, so I mean it seems that the platform does it because it wants to make sure that your comments or your edit makes sense before they need to do it, you know, without any limitations. So that would be... Yeah, yeah, so that's a good point. And so this is all at the point of, you know, like using the threshold rather than as an incentive device more like as a selection device, right? So this paper focuses more on the first role of selecting the delegation, like selecting the threshold, like potentially it's role in delegation, but clearly I agree with you that that could also have a selection mechanism that I still have not investigated in this paper. Great, thank you. Okay, so as I was saying, I'm using this data from this website for which I observe the full content. So all answers posted from 2013 to 2020 for all active users that are around 10,000 and of which I observe the full history of participation within the website. So how identify user types? So the reason, so one key idea of this paper is that I didn't want to identify type of users from actions that they make because potentially I wanted to observe exposed if the user types identified do differ in terms of actions. What I wanted to find, the idea of finding these types exactly was to recover from the data what the motives somehow that brought users to participate in the website. And with this idea in mind, the way I identify user types is by using information that they display on the user profiles. So I recollect all user profile pages of participating members. And after recovering information of what they displayed in that page, I use a data-driven approach combining a multiple color correspondence analysis with a K-means clustering to cluster these users in different groups. Okay, so this is really like a data-driven approach based on what type of information users display on their user profile page. So for the sake of time, I'm not gonna go in detail on the technique. But if you have a question, let me know. So the... Yeah, couple, sorry. At a high level, I have a question. At a high level, what information do you take into account to cluster individuals? Is it the text? Is it the pictures? I use kind of everything that the people can decide to display or not, which is the picture, which is the presence of a website, the presence of LinkedIn profile, the size of the biography and some text information about the biography they write, whether they have a full name or just a nickname and whether they have location, all this type of information that you're free to use or not when you register on the website. And the idea is that when I register on a website, I decide to disclose what I want based on what type of reason I'm registering. So after this process, I identify three type of users and naturally the identification of these groups is really based on what type of information the group display. So the first group that is the most numerous one, the largest group is what I call anonymous, which is a group that does not display information. So basically, as you can see, they mostly don't have information about themselves, they don't provide the LinkedIn or website information. There is then a second group of users, that is the second largest, which instead provide information, so normally they have a biographical description, they have a full name and website and so on, but do not have information about their external life outside the community. So they miss LinkedIn profile, external links and they have to less degree their personal website. And finally, the informative users are users that provide a lot of information about all dimensions, including information about their life outside of their community. So this is a bit how I identify these three groups and just to quickly give you an overview, these groups are very different also in behavior and anonymous users tend to be the one, so the informative users are the one mostly producing in the platform. So they are very few, but they participate a lot and they are having an important contribution to the final production on the website. And they are also the one that receive more medals and so on. So let's say that there is a correlation between more information and more activity and more productivity in the website. Okay, so let me move to the delegation system and what then variation I used that is affecting how much authority people have in the website. So as I mentioned before, people can accumulate the reputation points in this website and once they reach a threshold, they get more authority. So this is the main first variation that I use that allows me to compare users to histories before and after they reach this more authority level. And before February 25th, this threshold was set at 1,000 points. Anyway, the data allows me to use an extra layer of variation because this threshold has moved in the history of the website. And then I do observe not only people that gained authority before and after the threshold, but also people that potentially lost authority in particularly if they had between 1,000 and 2,000 points at the time of the change of the threshold. So these users that had passed the threshold under the previous and the first threshold set, then they lost authority because they had not yet reached the second threshold. So how the threshold affects user contribution. So the idea then of the structural model is to really try to study and to model the decision process that user make once they decide how much effort they make in the website. And the crucial factor is to incorporate the forward-looking behavior of users that are anticipating the possibility to reach this threshold in the future based on their current effort provision. So this is really modeled as a sort of investment decision with decreasing returns because people make effort today. This effort is gonna produce some reputation points tomorrow and then also in the next days after tomorrow about the decreasing. So this is how I model the investment decision of effort. And then this decision is really based on the idea that these points in the future may lead them to reach this higher level of authority. And clearly if people value reaching this higher level of authority, this preference is gonna affect their today action decision because they will be more willing to put higher effort and to pay higher cost of effort today to reach faster this threshold in the future. So the model then the structural model is like a simple dynamic discreet choice where the choice set is gonna be a discretization of the quantity and quality of effort in answering and editing. So more precisely I have three dimension included in the model like the A correspond to the quantity of answers, E corresponds to the quantity of edits and Q to the quality of answers. So for now, I don't include the quality of edits because since they are very small modifications in general it's hard to measure the quality. So I discretize these three variables and obtain a choice set of 21 options. And clearly combination of these options will lead to expected returns in terms of reputation points for the future. So I don't present here for the sake of time but in the paper I have the modellization of these returns and I estimated in a first stage the arrival of points in the future based on the all rate of arrival of points in the community. So people, users participating make this decision and choose what level of effort to make in these three dimensions such to maximize their expected discounted utility of their participation in the website. And again, let me stress this point like the key identification of the model is really like comparing the different options and the returns that each option is making is giving them in terms of points. So if people value and put high value in reaching the threshold tomorrow, they're gonna, I'm gonna observe today a higher effort level. If instead the people don't change significantly their effort level even if they have a chance to reach the threshold tomorrow, then it means that they put low value in reaching the threshold. So this difference across the, this different implication of the action today let me infer how much they value in reaching the threshold tomorrow. Yeah, people, Kevin has a question about if you can say more about measurement of quality so that Q parameter. Yeah, so I think maybe I don't have a slide. So the measurement quality that I use is to recover text measure of the text. So they generally I use the length of the text how many significant words so I discount from the number of words the number of stop words that are like non meaningful word. So I recover all these text measures and they regress these measures on the points that these posts have received on the first day of publication. And these give me an approximate measure of quality. Is that clear? So in particular, just to double check you don't take into account the votes from others to your answer. I take into account the votes on the first day of the answer. And I use some basically I regress text measure on the votes that the answer receives on the very first day of publication. Because then in the next day is potentially the answer is modified by other people. So I really wanted to capture the quality made by the author. So I'm just taking the exact text characteristics at the time of publication and I regress these variables on the number of up votes and votes that this answer got on the first day of publication and the predicted number of points is like the level of quality of this answer. Okay, so the user preferences that I assume for this model are the follow. So it's like a linear utility function that includes first of all, how many points people have at a given point in time. So in period T, R will capture how many reputation points people have accumulated so far. And then I include these two variables, C A and C E that capture the net utility from answering and editing. So this can be the cost of effort in answering and editing, net of some intrinsic utility of making those answer, those effort decisions. So what I want to display explicitly here is that potentially users may have some intrinsic utility from answering and editing which is then captured from the same variable of the cost. So this is gonna be a net cost of answering and editing. Then I include this variable C U M T that corresponds to the cumulative number of privileges that people can obtain on this website. So I didn't mention that people can get more authority but they can reach also either privileges in the website. And then this one to capture that potentially people just care about accumulating privileges and not specifically about authority. So that's why I want to include it in the utility function. And then I include the Adammi variable equal to one if the user has the reached authority at the time key. So this is gonna be exactly one if people have more authority and zero otherwise. And then I interact this dammi with a constant and with this net utility from answering and editing. C A and C E. And then the utility function is gonna be a random utility. So there is an interesting kind of preference shock, epsilon I T. So yes, I didn't specify but the way I model this net utility from answer is gonna be like a combination of the quality of the answer and the quantity on the answer weighted by the availability of questions. Because to answer is easier if there is a lot of questions out there out there that I can answer potentially but if there are very few potential it's harder for me to answer. So I wanted to capture that. And then I measure, I weight this the cost of answering by how many questions are available. Questions about the utility function? Okay, so, okay, no, I wanted to be more specific actually about that. So what are the parameters of the interest of this utility function? So first of all, this beta file that corresponds to the additional effect that the additional utility that answering will bring to the user once it will have reached this higher level of authority then somehow will measure the sensitivity of the user to the static incentive effect for answering. What I mean is that if this beta five for example is positive it means that people will be having a specifically higher utility in answering once they have reached higher authority on the website. If it's negative it means that it's becoming more costly for users to provide answers once they have more authority on editing. Similarly for beta six it will capture the sensitivity to the static incentive effect on editing. So if beta six again is different from zero significantly it means that the users will have specifically higher or lower utility in editing once they will have reached the higher level of authority. And then the dynamic incentive effect is captured by all the whole interaction without this dummy viable. Because clearly if this whole interaction is positive or negative then it means that the user will anticipate a positive or negative shocking utility once he will have reached the authority level. So this will impact his decision process before reaching the authority threshold because he's looking forward to this higher level or lower utility in the future. Okay, so if there is no question let me move to the results and estimation results. So first of all let me focus on the beta five which is the sensitivity of the static incentive effect on answering. So the estimates shows that only the anonymous users have a significantly different beta five which means that in general users do not, I mean, when users obtain higher authority on editing their utility in answering is not much affected. And in particular only anonymous users receive a higher cost of answering once they pass the threshold but the size of this effect is small. So this would suggest that when allocating authority on editing does not imply some specific negative utility on the other action. So there is no much spillover effect on the other actions that users can take. On the contrary, when people obtain authority on editing they have an increasingly higher value of editing. So as you can see for all users in this case and specifically for the anonymous users when they pass the threshold T then they receive a specifically higher utility in editing. So this suggests that they do have a specific higher preference of editing once they are endowed with authority on editing. And finally, the constant of the authority is kind of suggesting the direction of the dynamic the sensitivity of the dynamic incentive effect which shows that, so it shows that the anonymous users and informative users receives a positive utility from reaching the authority while identifiable users do not. So this already suggests that the dynamic incentive effect will be specifically relevant for the anonymous users. So let me be clear because of the size of the sensitivity. So to give a measure of how much users value to reach authority. So to say it in the word I used in the introduction the value of gaining authority the value for anonymous users correspond to around 250 points which is something that they could be obtaining with the three posts. And similarly for informative users the value that informative users put in acquiring authority is 329 points which is something that they will be able to get published in 28 posts. As suggested before, identifiable user do not really seems to be sensitive to that incentive they don't seems to care about obtaining authority so their value of gaining authority is quite low. So once I bring these preference estimates to the counterfactual simulation this will allow me to simulate how much production in the website will occur under different delegation scenarios. So in this slide I will show you just two types of scenarios. One in which the platform decides to not delegate authority based on performance but to delegate authority from the beginning to everyone. So this will be the extreme case of Wikipedia where people have full control over editing tasks from the very beginning of their participation. And the second scenario is instead more mimicking the stack exchange setup where people receive a full control over editing once they reach a certain performance threshold which is a here set of 500 points. So let me discuss the three types separately. So the most interesting types is the informative users. The informative users as I mentioned before are the ones that are the most sensitive to the dynamic incentive effect. This means that if I delegate authority to edit in editing from the very beginning of their participation they are not so basically I shut down the dynamic incentive effect system then I don't incentivize these users the green slash users, right? So the production in the website is much more slack. If instead I delegate on performance then the informative users are incentivized to produce more at the beginning to be able to reach faster this higher level of authority on editing. And this is why we see this spike of contribution at the very beginning of their history of participation. Differently the informant identifiable users as I mentioned before they are not sensitive to the dynamic incentive effect. So by shutting down the dynamic incentive effect I don't lose much of their participation. Finally. Yeah, for you have five minutes. Okay, yeah, thanks. Finally, you see this anonymous user on the zero line and this is because on the simplified setting of the simulation that the cost of participation for the anonymous users is too high. So in the simplified setting of the simulation they have just so high cost of participation that they don't participate. So even if they would potentially be incentivized by the dynamic incentive effect in practice they are not. For what instead relates to the edits I hear I want to talk about the static incentive effect of editing. So as I mentioned in the preference slide the estimate suggests that the people have a specifically higher value in editing once they reach at a higher level of authority on editing. This means that once I move to scenarios in which the delegation is delayed so I'm waiting time to, I mean people need to reach a certain threshold level of points to be able to have a full control over editing. Then clearly I lose editing activity because now people don't have full control from the beginning and then until they don't have full control they will produce less edits. So every scenarios that increase and delays the delegation would imply a reduction of the amount of edits that the pattern sees on its website. Okay so let me conclude the overall. So in this paper try to study how delegation of authority can be an incentive device and they try to investigate in particular two type of incentives. First the static incentive effect. So this is the idea that people may value to have authority and what the data shows is that indeed the people are more willing to participate in editing once they have more control on editing. At the same time doesn't seem to be important spillover effect on other tasks like answering. The other incentive effect that I study in this paper is the dynamic incentive effect that show and in the estimation shows that indeed people value to gain authority and in particular informative users and then they increase their participation in answering to be able to reach this threshold faster. As I said that this is in a way very much relevant relative to the type of users that the pattern is targeting. Informative users are very sensitive and they increase much their production to reach the threshold. But for example, identifiable users do not. And anonymous users, even if they would they have very high cost of production so they are not really responsive to the site use anyway. So just to draw some policy implication the optimal design of the website clearly depends on what type of users the platform is targeting and then on the composition of the community and then also on the objective of the platform. Clearly if the platform wants to maximize the quantity of edits on the website is better off in delegating from the very beginning because that will maximize the editing activity of the users. If instead the platform wants as well to maximize other tasks we may want to delegate the editing based on performance over other tasks. And this is the case of Spark Exchange. For example, Spark Exchange for sure cares about editing and improving the quality of answers but as Q and A's website are two sided platforms it really needs to incentivize production of answers because otherwise we don't have the dynamics that increase the production in the website. And then for that case, Spark Exchange may be a better off in trading off some editing activity to incentivize production of answers. And with this I conclude, thank you very much. Wonderful, you anticipated me. Thank you, Jacopo, Kevin, take it away. Great, thanks, Jacopo. Lovely presentation, you know, quite clear and the paper and the phenomena themselves quite rich and interesting. I'll have about a moment to discuss some of these things. I will disclose, I'm pretty, I'm somewhat biased towards the topic of, you know I've written on granting access and devolving control and looking at non-pecuniary motivations and their effects on network effects and industrial evolution in these contexts. So what I'd like to begin by stating is that I do believe this is a very important topic. It's a first order topic. There was a time when we studied platforms, systems competition in which we could model complementors as profit-seeking firms or in some sort of simple produced form. And over the past 20 years, really the many of the complementors on the on platforms, contributors on platforms have become behaviorally motivated, non-pecuniary motivated agents and certainly there's profit-seeking as well, but these are increasingly first order topics and what great scale and value in these contexts. A couple of elements of this I'd like to especially welcome. You know, I believe there's an element and this doesn't come out as much in the presentation as in the paper, but there's an attempt to kind of begin to frame this relationship between these complementors and the platform along the lines of what we've seen in the flavor of organizational economics and thinking about this as something analogous to a principal agent problem where there's a design of, in this case, control rights where that's, I think, a fresh and useful perspective that we can take. The industry organization buys us a lot of traction in these contexts, but I do believe the organizational economics perspective adds more. And also, although these are behavioral phenomena and in some sense difficult to fully model, I do believe there's no place for strong modeling in making sense of what's going on in these contexts. So these are all sort of important elements of the paper I wish to highlight. And just before, you know, there are only gonna be so many comments I can make, I would like to also say that Jacopo makes some efforts in the paper quite nicely to tie this to, you know, beyond platforms to literatures on delegation in their papers by Baker Gibbons, Murphy, Aguillon du Attrapin. There's this notion in that literature of a tradeoff in that agents somehow have some kind of special attribute. In much of that literature, it could be an issue of their privilege and information. Or in this case, we're looking at privileged or unique motivations that come through delegation. And I wish to just highlight that the tradeoffs in those literatures are in some sense, the decision to delegate or not, or under what conditions. And the heart of this paper is really a different kind of tradeoff. And I do think that, you know, I'll try to add suggestions as I go through this, but I think the paper could probably be clearer that in fact the tradeoff that Jacopo is intending to bring to the fore here is there will be delegation. This is not a question. There will be in some sense, the devolution of control to external agents to these complimenters. The tradeoff is whether to in some sense delay or create a threshold for delegating, or in fact, just to open it up more widely. And at the same time, which as you said, that even that is not sort of the first order goal of this paper as I read it. In some sense, the first order goal is really an establishment of existence that in fact the delegation of authority is a source of motivation that has economically consequential results. And there's an establishment that basic establishment of fact even precedes anything of describing this tradeoff. There's this question of, is there in fact an existence of this issue at all? And I'd like to just, you know, and I think the paper could probably go farther in distinguishing, you know, as we have all these thresholds that are being crossed, we're looking at one particular threshold that relates to editing privileges. And there could be any number of motivations related to crossing these thresholds or gaining these reputation points, as it were, it could be reputation and status. It could be in going through the system of various thresholds and gaining points. There's an element of gamification and intrinsic motivation. There are a lot of other possible explanations going on here. And I think sort of the core message of this being a motivation of achieving autonomy which has appeared in other literatures, whether that could be better established at its core. And I'll just make a couple of comments now on within the empirical analysis itself. In one minute, sorry. Yeah. The, there are, well, it goes straight to the structural model. The structural model gives a sense of dealing with endogeneity. However, it should be mentioned that, you know, in some sense it's allowing for there to be a distribution of responses to various variables, but there's enormous complexity in what brings somebody onto the platform, what makes them transition from sort of one observed type as it's defined to another observed type to what's leading them. There's just a lot of detail that I think is missing in the model itself which doesn't necessarily allow us to fully discern what is the source of the actual behavioral mechanisms and relating those to some source of exogenous variation. I'll share some more notes, but let's just do a quick kickoff of issues.