Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=

# Managers and Offices: Location-Based Determinants of Bureaucratic Effectiveness in India

Authors:   
Discussant: None  
Video: https://www.youtube.com/watch?v=PHYYOyrSItw&t=8600s


## Talk (02:23:20 – 03:10:19)

[02:23:29] Thank you everyone. uh I'm going to talk about uh uh other side of what effect managers can have on an organization from the public sector. Uh this paper is joined with Prajit from from UC Berkeley and Perminder from Indian Administrative  
[02:23:42] Service. Uh Aditya my co-author is also here and Matri Punjabi who is a PhD candidate at Michigan State and going to be on market next year. So please keep keep an eye if you're looking to uh recruit in development and and  
[02:23:54] environment. So let me start. Um we we try to uh unpack state capsity at at finer scale at micro level. Uh variation in in government performance uh across countries and also across like uh subnational uh like you know states or  
[02:24:09] like provinces is well documented. So here's a quote from um from Amaren and Shandre from their book is that some parts of India uh are more like islands of California in a sea of subsaharan Africa. So, so one one bit of data that  
[02:24:24] can clarify what what what they mean is that like they they mean two three different things but at least one of those is that uh for to get like public amities or government services uh you will see you see large variation for example to get birth certificate which seem like a very kind of like a  
[02:24:38] important document for for for for many purposes uh it the government may take 15 days in in Madhya Pradesh a northern state and and three days in in in Telangana southern state likewise these kind of government performance  
[02:24:51] differences also So transferred to actual outcomes such as infant mortality you will see large variation like 12 time variation between Kerala and UP for comparison like that variation in US is is three to four times. So u we uh we  
[02:25:05] attempt here to focus more on the variation in state capacity at finer scale like uh across districts uh uh subdists and and individual offices below subdists and and try to understand  
[02:25:18] its its approximate causes. So uh health and education those type of outcomes are are more usual proxies for state performance. Uh but they kind of sit downstream in the sense that like they are shaped by demand uh education  
[02:25:32] income um and also like they they materialize over long horizon. Uh it's also very difficult to attribute to a single office when these outcomes are affected by multiple schemes and programs. Uh states routine output is  
[02:25:44] service delivery. uh think about uh uh marriage certificate uh you know uh you know birth registration or or driving licenses or armed licenses and they are often produced by identifiable officer and staff. You can look at like who  
[02:25:58] signed the the certificate, who approved the certificate. Uh they are measurable at office level. They also are like highly frequent. So you can compare uh across different offices because the same service is being provided um uh in in in the whole region. Um governments  
[02:26:13] know that people value and they we also confirm that in our our our data that people value uh getting these services timely. So government wants make to make sure that offices are delivering it timely and they impose deadline. Um so  
[02:26:27] in in this context the data we are using uh the service delivery speed for these kind of certificates registrations uh that provide a clean proxy to to measure government performance and and and at at at like different office levels. Uh and  
[02:26:42] we are interested specifically in that like what drives the variation in service delivery across different offices. So the map here uh you can see a Punjabi state from India where we have a sub subd district shown there and uh this large variation uh in how much time  
[02:26:57] the government local offices take in processing application for cast certificate. Cast certificate is an important document to get uh benefits through affirmative action. Uh and there are large variation we find within a state uh in our data.  
[02:27:12] So one way and there can be multiple different ways uh but one particular way to explore these drivers of variation across uh you know different government offices is to look into manager versus office. Uh so manager here who provides leadership uh who runs the office uh  
[02:27:26] they are selected through an exam. They have a higher skill set. They also are incentivized to put more effort uh they have much higher incentives in terms of salary and and career progression. Uh and then then office here basically means everything else about the office  
[02:27:41] primarily the the staff uh who we observe are relatively immobile they are not transferred so frequently but also resources technology norms culture of the office organizational culture and and political economy factor. Uh manager is a responsible uh you know uh  
[02:27:56] government employee for service delivery. Uh in our in our our setting they are called approving authority. So if anything goes wrong they will be held responsible and staff reports to the manager. So they do all the ground work before these certificates or these these  
[02:28:10] licenses are issued and they report to the manager. Uh we do note that the offices may constitute many things. Uh as I mentioned earlier culture or political economy factor resources uh given our data we are able to focus more  
[02:28:22] on staff. Um and and then we are going to show show more results on that. So there are three questions. Uh first we ask uh how large is the office component related to the manager component in explaining these variations across  
[02:28:36] across in in government performance across government offices. Uh we here use uh egovernance data uh that uh allows us to measure office productivity across different offices in Punjab. We have more than 10 million observations here. Uh basically each observation is a  
[02:28:50] citizen application. Um and uh we observe productivity separately for the manager as well as staff because each application when citizen applies for a certificate or service uh that goes through a process and we can see the time stamp how much time a particular  
[02:29:05] person uh government employee took in processing that application. So that provides us a like a rich information on performance of these actors on on on this citizen application. uh we use AKM variance decomposition uh approach to  
[02:29:19] identify the primary source of variation like separating it between manager and and office specific effects.  
[02:29:26] Second uh question is uh we use this eventary design to see that how much managers do like they are able to affect the office productivity. So think about it like a like like a school principles like when a new principal is transferred  
[02:29:40] to a new school uh like whether the new principal will change outcomes significantly or the outcomes are defined by like how teachers work and what is the culture in that school. So uh here inferring causality is is is difficult like unless we can have a you know random experiment assigning  
[02:29:54] managers. So we leverage this transfers across different office locations and using a movers design uh we measure productivity change around around this manager transfers. Uh lastly we we asked this question how how do office and manager respond to exogenous and  
[02:30:08] permanent shock. So in collaboration with the state government we we ran this statewide RCT where we uh experimentally changed u um randomly changed deadline for some of these services which government is providing to citizens and we compare uh how this shock affects  
[02:30:23] manager and staff performance and specifically we look at that how whether office performance and staff performance whether it it it correlates with whether they have a good manager or high quality manager or low quality manager.  
[02:30:36] So here's preview of main findings. uh how large is the office component related to the manager component in explaining this variation across government offices. So we find that office effects dominate uh office accounts for about 83% uh of combined  
[02:30:50] manager office variance about five times of uh of variance due to the manager component and roughly 60% of total uh and how much do managers causally affect office productivity. So using the event  
[02:31:03] study design uh we find that like managers do not change destination office productivity at all. Uh in fact on transfers manager's own productivity also converges to uh destination office uh productivity in in a in a in in a in  
[02:31:16] a in a significant way and support staff time remains invariant to uh to the uh identity of the you know supervising manager. So these transfers are not really helpful in improving office productivity. That what that is what we  
[02:31:29] see in the data. And lastly from the RCT we see that this permanently permanently tightened delivery deadlines uh they did improve productivity in low low performing offices as one would expect but the performance improvement is  
[02:31:43] regardless is orthogonal to uh the manager quality. So that further confirms the point that uh validates the point that managers do not matter so much in uh government performance. Give me just one second. Let me just finish  
[02:31:53] overview. Um and um and yeah, so I this is the last slide I can I can yeah uh I can I can go very quickly. uh we we we we basically contribute this literature on government efficiency across  
[02:32:07] countries, subnational uh you know uh regions like uh uh there's the there long list of papers here like I think the particular paper which is very very very u uh I think um uh interesting like where people co-authors send this uh mails to different countries 150 plus  
[02:32:21] countries and try to see that whether the mail sent to an unexist non-existing uh address is is is returned uh to the to the to the to the sender or not. uh we here are more focused you know on the decomposing this variation in government  
[02:32:35] performance at subnational level. So two closest papers are one phenia 2022 econometrica and uh Michael best and his co-author's paper in AR in 2023. Uh they also do a similar kind of AKM decomposition. I think our advantage  
[02:32:50] here is that we can de we can we can see the performance or productivity of manager as well as staff using this timestamp data. So that allows us to say more on on on this uh role of personal in government performance. Uh you know  
[02:33:03] there's lot of papers on this. Uh we we we we try to focus on the staff versus manager and like our main finding here is kind of one contribution that we empirically show that staff inertia at the micro level. uh that is approximate source of this uh variation in  
[02:33:17] government uh performance across locations and and then we have this experimental evidence evidence on a office and staff both respond to this this shock. Yes, please. So uh my my experience is limited to the police but um if I'm not wrong I think you must be  
[02:33:32] getting most of your identification of the manager effects from managers who remain in this role for a long time. Uh and my concern would be that at the two extremes of productivity, you're getting people transferred to completely  
[02:33:45] different roles. So if someone proves to be totally incompetent in this particular type job, they may be assigned to like the archives division and then at which point they leave your data set. Or if someone proves to be incredibly competent, they're potentially assigned to state  
[02:33:59] headquarters with some job that gives them greater responsibility. So I might be a little concerned that in some sense your results that are showing relatively small scope for managerial you managerial contribution in some sense are limited because you're getting  
[02:34:13] selection out on both sides of the extremes right no that's a valid concern and I'm going to show you some evidence on that like we don't find evidence here that managers are transferred primarily on performance so think about this government setting that uh people are  
[02:34:25] transferred on like lot of different for lot of different reasons uh before elections after elections they are more transferred that is number one. Second thing is uh managers sometimes have electeno like you know government in general like top bureaucrats do not want  
[02:34:38] a person to stay in one place uh for a long time that that has its own political economy implications. Uh there are also some some uh some transfer because of like say house household family hardships that people want to move to a city where like their their  
[02:34:52] spouses and and and other reasons. We in our data do not see that performance are like you know like what you're basically indicating to like is some sort of matching like a assortative matching like where good managers are being sent to a a bad location bad office or or  
[02:35:06] vice versa. We do not find much evidence on that in our data. Yeah.  
[02:35:10] >> Yes. Mushri. >> For the first half of your research you were using um the word speed and um quality pretty interchangeably. But speed is not the only metric of quality like accuracy might matter for even the  
[02:35:24] example you gave on cast cards right so like how should we think about that sometimes speed might even be inversely related to quality right >> that is true so uh we we have some evidence on quality uh and it's not it's  
[02:35:38] kind of quality adjacent I would argue uh in the sense that like if uh if you are forced to work very fast uh you can either make errors or you can basically in this in this case officers have this option to send the application back to  
[02:35:51] citizens and so that counter can restart >> or or you can be responsive to bribes >> or you can be responsive to to bribe so we don't observe bribes for sure and and uh uh uh but we have some evidence on quality I'm going to show at least with  
[02:36:04] the RCT uh you know sample I yeah we have some results on that yeah we like just like we do not find that making offices run faster affects uh the send backs to citizens uh significantly.  
[02:36:18] Okay. So, uh now let me go in detail of setting and you know these results one by one. Uh Punjab is a state about 30 million people. Um and it's a very large it has very large professional uh state  
[02:36:32] bureaucracy. Uh they they started uh egovernance platform more than a decade ago. Um we focus on these five major department that process about 90% of citizen service applications. um revenue, health, uh social security, uh  
[02:36:46] rural development and local government. And uh as I described earlier, these are like often like different type of certificates, licenses um which people need to get like uh you know services from the government like for example,  
[02:36:57] you need arms license um um to to to have an ownership of own uh any kind of gun and and weapon. Um so one particular act here is highly relevant for us. the Punjab Transparency and Accountability  
[02:37:10] Act in 2018. It mandated uh statutory deadlines for specified services. So deadlines vary uh broadly uh widely uh like for simple certificates is deadline seven business days. Uh for more complicated certificates where like  
[02:37:24] again armed license where police verification is required it may take up to 45 it allows the government officers to take up to 45 days. Um so this this act also provides provides standardization uh with the egovernance data. uh it complements uh basically the  
[02:37:39] whole e-governance setup and it also helps uh measurement like how offices are performing and enforcing uh you know these rules more more more clearly uh it also like provides this deadline specifically as well as e-government data that comes from this act in a way  
[02:37:53] like you know helps us uh you know also design the experiment later uh processing time we argue is is a meaningful productivity measure uh we looked into citizen grievance data uh which is on an independent platform um like looking At a sample of 10,000  
[02:38:08] gances, we find that about half of them concerns the services we study. Um and uh in and for them like the most common complaint is a non-dely uh of the service by by by deadline. Uh deadlines are binding even even though we see  
[02:38:22] about 20% of applications not meeting the deadline. So this is something which officers do want to make sure but still like not in all cases and like basically that explains that like how much variation is there across like offices performing well and and and uh not so  
[02:38:36] well. So here is an example like how we should think about the setting like suppose uh in step one a a citizen uh you know requires um certain services. So uh they apply at a service center uh that  
[02:38:50] service uh uh that that application is sent to the office clerk uh who reviews the application and forwards to the manager. Manager thinks about like you know what all information is required to basically issue the certificate or service or not. uh if required the  
[02:39:03] manager will send it to the field staff who will go and verify like for example cash certificate or resident certificate or even like armed license uh it will require some field staff to go and inquire about it and once the information reaches the manager manager  
[02:39:17] can take take a you know a call on whether to uh reject or approve the application and issue the certificate.  
[02:39:22] Uh in this case cast certificate uh the deadline is about 17 days and in this example the real example we see uh the the office took about 16 days.  
[02:39:32] So who are our key actors here? So when I say manager versus staff. So managers are basically middle level bureaucrats uh like they are recruited through state uh government exams uh and they undergo like a lot of training. They have a lot of incentives for promotion. They have  
[02:39:46] like a whole ladder where they can you know eventually like promoted to higher levels uh in the government. Um and uh they are responsible for the same set of services within the same department.  
[02:39:57] They are frequently transferred. Uh so in contrast to uh the the highest cider in bureaucracy the Indian administrative service um uh the the the tenure here is much shorter. We see about 10 months average tenure u earlier documented uh  
[02:40:11] you know work by by uh IR and money uh they they talked they they they found about 16 months of tenure for is so we have a shorter tenure here. Um and we we we we see that like most of these tenur these transfers are driven by rotation  
[02:40:25] norms and political economy and not so much on performance which um you know I'm going to provide detail later. Yes team yes >> they they are definitely agents uh and  
[02:40:42] we don't have uh information on them. Yeah. And it is it is possible they facilitate u uh processing of applications faster by like asking for bribe. Uh yeah but we do not we do know  
[02:40:54] one thing that these offices do not face citizens directly. So uh these applications are submitted at a service kiosk which are one arm length away from from these offices. So there's some some some differentiation there but I I we  
[02:41:08] cannot remove the possibility that people are bribing to get the services.  
[02:41:15] So we have uh total about more than 100 services across five departments uh more than 10 million uh these uh timestamped applications. Uh we can split staff time and manager time using this uh time  
[02:41:26] information. Um total about uh transfer records for about thousand managers across 800 different offices. Uh we we focus on lateral transfers rather than promotions but results are very robust to that too. uh we we basically collapse  
[02:41:41] the data to manager weak label for AKM and movers design and manager service weak level for the RCT because the RCT basically uh is is a service level where we we change the deadline for different services.  
[02:41:55] >> Yes. >> So yeah yeah yeah. So I think in this in  
[02:42:07] this uh government system uh even though of these managers have to care about the service delivery on time uh it is very unlikely that making very faster service uh you know is going to affect their promotion. Promotions  
[02:42:22] usually are on based on number of experience uh number of years of experience and at a very like a regular schedule in the bureaucracy in India. Uh so uh they do have to care they have to respond to higher ups. If there are a lot of delays they have to do more  
[02:42:35] paperwork they will uh like the district magistrate may uh basically call them in a in a regular meetings but it is unlikely to affect uh their their their promotions um uh immediately. Yeah. Yes.  
[02:43:02] because you know that they can managers do  
[02:43:15] >> um so I I like we so we do not have data um on that like if they can uh uh like punish the staff or incentivize the staff to to better like what is going on within office between managers and staff  
[02:43:29] we do not see see much on that. So so that is one but you're right uh Alexandra's paper was uh like they have this mechanism where like managers can encourage staff who are close to retirement and encourage specifically the non-productive staff to retire  
[02:43:42] quickly so that they can recruit new people who are more productive in into the team. Yeah. But we do not see see much data on that. Yeah.  
[02:43:53] I mean my my my personal opinion on this is that like managers who are transient they have a limited uh you know like limited means to affect an office where there is a union and staff who are much more permanent. You have local people who live there and who have been  
[02:44:07] permanently there for a long time. If you're coming to that position for 10 months or or 12 months you're not going to uh basically jol the the part too much. Yes.  
[02:44:28] uh we we we we uh we we we include controls on that and does not that does not change our our result. Uh but you're you're right like so so some urban versus rural differences and all. So this this part I think we need to do a  
[02:44:39] little bit more on that. Yeah. Yes. FEMA >> related with the way you're talking about the office effects it's about the staff and they're entrenched and I guess I'm just wondering if there's just variation in  
[02:44:53] whether offices are underst staffed so the applications per staff member is that going to load onto the office of >> we use vacancy rate as a control in our regressions yeah so we we we do have information on that yeah >> but you're assuming that like the staff  
[02:45:06] allocation even if there's a vacancy is efficient like that you know that that the Yeah. It'd be nice to show us that >> Yeah.  
[02:45:16] >> Yeah. There just like that that there's modulo vacancies that they've allocated the right number of staff for the volume of work.  
[02:45:24] >> Okay. >> Yeah. I I don't have it right now, but yeah, we we can we have data on that.  
[02:45:28] Yes. >> Yes.  
[02:45:39] >> That office backlogs. Yeah.  
[02:45:45] >> Yeah. So, we we have a large enough window. You can see when I when I present results that it doesn't appear that manager moving to a new office uh his productivity or the office productivity goes down because there was a lot of backlog we we see pretty flat  
[02:45:59] uh you know effect there. Yeah. I I will get back to you when when I when I present the movers design result.  
[02:46:08] Okay. So, so our key outcome measure is the log of processing days. Uh how much time uh these uh like uh these uh applications are being processed first by the the office as as a total but also  
[02:46:21] we break it down by manager and staff. Uh uh like average is about 7 days. Um and uh uh and these are like processing days like legally meaningful because they basically managers like there's a stipulated deadline for this. They're  
[02:46:35] also high frequency. We see these services being you know pro like provided by these offices uh you know weekly basis like like a lot of applications and uh this is also as I earlier mentioned that citizen salient that citizens care about uh getting  
[02:46:48] these services on time. Uh on average we have about 47 weeks a manager uh stays in a in a given tenure um and um about 100 applications per week being being processed. uh about 11 days it takes to  
[02:47:02] process uh in like if you just take take the average not geometric mean uh and most of this is uh basically taken by the staff and relatively less time like and it there are two ways to to think about is that managers uh have less amount of work to do they just have to  
[02:47:16] verify and then approve it but also the other way is that uh managers have more incentive to make sure that these applications are are being processed on time that's why they take less time Yes.  
[02:47:30] >> Discretion do citizens have about which office they can go to? So for example, if I know this office is really bad, they're really slow, I I'll avoid it or do they have to go to their jurisdiction? I I think for some services they can get it from any office  
[02:47:44] but uh but for some services they have to stay in local office because think about it like if the resident certificate you apply in a district which is far from your district uh that is going to be rejected because of whatever like you know like the the the bureaucrat can write whatever comment  
[02:47:58] there because they are not able to verify directly. uh they can submit to different kiosk anywhere uh for most of services but the kiosk will transfer you to the the concerned local office and also it is very costly for you to go to other location just to get a service  
[02:48:12] when you already have a have an office in the subdist there. Yes.  
[02:48:37] Sorry, sorry, I didn't understand. You are saying there it's a kind of discrete like in a sense that like one you get one application at a time or you get a get a lot.  
[02:48:52] I see. Yeah. Uh you so in a way that if a staff is not working properly they will send incorrect applications to the manager  
[02:49:06] and um and uh and then manager will basically take more time on this right.  
[02:49:11] uh we we do I mean I think the evidence what we see rather is uh uh like when when when I just move they're also like the staff is also changing for them right like I have a staff in one office but I move to a new office I will have a  
[02:49:24] new set of staff right uh there we do not see such like uh clear evidence on that but I think we need to probably do more to understand uh what kind of corrections were were done send back rates do not respond so that is another  
[02:49:37] another way like if the send like if there are a lot incorrections uh like done and the manager will send it back to so we can definitely see the manager sending back to staff but for the purpose of regression we are we are aggregating it but if manager has to send it back to citizens we can we can  
[02:49:52] see that there's no impact on that to the  
[02:50:11] Yes. Because Yeah. Yeah. Yeah. Yeah. Thank you. Okay. So, uh here's a like uh our approach about this AKM variance decomposition which is like you know as you probably already know  
[02:50:25] that like it in the labor econ it's is a it's a it's a it's a it's a primary method for understanding why wages are different when uh and compare comparing like employee component and the firm component. we we bring it to uh this this data uh for on government  
[02:50:38] efficiency. So we have basically log of the number of process dish taken by manager I in a in a in a week t and we decompose a manager fixed effect and location fixed effect and include a set of controls which can basically uh you know explain like any any local factors  
[02:50:53] in that geography as well as staff vacancy uh etc. U there there are two main issues with the AKM variance decomposition. First is this the key assumption is that the error term in this equation uh that has to be orthogonal to the the whole uh you know  
[02:51:07] transfer history of of managers. So that is a relatively strong assumption. So we we we are we are aware of that. We show some evidence on that like uh managers do not start like in in the data like manager who are being transferred it's not like their performance is getting  
[02:51:21] worse or are getting better. Um and the second problem with this is limited mobility bias that we do not observe each manager in each office. So uh so that is that is that is uh because of that the plug-in variance components are going to be applied by us and there are  
[02:51:35] there are methods used in the literature uh we we we follow that um so here is one issue with this like if uh we don't have this for AKM uh estimation uh the largest connected component is important  
[02:51:48] because uh these effects uh manager and office effects are are jointly identified within this connected component. So we we basically end up uh using the largest connected component which has about 1500 managers uh across  
[02:52:00] 600 offices. Um and there are like two offices where we do not see manager transferring transferring from one type of department to other department and these two offices like remain isolated.  
[02:52:10] We also check for robustness with a with other approaches where uh you know multiple connected component with a smaller threshold and u and department specific uh largest component.  
[02:52:21] So here's the main result with AKM decomposition. Uh we see that the the the office variance is about 6 uh and that is about five times of uh of manager variance manager fixed effect variance. U so in in total office  
[02:52:35] explains about 60% of total variance. In contrast manager uh component explains about 12% of total variance and um and and uh manager and office effects are essentially uncorrelated. uh once we correct for the limited uh uh mobility  
[02:52:50] bias. Uh if if if we uh if you see the sketch plot here um so it does not like provide some evidence that  
[02:53:03] like manager are being uh you know uh selecting like which offices to go to etc. uh we also see that like manager have relatively like even manager's own time on processing these application they are affected by office fixed effect in the sense that office share is about  
[02:53:18] 66 uh it is more than what you will expect for the total but still like it's a it's a like manager has some some some control on their time but not as much as one would expect >> wasn't his question about the with  
[02:53:33] within your sample like how is this a a balance panel presumably not and I yeah it' be helpful to see the exits from the data set you know like was that correlated with the  
[02:53:46] >> exit of officers or managers or staff >> yeah I I don't know as much as Dan does but that like you know that they're not they're no longer in this cer >> okay yeah I I mean my my inspiration which is coming from the police is that  
[02:54:00] the main motivation for at least police officers is a fear of punishment is something really bad >> and this seems like kind of a mid-tier type job uh for these state government officials. I I could be wrong about that. Surely there's some worse position  
[02:54:13] out there that maybe the ones who really really annoy the local MLA or are incredibly incompetent get transferred to, let's call it the archives department or something like that, right? So, how many are going to archives and potentially there's some much better job like in charge of the  
[02:54:27] local transportation department uh where they the really most competent ones or the best connected ones are going to uh so that type of entry and exit the punishment and reward positions would be interesting.  
[02:54:39] >> I I think we need to look more into it and if we can present some results on like you know correlation between those between these two factors right yeah yes Yeah.  
[02:55:09] Um so I mean this this is not a choice in the sense that if you want to do like this AKM decomposition uh like basically these effects will be jointly identified only for within connected components  
[02:55:23] right so uh I think I think the solution to this would be uh we can compare some of this performance from the broader sample the full sample And from this comp uh uh this uh component uh this connected largest connect connected component I  
[02:55:38] we do have basically uh data on uh like we do compare uh in the paper I don't have on slides but we do have like how this uh performance as well as some of the characteristics are different from this samp this this this sample versus the full sample and we do not see much  
[02:55:53] much difference There we have multiple connected component  
[02:56:10] results. Yeah. So yeah, we use that. Um yeah, I maybe I can talk later if Yeah.  
[02:56:17] about this. Yeah. Please go ahead. No, we we we we we check for robustness  
[02:56:33] with uh you know just in labels as well. Um and that doesn't make difference.  
[02:56:37] Yeah. Okay. Good point. I I don't really remember but it's possible. Uh yeah, we can we can check that and and put that.  
[02:56:48] Yeah, that's true. Okay. Uh one more thing we do with this  
[02:57:26] AKM result like we decompose these variants of like this office fixed effect further into uh between districts um within districts and within tasil and uh we find that like this this most of this sits within district within  
[02:57:40] subdist. So basically think about within a subdist there are multiple type of offices providing services and this variance is primarily coming from like between office variation within subdist.  
[02:57:49] So that confirms our point that it is not so much because of the location like a district versus this district versus that district is that more because of the office specific things which are going on even within a subdist.  
[02:58:03] Um we look into the sorting. So as I promised earlier the the the correlation between alpha and s like the manager and and staff fixed effect is negative before we correct for limited mobility  
[02:58:16] bias. And once we we correct for uh for for this uh uh using uh this fixed like getting this fixed effect from separate regression where we have because we we have this outcome for staff time and the manager time uh this correlation becomes  
[02:58:29] positive. So but even even then uh like even even even negative or positive the the amount the the magnitude is very small. So that does not um uh you know provide any evidence in in favor of positive or negative sorting.  
[02:58:43] So on Yes. like you you mean here right  
[02:58:55] the the within within subdists >> you predict >> I mean uh yeah I I don't remember like we I think we did with urban ruler and there was  
[02:59:09] not much difference in urban ruler but we we probably need to do more and and put the result here too. Yeah, but when you think about til will usually have a little more semi-urban area and the rest of this is going to be  
[02:59:22] like yeah and most of these offices are in like you know in the in the center of the tile right like where you have the headquarter of the tessile so in that sense uh you are not going to see variation between super rural and super urban and these offices are also being located in a selected areas where like  
[02:59:37] you can get more higher population density and you know that's it uh so on mover line. Um we we we have the same outcome and here our kind of a treatment variable is basically difference in productivity at the  
[02:59:50] destination office uh with the with with the with the uh you know original original location with the manager is coming from and u and uh uh and in this equation like we have like u yeah the time and and manager fixed effect. So the main interpretation here we should  
[03:00:05] we should care about is on this uh uh you know uh TA uh if TA uh in the like right after the transfer if that is one what would mean is that uh the office productivity um despite the transfer it does not change and it converges  
[03:00:19] basically to destination office and if it is zero it means that manager is able to carry some of his or her managerial ability to the new office and and and is able to bring change there. So, so key assumption um um the same like this  
[03:00:33] transfer timing should be orthogonal u um and we we allow for this like a one time shock like if you transfer to a new office this your your productivity is going to go down um u like it doesn't really matter too much uh if we include  
[03:00:46] it or not include but we we we take care of that and uh we don't see much big I think it's it will be obvious with the main result as you can see here that uh before u a manager is transferred his performance does not like uh you know uh  
[03:01:01] you know vary too much. It's not like good performing manager like if it there was a dip there before transfer one can argue that uh managers are selected based on the good performance or vice versa and once they are transferred the productivity jumps to the destination  
[03:01:14] office right away. We see the same for uh office staff time as well as manager's own time on these applications. Uh and uh both basically uh tell us the same story uh that office staff time of course doesn't change. uh  
[03:01:28] that is the predominant like the main main factor uh uh describe like explaining the total time but also manager's own time uh is kind of converging towards uh what the destination office productivity is and that that basically confirms that like  
[03:01:41] they are constant uh which are office if office specific which the new manager is not able to overcome.  
[03:01:48] uh results are robust uh um and like with the wider windows that that like that provides like more more assurance that like this is not one-time shock and due to backlogs etc. um we um we also  
[03:02:02] see I think the fourth fourth bullet here is more interesting uh that we see symmetric effect uh when moves from high productive office to low productive office uh see immediate drop in manager productivity and vice versa. So that like further adds to this point that  
[03:02:16] managers are not able to bring that you know that change one would expect that a new manager is coming a new a high quality manager is coming and they would be able to change we don't see that uh and the third part uh of the of the paper about the RCT so we we uh  
[03:02:30] basically permanently shortened deadline for randomly selected uh this high volume services we basically uh fixed somewhere between 60th to 75 75th percentile in the pre-treatment processing time so on average we reduce the deadline for the services by by  
[03:02:44] about 3 days roughly uh 16% of the you know pre RCT deadline uh we we have a relatively small sur because we were focusing on the on the on the like you know high frequency sources so 30 sesses total 15 treated and 15 control um and u  
[03:02:59] and and we because we have a like large uh pre- period data so we we uh we can we can use different here we don't see any main effect of the deadline change and one would argue that if your office was already doing very well so we large  
[03:03:13] variance like the top 10 percentile offices were able to process an application in less than half of the time as the bottom 10 percentile offices. So if your office in the time top 10 percentile if anything uh you would probably you know uh uh take a break and and see that like you already  
[03:03:27] doing so much well and the deadline has changed it does not really affect you.  
[03:03:30] So uh on average we don't see much effect like not significant effect but once we uh once we you know bifurcated by low uh uh productive office and high productive office we do see that the low productive office are getting affected  
[03:03:42] because of this deadline change. Yes. >> Uh it's it's still over right?  
[03:03:51] >> Yeah. >> Yeah. Yeah. So I Yeah. So we uh so you can think about this like if there are like multiple services the managers are providing uh in the treated services their deadline has changed then you will probably be slow on the control services  
[03:04:04] right what we rather see in the data is that uh managers are becoming like the offices are becoming faster the low performing offices are being faster on the control surfaces as well so in in in that sense like it's >> time variation is causal right  
[03:04:18] >> yeah so here uh is one one cut of this result. Uh when we basically look at the low office productivity depending on like whether there it's a high quality manager and low quality manager defining that whether the manager and office  
[03:04:32] quality is uh one standard de deviation above the mean or below the mean. We see that the most of the effect is coming from the low offices like they are reducing their time taken by 3 days but this effect does not vary by whether they have a low manager or high manager.  
[03:04:47] So that further assures that uh that like manager quality does not matter so much. Uh most of this is coming from the staff time. As you can see here manager time here responds a little bit.  
[03:04:57] Managers do respond to this deadline change but overall uh this effect does not really change um by whether whether we have high quality or low quality manager. On the other hand, um if you look at that like if you are in a high office, uh managers here uh like um uh  
[03:05:11] high office uh in in the second row as well as last row uh managers do reduce their uh you know time uh time taken even further in response to deadline but staff actually swing on the other side.  
[03:05:22] They are starting to take more time because the new deadline is is is still like providing a lot of a lot of room to even even you know sh a little bit. So when staff realized that they were doing much better here even despite this new deadline already they are they started  
[03:05:37] taking even more time than um compared to before.  
[03:05:41] Here is event study uh results with that like how treatment effect varies in low lower versus high quality offices and and uh I think the results are very persistent after first few weeks on the quality I as as I mentioned  
[03:05:55] earlier like we we observe the sendbacks like where man staff and managers can send the the application back to citizens and that will reset the counter uh so it will not show up in their delay rate. So that was a big concern uh for the government as well when we when we were you know collaborating with them  
[03:06:10] and we do not see much effect on send back. So it's not the perfect quality measure we we know that but it at least provides some evidence that like they are not trying to manipulate uh the system in some other ways to just be fast to just being uh seen faster on on  
[03:06:24] the dashboard. So on spillover's point Simma mentioned uh so we see that 74% of manager process both treated and control services we uh initially of negative spillover that will be concerned because then we will  
[03:06:37] be you know in the in the RCT results we see positive spillover which basically uh basically gives us bias in the right direction um in the in the in the sense that it becomes it would become even more difficult to see treatment effect and rather we see that managers I I think one way to think about is that  
[03:06:51] like if they change the process for certain services in their office um when this deadline uh this external shock asks them to like you know uh become become faster that will uh transfer to other services as well within the same  
[03:07:04] office. Um yeah so just to wrap it up uh uh with this analysis we are able to see uh that like how much this office share explains  
[03:07:18] this variation across government in in government per performance across departments and our results like when we compare with these two existing papers uh in using Italian social security data she finds like about 76% it's quite  
[03:07:31] comparable and with Russian data uh best at all they find about 55%. uh our results also like because we have the staff performance too. So they kind of we are able to unpack at the local level like how manager and staff both are you  
[03:07:45] know contributing to the productivity and uh like overall it is consistent with this longheld belief that street level discretion for this frontier bureaucrats are like frontier staff uh they are they they do not really change  
[03:07:58] much because of transient supervisors. They cannot monitor permanent teams.  
[03:08:02] There are also cost like if you want to change the norm in the new office you may have to pay in different ways you have to put more effort and you have to you have to make enemies and the the organizational norms they persist over time and the managers are not able to  
[03:08:13] change much. So just to conclude um so we find that office effects dominate uh service delivery variation u um in in service delivery variation we we see that office effects are five times of the manager effects and they explain  
[03:08:27] more than half of the total variance uh productivity convergent immediately to the destination office when a manager is transferred. So unlike the private sector we do not see in the government sector such a high effect of of managers uh you know able to turn around things  
[03:08:41] in the these local offices. Um when I talked to an Indian administrative service officer about this result one time and like his one sentence application was that in government sector CEO CEO is just as good good as the company. So CEO cannot really change the company alone and his uh like that  
[03:08:56] person is is is not going to turn around as one would expect. Uh immo support staff who do not get transferred so frequently. Uh we pinpoint that as a specific location specific factor. Um there can be other factors too resources  
[03:09:09] uh computers um you know u you know stationary or or or norms or culture we do we are not we don't have much to say on that but we our data does allow us to say more about the staff um these mandates uh which which basically  
[03:09:23] tighten the deadline uh where we see that the low performing offices do respond to the top down mandates but their response is is is orthogonal to the manager quality. So policy implication I think a lot of a lot of  
[03:09:36] lot of uh you know incentives are there for the top bureaucrats in bureaucracies and um lot of policies are focused like if the government also tries to transfer the managers that as the first order thing first thing they would do if uh if they have to solve any problem that  
[03:09:50] works very well as earlier research by by um by by many people um um uh have documented that episodes or when there's a episodic or mission mode these bureaucrats are able to uh like you know be more productive But in routine task  
[03:10:04] we see that like office level intervention may be more you know relevant uh which can you know support staff capacity uh management practices or or like change the technology or or local culture in some way. Thank you.