col-xs-12
col-sm-12
col-md-12
col-lg-12
col-xs-12
col-sm-12
col-md-12
col-lg-12
videoType
dynamic-media
videoDmUrl
https://delivery-p136806-e1377785.adobeaemcloud.com/adobe/assets/urn:aaid:aem:341e8668-168c-4211-8b7e-85768ee3b947/play?assetname=Qianmiao-Chen-Public-Procurement-Agencies-Independence-Accountability-and-Transparency-II.mp4
keyFrameImage
videoDescription
Rigging the Scores: Corruption through Scoring Rule Manipulation in Public Procurement Auctions - Qianmiao (Michelle) Chen
timestamp

00:00 Start by introducing our uh presenters.

00:03 Uh,

00:03 first,

00:03 it will be Michelle Cheng.

00:06 She's a research economist in the Development Impact evaluation department,

00:10 DIME at here at the World Bank.

00:13 She earned her PhD in agricultural

00:16 and Research Economics from the University of California,

00:19 Berkeley

00:20 in 2024,

00:22 and her research focuses on exploring the intricate

00:25 relationship between policies and their impacts on firms

00:29 and the broader economy.

00:31 In particular,

00:31 she is interested in public procurement policy in developing countries.

00:36 Now this is the formal presentation.

00:38 For a more personal

00:39 note,

00:40 I would add that I worked with Michelle.

00:42 Uh,

00:43 she is really an awesome mind,

00:45 a brilliant mind,

00:47 and

00:48 I would add to all these skills the ability

00:50 to make a very complicated things look extremely easy.

00:55 Uh,

00:55 I was impressed since I had the honor to start working with her.

00:59 Hi,

01:00 everyone.

01:00 Uh,

01:00 thanks for,

01:01 uh,

01:02 thanks for uh Alex for this warm introduction of my paper.

01:06 Um,

01:06 yeah,

01:06 this is actually,

01:07 uh,

01:08 like,

01:08 uh,

01:08 from my PhD thesis,

01:10 um,

01:10 so I'm trying to study what's their,

01:13 like,

01:13 how we can use the granularly

01:16 big data that to detect corruption

01:18 and throw like a really data-driven perspective,

01:21 and I'm trying to

01:23 go through,

01:23 OK.

01:24 So we already like learned,

01:25 uh,

01:26 heard a lot of things like about how important And the public procurement is like,

01:30 what's the percentage of GDP and what's the percentage of government expenditure.

01:34 So,

01:35 um,

01:35 I don't need to like address that more

01:37 and particularly for the corruption and also mentioned

01:41 in some like keynote speakers like corruption in

01:43 public procurement is like more than half of

01:46 those corruption in the whole public sector.

01:49 So that's why I think like

01:51 trying to use the more and more arising data

01:54 sets to tangle this question is really important.

01:58 OK.

01:59 And um

02:00 so I think there's several um consensus like among all over

02:04 the world that what's the best practices we should adopt,

02:07 so one of them is like we need open auction.

02:10 That's the open bidding,

02:11 we encourage the competition and encourage the participation from

02:15 all over the country or like a region.

02:18 However,

02:20 Actually,

02:20 what I'm showing you is here,

02:22 open auction is not uh a key like that to address all the corruption,

02:27 and actually in anyway,

02:29 it's like a lot of firms can play around with these tools.

02:32 And one of the open auctions format,

02:35 now today I'm discussing is the open scoring auction.

02:38 And it's pretty general and pretty commonly used all over the world.

02:42 And what's the open scoring auction?

02:44 Um,

02:45 we learned some

02:46 first auction,

02:47 first prize auction,

02:49 2nd prize auction,

02:50 or any kind of price auction in textbook,

02:53 and that auction only bidding on the price.

02:56 However,

02:56 as a government,

02:57 we also care about quality,

02:59 right?

02:59 We,

03:00 we don't want just like,

03:01 OK.

03:01 They use the

03:02 uh either chases uh product but like with really bad quality.

03:08 So that's why the open scoring auction is kind of popular is it can give

03:12 uh government a formula that we can like summarize,

03:15 we can sum up the

03:17 quality score and the price score together,

03:20 that's all the contracts going to the firms winning the highest score.

03:25 And for sure,

03:26 there's a lot like description,

03:27 you can think,

03:28 there's like different ways you put on different things.

03:31 For example,

03:32 like here,

03:32 I show you the equations used in Chinese context,

03:36 we use uh two ways,

03:37 large weights,

03:38 one is on the price,

03:39 one is on the quality,

03:41 and the weight in the quality,

03:42 you definitely have a lot of other requirements,

03:45 you can break down to really detailed requirements there.

03:50 Here is a case study,

03:52 and this is a public procurement project happening in 2013 in countywide.

03:57 And that's our agriculture department trying to,

04:00 this is under a large

04:02 agriculture promotion activity to develop the rural area

04:06 and to distribute the fertilizer.

04:08 And what we can see in this call for tinder file,

04:12 we can see first,

04:13 they listed the budget,

04:14 so what's the maximum price they use for this product.

04:18 Second,

04:18 they list some like basic qualifications

04:21 and stating here,

04:22 we can.

04:23 See something weird,

04:24 right?

04:25 First,

04:25 they ask for registration capital should be like larger than 5 million RMB.

04:30 Back to 2013,

04:32 that's not a normal,

04:33 uh,

04:34 like size for firms,

04:35 for,

04:36 especially for like standardized uh fertilizer firms.

04:39 Second,

04:40 they ask for the product with some really

04:42 weird criteria bacteria embedded in the fertilizers,

04:46 and you can ask for like fertilizer.

04:48 uh

04:49 can be like more uh like highly qualified,

04:52 but not like some like special

04:54 bacteria like within us.

04:57 And then like we go into the scoring rule,

04:59 then there's a more interesting thing we can find

05:02 for the business and the technical rules,

05:04 they ask for,

05:05 you need to provide proof of the previous contracts that

05:10 those contracts also have thresholds

05:12 that at least 3 minutes.

05:14 RMB

05:15 for those contracts and also within really uh short time lengths,

05:19 like 22 years pass.

05:21 Second,

05:21 like to ask for a plan for the after service,

05:23 after sale services.

05:25 And I didn't show more details here.

05:27 In the plan for the after-sale services,

05:30 they ask for those firms can respond to their request within

05:34 30 minutes,

05:35 which means you need to relocated close to the uh local agency.

05:40 And uh in addition,

05:42 um,

05:42 there's a regulation in Chinese public procurement system is

05:46 they require at least 3 bidders in open auctions that you make the all,

05:51 all the auctions valid.

05:53 That's not uh like uh uncommon thing for a lot of contacts,

05:57 uh,

05:57 especially for like,

05:58 uh,

05:59 Canada,

05:59 India,

06:00 they all have this kind of like a requirement,

06:02 like minimum number of bidders to make this like

06:05 open auctions really valid for next step contracting stuff.

06:10 OK.

06:11 Then,

06:11 here's the results of this,

06:13 uh,

06:13 uh,

06:13 like auction.

06:14 And first thing,

06:15 we can see,

06:16 OK,

06:16 that's exactly 3 bidders,

06:18 uh,

06:18 good,

06:18 like we qualify for the like the next step contracting stuff.

06:23 Second,

06:24 we will see that if we look into the final score,

06:27 uh,

06:28 that's the

06:28 last second column,

06:30 we can see,

06:31 OK,

06:31 firm A,

06:32 this winner

06:33 got really high scores,

06:34 that's almost

06:36 100 out of 100.

06:38 And then like,

06:39 let's look into the B and C.

06:42 They are super close and super bad,

06:45 which means like,

06:46 why they even participate.

06:48 And then like we looked into the price score,

06:51 Amazingly,

06:52 a lot of those

06:53 price,

06:53 a lot of those like final score discrepancy

06:57 come from

06:58 not the price part,

06:59 but from the first part,

07:01 the business technical part we call like in total,

07:03 like quality part.

07:05 And then like,

07:07 this is

07:08 actually busted.

07:09 This is a real corruption case

07:11 that was documented in the judge documentation.

07:15 And firm A confessed in the documentation that,

07:18 like,

07:19 they got connection.

07:20 They know there's a large agriculture promotion program and

07:23 there are a lot of demand for the fertilizers.

07:26 So firm A got connection with the government,

07:29 uh,

07:29 with the,

07:30 uh,

07:30 director of the agriculture department.

07:32 And then like,

07:33 the director say like,

07:35 I can help you win,

07:36 and but I need like 5% kickback of the total,

07:39 uh,

07:39 like budget.

07:40 And firm B and C,

07:42 they are just like players,

07:44 and because like when you kind of like design all the rules as like call for tenders,

07:49 I just show you,

07:51 if you are like a pure or like innocent firm,

07:54 when you saw those documents,

07:55 what's in your mind is,

07:56 OK,

07:57 that's not my

07:58 game.

07:59 I just don't want to waste my time,

08:01 waste my money to participate in the game.

08:03 So

08:04 then like they still need to satisfy the three.

08:06 Bidders requirements.

08:07 What the firm A do is like they just invite firm B and C.

08:11 They use their name,

08:12 and also like prepare all the documentation for

08:15 firm B and C and like submit something

08:18 just like qualify for the 3 million bidders requirements.

08:22 And finally,

08:22 uh,

08:23 the,

08:23 the director was sent to jail for 7 years.

08:26 But

08:27 unfortunately,

08:28 like,

08:28 fortunately,

08:29 they find the case,

08:30 but unfortunately,

08:31 they,

08:32 when they found the case,

08:33 like the case already passed,

08:34 uh,

08:35 it's like

08:36 5 years later on,

08:37 so

08:38 It's only because

08:39 there's another anti-corruption investigation,

08:41 they can like check back and find this case.

08:44 So

08:45 it's,

08:45 there's a really no like real-time checking

08:48 or auditing things that can find those things

08:51 in the later,

08:52 in the earlier stage of the procurement.

08:54 OK.

08:55 Um,

08:56 so my question for the paper is too.

08:58 First,

08:59 how prevalent is corruption in open scoring auction within the Chinese context?

09:05 Second,

09:05 what kind of policy can help us to reduce the corruption

09:08 and also to reduce those distortions?

09:14 And before I move to,

09:15 uh,

09:15 like deeper into the paper,

09:17 I want to show you two stylized facts.

09:20 First,

09:20 uh,

09:20 I want to show you

09:21 what the distribution looks like when we look into like the winning margin.

09:26 How I define the winning margin is kind of follow the auctions of paper.

09:31 They,

09:31 when they measure the winning margin is

09:33 my score,

09:34 I'm the competitor,

09:35 my score minus the maximum score among other people.

09:40 So,

09:40 if I am the winner,

09:42 so I am the highest score,

09:43 and I'm minus the 2nd-highest score,

09:46 so that's the gap.

09:47 If I'm the loser,

09:48 so my score minus the winner's score.

09:50 So it's like how much,

09:52 what's the distance I compare to the winner.

09:55 So if I'm the loser,

09:56 I've got like a negative uh scores.

09:58 OK.

10:00 And if we look into the previous case,

10:02 and uh like to put them into the scale,

10:05 we will see,

10:05 OK,

10:05 firm A located

10:07 uh in the positive side,

10:09 firm BC and located in the negative side,

10:12 and B and C really close to each other.

10:14 And there's no data points between them.

10:18 And then we pull all those things together.

10:22 What looks like

10:24 in that by naturally,

10:25 we want to see something really smooth,

10:27 like the pic picture I show you,

10:29 we want some smooth,

10:30 there's no any like manipulations or like kind of like smooth distribution.

10:35 However,

10:36 when we look into 3 bidder cases,

10:39 then

10:40 things cause weird.

10:41 It's like there's

10:42 a big missing mass around the,

10:44 the 0,

10:45 which means

10:46 a lot like bidding

10:48 process or the outcomes.

10:49 It's like,

10:50 I'm the winner.

10:51 I just like to run further

10:53 really quick,

10:54 and all of the other people are just like thousands decent like meters away from me.

11:00 So that's why we saw this kind of a lot,

11:02 a lot missing mass around the zero.

11:05 But for sure,

11:05 I would not see like for 4 bidder on 4 or more bidder cases,

11:09 there's no corruption.

11:10 There's still corruption.

11:11 I can like later on show you.

11:13 It's just,

11:13 it's not kind of systematic things as compared to 3 bidder cases.

11:20 And then,

11:21 like,

11:21 let's look into the distribution of number of bidders,

11:24 and I compare the US data with Chinese data.

11:28 So,

11:28 one thing is like,

11:29 there's no restriction on number of bidders for US.

11:32 So we'll see like,

11:33 OK,

11:33 there's kind of like distribution

11:35 uh smoothly,

11:36 and also there's a lot of cases.

11:38 really competitive.

11:39 There's more than 7 bidder cases.

11:41 However,

11:42 for Chinese cases,

11:43 wow,

11:44 that's all around 3.

11:46 And like,

11:47 uh,

11:47 it's that's like a magic number that makes the distribution look like this.

11:51 And really less competition in the like uh 7 or more cases.

11:57 OK.

11:58 So,

11:58 uh,

11:58 then,

11:59 the

11:59 main context of paper has two parts.

12:01 First part is the Marri part.

12:03 Second,

12:03 I'm talk about what's the economic implications and the policies.

12:08 For the measurements part,

12:10 I collect a really rich and novel data set,

12:13 um,

12:13 from one province of China.

12:15 So covers,

12:16 uh,

12:16 11 years and

12:17 including all the speeding details about like what's their

12:21 uh final scores,

12:22 was their quality score and price score,

12:24 and then I link those procurements auction data

12:28 to firm registration data.

12:30 That's.

12:31 That way it can help us to know what's the identity of the firm,

12:35 what's their basic information,

12:37 and then I link those data to a firm uh tax survey,

12:41 so then that we can help us to better understand what's the productivity level,

12:45 what's the employee size of the firm.

12:48 So I kind of like a link all those data together

12:50 that you have a bigger picture about what's happening here.

12:54 Then I design a model base,

12:57 so it's come from like scoring auction theory,

13:00 that's,

13:00 but like,

13:00 there's not a lot like empirical strategies can really leverage those theories.

13:05 So,

13:06 I can try to link the theory with data together,

13:09 to distinguish or like to screen all this,

13:12 those,

13:13 those behaviors I just showed you.

13:16 OK.

13:16 For the statistic tests,

13:18 um,

13:18 I don't want to go to details about like this kind

13:21 of like auction theory and how game players play around,

13:25 but I just want to show you the basic,

13:27 uh,

13:27 equilibrium,

13:28 like in the auction,

13:29 the scoring auction model is,

13:31 when there's a winner,

13:33 They have the score.

13:35 They actually,

13:36 what their bidding strategy is looking all the competitors behind,

13:39 or like the expectation at least.

13:42 So what they need to try to do is not like,

13:44 I try my best on the 100%

13:47 because that way you just like put too much effort and

13:51 like your

13:52 uh profit is like so low.

13:54 And what you need to do is like look back to those competitors.

13:57 I only need to beat

13:59 the second best,

14:00 then I can win.

14:01 So,

14:02 it's always like,

14:04 I'm the winner,

14:05 and try to look the second best,

14:06 then

14:07 I'm done.

14:07 I'm just cut it and done.

14:09 But if you put too much effort,

14:11 you got like so,

14:13 so large score gaps,

14:14 which means

14:15 you just like put

14:16 waste a lot of effort,

14:17 money or time that make you

14:20 even win,

14:21 but like it's less uh like a profitable,

14:24 OK.

14:25 So,

14:25 uh,

14:26 what I did here is like,

14:28 uh,

14:28 pull all those data together,

14:30 and based on this,

14:31 um,

14:32 prediction of the model,

14:34 what I,

14:34 if for all the competitive scenario,

14:36 we should like see one distribution of those gaps,

14:40 I mean,

14:40 the gap between like the,

14:42 the,

14:43 um,

14:43 first,

14:43 uh,

14:44 the winner score and what's his

14:46 projection of the second best,

14:48 the highest score.

14:49 When we see that if there's all the compact cases,

14:52 we should see a,

14:53 a uniform distribution.

14:55 Uh,

14:55 I mean,

14:55 the one distribution.

14:57 However,

14:57 when you put all the data together,

14:59 we'll see,

14:59 OK,

14:59 clearly there are two distributions.

15:01 And this kind of separation can help us to estimate,

15:05 OK,

15:05 what's the percentage of the

15:07 auctions

15:08 has this kind of a huge gap that's non-reasonable behaviors

15:12 and And that is one

15:14 tool I use.

15:15 The second tool is more go into the auction level details,

15:18 like,

15:19 see,

15:19 OK,

15:20 given firm A,

15:21 we see like the scores are 98.

15:24 Then his projection is,

15:26 I'm competing with really strong competitors that pushed me to work really hard.

15:30 And then use his score as a benchmark to look back to the B and C.

15:35 We'll see,

15:35 OK,

15:36 B and C,

15:37 there is just really low,

15:39 and compared to

15:41 what projection on the A.

15:43 The distribution is to look like,

15:44 should look like what I just applied the distribution.

15:47 However,

15:48 B and C all form the really lower tile

15:51 and also clustered together,

15:53 then that's kind of like a reject the,

15:55 their main uh direct test.

15:57 So that's like a rejection of this non-hypothesis.

16:00 I label as there's like potential corruption here.

16:03 OK.

16:04 So I compare,

16:05 put them two,

16:07 those two text results together.

16:09 They kind of like mimic to each other

16:11 and all of those,

16:13 it just like gave us a huge number of corruption proportion

16:17 and 90

16:18 65% of the auctions

16:21 kind of like

16:23 feel the uh reject the

16:25 test and show some evidence of the corruption.

16:28 OK.

16:29 And also,

16:30 that number even like,

16:31 also really high for 4 bidder cases.

16:34 And I do some like,

16:36 uh uh textes about why that is possible,

16:38 but we didn't show this kind of gap in the 4 bidder cases.

16:42 And actually,

16:42 you separate those uh passing the test and not passing test.

16:46 For the non-passing set,

16:47 you still see this kind of like really huge missing gaps there.

16:51 OK.

16:54 And then

16:55 I,

16:56 you are like wondering,

16:57 OK,

16:57 well,

16:58 like all the task tests like huge gaps because corruption,

17:02 maybe not because they are all some other complicated things will happen,

17:06 right,

17:07 because some people maybe study the collusion,

17:09 maybe it's just.

17:10 Collusion like firm like play with them and not about the government side

17:14 or people will think,

17:15 OK,

17:16 there's not really competitive firms in the local market.

17:19 That's

17:20 so that's results in this kind of like

17:23 really bad like heroes.

17:25 But

17:26 to like further exclude those concerns,

17:29 what I did is I did all this study.

17:31 And

17:32 what I did is to hire 5

17:34 public procurement uh scoring experts.

17:36 So why I did that?

17:38 One thing is when I talked to the people,

17:40 it's like

17:41 all,

17:41 a lot of firms,

17:42 when they saw those call for tender files,

17:44 uh,

17:44 they actually have like since like,

17:46 OK,

17:46 there's a signal they sent to us,

17:48 the contract was already locked.

17:51 And then that tells us like really a lot like.

17:53 Like

17:54 local uh industry

17:56 policy,

17:57 or like a lot of experts,

17:58 they actually know what happens.

17:59 So what I did is I randomly sample 500 call for Tinder files from all those files,

18:06 and

18:07 without letting them know what's the outcome,

18:08 and let them just read it all the

18:11 rules and highlight anything they think that's suspicious,

18:15 or they think there's like non-necessary rules,

18:19 and they will put explanation on the uh uh margins.

18:23 And then,

18:24 here's the like example I show you,

18:26 that's the expert audio survey.

18:28 And this is one particular um example.

18:32 This is try to procure the circulated television

18:36 system for one city.

18:38 And it's a huge contrast,

18:40 and

18:41 in the business score we can say the first category

18:43 they ask for software development and level 2 and involved,

18:48 and then like they have some like standards they ask for,

18:51 and experts see like

18:52 those rules are not.

18:54 There's industry requirements

18:57 and they think it's kind of like specific for suppliers

19:00 and the second they ask for a lot of certificates

19:03 and they're not,

19:04 not really spacious about like the certificate but expert put they say like

19:09 usually there are 3 come together.

19:15 The expert,

19:15 yeah,

19:16 that's a good question.

19:16 I didn't talk more about like the what's the scoring evaluation process,

19:21 because

19:22 um

19:23 without this kind of like uh scoring process,

19:26 previously,

19:26 they just like,

19:27 OK,

19:28 I

19:28 can find 5 experts to stay there,

19:31 give score for each person.

19:33 And now they have like a random selection of this uh like expert.

19:37 They just like those experts are like in

19:39 every industry or sectors that have their own

19:42 uh like uh expertise,

19:44 like engineer or are they kind of like

19:46 um manufacturer,

19:48 uh,

19:48 like their expertise,

19:49 those persons will be randomly selected to stay there to

19:52 give the score for each proposal.

19:54 So I am like invite those experts to do it.

19:58 OK.

19:59 Um,

20:00 then the second group of requirements,

20:02 it's just like a purchase like the circulated television,

20:05 but they ask for,

20:06 OK,

20:07 the manager should have at least a master's degree

20:10 and ask for some like information system,

20:13 project manager,

20:15 a certificate,

20:16 any like a lot of a certificate that's

20:20 actually not necessary at all,

20:22 and like that's like expert on market here.

20:25 And interestingly,

20:26 when they finished all this,

20:27 all this study,

20:28 I asked the expert,

20:29 like the expert told me,

20:31 he exactly know

20:32 what's the supplier of this project is

20:35 and told me the name,

20:36 and then I checked back with the contract is exactly the that firm.

20:39 So,

20:41 And then like I pull all those 500 copies together

20:45 and do some uh kind of like 2x2 table

20:48 on the diagonal you can see that's like

20:50 kind of consistent of my model prediction and the

20:54 uh

20:54 expert there judge

20:56 and on the other side there's some like uh mismatch there,

21:00 but

21:00 we kind of see like expert evaluation is ground truth,

21:03 right.

21:04 Right,

21:04 because they also have their own bias in sometimes.

21:07 So,

21:07 but this is basically want to provide some evidence to see like at least most of

21:12 what the model captures

21:14 is really like due to the corruption or like this kind of customized scoring rule.

21:20 OK

21:21 Then,

21:22 uh,

21:22 I will move to the,

21:23 um,

21:24 economic implications of policies.

21:27 And,

21:27 um,

21:27 for sure,

21:27 people want to say,

21:28 OK,

21:29 what those corrupt firms look like.

21:31 And I didn't put the results here,

21:33 but I just like to do the lasso,

21:35 just put all those firm characteristics to say,

21:37 OK,

21:37 what can explain like their corrupt label.

21:41 And what I find is,

21:42 most of them are local,

21:43 local to the buyer agency.

21:45 Second is they have like

21:48 state-owned connected in some layers of their ownership.

21:52 Third is they are like less productive.

21:56 Then I asked,

21:57 uh,

21:57 ask,

21:58 can we just really,

22:00 uh,

22:00 reduce the corruption through any policies we can like to try to reduce this.

22:05 And I,

22:05 using the anti-corruption campaign,

22:07 that's a,

22:07 a huge campaign in China starts from 2012.

22:11 It targets a lot like different

22:13 Officials in different departments and different levels,

22:15 just like look back into their cases and to find who is corrupt or not.

22:20 And what I find is

22:21 exactly after the corruption or like the person was under investigation,

22:25 we saw more bidders or more firms come into competition

22:29 and

22:31 And also we saw this like reduce of the probability of the corruption,

22:35 but it's only like kind of delayed because we have like

22:38 a long time like a planning of the procurement process.

22:40 It's not like direct effect you will

22:43 expect after the person was under investigation.

22:46 And then like we break down to the investigation to a higher level

22:50 officials and the lower level officials.

22:52 What we find is

22:53 most of the results come from higher-level officials.

22:57 If you adjust like a front desk,

22:59 uh,

22:59 police to do something,

23:01 not to have the powerful

23:03 influence on who you want to allocate the contract,

23:06 then you have less influence on this kind of allocation process.

23:12 But,

23:12 uh,

23:12 one thing I need to like,

23:13 uh,

23:14 the point out is

23:15 a lot,

23:16 all of those effects are pretty much like really short term.

23:20 And if we talk to people,

23:21 it's like if there's no systematic change in the system,

23:25 then

23:26 um people just like let people out and new people came,

23:28 and all new connection just like formed uh within the new people.

23:33 So there's not really

23:35 deeply changes of those behaviors.

23:38 And

23:39 then like,

23:39 what kind of policy can we really think to tangle this?

23:43 One experiment,

23:44 uh,

23:44 not experiment,

23:45 the counterfactual structural estimation I did is,

23:47 oh,

23:47 how about just like we

23:49 invite those experts to do another round of evaluation.

23:53 If they think those

23:55 rules are not necessary,

23:55 we just take them out.

23:57 And I did this counterfactual analysis just based on those 500 audit study I did.

24:02 So what I did is,

24:03 OK,

24:04 let's remove all those rules,

24:05 rescale their.

24:07 Quality scores

24:08 and then like say,

24:09 OK,

24:09 there's no unfair rules.

24:12 Let them compete together.

24:13 What's the new results.

24:15 In addition,

24:16 I add more entry because we see that if there's less corrupt

24:20 um requirements,

24:21 then there are more entries.

24:23 So I combine those two together

24:25 and what I show you is

24:26 if we just remove those non-necessary rules,

24:30 the price can decrease by 12%

24:33 and the quality can increase.

24:35 And we introduce more entry,

24:37 even though it's like uh the price maybe like in decrease more,

24:42 but uh the overall like

24:44 all the major contribution comes from the

24:46 competitive like without like really unfair requirements.

24:51 OK.

24:52 Uh,

24:52 so I'm trying to wrap up.

24:53 Um,

24:54 the takeaway of this paper,

24:56 first,

24:56 the merriment part.

24:57 I think also like mentioned by a lot of people,

25:00 we now have a lot of like really rich data sets,

25:03 and that's really provide our opportunities to

25:06 think about what kind of indicators,

25:08 what about like a methodology,

25:11 we can capture

25:12 those kind of misbehaviorals from the data set.

25:15 Second,

25:15 I want the,

25:16 uh,

25:17 a lot of those,

25:18 um,

25:18 methodology can be generally applied to other contexts if using the same

25:23 uh auction format or,

25:25 uh,

25:25 there's uh any like similar um context there.

25:30 In terms of economic implications and the policies,

25:34 what we really want is

25:35 not just like anti-corruption,

25:37 that's like,

25:38 OK,

25:38 we

25:39 Investigates person,

25:41 but more about thinking about what's what's wrong,

25:44 what procedure we need to target,

25:46 and what's their logic and the incentive to do that.

25:49 So

25:50 as long as we can figure out which procedure

25:53 we can like really put effort on,

25:55 then like we can like change it so we can target more accurately

25:58 to change those regulation

26:00 and change the people's behaviors,

26:02 um.

26:03 For sure,

26:03 I think there's more and more research can be done

26:05 in this area,

26:06 especially has like large language model and AI powerful things

26:10 we can

26:11 use,

26:12 and I think public procurement is really perfect setting

26:15 for those tools because we have a lot of documents

26:18 that one people or like data,

26:20 the,

26:21 the traditional data structure cannot capture those things,

26:24 but AI can help us do a lot of things to do like.

26:26 Like tax analysis

26:28 or and even like we can help us to identify which rules that are

26:32 more like necessary as core for this competition and the words we cannot put them

26:37 so it can help those audits we who do not have background

26:41 knowledge to do better regulation and also put firms to do a real

26:46 uh competition induction.

26:47 Yeah,

26:48 thank you all.

26:48 That's my presentation.

showAllTimestamps
no
transcript
Start by introducing our uh presenters. Uh, first, it will be Michelle Cheng. She's a research economist in the Development Impact evaluation department, DIME at here at the World Bank. She earned her PhD in agricultural and Research Economics from the University of California, Berkeley in 2024, and her research focuses on exploring the intricate relationship between policies and their impacts on firms and the broader economy. In particular, she is interested in public procurement policy in developing countries. Now this is the formal presentation. For a more personal note, I would add that I worked with Michelle. Uh, she is really an awesome mind, a brilliant mind, and I would add to all these skills the ability to make a very complicated things look extremely easy. Uh, I was impressed since I had the honor to start working with her. Hi, everyone. Uh, thanks for, uh, thanks for uh Alex for this warm introduction of my paper. Um, yeah, this is actually, uh, like, uh, from my PhD thesis, um, so I'm trying to study what's their, like, how we can use the granularly big data that to detect corruption and throw like a really data-driven perspective, and I'm trying to go through, OK. So we already like learned, uh, heard a lot of things like about how important And the public procurement is like, what's the percentage of GDP and what's the percentage of government expenditure. So, um, I don't need to like address that more and particularly for the corruption and also mentioned in some like keynote speakers like corruption in public procurement is like more than half of those corruption in the whole public sector. So that's why I think like trying to use the more and more arising data sets to tangle this question is really important. OK. And um so I think there's several um consensus like among all over the world that what's the best practices we should adopt, so one of them is like we need open auction. That's the open bidding, we encourage the competition and encourage the participation from all over the country or like a region. However, Actually, what I'm showing you is here, open auction is not uh a key like that to address all the corruption, and actually in anyway, it's like a lot of firms can play around with these tools. And one of the open auctions format, now today I'm discussing is the open scoring auction. And it's pretty general and pretty commonly used all over the world. And what's the open scoring auction? Um, we learned some first auction, first prize auction, 2nd prize auction, or any kind of price auction in textbook, and that auction only bidding on the price. However, as a government, we also care about quality, right? We, we don't want just like, OK. They use the uh either chases uh product but like with really bad quality. So that's why the open scoring auction is kind of popular is it can give uh government a formula that we can like summarize, we can sum up the quality score and the price score together, that's all the contracts going to the firms winning the highest score. And for sure, there's a lot like description, you can think, there's like different ways you put on different things. For example, like here, I show you the equations used in Chinese context, we use uh two ways, large weights, one is on the price, one is on the quality, and the weight in the quality, you definitely have a lot of other requirements, you can break down to really detailed requirements there. Here is a case study, and this is a public procurement project happening in 2013 in countywide. And that's our agriculture department trying to, this is under a large agriculture promotion activity to develop the rural area and to distribute the fertilizer. And what we can see in this call for tinder file, we can see first, they listed the budget, so what's the maximum price they use for this product. Second, they list some like basic qualifications and stating here, we can. See something weird, right? First, they ask for registration capital should be like larger than 5 million RMB. Back to 2013, that's not a normal, uh, like size for firms, for, especially for like standardized uh fertilizer firms. Second, they ask for the product with some really weird criteria bacteria embedded in the fertilizers, and you can ask for like fertilizer. uh can be like more uh like highly qualified, but not like some like special bacteria like within us. And then like we go into the scoring rule, then there's a more interesting thing we can find for the business and the technical rules, they ask for, you need to provide proof of the previous contracts that those contracts also have thresholds that at least 3 minutes. RMB for those contracts and also within really uh short time lengths, like 22 years pass. Second, like to ask for a plan for the after service, after sale services. And I didn't show more details here. In the plan for the after-sale services, they ask for those firms can respond to their request within 30 minutes, which means you need to relocated close to the uh local agency. And uh in addition, um, there's a regulation in Chinese public procurement system is they require at least 3 bidders in open auctions that you make the all, all the auctions valid. That's not uh like uh uncommon thing for a lot of contacts, uh, especially for like, uh, Canada, India, they all have this kind of like a requirement, like minimum number of bidders to make this like open auctions really valid for next step contracting stuff. OK. Then, here's the results of this, uh, uh, like auction. And first thing, we can see, OK, that's exactly 3 bidders, uh, good, like we qualify for the like the next step contracting stuff. Second, we will see that if we look into the final score, uh, that's the last second column, we can see, OK, firm A, this winner got really high scores, that's almost 100 out of 100. And then like, let's look into the B and C. They are super close and super bad, which means like, why they even participate. And then like we looked into the price score, Amazingly, a lot of those price, a lot of those like final score discrepancy come from not the price part, but from the first part, the business technical part we call like in total, like quality part. And then like, this is actually busted. This is a real corruption case that was documented in the judge documentation. And firm A confessed in the documentation that, like, they got connection. They know there's a large agriculture promotion program and there are a lot of demand for the fertilizers. So firm A got connection with the government, uh, with the, uh, director of the agriculture department. And then like, the director say like, I can help you win, and but I need like 5% kickback of the total, uh, like budget. And firm B and C, they are just like players, and because like when you kind of like design all the rules as like call for tenders, I just show you, if you are like a pure or like innocent firm, when you saw those documents, what's in your mind is, OK, that's not my game. I just don't want to waste my time, waste my money to participate in the game. So then like they still need to satisfy the three. Bidders requirements. What the firm A do is like they just invite firm B and C. They use their name, and also like prepare all the documentation for firm B and C and like submit something just like qualify for the 3 million bidders requirements. And finally, uh, the, the director was sent to jail for 7 years. But unfortunately, like, fortunately, they find the case, but unfortunately, they, when they found the case, like the case already passed, uh, it's like 5 years later on, so It's only because there's another anti-corruption investigation, they can like check back and find this case. So it's, there's a really no like real-time checking or auditing things that can find those things in the later, in the earlier stage of the procurement. OK. Um, so my question for the paper is too. First, how prevalent is corruption in open scoring auction within the Chinese context? Second, what kind of policy can help us to reduce the corruption and also to reduce those distortions? And before I move to, uh, like deeper into the paper, I want to show you two stylized facts. First, uh, I want to show you what the distribution looks like when we look into like the winning margin. How I define the winning margin is kind of follow the auctions of paper. They, when they measure the winning margin is my score, I'm the competitor, my score minus the maximum score among other people. So, if I am the winner, so I am the highest score, and I'm minus the 2nd-highest score, so that's the gap. If I'm the loser, so my score minus the winner's score. So it's like how much, what's the distance I compare to the winner. So if I'm the loser, I've got like a negative uh scores. OK. And if we look into the previous case, and uh like to put them into the scale, we will see, OK, firm A located uh in the positive side, firm BC and located in the negative side, and B and C really close to each other. And there's no data points between them. And then we pull all those things together. What looks like in that by naturally, we want to see something really smooth, like the pic picture I show you, we want some smooth, there's no any like manipulations or like kind of like smooth distribution. However, when we look into 3 bidder cases, then things cause weird. It's like there's a big missing mass around the, the 0, which means a lot like bidding process or the outcomes. It's like, I'm the winner. I just like to run further really quick, and all of the other people are just like thousands decent like meters away from me. So that's why we saw this kind of a lot, a lot missing mass around the zero. But for sure, I would not see like for 4 bidder on 4 or more bidder cases, there's no corruption. There's still corruption. I can like later on show you. It's just, it's not kind of systematic things as compared to 3 bidder cases. And then, like, let's look into the distribution of number of bidders, and I compare the US data with Chinese data. So, one thing is like, there's no restriction on number of bidders for US. So we'll see like, OK, there's kind of like distribution uh smoothly, and also there's a lot of cases. really competitive. There's more than 7 bidder cases. However, for Chinese cases, wow, that's all around 3. And like, uh, it's that's like a magic number that makes the distribution look like this. And really less competition in the like uh 7 or more cases. OK. So, uh, then, the main context of paper has two parts. First part is the Marri part. Second, I'm talk about what's the economic implications and the policies. For the measurements part, I collect a really rich and novel data set, um, from one province of China. So covers, uh, 11 years and including all the speeding details about like what's their uh final scores, was their quality score and price score, and then I link those procurements auction data to firm registration data. That's. That way it can help us to know what's the identity of the firm, what's their basic information, and then I link those data to a firm uh tax survey, so then that we can help us to better understand what's the productivity level, what's the employee size of the firm. So I kind of like a link all those data together that you have a bigger picture about what's happening here. Then I design a model base, so it's come from like scoring auction theory, that's, but like, there's not a lot like empirical strategies can really leverage those theories. So, I can try to link the theory with data together, to distinguish or like to screen all this, those, those behaviors I just showed you. OK. For the statistic tests, um, I don't want to go to details about like this kind of like auction theory and how game players play around, but I just want to show you the basic, uh, equilibrium, like in the auction, the scoring auction model is, when there's a winner, They have the score. They actually, what their bidding strategy is looking all the competitors behind, or like the expectation at least. So what they need to try to do is not like, I try my best on the 100% because that way you just like put too much effort and like your uh profit is like so low. And what you need to do is like look back to those competitors. I only need to beat the second best, then I can win. So, it's always like, I'm the winner, and try to look the second best, then I'm done. I'm just cut it and done. But if you put too much effort, you got like so, so large score gaps, which means you just like put waste a lot of effort, money or time that make you even win, but like it's less uh like a profitable, OK. So, uh, what I did here is like, uh, pull all those data together, and based on this, um, prediction of the model, what I, if for all the competitive scenario, we should like see one distribution of those gaps, I mean, the gap between like the, the, um, first, uh, the winner score and what's his projection of the second best, the highest score. When we see that if there's all the compact cases, we should see a, a uniform distribution. Uh, I mean, the one distribution. However, when you put all the data together, we'll see, OK, clearly there are two distributions. And this kind of separation can help us to estimate, OK, what's the percentage of the auctions has this kind of a huge gap that's non-reasonable behaviors and And that is one tool I use. The second tool is more go into the auction level details, like, see, OK, given firm A, we see like the scores are 98. Then his projection is, I'm competing with really strong competitors that pushed me to work really hard. And then use his score as a benchmark to look back to the B and C. We'll see, OK, B and C, there is just really low, and compared to what projection on the A. The distribution is to look like, should look like what I just applied the distribution. However, B and C all form the really lower tile and also clustered together, then that's kind of like a reject the, their main uh direct test. So that's like a rejection of this non-hypothesis. I label as there's like potential corruption here. OK. So I compare, put them two, those two text results together. They kind of like mimic to each other and all of those, it just like gave us a huge number of corruption proportion and 90 65% of the auctions kind of like feel the uh reject the test and show some evidence of the corruption. OK. And also, that number even like, also really high for 4 bidder cases. And I do some like, uh uh textes about why that is possible, but we didn't show this kind of gap in the 4 bidder cases. And actually, you separate those uh passing the test and not passing test. For the non-passing set, you still see this kind of like really huge missing gaps there. OK. And then I, you are like wondering, OK, well, like all the task tests like huge gaps because corruption, maybe not because they are all some other complicated things will happen, right, because some people maybe study the collusion, maybe it's just. Collusion like firm like play with them and not about the government side or people will think, OK, there's not really competitive firms in the local market. That's so that's results in this kind of like really bad like heroes. But to like further exclude those concerns, what I did is I did all this study. And what I did is to hire 5 public procurement uh scoring experts. So why I did that? One thing is when I talked to the people, it's like all, a lot of firms, when they saw those call for tender files, uh, they actually have like since like, OK, there's a signal they sent to us, the contract was already locked. And then that tells us like really a lot like. Like local uh industry policy, or like a lot of experts, they actually know what happens. So what I did is I randomly sample 500 call for Tinder files from all those files, and without letting them know what's the outcome, and let them just read it all the rules and highlight anything they think that's suspicious, or they think there's like non-necessary rules, and they will put explanation on the uh uh margins. And then, here's the like example I show you, that's the expert audio survey. And this is one particular um example. This is try to procure the circulated television system for one city. And it's a huge contrast, and in the business score we can say the first category they ask for software development and level 2 and involved, and then like they have some like standards they ask for, and experts see like those rules are not. There's industry requirements and they think it's kind of like specific for suppliers and the second they ask for a lot of certificates and they're not, not really spacious about like the certificate but expert put they say like usually there are 3 come together. The expert, yeah, that's a good question. I didn't talk more about like the what's the scoring evaluation process, because um without this kind of like uh scoring process, previously, they just like, OK, I can find 5 experts to stay there, give score for each person. And now they have like a random selection of this uh like expert. They just like those experts are like in every industry or sectors that have their own uh like uh expertise, like engineer or are they kind of like um manufacturer, uh, like their expertise, those persons will be randomly selected to stay there to give the score for each proposal. So I am like invite those experts to do it. OK. Um, then the second group of requirements, it's just like a purchase like the circulated television, but they ask for, OK, the manager should have at least a master's degree and ask for some like information system, project manager, a certificate, any like a lot of a certificate that's actually not necessary at all, and like that's like expert on market here. And interestingly, when they finished all this, all this study, I asked the expert, like the expert told me, he exactly know what's the supplier of this project is and told me the name, and then I checked back with the contract is exactly the that firm. So, And then like I pull all those 500 copies together and do some uh kind of like 2x2 table on the diagonal you can see that's like kind of consistent of my model prediction and the uh expert there judge and on the other side there's some like uh mismatch there, but we kind of see like expert evaluation is ground truth, right. Right, because they also have their own bias in sometimes. So, but this is basically want to provide some evidence to see like at least most of what the model captures is really like due to the corruption or like this kind of customized scoring rule. OK Then, uh, I will move to the, um, economic implications of policies. And, um, for sure, people want to say, OK, what those corrupt firms look like. And I didn't put the results here, but I just like to do the lasso, just put all those firm characteristics to say, OK, what can explain like their corrupt label. And what I find is, most of them are local, local to the buyer agency. Second is they have like state-owned connected in some layers of their ownership. Third is they are like less productive. Then I asked, uh, ask, can we just really, uh, reduce the corruption through any policies we can like to try to reduce this. And I, using the anti-corruption campaign, that's a, a huge campaign in China starts from 2012. It targets a lot like different Officials in different departments and different levels, just like look back into their cases and to find who is corrupt or not. And what I find is exactly after the corruption or like the person was under investigation, we saw more bidders or more firms come into competition and And also we saw this like reduce of the probability of the corruption, but it's only like kind of delayed because we have like a long time like a planning of the procurement process. It's not like direct effect you will expect after the person was under investigation. And then like we break down to the investigation to a higher level officials and the lower level officials. What we find is most of the results come from higher-level officials. If you adjust like a front desk, uh, police to do something, not to have the powerful influence on who you want to allocate the contract, then you have less influence on this kind of allocation process. But, uh, one thing I need to like, uh, the point out is a lot, all of those effects are pretty much like really short term. And if we talk to people, it's like if there's no systematic change in the system, then um people just like let people out and new people came, and all new connection just like formed uh within the new people. So there's not really deeply changes of those behaviors. And then like, what kind of policy can we really think to tangle this? One experiment, uh, not experiment, the counterfactual structural estimation I did is, oh, how about just like we invite those experts to do another round of evaluation. If they think those rules are not necessary, we just take them out. And I did this counterfactual analysis just based on those 500 audit study I did. So what I did is, OK, let's remove all those rules, rescale their. Quality scores and then like say, OK, there's no unfair rules. Let them compete together. What's the new results. In addition, I add more entry because we see that if there's less corrupt um requirements, then there are more entries. So I combine those two together and what I show you is if we just remove those non-necessary rules, the price can decrease by 12% and the quality can increase. And we introduce more entry, even though it's like uh the price maybe like in decrease more, but uh the overall like all the major contribution comes from the competitive like without like really unfair requirements. OK. Uh, so I'm trying to wrap up. Um, the takeaway of this paper, first, the merriment part. I think also like mentioned by a lot of people, we now have a lot of like really rich data sets, and that's really provide our opportunities to think about what kind of indicators, what about like a methodology, we can capture those kind of misbehaviorals from the data set. Second, I want the, uh, a lot of those, um, methodology can be generally applied to other contexts if using the same uh auction format or, uh, there's uh any like similar um context there. In terms of economic implications and the policies, what we really want is not just like anti-corruption, that's like, OK, we Investigates person, but more about thinking about what's what's wrong, what procedure we need to target, and what's their logic and the incentive to do that. So as long as we can figure out which procedure we can like really put effort on, then like we can like change it so we can target more accurately to change those regulation and change the people's behaviors, um. For sure, I think there's more and more research can be done in this area, especially has like large language model and AI powerful things we can use, and I think public procurement is really perfect setting for those tools because we have a lot of documents that one people or like data, the, the traditional data structure cannot capture those things, but AI can help us do a lot of things to do like. Like tax analysis or and even like we can help us to identify which rules that are more like necessary as core for this competition and the words we cannot put them so it can help those audits we who do not have background knowledge to do better regulation and also put firms to do a real uh competition induction. Yeah, thank you all. That's my presentation.
showAllTranscripts
no
duration
PT26M51S
scene7File
worldbank/Qianmiao-Chen-Public-Procurement-Agencies-Independence-Accountability-and-Transparency-II
scene7Domain
https://worldbank.scene7.com/
scene7FileAvs
worldbank/Qianmiao-Chen-Public-Procurement-Agencies-Independence-Accountability-and-Transparency-II-AVS
title
Qianmiao-Chen-Public-Procurement-Agencies-Independence-Accountability-and-Transparency-II
description
Qianmiao-Chen-Public-Procurement-Agencies-Independence-Accountability-and-Transparency-II
showTimestampAndTranscript
yes
col-xs-12
col-sm-12
col-md-3
col-lg-3
col-xs-12
col-sm-12
col-md-7
col-lg-7
  • add-style
  • lp-body-content
col-xs-12
col-sm-12
col-md-2
col-lg-2