«CHINA'S LAND MARKET AUCTIONS: EVIDENCE OF CORRUPTION Hongbin Cai J. Vernon Henderson Qinghua Zhang Working Paper 15067 ...»
two stage auction if she cares about her corruption income sufficiently. 14 Therefore, if the government official is more corrupt, then she is more likely to choose a two stage auction. If she cares more about the fiscal revenue of the city, she is more likely to choose an English auction. Such “caring” of course will not in general be exogenous. One could consider the probability of corruption being detected and punished by government auditing and a penalty function, which may also depend on the tolerance of the local government for lost revenue. In such a context, English auctions are much more visible in the media and may be subject to greater scrutiny.
With corruption, the role of hot versus cold properties also changes. Note that the probability of not having a non-corrupt bidder enter is F N −1 (V−1 ) in an English auction ˆ and F N −1 (V0 ) in a two stage auction. Since V0 V−1, the greater the number of potential ˆ bidders, N, the larger the difference in the likelihood that the corrupt bidder wins between the two auctions. Thus, motivated by helping her partner bidder win the auction, a corrupt government official prefers two stage auctions for hotter properties. Moreover, if the bribes which corrupt developers are willing to pay are related to their profits, corrupt government officials have additional incentives to assign hot properties to two stage auctions. Therefore, our analysis suggests positive selection on unobservables into two stage auctions under corruption.
Other auction choice considerations In an auction setting like ours, collusion among bidders (a group of developer forming a bidding ring) is quite plausible. In the existing literature, scholars have studied collusion in other auction settings (e.g., McAfee and McMillian, 1992, Bajari and Ye, 2003, and Athey, Levin and Seira, 2008). While in China, a group of developers may be attempted to rig an auction, there are several reasons we don’t focus on it in this setting.
First, the government’s focus on corruption in land markets has not been on collusive bidding, but rather on corruption among officials. In China it may be less appealing (more dangerous) for individuals to collude against the state per se, as opposed to involving local officials. Also in our context, there seems to be no reason why collusion Moreover, if the bribery payment to the government official increases with the partner developer’s net profit, then they both are more likely to favor a two stage auction, since the price the corrupt developer pays is likely to be lower.
among bidders would be more successful in a two stage auction than in an English auction, so collusion among bidders would not explain the substantial difference in the likelihood of non-competitive bidding between the two stage and English auctions observed in our data. Thus, it is unclear how the concern about collusion among bidders would affect the land bureau’s choice of auction format. As we show later, our instrumental variables that are related to corruption and city budget needs are strongly correlated with the choice of auction format and there is positive selection on unobservables into two stage auctions. It seems land bureau corruption is at the heart of any explanation of the positive selection. For these reasons we focus on corruption in this paper, and ignore the issue of bidder collusion.
Another factor that may affect the comparison between English and two stage auctions is the riskiness of the land to be auctioned. For different properties, the variance of the private value components across bidders could differ. For a given reserve price, absent corruption, the land bureau might want to assign high variance properties to two stage auctions. The reason is that, when there is a fat left tail of the distribution of Vi, the solution to equation (1) may be quite large, resulting in a low chance of a sale in the English auction. Thus revenue-maximizing officials would choose two stage auctions for risky lands. This would suggest that two stage auctions are associated with a higher probability of no sales; we observe the opposite in the data. Nevertheless, below we control for a number of observables which could be related to variance of valuations such as property type, size, and distance from the city center.
One additional issue we choose to ignore is the sequence of land sales in a city. Our decision is based on the following considerations. First, while it is certainly true that developers can always bid on the next available land, land auctions differ from on-line auctions of staple goods in that land available for development in a particular city within a particular period of time is limited. Considering the heterogeneity of land and the heterogeneity of developers, a developer may not easily find many perfectly substitutable pieces of land and thus has to treat each auction seriously. Second, it does not seem to us that the issue of the sequence of land sales would fundamentally change our arguments about auction choices with or without corruption. Thirdly, a formal modeling of the sequence of land auctions is clearly not tractable in our current framework.
What we see in the data concerning potential signaling In our data, in general, we know only sales and reserve prices and nothing about the bidding process itself—sequence of bids and number of bidders. However for Beijing we have a sample of 195 two stage auctions, where we know the number of bidders as well as the date when the first bid is submitted. From that data we learn several things. First, and most critically from Table 1, bidders do not signal valuations. In all auctions with just one bidder, almost all bids are within 0.5% of reserve price. This is consistent with our corruption story, but inconsistent with the separating equilibrium of the two stage auction without corruption. Once we have 2 or more bidders then a spread develops. This is the basis for later defining whether an auction is competitive (has more than one bidder) or not, based on spread. Note auctions can be highly contested: in 26 of the cases with 3 or more bidders, there are reported to be over 65 bids in each of the auctions.
Columns 1 and 2 of Table 2 show that, conditional on property characteristics, having a first day bidder reduces the number of bidders, despite the positive bias (having a first day bid, given 10 days to bid, should mechanically raise the expected number of bidders). Similarly, in columns 3 and 4, having a first day bidder makes it more likely that the auction will be non-competitive. Again this is consistent with the corruption story. But the effects in columns 3 and 4 are weak. It turns out that in Beijing sometimes properties are sold which, contrary to national policy on auctions, have not been cleared for redevelopment; and, in Beijing, we have good data on clearance or not. Not being cleared reduces the number of bidders and increases the chances of non-competition.
Controlling for this variable with a sample size of 155, sharpens the first day bidder variable in both formulations, with t-statistics in columns 3 and 4 on the variable rising to
1.99 and 1.83 respectively.
3. The data and basic patterns For our econometric analysis, we have data for 15 cities from 2003-2007, 15 from the Land Bureau of China (or its branches at the city-level).16 For each auction, the land These are Xiamen, Guangzhou, Shenzhen, Nanning, Changchun, Suzhou, Wuxi, Nanchang, Shenyang, Taiyuan, Chengdu, Tianjin, Hangzhou, Ningbo, and Chongqing.
bureau provides detailed information and posts it on its official website www.landlist.cn.
Information includes: the address, the area (in square meters), the use restriction (business, residential, mixed), the type of auction, the reserve price, the sales price if the sale is complete, the post date which is the first date bids are accepted, the sale date, and the buyer’s identity. Sometimes additional information is given, such as the maximum floor-to-area ratio, the building-density, the green coverage rate, and whether the property is cleared or not. For some items including the last, explicit information is only provided in a limited number of cases.
We also obtained the geo-economic characteristics of each piece of land for sale through bendi.google.com. Specifically, we locate each piece of land in the digital map of bendi.google.com using its street address. We then measure the line distance between the land and the CBD of the city where it is in. For the Chinese cities in the sample, we have no difficulty in identifying one central business district. We also create some dummy variables to indicate, whether within a 2.5 km. radius of the center of the property for sale, there is railway (including light rail and subway) or highway.
Our base data consists of 4016 listings, where a listing is a property put up for auction whether the auction is completed and results in a sale, or not. Our 4016 listings exclude industrial use land (about 7% of total listings). As in the USA, industrial land use has a low and highly variable unit price; regressions using USA data which examine the determinants of sales prices for industrial land have low explanatory power (DiPasquale and Wheaton, 1996). More critically in China, such properties are often sufficiently far from the city center stretching into peri-urban areas, that we couldn’t get location characteristics from bendi.google.com.
Of the 4016 listings, 607 remain unsold. Another 1107, while sold, do not have information on either reserve price or sales price, or both. We focus on the remaining 2302 which are completed auctions with full price information. In the Appendix we explore the effect of focusing just on this sample. Here we note some key findings from the Appendix. First, for properties that sell, those with full versus deficient price information have similar unit and reserve sales prices where information does exist on We exclude Shanghai, Beijing and Nanjing. Shanghai has no English auctions; Beijing has 1; and Nanjing 3 (which are a tiny fraction of sales). In all specifications we utilize city fixed effects, so within city variation in the data (in particular in auction formats which is our focus) is essential.
one or the other and only differ in observables in two minor ways: properties without full price information tend to be older listings and nearer the city center. The differences in samples for sales with full versus limited price information seem to be “innocent.” However, unsold properties compared to our working sample of 2302 show distinct differences. For example, unsold properties are more likely to have been offered at English auction potentially evidence of positive selection into two stage auctions; and, not surprisingly, to have been listed more recently. In terms of sales dates, we suspect unsold properties (i.e., those which didn’t sell 10 days after posting) are eventually removed from public listing on the internet, perhaps rebundled, and then relisted, which makes statistical analysis of sale versus no sale difficult, since we don’t know which properties are being offered for the first versus second time.
Differences across auction types Table 3 is summary of basic statistics about the data, for completed transactions by auction type. In Part a, compared to English auctions, two stage auctions have significantly lower mean unit sales prices and are significantly less likely to sell competitively (have a spread greater than 1.005). However they have no significant difference in unit sales price, conditional on a competitive sale. This suggests that the main effect of two stage auctions is to affect price by deterring entry and competition.
We note two stage auctions have a greater proportion of commercial properties.
However, we decided that whether a property was designated as commercial was not an element on which we wanted to focus. As Part b of the table reveals, commercial relative to residential and mixed use (which are fairly similar) properties are more likely to be sold through two stage auction and without competition (60% sold non-competitively versus 40% for residential and mixed use). However unit sales prices across uses are similar, both for those that are sold competitively and for those that are not.
4. Baseline effect of auction type on sales prices In this section we explore the overall effect of auction type on unit sales prices. As we will see in Sections 5 and 6, we are in essence estimating a reduced form price equation.