«CHINA'S LAND MARKET AUCTIONS: EVIDENCE OF CORRUPTION Hongbin Cai J. Vernon Henderson Qinghua Zhang Working Paper 15067 ...»
Based on the conceptual section, consider the specification ln sale price = ln common value + f ( potential number of bidders, auction type, e ) (6) This specification follows the notion that there is a common value component to any bidders’ valuation. Given this common value, ex ante sales price then depends on the number of potential bidders and potentially the auction format, with the ex post sales price dependent on the actual drawings of private valuations (which e encapsulates). In the data, the potential number of bidders and certain determinants of the potential number of bidders (e.g. certain property characteristics) are unobserved. Choice of auction format may be related to these unobservables in one of two ways, which we hope to distinguish.
In the absence of corruption, the land bureau might choose to sell “cold” properties by two stage auction, since, relative to an English auction, they may be more likely to draw in at least one bidder. This implies negative selection on properties sold by two stage auction. However two stage auctions may be more corruptible, and part of that may involve positive selection— setting aside the most “delectable morsels” for corrupt participants.
In equation (6), we assume reserve price is proportional to common value, with an added error component that is unrelated to any particulars of the sale (“evaluator error” in e ). As noted above, reserve price is set by an outside committee, using a formula based upon the valuation of the land parcel carried out by an independent private land appraiser.
In that sense reserve price is an exogenous valuation of property based on observed and unobserved (to us) aspects of the property and general local market conditions. For the same common values to two different properties, the number of potential bidders will vary with the city in question (number of active land developers, controlled below by city and time fixed effects) and aspects of the property. For example, the potential number of bidders may differ for certain types of uses or properties near or further from the city center.
We implement equation (6) with
for property i in city j which is sold at time t. X ’s are observed property characteristics such as use restriction, area, and distance to the city center. Auction type, dijt, is whether the land bureau chooses a two stage auction (=1) or not (=0), so that D is the effect of auction type on sales price, which we would like to identify. The dummy terms u j and δ t capture city and yearn fixed effects. The arguments in ε ijt are unobserved timevarying city conditions or property characteristics, which controlling for common value (reserve price), may increase the number of potential bidders. These conditions may affect auction choice. The land bureau’s selection of properties into two stage auctions may involve either negative or positive selection, as noted above.
4.1 Selection problem and instruments.
To deal with the auction selection issue, for our baseline results, we estimate a Heckman (1978) endogenous dummy variable model, with a selection control function based on the inverse Mill’s ratio of a probit on auction type, 17 as well as the MLE version of that. Also we do IV estimation. We experimented with adding interactions of auction type with covariates to the IV specification, allowing auction effects to vary with covariates but the effects are not instructive, especially given we already have a reduced form specification. 18 Instrumental, or control function variables are ones which we think affect selection of auction type by the land bureau, but not sales values conditional on our covariates.
We have several instruments which appear to influence choice of auction type.
We generally use two sets of two instruments each. Most arise from a pattern in the data which is illustrated for the first set, as follows. In the month before a new party secretary takes office in a city, the land bureau switches more to using English auctions and then a month later it switches back, in fact switching away from English auctions (in effect, catching-up to its usual mix). We view this as the land bureau being cautious: “cleaningThe selection terms are respectively φ ( Z ijtγˆ ) / Φ ( Z ijtγˆ ) and -φ ( Z ijtγˆ ) / (1 − Φ ( Z ijtγˆ )) where Z ijt, γˆ are the covariates and estimated parameters from the probit on auction type.
We experimented with allowing treatment effects to vary with observables, by adding variables for auction type interacted with the deviation of property characteristics from their means. In OLS the interactions are not significant. In the IV (2SLS) results, the interactions are somewhat statistically stronger and the average treatment effect rises from -.53 (with 7 instruments) to -.81. However there is little variation in treatment responses as covariates go from low to high values.
up” temporarily in the face of uncertainty about the new party secretary’s views on land market corruption; and then returning to business as usual. The same phenomenon occurs with the second set, land corruption cases, although the timing is different.
We have the number of cases per month that relate to real estate corruption in any city j, reported on Google China. Such cases could involve the removal of a major local government official, the indictment of officials, the execution of officials, or a criminal investigation on land transactions. During this month when a case occurs, officials are more careful and schedule more English auctions. A month later they again revert and catch-up to business as usual. A few months after the case, a sanitized report on the case (the average is about.03 reports per city per month) is announced on state run news agencies and picked up by Google China. The announcements on Google China appear to occur 3 months after the case, in the sense that 3 months earlier English auctions jump up, followed in the next month by a drop down.
We have two other types of instruments as well and use them in some robustness experiments. We have a source on corruption investigations more generally, which is the number of news reports per month by the main news agency in China, Xinhua, on corruption in any city j. Xinhua is a state run news agency. Our hypothesized scenario is the city government, the local party, or the National Audit Office conducts an enquiry into local corruption, of which the local land bureau is fully aware. Again, during this month, officials are more careful and schedule more English auctions. A month later they again revert and catch-up to business as usual. A couple of months after the investigation, Xinhua reports on the investigation (the average is about.9 reports per city per month).
Thus English auctions increase 2 months before the month of the news report and decrease the next month. This timing of the pattern of one month up followed by one month down is found by experimentation in the data, but it is a clear pattern in all three situations—new party secretary, real estate corruption cases, and corruption investigations.
Finally, we have a measure of the pressure on the land bureau to raise more money through land sales, coming from the city government. We measure the gap between budgetary expenditures of the city E and on-the books revenue R. An
instrument would be the lagged growth in the relative deficit:
( E jt −1 − R jt −1 ) / R jt −1 − ( E jt − 2 − R jt − 2 ) / R jt − 2. With city fixed effects we would effectively be instrumenting with the lagged rate of change in this gap and are treating this growth rate as somewhat idiosyncratic and not connected to city demand conditions that would affect the current and future housing market (given city and year fixed effects). Higher lagged deficit growth rates induce more English auctions.
In summary, in the tables in the paper, we use just the first two sets of instruments: party secretary change and real estate corruptions cases. Thus our vector of instruments Z consists of dummy variables for any listing which occurs when a new party secretary takes office (one month lead and one month lag) and dummy variables for any listing which occurs when Google reports a land use corruption case (three months lead and two months lead). These are the strongest instruments; and the Google reporting of corruption is directly connected to real estate corruption. Growth in the relative city fiscal on-the-books deficit in the year before the listing is also a strong instrument at times but is potentially objectionable with only annual variation over 2003-2007 and the potential to be related to real estate prices. We will report (the almost identical) results for key situations, using all 7 instruments. And for few experiments using sub-samples in the paper which are reported just in the text, we use all 7 instruments in order to have sufficient variations within sub-sample cities and time periods for instruments to have some strength. Next we examine the strength of the instruments and later we report results on tests of their validity.
Choice of auction type Before turning to the sales price estimation we examine the choice of auction type, both to see the role of the instruments and to examine the choice itself. Results are in Table 4, for the situation where we include 4 instruments and where we include all 7. The effect of reserve price on auction type is essentially zero, which is consistent with the idea that reserve price setting is independent of auction choice. Choice of auction type is significantly influenced by land use, where the base case, commercial land, is much more likely to be sold in two stage auction, consistent with Table 3. Commercial land consists of smaller plots, which may be of more interest to specialized neighborhood developers within the city and may have fewer potential bidders. Also, more likely to be sold at two stage auction is land near rails (probably land urbanized in the Maoist era) but not near highways (land urbanized more recently).
Of particular interest is how instruments influence auction choice. In column 1, the variables for the change in party secretary and for announcements of land corruption cases have the hypothesized patterns and are generally significant. In column 2 the other three instruments have the hypothesized effects as well. For four instruments the Fstatistic based on the change in the value of the LLF from adding instruments to the probit is 8.1. This is not as high as one would like, but it is reasonable in a context where we have city fixed effects. Going to seven instruments lowers the 1st stage F-statistic, one reason for settling on four instruments.
4.2 Sales Price Results Sales price results are in Table 5. In all specifications, a 1% increase in reserve price raises sales price by just over 0.9%. Why is the elasticity less than 1? A higher reserve price also contains an effect to discourage entry of potential bidders (where we assume appraisers set a reserve price that is common value plus an idiosyncratic error component). Property characteristics are interpreted to affect the number of potential bidders, conditional on reserve price. Sales prices are distinctly lower for larger plots which may be less manageable, or have fewer experienced developers who would try to utilize them.
The key variable concerns choice of auction type. In OLS estimation, prices are lower for two stage auctions by 17%, as an assumed average effect. With correction for selection, the coefficient has a much larger negative value. The Heckman MLE estimate is about -0.70, about 4 times larger in absolute value. The IV coefficient (standard error) when the first stage simply uses the 4 instruments (i.e., linear probability) is similar, -.646 (.267) for LIML. Coefficients are fairly precisely estimated.
While the Heckman MLE is the preferred specification, we note the result is sensitive to some alternatives: the 2-step Heckman estimate which is less efficient and the LIML estimate with predicted probabilities from a first stage Probit as instruments yield coefficients that are smaller in magnitude: