Environmental Economics

Conducting hypothesis tests with a Turnbull estimator with pooled data drives me crazy

I wrote a referee comment to the effect of:

Many contingent valuation method researchers use the nonparametrice Turnbull WTP estimates for hypothesis testing. This is inappropriate when the data must be "pooled" to get the willingness to pay (e.g., the "vote in favor" variable) to decrease with the cost amount. Sometimes, due to small samples, poorly chosen cost amounts or respondent inattentiveness, the percentage of “vote in favor” responses is not monotonically decreasing with the cost amount. The Turnbull estimator requires that the “vote in favor” responses are pooled over prices until the pooled responses are monotonically decreasing. This is, in effect, a recoding of the dependent variable. This makes the WTP estimates inappropriate for hypothesis testing. 

The authors halfway defended their practice because everyone does it. Do I have to be your parent? If everyone does it, does that make it right?

Are there any other examples in the literature where we allow researchers to recode their dependent variable so that it conforms to theory and then use the recoded data for hypothesis testing?

Understanding Bidding Behavior in Multi-unit Experimental Auctions with Latent Class Models

Tanga Mohr and John Whitehead [1]
Department of Economics 
Appalachian State University

Introduction

The Regional Greenhouse Gas Initiative (RGGI) is a cap-and-trade program that covers the electric power sector in more than 10 northeastern states. The cap-and-trade program creates markets for a limited number CO2 allowances, reducing greenhouse gases. Laboratory experiments were used to inform RGGI about the most efficient design for the primary auction and the secondary markets (e.g., Shobe et al. 2010). These experiments were single unit auctions but RGGI conducts multi-unit auctions. The purpose of this research is to explore the efficiency of multi-unit auction designs in the RGGI context.

Auctions

In first price auctions, bidders pay their bid. Theory predicts that bidders in first price auctions of a single unit will shade their bids. In second price auctions, all winning bidders pay the same market clearing bid. Theory predicts that bids will be equal to value in second price auctions of a single unit. Theory is not so clear in first and second price multi-unit auctions (Khezr and Cumpston 2022).

Real auctions are incentivized; i.e., subject earnings are real and depend on bidding behavior. Hypothetical auctions are not incentivized; i.e., subject earnings are fixed and do not depend on bidding behavior. We expect incentivized subjects to make bids closer to theoretical predictions (noting that theoretical predictions are not sharp in multi-unit auctions) (see Mohr and Whitehead 2023).

Methods

We conducted multi-unit induced value auctions using the VECONLAB platform. In induced value auctions, subjects are told how much an item is worth and then make a bid for that item. Each subject has demand for three units and the induced value for each unit differs in each round and over 6 rounds of bidding. We have 74 subjects in four treatments:

1. Real, 1st price auction (n=25)
2. Hypothetical, 1st price auction (n=14)
3. Real, 2nd price auction (n=17)
4. Hypothetical, 2nd price auction (n=18)

The randomly assigned induced values from from $51 to $150. The means of the bids and induced values suggest that experimental subjects shade their bids in each treatment. 

Table 2. Variable means

	1st price, real	1st price, hypo	2nd price, real	2nd price, hypo
Bid	70.90	71.28	77.11	61.95
Induced Value	98.75	99.61	97.65	97.19
Sample size	450	306	324	324

Results

We use latent class regression models to explore various bidding strategies that were used by subjects. The model is:

Bid = a + b*Value + c*Hypointr

where Bid is the Bid, Value is the induced value and Hypointr is the interaction between the hypothetical treatment and the induced value.

Using naive OLS models with clustered standard errors (assuming that all subjects behave in the same way), we find no differences in bidding behavior in the first price auctions between real and hypothetical treatments (see Table 2). In the second price auction, hypothetical bids are 16% below real bids. 

Table 2. Ordinary least squares regression with clustered standard errors
(Dependent variable = Bid)

Using latent class models we identify two different types of bidding behavior for both auctions. In the first price auctions one class suggests that subjects in the hypothetical session bid their value and shade their bids by 75% in the real sessions. In the other class, all subjects shade their bids by 65%.

In the second price auctions one class suggests that real and hypothetical subjects shade their bids by 84% and 88%, respectively. In the other class, real and hypothetical subjects shade their bids by 72% and 53%, respectively.

Table 2. Latent class ordinary least squares regression with clustered standard errors
(Dependent variable = Bid)

Conclusions

We find some evidence that real and hypothetical bidding behavior is different in first and second price auctions. Hypotheical bids are greater than real bids in one class in the first price auction. Hypothetical bidders tend to bid their value which would lead to zero profits. In the 2nd price auctions where we might expect bidders to bid their value, subjects in both real and hypothetical auctions shade their bids. In one class of behavior, subjects in the hypothetical treatments shade their bids by significantly more than subjects in the real treatment. Latent class models can lend additional insights to experimental auction behavior. We plan to conduct more real (i.e., incentivized) first and second price auctions in the future. [2] 

References

Khezr, Peyman, and Anne Cumpston. "A review of multiunit auctions with homogeneous goods." Journal of Economic Surveys 36, no. 4 (2022): 1225-1247.

Mohr, Tanga, and John C. Whitehead. "External validity of inferred attribute non-attendance: Evidence from a laboratory experiment with real and hypothetical payoffs." Department of Economics Working Paper No. 23-05. 2023.

Shobe, William, Karen Palmer, Erica Myers, Charles Holt, Jacob Goeree, and Dallas Burtraw. "An experimental analysis of auctioning emission allowances under a loose cap." Agricultural and Resource Economics Review 39, no. 2 (2010): 162-175.

Note

[1] This study was funded by the Walker College of Business Dean's Club. It was conducted with a student at Appalachian State University who was going to use it for an Honors Thesis. The student ghosted on us and we're left with the responsibility of producing a poster for the Dean's Club poster session (a requirement for securing Dean's Club funding). 

[2] Updated October 21

A new working paper shows that everything isn't fine

In which we* use old-timey contingent valuation willingness to pay for a recreation trip questions. After this paper and others (in the past and in the future), I'm thinking that attribute non-attendance mitigates hypothetical bias, fat tails, scope insensitivity, etc. I'm not sure why it hasn't caught on 100% yet. Everyone seems to think that if we can only use stated preference "best practices" then everything is going to be fine. I think no, everything isn't fine (and that doesn't even factor in the enormous cost of stated preference "best practices."

Here is the link: https://econpapers.repec.org/paper/aplwpaper/23-09.htm

*Authors: John C. Whitehead, William P. Anderson, Jr., Dennis Guignet, Craig E. Landry and O. Ashton Morgan

**Not that we had enough money for "best practices" for this paper.

This looks like something that a real environmental economist might be able to use in their research*

From Data are Plural (10/11): 

Michigan air permit violations. For local news organization Planet Detroit, freelance journalist Shelby Jouppi has built a daily-updating dashboard of air quality permit violations cited by Michigan’s Department of Environment, Great Lakes and Energy. The dataset lists 1,500+ violation notices since 2018; for each, it provides the notice date and findings, facility name and location, and more. To construct it, Jouppi had to scrape individual notice PDFs from the department’s website and then extract the information from those documents. Read more: “Southwest Detroit steel slag processor receives 12th air quality permit violation for fallout since 2018,” an article by Jouppi based on the data.

Here is a screenshot of the map:

*Not someone who uses stated preference data.

Daily demand and supply: olive oil

From the WaPo (Olive oil prices reach record highs as Spain’s harvest is halved):

Extreme heat, wildfires and drought have decimated much of the world’s olive oil harvest yet again, driving prices to a record high of $9,000 per metric ton.

Most home cooks aren’t buying olive oil by the ton. But retail olive oil prices in the United States have risen in recent years because of extreme weather in olive-oil-producing countries, growing 12.5 percent this year atop an 8.8 percent increase in 2022, according to Circana, a Chicago-based market research firm.

Spain, the source of half the world’s olive oil supply and the global price setter, in May reported a drop in production of 48 percent compared to last year. Concerns intensified following the release of the most recent olive oil report from the Spanish government, which showed dwindling supplies in August.

"Oh, dear!"

 

Errata: Estimating recreation benefits through joint estimation of revealed and stated preference discrete choice data

I recently received a request for the NLogit code for this article:

Whitehead, John C., and Daniel K. Lew. "Estimating recreation benefits through joint estimation of revealed and stated preference discrete choice data." Empirical Economics 58 (2020): 2009-2029.

I was happy to oblige but it took a second because, since we estimated those models, I had gone into the program and tried a bunch of attribute non-attendance models. The program was a mess. So, I had to hunt the different models down and rerun everything to make sure it was working and ... discovered a minor error in Table 6. The scaled multinomial logit model was estimated with both the revealed and stated preference data so the number of time periods should be 8 instead of 4. Ugh.

 

Breaking news: "State parks won’t be closed because of the shutdown" (WSJ)

From the WSJ (National Parks Will Close if Government Shuts Down):

National parks will close their gates if lawmakers don’t pass legislation to keep the federal government funded by the end of this week.

The Biden administration announced Friday morning that sites run by the National Park Service will close if government funding lapses on Sunday. The administration will release a full contingency plan later Friday. The closures would roll out over the weekend and into Monday if a shutdown does occur, a senior Interior Department official said.

The closures will affect national parks, including sites like Yosemite and Yellowstone, and other monuments and sites like the National Mall and memorials in Washington, D.C. During the shutdown, thousands of park rangers will be furloughed.

Visitors will still be able to access some parks during the shutdown. While some parks have entry points that can be closed to guests, visitors could go to many other federally run destinations that are easier to access. State parks won’t be closed because of the shutdown.

The nonprofit National Parks Conservation Association, citing government data, projects that the parks could see nearly one million fewer visitors and an economic loss of as much as $70 million for every day that the destinations are closed in October.

Here is one reason to close them down:

During the most recent government shutdown, the Trump administration kept national parks open with lower staffing levels. As travelers continued to visit, trash and toilet facilities overflowed at some locations. Visitors also caused damage to some locations, including Joshua Tree National Park.

...

The Government Accountability Office rebuked the Trump administration for keeping the parks up and running. In a 2019 legal opinion, the GAO said the administration ran afoul of federal rules that dictate how money can be spent during a lapse in appropriations. The GAO also warned that similar moves in the future would be considered “knowing and willful violations” of the law.

Thanks Kevin!

Oh, and the WSJ felt obliged to add this:

State parks won’t be closed because of the shutdown.

Because ... um, they're state parks?

I presented a paper in both Alaska and Hawaii this year

It was the same paper. I was excited to boast that I might be the only person who has ever presented the same paper in the same year in the 49th and 50th states. As it turns out, I wasn't the only one who has ever done that. Another person presented the same paper at the same two places at almost the same times! Ugh.

In June we attended the 2023 Summer Workshop at the University of Alaska Anchorage and presented during the same session.

In September we both presented in the Workshop on Energy and Environmental Research at the University of Hawaii just one week apart. Also, my presentation was virtual (not sure about Renato's). 

But still, the population of economists who have presented a paper in both Alaska and Hawaii within a 4 month window has to be small. Maybe n=2. In that way, I'm special.

Here is the abstract and the presentation (PDF):

We estimate economic benefits of avoiding reductions in drinking water quality due to sea level rise accruing to North Carolina (NC) coastal tourists. Using stated preference stated preference methods data with recent coastal visitors, we find that tourists are 2%, 8%, and 11% less likely to take an overnight trip if drinking water tastes slightly, moderately, or very salty at their chosen destination. The majority of those who decline a trip would take a trip to another NC beach without water quality issues, others would take another type of trip, with a minority opting to stay home. Willingness to pay for an overnight beach trip declines with the salty taste of drinking water. We find evidence of attribute non-attendance in the stated preference data, which impacts the regression model and willingness to pay for trips. Combining economic and hydrology models, annual aggregate welfare losses due to low drinking water quality could be as high as $401 million, $656 million and $1.02 billion in 2040, 2060 and 2080.

What the abstract doesn't describe is the funnest part. We get to use old-fashioned contingent valuation for valuing these trips ... and it works (it always works!)! The literature review starts with Brown and Hammock (REStat 1973), spends some time in the 1980s and 1990s and then skips to 2017. 

A working paper should be out in about a week (incorporating comments from the paper posted at WEER). 

Footnote:

And hey, if you are doing stated preference research you probably have attribute non-attendance on your cost variable which biases your WTP estimates upwards. Even if you follow all of the guidelines in Johnston et al. (JAERE 2017) (which would make your study a multi-year, multi-million dollar endeavor).

Teaching consumer surplus with the Becker-DeGroot-Marschak method in an environmental and resource economics course

The other day in my intro environmental and natural resource economics class we used the BDM method to elicit willingness to pay (WTP) values for an App State travel tumbler (I paid $26 with a faculty discount at the university bookstore). I explained the BDM with these slides [Download BDM] and each student had a "payment card" for revealing their WTP [Download BDM-WTP]. The average WTP was $6.23 (n=26). I entered the WTP values into Excel, sorted them from highest to lowest and plotted them along with the randomly chosen price ($7). At this price, 12 of 26 students would have purchased the tumbler and enjoyed a consumer surplus of $45 (CS = WTP - price).

As a preview of what we are going to do later in the semester, I simulated a dichotomous choice stated preference exercise. For each WTP value I randomly chose a price that ranged from $2 to $14 (average = $5.69) and simulated whether the consumer would purchase the tumbler or not. Fifty percent of the consumers would purchase the product. I then estimated a linear probability model: Pr(purchase=1) = 0.711 - 0.0372 x Price. Plotting this line and calculating the consumer surplus area of the triangle yields a CS estimate of $6.81 -- very close to the actual CS average. I told them that this valuation approach is called the dichotomous choice WTP survey approach and is used in E&R economics (and marketing) to estimate WTP values for environmental amenities and consumer goods.

I randomly chose one of the student's WTP sheets and the WTP was less than $7. Then I chose another and the WTP was $15. This student paid me $7 and is, apparently, enjoying $8 of consumer surplus. 

I think I've convinced many of the students in this class that consumer surplus is equivalent to getting a "good deal". Now the trick is to convince them that WTP measures the value of nonmarket goods ... too be continued.

Reference

Schubert, Jens. "Using the Becker-DeGroot-Marschak Mechanism to Teach Willingness to Pay and Consumer Surplus." Journal of Economics Teaching 8, no. 1 (2023): 1-11. [URL]

How hard is it to proofread your paper after you use ChatGPT to write it?

From Nature News (Scientific sleuths spot dishonest ChatGPT use in papers) via Retraction Watch Weekend Reads [brackets added below]:

Searching for key phrases picks up only naive undeclared uses of ChatGPT — in which authors forgot to edit out the telltale signs — so the number of undisclosed peer-reviewed papers generated with the undeclared assistance of ChatGPT is likely to be much greater. “It’s only the tip of the iceberg,” ["scientific sleuth Guillame"] Cabanac says. (The telltale signs change too: ChatGPT’s ‘Regenerate response’ button changed earlier this year to ‘Regenerate’ in an update to the tool).

Cabanac has detected typical ChatGPT phrases in a handful of papers published in Elsevier journals. The latest is a paper that was published on 3 August in Resources Policy that explored the impact of e-commerce on fossil-fuel efficiency in developing countries. Cabanac noticed that some of the equations in the paper didn’t make sense, but the giveaway was above a table: ‘Please note that as an AI language model, I am unable to generate specific tables or conduct tests …’

I wanted to see it for myself and I bet you do too (at the bottom of the screenshot):

Let's say you want to use ChatGPT to help write your papers (because writing you own papers is soooooo hard). First, you must admit it in your acknowledgements or your risk retraction. Second, actually read your own paper and edit this nonsense out.