knowledge is powder: risk taking behavior, forecasting services … · 2019. 12. 31. · knowledge...

Knowledge is Powder: Risk Taking Behavior, Forecasting Services and

Information Sharing in Natural Hazard Management

Perry Ferrell∗

Department of Economics, West Virginia University

[email protected]

Abstract

When making decisions under uncertainty or engaging in risky behaviors, people rely on

their heuristics as well as information collected from outside sources. Does the introduction

of additional external information lead to better outcomes for the relevant agents? Using

data from reported avalanche incidents in Colorado, this paper shows modest effects that

avalanche forecasting services reduce dangerous incidents. Time variation in the initiation

and termination of daily forecast publishing is used to estimate the effect of the forecast on

unintentional human involved incidents. Individuals in this setting also have direct access to

inputs for the forecasting services, field observation forums on the forecasting center’s web

site, where observations are shared by forecasters and voluntarily contributed by backcountry

users. In addition to effects of the expert’s aggregation of information in the forecast, this

paper investigates the impact of disaggregated information shared by peers and forecasters on

outcomes.

Keywords Risky Behavior, Forecasting, Uncertainty, Avalanches

JEL Codes: D83 H44 Q58

∗1601 University Av, Morgantown, WV Acknowledgments: This research is supported by the Property and

Environment Research Center. The author would like to thanks Josh P. Hill, Wally Thurman, Randy Rucker, and the

many helpful individuals at PERC. Additionally, I would like to thank the fellow graduate scholars, Andres Mendez,

Casey Rozowski, and Henry Holmes for kindly providing feedback on the most recent thought in my head. I owe

much gratitude for the knowledge, history, and data sharing from the avalanche forecasting community, especially

Mike Cooperstien, Brian Lazar, Alex Marienthal, and Karl Birkeland.

1

mailto:[email protected]

Ferrell-Knowledge is Powder

I. Introduction

Taking on risk is a necessary part of everyday life. Traveling anywhere requires an individual to

subsume the underlying risks associated with their chosen method of travel. And it is natural for

people to want to minimize the risk they are exposed to, but this risk may be part of the enjoyment

of the activity. Sometimes individuals do not utilize options available to them to decrease the

risks they are undertaking, like not wearing a seat belt in an automobile or riding a motorcycle

without a helmet. Often, policies rely on coercion to incentivize individuals to decrease the amount

of risk they undertake in a given activity, such as mandatory seat belt and helmet laws in the

aforementioned case. While coercive sanctions are often justified because negative externalities

emanate from the risky behavior in question, such as drunk driving or maternal substance abuse,

they are also used in cases where the negative externalities are minimal to non-existant. Motorcycle

helmet laws are a prime example, where the additional risk from riding a motorcycle without a

helmet cannot spill over to other individuals.

Other policy options exist beyond coercion to incentivize optimal risk taking. Instead of some

negative incentive, sometimes programs as simple as providing individuals more information

about the relevant risky behavior can discourage undue risk taking. For example, extensive public

heath campaigns to inform smokers about the long-term health consequences of smoking aided

in reducing smoking rates in Western countries. Often these policies are implemented together

introducing negative incentives as well as increasing efforts to better inform individuals about the

risk. The Surgeon General’s Warnings are used along with increased tobacco taxes. In this paper I

investigate the extent to which information alone impacts risk taking behavior, using a unique

setting free from negative incentives and information consumption is entirely voluntary, avalanche

hazard management in winter backcountry recreation.

There are a number of ways to postulate increases in an individual’s information set could

license more risk taking behavior or may lead to observing more incidents. If uncertainty about

the underlying risk deters people from engaging in a risky activity, then more more information

might lead to more incidents by increasing the extensive margin. For example, news informing

people of road plowing in winter weather conditions might elicit risk averse types to travel who

would not have otherwise without seeing cleared roads and snow plows on the news. Additional

2


information may also increase observed incidents along the extensive margin, licensing less risk

averse types to take on more risk.

Backcountry recreators use information provided by a US Forest Service (USFS) center, where

they publish a daily bulletin during peak season with danger ratings and information about the

main avalanche problems of concern. Forecasters create the daily bulletin based on weather, expe-

rience, and observations from the field, which are provided by USFS observers and backcountry

users. Snow stability tests and field observations are conducted by employees of the avalanche

centers, as well as other professionals such as guides and instructors, and unaffiliated backcountry

travelers. Most USFS avalanche centers have a forum on their website where the general public

can submit reports and observations in addition to the forecasters making their specific field

observations available to the public. This creates a setting where backcountry users can utilize two

information sources in their decision making, the centralized daily bulletin and the jointly sourced

observation forum.

Identifying the impact of the forecast and information on outcomes in the backcountry requires

careful examination of observational data. A randomized control trial with the forecaster products

is not possible for obvious reasons. Additionally, simply regressing outcomes on forecaster oper-

ations is rife with endogeneity because forecasters operate when the underlying danger level is

highest. I use differences in dates that forecasters commence and cease daily forecasting operations

to identify the effects of the published advisory bulletin on incidents. Incidents are considered any

time a backcountry traveler reports unintentionally triggering an avalanche. This includes reports

from skiers, snowboarders, snowmobilers, as well as alpine climbers and hikers. A daily panel

across the 10 forecasted zones in Colorado shows a modest negative impact of daily advisory

publishing on reported incidents.

Estimating the effects of disaggregated information supplied by the USFS forecasters and from

an individual’s peers, and separating that from the forecast, is an equally difficult task. The same

underlying factors that would increase observed incidents would also increase reporting of other

observations, and these latent factors are considered in the forecast. To net out the impact of daily

forecasts, I exploit policy changes in the 2018-19 season that required the Colorado forecasting

office to remain open at full capacity two weeks longer than normal, and continue to operate on a

daily but reduced capacity for four weeks beyond that. Utah is used as a control group.

3


The result from a difference in difference specification with the CAIC season extension shows

that additional forecasting services does reduce reported incident rate. Unsurprisingly, providing

risk takers with more information about the risky behavior they choose to engage in results in

fewer incidents.

II. Background

i. History of Avalanche Forecasting

Formal avalanche forecasting began in Switzerland between the world wars.1 Avalanches

during battles in the alpine during World War One added to the hazards of war. The Swiss

military started training units on avalanche safety and advising them on current danger levels.

This operation eventually morphed into the Swiss Institute for Snow and Avalanche Research

which began issuing bulletins for non-military users. The Swiss center has one central office

which provides bulletins for zones covering the whole country. This centralized model would later

influence the formation of avalanche centers in North America.

Avalanche forecasting in the United States (and arguably recreation in the alpine) started

with the return of 10th Mountain Division soldiers from WWII. Monty Atwater, a 10th Mountain

veteran began avalanche mitigation outside of Salt Lake City, UT in Little Cottonwood Canyon in

the 1940s. There he pioneered the use of explosives and military artillery to purposefully trigger

slides (Atwater, 1968). For several decades, avalanche work focused on mitigation in specific

slide paths or passes to keep highway corridors, ski areas, and industry in the alpine open. The

first program focused on advising backcountry recreationists started in 1973 with the US Forest

Service’s Colorado Avalanche Warning Program (Williams, 1998).

ii. Funding

Avalanche centers in the US all have an attached non-profit that raises funds to support forecast

production and other activities like outreach. Their annual operating budget is partially supplied

1Origins of the avalanche bulletin – history and background.

4


Figure 1: Budget sources for the Utah Avalanche Center in 2018/19 season. Soure: UAC 18/19 Annual report

by government funds and by private donations through the attach 501(c). The share of revenues

coming from government sources vs private contributions varies from center to center. The Utah

Avalanche Center (UAC) received 48% of its annual operating budget from contributions and

fundraising in 2018/19, while 18% was contributed from the USFS and 11% from the State of

Utah.2 In addition to these sources the UAC sells merchandise and lift tickets, which are donated

by ski resorts, as well as contributions from local government organizations like fire and rescue

services and earns revenue from .

iii. Colorado

Avalanche forecasting in Colorado evolved differently from the other centers in the United

States. Being one of the first centers in the United States, Colorado and the Pacific Northwest Center

started based around the European model of one central office that received information from all

over the forecasted zones to issue hazard ratings, as opposed the newer, more localized offices

like the Gallatin National Forest Avalanche Center in Bozeman, MT. The Colorado Avalanche

Information Center (CAIC), started in 1973, is not a direct subsidiary of the US Forest Service

like the other centers across the country, though it is affiliated. It is a joint operation between

the Colorado Department of Natural Resources and the Colorado Department of Transportation.

While most states with highways exposed to avalanche terrain have a DOT wing responsible for

2The total revenues for the UAC totaled $1,019,000 for the 2018/19 season.

5


Figure 2: Budget sources for the CAIC in the 2018/19 season. Soure: CAIC 18/19 Annual report

keeping the roads open, and may share information with the USFS forecasters covering public

lands, no other centers formally work with their DOT counterparts like Colorado.

The CAIC’s revenue for operations in the 2018/19 season was slightly over $1.5 million.

Since the CAIC has a direct partnership with the state’s DOT, their largest funding source is

disbursement from the state government at 87 % in 20189-19 and only 3% from USFS sources. The

5% of their budget for forecasting activities is supplied by their attached non-profit. The Friends

of the CAIC funds much of the website and app development for the center directly, outside of

the CAIC operational budget. Unlike Utah, much of the CAIC’s operational expenses like staff

salaries are supplied from government revenue. The funding to broadcast the information to the

public, and educate people on how to use it, is largely supplied by their non-profit. Because of the

CAIC’s relationship with the State of Colorado, and because Colorado has much higher elevation

avalanche-prone terrain, the CAIC contract with the State of Colorado was adapted in the 2018/19

season requiring them to issue daily recreational forecasts through the end of May every year, and

then issue a daily forecast for the month of May where the 10 zones in figure 3 are aggregated

into 3, northern, central, and southern mountains. Most other USFS centers cease daily forecasting

sometime in April depending on weather and funding.

6


Figure 3: CAIC Forecast Regions

iv. Forecasting Process

Forecasters produce a daily advisory bulletin to advise backcountry travelers of the avalanche

danger in the backcountry. A sample advisory is included in the appendix, see A1. Forecasters

in North America use some variant of the ’Conceptual Model of Avalanche Hazard’ to arrive at

the daily bulletin (Statham et al., 2010). They will attribute a danger level to various aspects and

elevations based on the danger scale shown in Figure 4. Colorado issues a danger rating for 3

elevation bands, above tree line, near tree line, and below tree line.

At = F(At−1, E(Wt), QFt−1 , QPt−1) (1)

For the purposes of this paper I will model the advisory bulletin production function as

Equation 1. The daily advisory bulletin, At, is a function the previous bulletin which is augmented

and updated by the experience and knowledge of the forecaster using current information. This

current information is a set of inputs including the expectation of the day weather, E(Wt), and

the recent field observations gathered by the forecasters, QF, and submitted by the public, QP. I

assume that forecasters objective function is to minimize the difference between their published

advisory bulletin, A, and the true underlying riskiness of the snowpack, A, which is unobserved.

7


Figure 4: North American Avalanche Danger Scale (Statham et al., 2010)

QFt = FF(At, Wt, QFt−1 , QPt−1) (2)

QPt = FP(At, Wt, QFt−1 , QPt−1 , t) (3)

Information is supplied by the forecasters in equation 2 and by the public according to equation

3. Both parties make travel decision based on the daily advisory, At, and the weather on that day

which both affect terrain travel decision by each party and what they can observe. Additionally,

demand for current information in t is a function of the quantity and type of information gathered

in t − 1 by both parties. I include t in equation 3 because the number of backcountry travelers is

greatly affected by the draw of t from the set of days in the week. There will be a lot more people

in or near avalanche terrain on weekends relative to weekdays. The forecasting offices run at a

relatively stable staffing level 7 days a week throughout the season so I assume the draw of t does

not affect equation 2.

8


v. Backcountry Users and the Avalanche Bulletin

I will lay out a generalized decision sequence for an avalanche savvy backcountry user as

follows for reference.

1. When deciding whether to head to the backcountry and in or around avalanche terrain, read

the avalanche advisory for the day.

2. If a decision is made to enter avalanche terrain, consider a route that minimizes risk based on

the information in the advisory, and possibly additional sources like the observation forums

depending on skill level.

3. Decide what additional information to collect while on the route, such as stopping to dig

pits and conduct formal stability tests.

4. Continue to adjust (or not) the route based on the information from the advisory and personal

observations in the field.

There are two important trade-offs faced by the BCU which make the advisory bulletin relevant.

Most people do not enter avalanche terrain just for the thrill of risk exposure alone, but because

ideal slope angle for skiing is also prime avalanche terrain3 In addition to this, avalanche danger is

also increasing with snow quality. For example, most skiers would prefer fresh, soft snow from a

recent storm opposed to old, refrozen snow on which it is harder to control skis or a snowmobile.

The old refrozen snow is likely to have a stronger bond to other snow layers compared to the

recent storm snow which is more likely to slide, ceteris paribus. Without overgeneralizing, the

more enjoyable the snow and terrain conditions are for the backcountry user, the higher the danger

associated with avalanches.

These trade offs require the backcountry user to venture out on days where the danger level is

moderate or considerable if she is searching for better snow conditions. The avalanche advisory

danger levels in figure 4 are a pretty coarse signal. Two considerable days may be rated as such for

separate reasons and a specific location might carry a high risk one day, but be a safer alternative

under separate circumstances. Additionally there is substantial variation in prescribed danger

3Most avalanches occur on slope angles between 30◦ and 45◦.

9


ratings between forecasters when operating on the same information sets (Lazar et al., 2016). In

the spirit of mountaineering, individuals are left to take the level of risk they are willing to accept.

This is where supplemental information beyond the advisory bulletin becomes useful for more

advanced backcountry users when making decisions to minimize risk within a given advisory

danger rating. Backcountry users are then faced with the decision to enter information collected

while in the field into the public domain.

vi. Field Observation Reporting

Most USFS avalanche centers collect information through an online forum, which is available on

the center’s website on a separate page from the avalanche advisory bulletin. Public backcountry

users can submit information through a portal on the center’s website and it appears side by side

with the field observations posted by the forecasters, though an affiliation is usually included so

people can know if the source is an amateur or professional. Some of these forums appear in

a more blog like style, while others are in a more formal table linking users to a separate page

containing the full report.4 The forecasters then use these observations and other information like

weather to produce their end product, the avalanche advisory bulletin.

Field observations that are reported may be of several types including snowpits,5 general

observations, avalanches, and incidents.6 An example of a publicly submitted report is included

in the appendix figure A2. Observations all have the date and source of the report, private

backcountry travelers, professionals such as guides and ski patrol, or forecasters. The public has

the option to submit anonymously. The example report is from an unintentionally triggered slide

where one snowboarder was caught. An assessment of the weather, snowpack, and account of the

incident plus photos of the fracture line and path are included. Observations that do not involve

4See Mount Washington Avalanche Center for an example of a more blog like forum. See CAIC or Sawtooth

Avalanche Center for examples of more formal table like forums.5Snowpits refer to formal investigations of cohesion in the snowpack where a person digs into the layers and

performs stability tests to asses slide potential. See https://avalanche.org/avalanche-encyclopedia/stability-test/ for

more information.6I will refer to avalanches as naturally occurring slides or intentionally triggered by a backcountry traveler. Incidents

are slides where a backcountry traveler was caught or unintentionally triggered a slide.

10

https://www.mountwashingtonavalanchecenter.org/mwac-observations/

https://avalanche.state.co.us/observations/field-reports/

https://www.sawtoothavalanche.com/view-obs.php

https://www.sawtoothavalanche.com/view-obs.php

https://avalanche.org/avalanche-encyclopedia/stability-test/


slide activity might include snowpit information either verbally, from photographs, or from an

online application Snow Pilot for formally documenting snowpit tests.

One unique feature of studying avalanche field observations is the extremely high rate of decay

of the information. The constantly changing state of the snowpack subject to weather and climate

effects, both interday and intraday, makes the information content of field observations extremely

perishable. Especially depending on the recency of major snowfall, wind, or rain, information

included in an observations loses much of its value beyond a 48 to 72 hour window.7 Thus

investment in supplying information is a constant and ongoing process throughout avalanche

season, as opposed to other crowd sourced activities like fundraising drives for local public radio

stations which can achieve economies of scale by conducting fundraisers on an annual or quarterly

basis.

The incentives for a backcountry user to contribute information to this local public good may

come from two sources, which are not exclusive. The first is that they value an accurate daily

avalanche bulletin, for their own use or for others, and are willing to exert effort to increase the

information available to forecasters generating the bulletin. Second, a private individual may

report out of some form of pro-social warm glow, not motivated by having direct input into the

forecaster’s information set, but instead to communicate that information directly to peers using

the forum for the avalanche center. This effect may come from two sources.

Beyond the direct costs of time and effort, other incentives faced by the general public may lead

to under reporting. Fresh untracked snow on public land is a common resource good, rivalrous

in consumption but not excludable. This can create a tight lipped culture about sharing favorite

locations in certain backcountry zones. Skiers can be disincentivized from sending in field reports

because it could publicize private knowledge of good skiing areas, bringing others to track out

fresh snow.8 This effect does not operate through the forecast, but instead from other backcountry

users looking directly at posted observations. If there is specific location information in the report

then it may advertise that area to others that would not have considered going there. Just like

7There are some exceptions to this, like information that signals the presence of certain problem types that may be

persistent. I argue that this is an exception and not a norm. These types of persistent problems forecasters can usually

detect from weather and observations just serve to confirm their inference from the weather.8When a skier descend the mountain, she disturbs the surface of the snow. This decreases the quality of the snow for

the next user.

11

https://snowpilot.org/


surfers protect good surf spots, winter backcountry travelers may be wary of sending in a report

that discloses their ’secret stash’ to non-locals.

Additionally, incidents may be under reported because people do not want to admit mistakes.

Particularly among low consequence incidents where there is not a full burial, severe injury, or

death, those involved may not want to add to the public record an account possibly caused by

poor decision making, inexperience, or hubris.

I do not assume all reports from the general public as perfect substitutes with reports from

professional forecasters. While certain types of information, such as photographs of recent slide

activity are more directly substitutable, more involved assessments such as snow pits and column

tests likely vary in type depending on the source. Stability tests from the general public are not

going to receive as much weight as those same assessments reported by avalanche professionals,

both by the forecasters when making the daily forecast and by other backcountry users when

consuming information in the observation forums.

III. Methods

Inczt = B1Azt + B2Xzt + B3Pt + τt + γzt + εzt (4)

I estimate the number of reported incidents as a function of forecasting operations. The unit

of observation is Inczt, the number of reported incidents where a human was unintentionally

involved in zone z on day t as a function of Azt, the operations of the forecasters and controls

in Xzt including natural avalanche reports, weather, and climate. Zone by season fixed effects

are included with γzt as well as a time trend variable τ. The variables of interest are Pt which

is an indicator variable representing the 10 days leading up to the start of daily forecasting, and

separately the first ten days of the season. Similarly this is replicated around the closing of daily

forecasting.

12


Figure 5: Average reported avalanches and incidents across season

0

1

2

0 100 200

Days into Forecasting Season

Ave

rage

Eve

nts/

day

Event

Avalanches

Accidents

Figure 6: Reported avalanches and incidents around forecasting open and close

0

20

40

60

−20 −10 0 10 20

Days

Eve

nts Event

Avalanches

Accidents

13


IV. Data

Daily avalanche bulletin information and reported avalanche data are obtained from the CAIC.

This includes all variables in a daily forecast9 such as danger level for each elevation, the specified

main avalanche hazards for the day, and on what aspects to expect those hazards, as well as the

predicted frequency and size. Information on avalanches comes from the CAIC database where

they log all avalanches that the forecasts observe or they receive reports of. Information on the

trigger, size, and destructive power, in addition to other variables is included. Of particular interest

for this project is the trigger information. Major avalanche incidents where a person is fully buried

or severely injured is fortunately infrequent. There are only 71 major incidents in my sample

from 2014 to the 2018/19 season in Colorado. I can identify incidents where a human triggered

an avalanche unintentionally, however, based on a combination of variables. It is important to

separate these from intentionally triggered slides, a common practice to asses stability.10 While

many of these incidents did not have major consequences, under slightly different circumstances it

is possible outcomes could be different. The example observation in figure A2 shows one such

example. There are 711 instances of avalanches where a person was unintentionally involved in

my sample. My main outcome variable of interest, lumps both major and minor incidents together.

The sample covers the 2013/14 to 2018/19 season. I consider a season to run from October 15th to

June 15th.

Data on observations were obtained from the CAIC website using the Rvest package for R to

crawl the web pages. I consider an observation any report that is submitted to the CAIC forum

and can be placed in one of the 10 zones covered by CAIC forecasts. Reports from outside the

forecasted areas are not considered in the data. I aggregated these at the daily level between

CAIC employees, general public, and other snow professionals.11 On average CAIC-employed

forecasters will post observations every other day in a given zone. The general public sends about

9An example can be seen in figure A1.10Backcountry users may try to cause an avalanche on a slope from a safe spot before venturing onto it. Common

methods used include ski or slope cutting and cornice drops.11Any observer who reports their name or organization (see figure A2) as part of ski patrol, a guide service, or any

avalanche education courses (because these are done under the supervision of guides) as professional. I keep them

separate from the public because the information likely higher quality and probably not used in the same way by

forecasters.

14


Table 1: Summary Statistics for Colorado, Zone day

Statistic N Mean St. Dev. Min Pctl(25) Pctl(75) Max

Unintentional Trigger 14,008 0.055 0.286 0 0 0 7

Accident 14,008 0.005 0.073 0 0 0 2

Incident 14,008 0.056 0.287 0 0 0 7

Other Avy 14,008 0.591 2.212 0 0 0 65

Open 14,008 0.629

Forecaster Obs 14,008 0.687 1.500 0 0 1 24

Public Obs 14,008 0.345 0.841 0 0 0 11

one fewer observation a week on average and snow professionals contribute considerably less

data but they are also a smaller group relative to the public and not doing observations full time

like CAIC forecasters. The data set is truncated at the avalanche forecasting season running from

November to May. The out of season months are not included.

Figure 5 shows the average number of reported incidents and other avalanches on a given day

across Colorado for the winter seasons in the sample. This graph is normalized to 0 representing

the start of daily forecasting for the season. Figure 6 shows the total number of incidents and other

reported avalanches.

Table 2 shows a simple dummy variable representing forecasting operations on incidents and

the subset of incidents that are major accidents. As expected, incidents rates are higher during the

forecasting season because the forecasters are going to operate during the highest danger time

periods. Because of that, these coefficients should not be interpreted as forecasters increasing

incidents. To investigate the impact of forecasters on incidents, I examine 10 before opening and

after closing to the nearest 10 days in the forecasting season. These results are included in table 3.

A dummy variable is included for the peak season. The post open coefficients are suggestive that

an increase in incidents occurs along with the start of daily forecasting. I would not interpret this

as causal because the forecasters likely anticipate increases in danger from oncoming storm cycles

15


Table 2: Incidents during forecasting season

Dependent variable:

Incident Accident

(1) (2) (3) (4)

Forecasting 0.049∗∗∗ 0.017∗ 0.004∗∗ 0.004

(0.005) (0.010) (0.001) (0.003)

Other Avalanche 0.017∗∗∗ 0.013∗∗∗ 0.002∗∗∗ 0.002∗∗∗

(0.001) (0.001) (0.0003) (0.0003)

Zone FEs N Y N Y

Season FEs N Y N Y

Month FEs N Y N Y

Day of Week FEs N Y N Y

Observations 13,157 13,157 13,157 13,157

Adjusted R2 0.026 0.054 0.004 0.008

Residual Std. Error 0.292 (df = 13154) 0.287 (df = 13127) 0.075 (df = 13154) 0.075 (df = 13127)

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

16


which is factored into their decision to open.

Replacing the forecasting dummy variable with a set of control variables for the daily forecast

yields different results. Table 4 shows these models. There is no significant increase in incidents

after opening when controlling for the zone’s daily prescribed danger levels. However, a modest

decrease in incidents appears before closing for the season. Across this 10 day window, that would

mean 2.8 fewer incidents in Colorado. The post close coefficient is neither negative nor significant.

This suggests that the zone forecast decreases incidents along some margin.

These estimates for the impact of daily forecasting on reported incidents are likely to be

extremely biased downward for several reasons. First the forecasters begin operations before

the danger level has substantially increased, and cease operations after the snowpack begins to

consolidate for the spring. Second, there simply are not that many people out in the backcountry

around the opening and closing of forecasting operations relative to peak season mid-winter. At

the beginning and end of the season, there is minimal snow coverage at elevations around access

points. This almost entirely prevents snowmobile access, and hiking through thin coverage deters

many skiers and snowboarders. Third, the majority of incidents that comprise the data set are

self-reported. I argue that backcountry users are more likely to report information to the center

when that information has more value, i.e potential to affect the advisory bulletin, not just directly

inform other backcountry users.

It is reasonable to argue instead that private backcountry users might increase reporting in

absence of forecaster activity. When a local public good is jointly provided by a government agency

and private voluntary contributions, removing the government provision might increase voluntary

contributions through increased impure altruism or reputational effects (Andreoni, 1990; Bénabou

and Tirole, 2006). To test this I look at other privately reported observations around the start and

end dates. These include people submitting information about naturally occurring avalanches,

intentionally triggered avalanches, snow pits, and other general field observations.

Table 5 replaces reported incidents with other general observations. A modest increase in

reporting occurs when the center opens for the season. This is likely because opening is an endoge-

nous decision determined by the forecasters. They likely commence daily operations for the season

in anticipation of the first major winter storm cycle. This would increase incidents and observations.

17


Table 3: Differences in incident rate around open and close

Dependent variable:

Incidents

(1) (2) (3)

Other Avalanches 0.014∗ 0.014∗ 0.014∗

(0.008) (0.008) (0.007)

Pre Open 0.022∗ 0.020∗ 0.015

(0.012) (0.012) (0.011)

Post Open 0.051∗∗ 0.049∗∗ 0.042∗∗

(0.021) (0.021) (0.018)

Mid Season 0.061∗∗∗ 0.060∗∗∗ 0.046∗∗∗

(0.017) (0.017) (0.014)

Pre Close 0.010 0.010 0.001

(0.007) (0.007) (0.008)

Post Close 0.007 0.007 0.0003

(0.007) (0.007) (0.008)

Season Trend N Y Y

Season Trend2 N N Y

Zone by Season FEs Y Y Y

Observations 14,008 14,008 14,008

R2 0.054 0.054 0.055

Adjusted R2 0.050 0.050 0.050

Residual Std. Error 0.280 (df = 13942) 0.280 (df = 13941) 0.280 (df = 13940)

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

18


Table 4

Dependent variable:

Incidents

(1) (2) (3)

Other Avalanches 0.010 0.010 0.010

(0.008) (0.008) (0.008)

Pre Open 0.021 0.022 0.019

(0.014) (0.014) (0.013)

Post Open 0.005 0.006 0.010

(0.011) (0.012) (0.013)

Pre Close −0.028∗∗∗ −0.028∗∗∗ −0.025∗∗∗

(0.008) (0.008) (0.007)

Post Close 0.008 0.007 0.003

(0.007) (0.008) (0.008)

Forecast Controls Y Y Y

Season Trend N Y Y

Season Trend2 N N Y


Observations 14,008 14,008 14,008

R2 0.067 0.067 0.067

Adjusted R2 0.062 0.062 0.062


Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

19


Table 5

Dependent variable:

General Public Observations

(1) (2) (3)

Incidents 0.549∗∗∗ 0.549∗∗∗ 0.546∗∗∗

(0.064) (0.064) (0.065)

Pre Open 0.131∗ 0.075 0.051

(0.072) (0.061) (0.059)

Post Open 0.168∗∗∗ 0.116∗∗ 0.080∗

(0.055) (0.047) (0.041)

Mid Season 0.427∗∗∗ 0.406∗∗∗ 0.339∗∗∗

(0.089) (0.084) (0.065)

Pre Close 0.080∗∗ 0.091∗∗ 0.049

(0.035) (0.036) (0.043)

Post Close 0.065∗∗ 0.081∗∗ 0.048

(0.031) (0.034) (0.041)

Season Trend N Y Y

Season Trend2 N N Y


Observations 14,008 14,008 14,008

R2 0.202 0.203 0.204

Adjusted R2 0.199 0.200 0.201


Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

20


Table 6

Dependent variable:

General Public Observations

(1) (2) (3)

Incidents 0.471∗∗∗ 0.472∗∗∗ 0.471∗∗∗

(0.084) (0.085) (0.080)

Pre Open 0.125∗ 0.080 0.067

(0.076) (0.065) (0.060)

Post Open −0.164∗∗∗ −0.189∗∗∗ −0.172∗∗∗

(0.046) (0.048) (0.042)

Pre Close −0.220∗∗∗ −0.198∗∗∗ −0.185∗∗∗

(0.047) (0.043) (0.038)

Post Close 0.068∗∗ 0.081∗∗ 0.061∗

(0.034) (0.037) (0.036)

Forecast Controls Y Y Y

Season Trend N Y Y

Season Trend2 N N Y


Observations 14,008 14,008 14,008

R2 0.243 0.244 0.244

Adjusted R2 0.239 0.240 0.240


Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

21


Replacing the forecasting dummy with the forecast variables in Table 6 show noticeable de-

creases in observations before the end of the season. While this casts some doubt in the effects of

forecasting on close calls at the end of the season show in table 4, it does not rule out the impact

of forecasting activities on outcomes for backcountry users. Reporting of other avalanches and

reported close calls and accidents are going to be very highly correlated with each other, and

teasing these apart is a difficult task.

i. Colorado Season Extension

To better isolate the impact of avalanche forecasting services on outcomes I exploit the forecast-

ing season extension in Colorado for 2019 and use Utah as a control. As discussed earlier, CAIC’s

contract with the State required them to continure forecasting each zone until the end April and

then continue on a reduced format through May. Utah ceased daily forecasting activities on April,

20th, 2019 per their usual method of operations.

Inczt = B1Azt + B2Postseasont + B3(Postseasont ∗ COz) + B4(Postseasont ∗ 2019t)

+B5(Postseasont ∗ 2019t ∗ COz) + Xzt + γzt + εzt

(5)

I estimate a difference in difference model using Utah as control when Colorado extends fore-

casting in 2019. I use the highest forecasted danger level at any elevation or aspect for the days

forecast, because Utah and Colorado report daily avalanche forecasts in slightly different formats.

This model is varies slightly from the standard DiD framework. I use low danger days during the

forecasted season as my omitted category, allowing Postseasont to be included and not be colinear

with Azt, controls for the daily forecasted danger level. A zone day receives a published advisory

bulletin or the forecast center is closed, picked up by the post season variable. I want to estimate

what happens to incidents when Colorado stays open longer in 2019. I omit the forecasted danger

levels for the zone days in Colorado after the UAC closes for 2019, and replace with the postseason

dummy to get an average treatment effect for Colorado’s season extension. All models include a

zone by season fixed effect and standard errors are clustered on zone and day.

22


Model 1 in Table 7 shows the results from the model specified in equation 5. The coefficient

of interest is the interaction of the postseason, with Colorado in 2019. The results from model

1 show there are fewer accidents, on average, after the forecasters close for the season. Within

the postseason, there is a higher average accident rate in Colorado. Model 2 is the same exact

specification as Model 1, however the sample is paired down to the months of April, May and the

first two weeks of June. The Postseason 2019 interaction shows higher incident rates on average

in Colorado and Utah late in the season compared to previous seasons. This is to be expected.

The 2018-19 season brought snow totals significantly above the historical average in both of these

states12 and subsequently avalanche cycles that had not been seen in decades. Model 2 suggests

that Colorado’s incident rate was .08 lower when the CAIC continued daily forecasting into May

unlike years prior.

The drawback with data in this setting, like many other areas of economics research, it the

issue of selection in reporting. Forecasters only know about avalanches that they see or have

reported to them. Incidents in this context are generally reported by the victims. Traditionally

this biases the estimates toward zero. I argue that, because the forecast is a function of reported

observations, because the CAIC operates longer, they are more likely to receive incident reports

than less, further biasing estimates downward. An alternate story could be told, however. If

private individuals try to make up for the lower provision of the public good when the forecast-

ers close shop for the season, negative effects of incidents from forecasting longer may be the

result voluntary contributors viewing the forecast as substitute for their efforts. To investigate

this alternate pathway, I run the same specifications in Model 1 and 2, but replace the reported

incidents with the set of all other public observations. In the reduced sample in Model 4, estimates

show no significant displacement of other public observations during the CAIC season extension.

Therefore, if there is some substitution effect that might create misleading results, it would have

to operate only through the incentives to report incidents and not other observations, and vary

differently in 2019 from previous seasons.

Table 8 is identical to the previous models, except it replaces the indicator for treatment with

the continued forecasts. It shows most of the reduction in incidents occurring on considerable

12Alta Resort in Utah had base snow height increase after their USFS lease required them to close for the season.

Several resorts in Colorado and Utah ran lifts on July, 4th, 2019.

23


Table 7: CAIC Season Extension Differences Model

Dependent variable:

Incident Public Reports

(1) (2) (3) (4)

Other Avy 0.019∗∗ 0.043∗∗∗

(0.008) (0.009)

Incident 0.281∗∗∗ 0.321∗∗∗

(0.079) (0.091)

Moderate 0.086∗∗∗ 0.029 0.155∗∗∗ −0.010

(0.006) (0.026) (0.055) (0.061)

Considerable 0.163∗∗∗ 0.039 1.051∗∗∗ 0.967∗∗

(0.013) (0.077) (0.161) (0.481)

High 0.256∗∗∗ 0.120 3.195∗∗∗ 3.674∗∗∗

(0.043) (0.182) (0.555) (1.355)

Extreme −0.110∗∗∗ 5.140∗∗∗

(0.025) (1.456)

Postseason −0.070∗∗∗ −0.148∗∗∗ 0.312∗∗∗ −0.054

(0.005) (0.031) (0.074) (0.044)

Postseason:CO 0.101∗∗∗ 0.151∗∗∗ −0.450∗∗∗ −0.228

(0.015) (0.023) (0.075) (0.161)

Postseason:2019 −0.011∗∗ 0.105∗∗∗ 0.074∗∗ 0.108∗∗

(0.005) (0.027) (0.032) (0.054)

Postseason:CO:2019 0.023 −0.084∗∗∗ −0.406∗∗∗ −0.073

(0.016) (0.024) (0.056) (0.088)

Zone by Season FEs Y Y Y Y

Observations 21,264 7,349 21,264 7,349

Adjusted R2 0.115 0.088 0.185 0.146


Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

24


Table 8

Dependent variable:

Incident Public Reports

(1) (2) (3) (4)

Postseason −0.073 −0.116 0.380∗∗ −0.011

(0.050) (0.087) (0.174) (0.051)

Postseason:CO 0.109 0.132 −0.544∗∗ −0.233∗

(0.078) (0.110) (0.234) (0.133)

Postseason:2019 0.008 0.056∗ −0.166∗ 0.064

(0.011) (0.033) (0.097) (0.103)

Low Danger:Post CO 2019 0.005 −0.019 −0.205 −0.026

(0.032) (0.051) (0.287) (0.176)

Moderate Danger:Post CO 2019 −0.073∗ −0.041 −0.389∗∗∗ −0.166

(0.039) (0.053) (0.132) (0.150)

Considerable Danger:Post CO 2019 −0.232∗∗∗ −0.177∗∗ 0.538 0.768

(0.067) (0.073) (0.420) (0.546)

Forecast Controls Y Y Y Y

Zone by Season FEs Y Y Y Y

Observations 21,264 7,349 21,264 7,349

Adjusted R2 0.115 0.087 0.185 0.147


Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

25


danger days, the highest danger rating that the CAIC used after mid-April.13 This fits the broader

picture here anecdotally. Backcountry users are much more likely to disregard standard peak sea-

son safe travel procedures, such as leaving behind avalanche safety gear and traveling underneath

avalanche prone slopes for extended periods of time when the snowpack consolidates to spring

time conditions. Warmer spring weather brings on a different set of avalanche problems that

still pose danger, and new snow accumulation at higher elevations in April and May is common,

bringing added risk. Broadcasting a danger forecast reminding backcountry users that springtime

and low danger does not mean no danger should decrease incidents on some margin, and the

empirics support this.

V. Observation Markets and Outcomes

Does the information in the forums have an affect on outcomes independent from the forecast?

And if so, is the effect driven by the forecasters or by the voluntary private contributors? I plan use

the website roll out of observation forums looking back in time to test this. Idaho centers did not

use open forums until the 2015/16 seasons. While the website did not publicly post observations,

they had a submission page. The main concern with this approach is that the outcome variable is

largely self reported, and if the introduction of the public forums changes reporting rates then it

violates even a lax parallel trends assumption.

VI. Discussion and Conclusion

The CAIC extending their daily forecasting operations into May in 2019 prevented roughly 2.5

incidents per zone by a very conservative back of the envelope calculation. Historically in April

and May in Colorado, the portion of avalanche incidents that result in serious accidents is right at

15%. This translates to 3 fewer accidents across Colorado in the 8 zones that hold enough snow to

13High and extreme danger days often follow large amounts of snowfall that produce significant natural avalanche

cycles. These storm cycles tend not to occur after March

26


draw backcountry recreators in the spring and summer months.14

This decreased accident rate resulting from forecasting services suggests that access to addi-

tional information about a risky behavior reduces incidents and outweighs any effects from any

sort of pseudo moral hazard or adverse selection. The CAIC achieved their desired outcome of

reducing incidents by offering additional forecasting services and completely offsets and outweighs

any increase in observed accidents from providing novice backcountry recreators a false sense of

security, or licensing more risk taking behavior.

One of the main hindrances to using data from this unique setting is the selection into reporting.

It should be restated that the results should all be interpreted as conditional on reported. This is a

frequent issue in the social sciences; every paper using crime data only knows reported crimes, not

the full universe of crimes. An additional critique may be external validity. Winter backcountry

recreators may be very different from the rest of the population. After all, it takes a unique type of

person to want to spend 3 hours climbing up a mountain to ski down in 10 minutes, but motorcycle

riders have also displayed different preferences as well. I argue that this helps for identification

purposes. While the populations of Colorado and Utah are very different, the extreme selection

into winter time backcountry recreation likely results in two communities that are very similar.

Avalanche forecasting services have significant impacts on outcomes. The estimates here come

from time periods when avalanche conditions are at their safest during the year, and yet this

program still reduces accidents. There is a 23 percent reduction in incident rates that can be

attributed to forecasting services. It is important to note that the ideal incident rate it likely not

zero, accidentally triggering a slide is sometimes unavoidable in the backcountry. Outcomes can

drastically differ, however, depending on the choices made by that individual. Forecasters want

to discourage people from being below the danger, not on top of it. These results suggest that

providing people with additional information about risks allows them to optimize accordingly

and improve outcomes.

14Not to stretch a back of the envelope calculation too far, but using the potion of accidents that are fatal, would mean

1 less fatality approximately every 3 years.

27


References

Andreoni, James (1990). “Impure altruism and donations to public goods: A theory of warm-glow

giving”. In: The Economic Journal 100.401, pp. 464–477.

Atwater, Montgomery Meigs (1968). The Avalanche Hunters. Macrae Smith Co.

Bénabou, Roland and Jean Tirole (2006). “Incentives and prosocial behavior”. In: American Economic

Review 96.5, pp. 1652–1678.

Lazar, Brian et al. (2016). “North American avalanche danger scale: Do backcountry forecasters

apply it consistently”. In: Proceedings ISSW, pp. 457–465.

SLF Switzerland. Origins of the avalanche bulletin – history and background. https://www.slf.ch/en/about-

the-slf/portrait/history/origins-of-the-avalanche-bulletin.html.

Statham, Grant et al. (2010). “The North American public avalanche danger scale”. In: 2010

International Snow Science Workshop, pp. 117–123.

Williams, Knox (1998). “An overview of avalanche forecasting in North America”. In: Proceedings

of the international snow science workshop, Sunriver, OR, ISSW Workshop Committee, pp. 161–169.

A. Appendix

28


Figure A1: An example bulletin from the CAIC

6/13/2019 CAIC Forecast

https://www.avalanche.state.co.us/caic/pub_bc_avo_fx_print.php?bc_avo_fx_id=10678 1/1

Backcountry Avalanche Forecast Sawatch Range

Summary Recent snowfall and steady northwest winds continue to load buried weak layers creating lingering dangerous avalancheconditions. Backcountry travelers can easily trigger large and potentially deadly avalanches on slopes below corniced ridgelines,steep rollovers in open areas, and the drifted sides of gullies. If you trigger an avalanche in the freshly drifted snow it may stepdown into older weak layers entraining much more snow and gathering more destructive force. Avalanches will be largest andmost dangerous on northeast through east to southeast-facing aspects where the slabs are thickest.

Even in wind-sheltered terrain, consider that you can trigger avalanches remotely, or from far away. A group of skiers triggeredtwo large avalanches breaking a couple feet deep near Cottonwood Pass on Wednesday on northerly terrain near treeline. Thesimplest approach right now is to stick to slopes less than about 30 degrees, without steeper terrain overhead.

Weather Forecast for 11,000ft Issued Thursday, Jan 24, 2019 at 6:45 AM by Ben Pritchett

Thursday Thursday Night Friday

Temperature (ºF) 28 to 33 15 to 20 32 to 37

Wind Speed (mph) 15 to 25 7 to 17 12 to 22

Wind Direction WSW WSW WSW

Sky Cover Mostly Cloudy Partly Cloudy Mostly Cloudy

Snow (in) 2 to 4 0 to 1 0 to 1

Avalanche conditions can change rapidly during snow storms, wind storms, or rapid temperature change. For the mostcurrent information, go to www.colorado.gov/avalanche.

© 2008-2018 Colorado Avalanche Information Center. All rights reserved.

29


Figure A2: A sample (high quality) publicly submitted field report to the CAIC website

6/12/2019 CAIC

https://avalanche.state.co.us/caic/obs/obs_report.php?obs_id=56531&display=printerfriendly 1/4

BC Zone Observation Report

Wednesday, May 1, 2019 at 12:00 AM Vail & Summit County

Details

Date: 2019/05/01

Observer: Patrick GephartOrganization: Public

Location

BC Zone: Vail & Summit CountyArea Description: N/NW facing couloir off east ridge of Pacific PeakRoute Description: Spruce Creek trailhead to Mohawk Lakes to base of couloir

Weather

Weather Description: Calm below 11,500. High winds above 12,000 ft. to base of couloir. Seemed likewind was consistently S/SW. Overcast to broken cloud cover with short windows of sun.

Snowpack

Snowpack Description: High SWE in new snow was noticed from the trailhead and throughout the day.New snow ranged from 6 inches to what seemed like 1-2 feet in couloir proper. Whumpf was noted insnow pack below treeline in flat but was concluded to be weak freezing in lower elevations. Wind slabwas the key issue we were looking for during they day.

Avalanches

Avalanche Description: Myself and 1 partner (splitboarder) switched over to crampons in a safe zonefrom both wind and the slop itself below a large rock face. From the start of the boot back it was clearthat the snow was deep but no weak layers were evident, including wind slab. The new snow appearedto be well bonded and had came in entirely right side up. Once we got above 13,000 ft. and toward thetop of the couloir we noticed a small slab, but it seemed very well bonded and showed no signs ofpropagation when breaking through this. Towards the top out this wind slab became thicker, but againmanageable due to its structure, or so we thought (heuristic in hindsight). A small but notable convexityright before the top out was noted which would be key later. We transitioned and set up for the descent.I was dropping first. Upon gaining speed and making the first turn I noted the wind slab was a bit tickerwhere I was riding then where we booted. As I turned by the small convexity the wind slab broke,flowing skier's left and propagating a bit towards the skier's left, eastern aspect of the apron below thesheltered choke. I was able to turn hard skier's right and drive my hands in the consolidated snow in thebed surface, arresting myself and getting out of the slide path. I would say I was caught for 20-30 feetbefore exiting the slide (GoPro video that I will submit later may be helpful here). The avalanche ran tothe bottom of the couloir through the apron. Small debris pile and the wind slab portion that slid seemedto be isolated only to where we noted it, upon inspection of the flanks. If we had booted up the climber's

6/12/2019 CAIC


right, most eastern facing aspect I think we would have noted this as more of a weak layer much quickerand turned around. I would classify the avalanche as a R1-2 D1.5. If I had been taken by it I would havegotten take over a few rocks, but would not have been buried. Heuristics certainly played a big parthere. Both myself and partner consider ourselves conservative backcountry snowboarders withadvanced snow science knowledge, safe practices, constant discussion, continuing observations, etc.We have safely navigated isolated pockets of wind slab before in isothermic, spring snow packs, andmanaged them safely. We deemed what we found on the climb up "safe" due to its structure and notedresistance to propagate when punching through, although it was still a wind slab. The classic "I've beenin these conditions before and we managed them safely" was at play here. It only took one turn to findthe shallow weak spot that propagated to a thicker slab to skier's left to create a wind slab avalanche.Luckily I was able to exit it quickly. Spatial variability of the slab's thickness was also a factor, as wewere climbing in the thinner part on climber's left, but the thicker and more dangerous part was to ourclimber's right and out of observation on the way up. This was a good wake up and will keep our headson more of a swivel in the future. Upon reading the avalanche forecast again for 4/30/19 we deemed itspot on and exactly what we encountered. Great reporting and a user error on our end.

Date Location/Path # Elev Asp Type Trig SizeR SizeD

2019/05/01 † 10-mile Range 1 >TL E SS AR R1 D1.5 Date: 2019/05/01 (Estimated)Observer: Patrick GephartOrganization: PublicArea Description: Pacific Peak Landmark: 10-mile Range

Media

Images

6/12/2019 CAIC


Figure 1: Crown looking towards skiers left. I triggered skier's right

Figure 2: Avalanche propagation, flank and bed surface in first apron before second choke

6/12/2019 CAIC


Figure 3: Avalanche propagation and flank

Figure 4: Looking up at avalanche propagation and flank from towards the bottom of the couloir. Note thefew rocks that would cause injury if taken over. Couloir goes looker's left past view.

30

knowledge is powder: risk taking behavior, forecasting services … · 2019. 12. 31. · knowledge...

Documents