Lessons from Tohoku: why earthquake hazard maps often fail and what to do about it Tohoku, Japan March 11, 2011 M 9.1 NY Times CNN Seth Stein, Northwestern University Robert Geller, University of Tokyo Mian Liu, University of Missouri

Tools in preparing for natural disasters include - Long term forecasts: yr (earthquakes), 100 yr (climate change), 1-10 yr (hurricane, volcano) - Short term predictions: days (hurricanes), days to months (volcano), hours (tornado) - Real time warnings: (hours to minutes) tsunami, earthquake shaking, hurricane, tornado, flood Sometimes these work, sometimes they fail

FAILURES False negative - unpredicted hazard Loss of life & property False positive - overpredicted hazard Wasted resources, public loses confidence Authorities typically ignore, deny, excuse, or minimize failure More useful to analyze failures to improve future performance

2008: Hurricane Ike predicted to hit Miami

Ike’s actual track

Ike predicted to bring certain death

Actual deaths: < 50 of 40,000 Error 800x

If it had been a weekday, Major cost

K. Emanuel CNN 8/26/11 NYT 8/28/11

Economic loss ? What if weekday?

Science News 6/15/91 The local economy collapsed, said Glenn Thompson, Mammoth Lakes' town manager. Housing prices fell 40 percent overnight. In the next few years, dozens of businesses closed, new shopping centers stood empty and townspeople left to seek jobs elsewhere. (NYT 9/11/90)

NY Times 3/31/2011 Expensive seawalls - longer than Great Wall of China - proved ineffective 180/300 km swept away or destroyed In some cases discouraged evacuation

Japan spent lots of effort on national hazard map, but 2011 M 9.1 Tohoku, 1995 Kobe M 7.3 & others in areas mapped as low hazard In contrast: map assumed high hazard in Tokai “gap” Geller 2011

Hazard map crucial for mitigation strategy Optimal level of mitigation minimizes total cost = sum of mitigation cost + expected loss for given mitigation level Expected loss = ∑ (loss in i th expected event x assumed probability of that event) Less mitigation decreases construction costs but increases expected loss and thus total cost More mitigation gives less expected loss but higher total cost Because assumed probability taken from hazard map, inaccurate map biases mitigation - too little or too much Stein & Stein, 2012

Including uncertainty Consider marginal costs C’(n) & benefits Q’(n) (derivatives) Stein & Stein, 2012 More mitigation costs more But reduces loss Optimum is where marginal curves are equal, n* Uncertainty in hazard model causes uncertainty in expected loss. We are risk averse, so add risk term R(n) proportional to uncertainty in loss, yielding higher mitigation level n** Crucial to understand hazard model uncertainty cost Benefit (loss reduction )

NY Times 11/2/2011 Choosing policy involves politics, economics, geoscience Too expensive to rebuild for 2011 sized tsunami >100 $B for new defences only slightly higher than old ones “In 30 years there might be nothing left there but fancy breakwaters and empty houses.”

Hazard maps fail because of - bad physics (incorrect description of earthquake processes) -bad assumptions (mapmakers’ choice of poorly known parameters) - bad data (lacking, incomplete, or underappreciated) - bad luck (low probability events) and combinations of these (Tohoku!)

Expected Earthquake Sources 50 to 150 km segments M7.5 to 8.2 (Headquarters for Earthquake Research Promotion) Off Sanriku-oki North ~M8 0.2 to 10% Off Sanriku-oki Central ~M to 90% Off Fukushima ~M7.4 7% Off Ibaraki ~M6.7 – M7.2 90% Detailed model of segments with 30 year probabilities Sanriku to Boso M8.2 (plate boundary) 20% Sanriku to Boso M8.2 (Intraplate) 4-7% Off Miyagi ~M7.5 > 90% J. Mori Assumption: No M > 8.2

Giant earthquake broke five segments 2011 Tohoku Earthquake 450 km long fault, M 9.1 (Aftershock map from USGS) J. Mori Expected Earthquake Sources 50 to 150 km segments M7.5 to 8.2 (Headquarters for Earthquake Research Promotion)

Tsunami runup approximately twice fault slip (Plafker, Okal & Synolakis 2004) M9 generates much larger tsunami Planning assumed maximum magnitude 8 Seawalls 5-10 m high CNN NYTStein & Okal, 2011

Didn’t consider historical record of large tsunamis NYT 4/20/11

Lack of M9s in record seemed consistent with model that M9s only occur where lithosphere younger than 80 Myr subducts faster than 50 mm/yr (Ruff and Kanamori, 1980) Disproved by Sumatra 2004 M9.3 and dataset reanalysis (Stein & Okal, 2007) Short record at most SZs didn’t include rarer, larger multisegment ruptures Stein & Okal, 2011

NY Times 3/21/11 Why?

Hazard maps are hard to get right: success depends on accuracy of four assumptions over years Where will large earthquakes occur? When will they occur? How large will they be? How strong will their shaking be? Uncertainty & map failure often result because these are often hard to assess

2001 hazard map M7 earthquake shaking much greater than maximum predicted for next 500 years Including GPS would have predicted higher hazard

2008 Wenchuan earthquake (Mw 7.9) was not expected: map showed low hazard based on lack of recent earthquakes Didn’t use GPS data showing 1-2 mm/yr (~ Wasatch) Earthquakes prior to the 2008 Wenchuan event Aftershocks of the Wenchuan event delineating the rupture zone

Maps are like ‘Whack-a-mole’ - you wait for the mole to come up where it went down, but it’s likely to pop up somewhere else.

What to do Continue research on fundamental scientific questions (geoscience community’s job!) Realistically assess map uncertainties and present them to help users decide how much credence to place in maps & how to use them given uncertainties Develop methods to objectively test hazard maps - which hasn’t been done despite their wide use - and thus guide future improvements

Comparing map predictions shows the large uncertainties (~3X) resulting from different assumptions Stein et al, 2012 Newman et al, 2001

Testing analogy: evidence-based medicine objectively evaluates widely used treatments, often with embarrassing results Although more than 650,000 arthroscopic knee surgeries at a cost of roughly $5,000 each were being performed each year, a controlled experiment showed that "the outcomes were no better than a placebo procedure."

Bad luck or bad map? Compare maximum acceleration observed to that predicted by both map and null hypotheses. A simple null hypothesis is regionally uniformly distributed hazard. Japanese map seems to be doing worse than this null hypothesis. Geller 2011 Test maps by comparison to what happened after they were published.

Avoid biases from new maps made after a large earthquake that earlier maps missed. Frankel et al, 2010 Before 2010 Haiti M7After 2010 Haiti M7 4X

A posteriori changes to a model are "Texas sharpshooting:” shoot at the barn and then draw circles around the bullet holes.

Possible problem: Overparameterized model (overfit data): Given a trend with scatter, fitting a higher order polynomial can give a better fit to the past data but a worse fit to future data Analogously, a seismic hazard map fit to details of past earthquakes could be a worse predictor of future ones than a less detailed map How much detail is useful? Linear fit Quadratic fit

Challenge: Users want predictions even if they’re poor Future Nobel Prize winner Kenneth Arrow served as a military weather forecaster. As he described, “my colleagues had the responsibility of preparing long-range weather forecasts, i.e., for the following month. The statisticians among us subjected these forecasts to verification and found they differed in no way from chance. The forecasters themselves were convinced and requested that the forecasts be discontinued. The reply read approximately: "The commanding general is well aware that the forecasts are no good. However, he needs them for planning purposes." Gardner, D., Future Babble: Why Expert Predictions Fail - and Why We Believe Them Anyway, 2010

What to do given that earthquake hazard maps often fail Tohoku, Japan March 11, 2011 M 9.1 NY Times CNN Seth Stein, Northwestern University Robert Geller, University of Tokyo Mian Liu, University of Missouri