1-s2.0-s0001457513003242-main
TRANSCRIPT
-
8/10/2019 1-s2.0-S0001457513003242-main
1/7
Accident Analysis and Prevention 60 (2013) 5056
Contents lists available at ScienceDirect
Accident Analysis and Prevention
journal homepage: www.elsevier .com/ locate /aap
The killing zone revisited: Serial nonlinearities predict general
aviation accident rates from pilot total flight hours
William R. Knecht
Civil AerospaceMedical Institute, Federal Aviation Administration,MMAC/AAM-510, 6500S.MacArthurBoulevard,OklahomaCity,OK 73079, UnitedStates
a r t i c l e i n f o
Article history:
Received 17 January 2013
Received in revised form 24 July 2013Accepted14 August 2013
Keywords:
General aviation
Accidents
Risk
Flight hours
Modeling
Nonlinear
a b s t r a c t
Background: Is there a killing zone (Craig, 2001)arange ofpilot flight time overwhichgeneral aviation
(GA)pilots areatgreatest risk?Morebroadly,canwepredictaccidentrates, givenapilots totalflighthours
(TFH)? These questions interest pilots, aviation policy makers, insurance underwriters, and researchers
alike.Most GA research studies implicitly assume that accident rates are linearly related to TFH, but that
relation may actually be multiply nonlinear. This work explores the ability ofserial nonlinearmodeling
functions to predict GA accident rates from noisy rate data binned by TFH.
Method: Two sets ofNational Transportation Safety Board (NTSB)/FederalAviation Administration (FAA)
data were log-transformed, then curve-fit to a gamma-pdf-based function. Despite high rate-noise, this
produced weighted goodness-of-fit (R2w) estimates of .654 and .775 for non-instrument-rated (non-IR)and instrument-rated pilots (IR) respectively.
Conclusion: Serial-nonlinearmodels could be useful to directly predict GA accident rates from TFH, and
as an independent variable or covariate to control for flight risk during data analysis. Applied to FAA
data, these models imply that the killing zone may be broader than imagined. Relatively high risk for
an individual pilot may extendwell beyond the 2000-hmark before leveling offto a baseline rate.
Published by Elsevier Ltd.
1. Introduction
Totalflighthours isoftenused torepresentpilot flightexperience
and/or risk exposure (e.g., Dionne et al., 1992; Guohua et al., 2003;
Mills, 2005;Sherman,1997). TFHhasoftenservedas eitheran inde-
pendent variable (IV) in itsownright or as a statistical covariate, to
control for the effects of experience or risk.
However, many investigators have found no reliable relations
between TFH and their dependent variables-of-interest (DV). This
may be because most typically assume an underlying linear rela-
tion( y = a+ bx) between IVandDVfor instancebetweenTFHand
accident frequency and/or rate.
Strong evidence is beginning to emerge that some TFH rela-
tionsmaybe nonlinear. For instance, Bazargan andGuzhva (2007)
reported that the nonlinear logarithmic transform log(TFH) signif-
icantly predictedGA fatalities in a logistic regression model.
In his popular 2001 book, TheKilling Zone, Paul Craig presented
extensive evidence that GA pilot fatalities and accident frequen-
cies relate nonlinearly to TFH. Fatalities, for instance, occur most
frequently at middle ranges of TFH (50350), perhaps because
pilots are overconfident yet lack actual experience and skill with
Tel.: +1 405 954 6848.
E-mail address:[email protected]
rare, challenging events. Craig supported this hypothesis with his-
tograms of GA accident frequencies, although not with a formal
model. In group after group, a similar pattern emerged, one of a
skewed camel hump with a long tail at higher TFH.
Following Craig, I investigated a nonlinear model for accident
frequencycounts (Knecht, 2012). Derived fromNTSB andFAAdata,
Fig. 1 typifies thelookandfeelof TFHversusfatal accident counts
for a group of 831 U.S. instrument-rated (IR) GA pilots from 1983
to 2011.
This skewed hump is quite durable, persisting whether we
break out different data categories (e.g., IR vs. non-IR or serious
vs. fatal accidents) or even combine them. For example, Fig. 2
shows GA accident data for 832 pilots from2003 to 2007 for NTSB
SERIOUS+ FATAL accident categories that were combined to boost
frequency counts.
These data were successfully modeled by a gamma probabil-
ity density function (pdf) operating on log-transformed TFH. We
conclude that frequency count histograms support the notion of a
nonlinear killing zone somewhere in the lower range of TFH.
This is useful information. But, it now raises a new question:
How can webe certain that this relation is not an artifact,merely
a side-effect of the fact that there are just fewer pilots at higher
TFH, hence fewer accidents? Fig. 3 shows how frequency counts
for amatched sampleofnon-accidentpilotsalso drop offmarkedly
as TFH increase.
0001-4575/$ seefrontmatter.Published by Elsevier Ltd.
http://dx.doi.org/10.1016/j.aap.2013.08.012
http://localhost/var/www/apps/conversion/tmp/scratch_1/dx.doi.org/10.1016/j.aap.2013.08.012http://localhost/var/www/apps/conversion/tmp/scratch_1/dx.doi.org/10.1016/j.aap.2013.08.012http://www.sciencedirect.com/science/journal/00014575http://www.elsevier.com/locate/aapmailto:[email protected]://localhost/var/www/apps/conversion/tmp/scratch_1/dx.doi.org/10.1016/j.aap.2013.08.012http://localhost/var/www/apps/conversion/tmp/scratch_1/dx.doi.org/10.1016/j.aap.2013.08.012mailto:[email protected]://crossmark.crossref.org/dialog/?doi=10.1016/j.aap.2013.08.012&domain=pdfhttp://www.elsevier.com/locate/aaphttp://www.sciencedirect.com/science/journal/00014575http://localhost/var/www/apps/conversion/tmp/scratch_1/dx.doi.org/10.1016/j.aap.2013.08.012 -
8/10/2019 1-s2.0-S0001457513003242-main
2/7
W.R. Knecht/ Accident Analysis andPrevention 60 (2013) 5056 51
Fig. 1. Frequencyhistogram ofU.S. GA IR fatal accident counts, 19832011 (y-axis)
as a function of total flight hours (x-axis, binwidth= 100TFH).
Fig.2. Frequency histogramof GA IRSERIOUS+ FATAL accident counts as a function
ofTFH (bin width= 100TFH).
Figs 2 and 3 look suspiciously similar. In fact, the Pearson cor-
relation between them is .845 (p< .001). This implies that most
(r2 .71) of the killing zone seen in accident frequency counts
can beexplained simply by the fact that the numbers of pilots tend
to decrease as TFH increases.
With that inmind,whatsupportnow remains forthehypothesis
of an actual killing zone? And, if it exists, what form does it likely
Fig. 3. A histogramof GA IRnon-accident pilot countsfromthe same data cohortas
Fig. 2, showinghow numbers of pilotsdecrease rapidly as TFHincrease.
Fig. 4. Anoisyhistogramshowingcombined serious+ fatal accident ratesfor all
GA pilots as a function of TFH (binwidth= 100).
Data fromKnecht andSmith (2013).
take? We can address these questions by analyzing accident rate
histograms, in which the ith data bins rate is
ratei = accidentsi
accidentsi + nonaccidentsi(1)
Rateswill control forthe fact that ourfrequency countsdecrease
as TFH increases. If a killing zone continues to persist with rates,
thenwe have gained new insight intoGA flight risk.
1.1. The problem with rates
Unfortunately, rate data can be quite noisy, as Fig. 4 depicts.
Rate-noise results fromsamplingerrorrandomfluctuations in
rate from one bin to the next. It is a vexing problem at higher TFH
where Eq. (1)s denominator is small compared to its numerator,
and even small changes in frequency count can cause substantial
rate variations.It is nothard toappreciate theusefulness of amodeling function
here. A good model would smooth the noise in the data, allow-
ing investigators to better predict rates, given TFH. This would
be directly useful in areas such as determination of insurance
premiums,allocation of resources forpilottraining,accident inves-
tigation staffing, and public relations. To safety researchers, it
would be useful as an IV or statistical covariate, to either factor
in or out risk as a function of flight experience. This would have
broad application in numerous types of aviation research.
With this inmind, the present workmoves fromfittingGAacci-
dent frequency countdata to fitting accident rates as a function of
GA pilot TFH. The calculation of rates is a key step in risk analy-
sis because rates will control for the actual number of individuals
within a given data bin.
2. Method
2.1. The modeling function
Ideally, modeling functions should be motivated by theory
involving causal processes inherent to the data. Unfortunately, in
aviation accidents, causes tend to be numerous and hard to disen-
tangle, making theory-basedmodeling difficult.
We therefore proceed with a less ambitious goal of simply try-
ing to fit a well-behaved, continuous mathematical function to
empirical GA accident rate data. Standard techniques are used,
namely minimization of least-squares residuals between a model
and empirical data.
-
8/10/2019 1-s2.0-S0001457513003242-main
3/7
52 W.R. Knecht/ Accident Analysis andPrevention60 (2013) 5056
Fig. 5. pdfwith various values of and .
The model itself begins with the versatile probability density
function pdf (Spanier and Oldham, 1987). The x-axis representsTFH. They-axiscanrepresenteither frequencycountor proportion.It contains a shape parameter, >0,and a scale parameter, >0.
pdf.c(x;,) =ex/ x1
() (2)
Thegammafunction() itselfis describedas theEuler integral,defined for >0
() =
0
t1etdt (3)
Shown in Fig. 5, pdf has the attractive feature of being able
to represent overdispersion (variance> mean) or underdispersion
(variance1 =100i50.Note that, in calculating a series of binwise accident rates, we
are not technically constructing a pdf. Rather, we are constructing
a rate histogram and using Eq. (4) (which is based on a pdf) to fit
that histogram.
Note also that we must openly address the issue of whether
these data, being a sample of convenience, may contain hidden
biases that could affect parameterization of Eq. (4). That is not
impossible. By definition, hiddenbiases areunknown andtypically
revealed only after more comprehensive data become available.
Until then, however, we argue there is substantial value in dis-
cussing serial nonlinearities, and layingoutamethodology for their
future use.
-
8/10/2019 1-s2.0-S0001457513003242-main
4/7
W.R. Knecht/ Accident Analysis andPrevention 60 (2013) 5056 53
2.3. Parameter estimation
The NonlinearModelFit function ofMathematica 7.0 (Wolfram,
2008) is used for parameter estimation. Specifically, since param-
eters b, , and of Eq. (4) are constrained to remain >0, theKarushKuhnTucker (KKT) method is used.
2.4. The problem of noise
NoisydatasuchasFig.4s posegreatdifficulty forstandardleast-
squares minimization. Parameters often fail to converge in rugged
errorscapes, oscillating around a saddlepoint, andmay converge to
a local minimum rather than a global one.
To confront the problem of noise, we first tried the simplest
approach of usingwider data bins. Surprisingly, this hadalmostno
effect. Accidents within these data tended to bunch up in several
adjacent bins, leaving noway to smooththemwithout resorting to
arbitrary bin widths of convenience.
We next tried dropping outliers greater than 2.53 s (standard
deviations) from thegroupsamplemeany. Thismethodwas finallyabandoned because it relied on an invalid assumption of distri-
butional normality, destroyeduseful information, andbecause the
extremely long tail of the fitting function resulted in an s so small
that very few of the very noisy values in the long tail were being
dropped anyway.
A thirdapproachwasnexttried, involvingdata-smoothingusing
moving averages.Unfortunately, this also hadtheeffectof destroy-
ing information by smearing.
Simulatedannealingwasthe fourthapproach(Kirkpatrick et al.,
1983)adding noise to each parameter start value before engag-
ing in error-gradient descent, then cooling the noise, allowing
parameter estimates to bounce around and out of local minima
until a global near-minimum could be found. Nonetheless, even
this approach would unfortunately still leave one major logical
problem:Allourdata binsincluding thenoisywould have equal
opportunity to influence the residual sum-of-squares during the
data-fit. In fact, any one bin, one vote approach would implic-
itly allowall bins equal influence,nomatter howmany individualseach binrepresentedor what that binsexpectedvariancewas. This
would confer inordinately large influence to sparsely populated
bins and those with inherently low reliability.
After weighing the above considerations, we finally settle on
the procedure ofweighting empty bins by 0 while weighting non-
empty bins accident rate ri by the inverse of its sample variance
(1/s2i). This is easily done in Mathematica by supplying a vector
Weightsw to NonlinearModelFit, wherew= {w1,w2, . . .,wi}.Since our data are rates are proportions, 0 ri1, the ith bin
weight wibecomes
wi =1
s2i
=ni 1
ri(1 ri)(1 ni/Ni) (5)
ni being the ith bins total frequency count (accidents+ non-accidents) used to derive ri, and Ni being the number of pilots in
the general populationwith that rangeof TFH.
Inthisparticularinstance,wedo notknowNi,soweshallassume
a constant sampling ratio ni/Ni, making the term 1 (ni/Ni) con-
stant for all bins. Since weights only need to be expressed relative
to each other during curve-fitting, we can eliminate this constant
term, leading to a functional weight of
wi =ni 1
ri(1 ri) (6)
Unfortunately, Eq. (6) is ill-behaved when ri =0 or 1, leading to
divisionbyzero. This canarisewhen eitherTFHbins aretoo narrow
orwhenTFH is large,making the bin sample sizenitoo small, since
few pilots have high TFH. Fig. 6 (top) illustrates the problem.
Fig. 6. (red, solid lines) With Eq. (6), as r0 or 1, 1/(r(1 r)); (top, dashed,
blue line) With =001, Eq. (7) gives a good approximation across most ofr, from
around .006< r< .994, but gracefully self-limits at r=0 and 1. (For interpretation of
the references tocolorin this figurelegend,the readeris referredto thewebversion
ofthe article.)
To address this issue, Eq. (6) is modified slightly by adding a
term designed to constrain behavior at 0 and 1:
wi =ni 1
(ri + )(1 ri + ) (7)
As Fig. 6 shows,when is small, Eq. (7) closelyapproximatesEq.(6) over most of its range but self-limits gracefully at ri =0 and 1.
2.5. Parameter start values and evolution
Inmultidimensionalparameterspaceswithunknownlocalmin-
ima, the choice of start values can be critical. A hybrid approach is
used here. Start values are first selected that roughly approximate
the shape of the binned data, namelyA=1, =6, =0.5, =b=0.Toallowparametersto evolveinamannerthatemulatesanneal-
ing, NonlinearModelFitis setup touse weighted binneddata and to
runmultiple times insidea For[i=1, >1.0, i= i+1] loop. Inside theFor[] loop, to emulate noise injected at each iteration, the value
of each parameterpjismultiplied by a randomreal numberdrawn
fromtherange1.0withstartingat0.02.Duringeach (ith)iter-
ation of the For[] loop,NonlinearModelFitis itself allowed to run for
i iterations, afterwhich i is increased by1while is decreasedby asmall fixed amount (.0002).thus emulates thecooling rate. Even-tually,fallsbelow1.0, halting theinjectionofparameternoiseand
terminating the For[] loop after one final gradient descent.
Running this method repeatedly on a given dataset, we can
broadly sample the parameter space. After inspection of the
graphic data-fit plots verifies a good fit, and multiple runs shows
parameters similar to at least three significant figures, we can be
confident that those parameter values represent a near-optimal
data-fit.
-
8/10/2019 1-s2.0-S0001457513003242-main
5/7
54 W.R. Knecht/ Accident Analysis andPrevention60 (2013) 5056
Fig. 7. Annualized non-IR GA accident rates (y=predicted rate for x50 TFH,
medianTFH=250.5).
2.6. Model evaluation
Goodness-of-fit can bemeasured by the coefficient of determi-
nation, R2. The weighted form is usually taken as
R2
w = 1
SSw,error
SSw,total = 1ni=1wi(yi fi)2
ni=1
wi(yi yw)2 (8)
with weightedmeans, for instanceYw , oftheform
yw =
niwiyiniwi
(9)
Inour case, since Eq. (8) sometimesthrows values
-
8/10/2019 1-s2.0-S0001457513003242-main
6/7
W.R. Knecht/ Accident Analysis andPrevention 60 (2013) 5056 55
4.2. A point-estimate of relative GA flight risk
To estimate an annualized accident rate for a given value of
TFH=x50, simply populate Eq. (4)s parameters and insert x.2
To illustrate, for non-IR pilots at the median value of TFH =xNIR =250.5, Eq. (4) becomes
.00252081+ .0130177
e(ln(250.5).364805)/.0923902(ln(250.5) .364805)64.37621.092390264.3762
(64.3762)
Theresult is approximately.0068(1 in150)expectedaccidents
per 100 TFH,which we can see is correct fromFig. 7.
This point-estimatecanbeusedaseitheranIV orcovariatewhen
our data contain only a single value of TFH for each pilot. Keep in
mind that, in using such a relative estimate, we implicitly assume
that (a) all pilots fly exactly 100 TFH per year, and (b) their acci-
dent rate itself does not change during that time. Naturally, both
these assumptions are false for most pilots and can only be safely
ignored where large sample sizes dampen random fluctuations.
Thismay seem like a grave restriction.But, theexact same assump-
tionsusuallyunderliethelinearmodel,and aremerely rarely stated
so explicitly.
4.3. A definite-product estimate of annual GA flight risk
In thefortunatecircumstancewhere ourdata contain twoaccu-
rately time-stamped values of TFH for each pilot,3 we can better
estimate each pilots absolute predicted accident rate based on
the general form:
risk = 1
TFH2x=TFH1
(1hourly) (11)
where risk estimates the chance of having an accident over thetimespan TFH
2TFH
1. We calculate the definite product (),
which simply subtracts from 1.0 the multiplied individual esti-
mated hourly probabilities ofnothaving an accident (1hourly),with hourlydefined as
hourly =1
100
b+A
e(ln(x))/(ln(x) )1
()
(12)
Here, the constant 100normalizes for each of our accident-rate
bins having been 100TFHwide when parameterized.
riskavoidsamajor deficiencyof thepoint-estimatebypredict-ing cumulative accident probability over any desired number of
TFH rather than assuming risk over an arbitrary span of 100 TFH.
To illustrate, suppose that an IR pilot reports having flown 310
TFH on6 September,2012, and 433 TFH on29August, 2013(a span
of 357 days, or .978 years). Eqs. (11) and (12) become
risk = 1
433x=310
1
1
100
.00112073+ .0173821
e(ln(x)2.24346)/.0890114(ln(x) 2.24346)51.34721.089011451.3472
(51.3472)
= 1
433x=310
.9999887927
.0001132895
(ln(x) 2.24346)50.3472
x11.23451603
2 (64.3762)9.460741087. (51.3472)1.188441065. Those using SPSS
may find only the natural log of, so re-transformas eln(()) .3 Alldatashould befilteredto ensurethatt2> t1and TFH2> TFH1 . Experiencewith
FAA data provesthat this is usually, butnot always, thecase.
Here, risk represents a cumulative .0067 chance of having atleast one accident over the TFH range of 310433, with data accu-
mulated over .978 years. This can be used as-is for some purposes.
For use as an IV or covariate, it can be annualized by multiplying
.0067 (1/.978)= .0069.
5. Discussion
Can total flight hours predict general aviation accident rates? If
so, what does that relation look like? Is there a killing zonea
range of TFH over which GA pilots are at greatest risk? These
questions are of great interest to pilots, aviation policy makers,
researchers, and insurance underwriters alike.
Using accident frequency counts, Craig (2001) proposed that
such a zone spans a range of approximately 50 to 350 total flight
hours. The current work revisits the question, using accident rates
to incorporate the essential additional information of numbers of
pilots at each level of TFH. Rates, of course, introduce their own
difficulties, andourmethodology attempts to address these.
If such a killing zone truly exists, what does it look like? Many
aviation research studies implicitly assume that accident rates are
linearly related to TFH. In fact, that relation appearsmarkedly non-
linear, and may even be a function of serial nonlinearitiesforexample, a log transform of TFH followed by a gamma probabil-
ity density function-like transform incorporating a base rate (see
Eq. (4)).
The present work uses such a serial nonlinear function rateto predict GA accident rates from TFH, given extremely noisy
real-world data. Two sets of 20032007 NTSB/FAA data pro-
duce weighted goodness-of-fit (r2w) estimates of .654 and .775 for
non-instrument-rated and instrument-rated pilots, respectively,
considered moderately good by convention.
Shownpreviously, Figs. 7 and 8 illustrate therawrate data with
the modeling function rate =y=f(x) superimposed. Each model
y-value represents the estimated probability of having a serious-
to-fatal accident for a pilot flying 100 TFH (x50) over the course
of oneyear. Theyellowregionsurrounding themodel curve repre-sents the .95 confidence interval.
Consistent with our intuition and previous frequency count
studies, these models suggest that a killing zone may indeed
exist. Accident rates may increase for GA pilots early in their
post-certification careers, reach a peak, decline with greater flight
experience, then asymptote to some baseline level.
However, thekilling zonemaybe farbroader than earlier imag-
ined. Relatively high risk for an individual pilot may extend well
beyond the 2000-hmark before leveling off to a baseline rate.
For now, we should consider these conclusions tentative. First,
the 67,687 pilots in these datasets constitute only slightly more
than 10% of those currently licensed in the U.S. While this might
represent an excellent dataset size in other venues, the relatively
tiny number of pilots at high TFH leads to considerable noise in
high-TFH rates. Data-weighting can overcome part of the problem
but leads to wide confidence intervals at high TFH, reflecting the
lower statistical reliability of high-TFH data.
The net effect is that, at this point in time, we have high con-
fidence in our data only at relatively low values of TFH. Better
results await better cross-communication between the FAA and
-
8/10/2019 1-s2.0-S0001457513003242-main
7/7
56 W.R. Knecht/ Accident Analysis andPrevention60 (2013) 5056
NTSB databases, which will result inmuch larger sample sizes and
much greater reliability across the full spectrum of TFH.
Weendwithanappeal tothe FAA andNTSB. The dataneeded to
compute accident rates such as these come frommultiple sources
that are hard to access in the U.S. These include the FAA CAIS and
DIWS databases, as well as the NTSB Aviation Accident Database.
These databases do not intercommunicate well, and could be
greatly improved by better record-keeping, andmost particularly
by sharing a common pilot identifier code. The NTSB has previ-
ously made similar recommendations (NTSB, 2005) that remain
outstanding.
This studyhas exploredamethodologythatwill hopefullyprove
useful, both in its own right and as a stimulus to further research.
Many researchers have olddata that canbe easily reanalyzed. And,
the sooner we challenge our nave assumptions that all indepen-
dentvariablesrelate linearly to theirdependentvariables, themore
accurate anduseful aviation risk research will become.
Acknowledgements
Deep appreciation is extended to Joe Mooney, FAA (AVP-210),
for his help reviewing the NTSB database queries that provided
the accident data used here, to David Nelms, FAA (AAM-300), for
providing TFH from the DIWS database, and to Dana Broach, FAA
(AAM-500) forvaluable commentary on this seriesofmanuscripts.
References
Bazargan,M., Guzhva,V.S.,2007.Factorscontributingto fatalitiesin generalaviation.World Rev. Intermodal Transp. Res. 1 (2), 170181.
Chiew, F.H.S., Srikanthan, R., Frost, A.J., Payne, E.G.I., 2005. Reliability of daily andannual stochastic rainfall data generated from different data lengths and data
characteristics. In: MODSIM 2005 International Congress on Modeling andSimulation, Modeling and Simulation Society of Australia and New Zealand,Melbourne, December2005, pp. 12231229.
Craig, P.A., 2001.The Killing Zone. McGraw-Hill, NewYork.Dionne, G.,Gagne, R., Vanasse, C.,1992.A Statistical Analysis of Airline Accidentsin
Canada, 19761987 (Report No. CRT-811). Center for Research on Transporta-tion,Montreal.
Guohua, L., Baker, S.P.,Grabowski, J.G., Qiang, Y.,McCarthy, M.L.,Rebok, G.W., 2003.Age, flight experience, andrisk of crash involvement in a cohortof professionalpilots. Am.J. Epidemiol. 157 (10), 874880.
Hogg, R.V., Klugman, S.A., 1984. LossDistributions. Wiley, NewYork.
Kirkpatrick, S.,Gelatt, C.D., Vecchi,M.P., 1983.Optimizationby simulatedannealing.Science 220,671680.
Knecht, W.R., 2012. Predicting General Aviation Accident Frequency from PilotTotal Flight Hours (Technical Report No. DOT/FAA/AM-12/15). Federal AviationAdministration,Office of AerospaceMedicine,Washington, DC.
Knecht, W.R., Smith, J.,2013.Effects of Training School Type andExaminer Type onGeneralAviationFlightSafety (TechnicalReportNo.DOT/FAA/AM-13/4). FederalAviationAdministration, Office of AerospaceMedicine,Washington, DC.
Mackey, J.B., 1998. Forecasting wet microbursts associated with summertime air-mass thunderstormsover the southeastern United States. Air Force InstituteofTechnology (UnpublishedMS Thesis).
Mills, W.D., 2005. The association of aviators health conditions, age, gender, andflight hours with aircraft accidents and incidents. University of Oklahoma,Health Sciences Center Graduate College, Norman, OK (Unpublished Ph.D. dis-sertation).
National Transportation Safety Board, 2005. Current procedures for col-lecting and reporting U.S. general aviation accident and activity data(Safety Report No. NTSB/SR-05/02). NTSB, Washington, DC. Available from:www.ntsb.gov/doclib/reports/2005/sr0502.pdf
NationalTransportation SafetyBoard, 2011.User-downloadabledatabase. Availablefrom: www.ntsb.gov/avdata
Sherman, P.J., 1997. Aircrews evaluations of flight deck automation training anduse: measuring and ameliorating threats to safety. The University of Texas atAustin (Unpublished doctoral dissertation).
Spanier, J.,Oldham, K.B., 1987.AnAtlas of Functions. Hemisphere,New York.Winkelmann, R., 2008.EconometricAnalysisof Count Data,5th ed.Springer-Verlag,
Berlin.WolframMathematica DocumentationCenter,2008. Somenotes on internal imple-
mentation. Available from: http://reference.wolfram.com/mathematica/note/SomeNotesOnInternalImplementation.html#20880
http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0015http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0030http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://www.ntsb.gov/doclib/reports/2005/sr0502.pdfhttp://www.ntsb.gov/avdatahttp://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0075http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://reference.wolfram.com/mathematica/note/SomeNotesOnInternalImplementation.html#20880http://reference.wolfram.com/mathematica/note/SomeNotesOnInternalImplementation.html#20880http://reference.wolfram.com/mathematica/note/SomeNotesOnInternalImplementation.html#20880http://reference.wolfram.com/mathematica/note/SomeNotesOnInternalImplementation.html#20880http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0080http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0075http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0075http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0075http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0075http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0075http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0075http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0075http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0070http://www.ntsb.gov/avdatahttp://www.ntsb.gov/doclib/reports/2005/sr0502.pdfhttp://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0055http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0050http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0045http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0040http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0035http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0030http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0030http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0030http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0030http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0030http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0025http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0020http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0015http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0015http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0015http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0015http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0015http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0015http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0010http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005http://refhub.elsevier.com/S0001-4575(13)00324-2/sbref0005