AN ASSESSMENT OF FLIGHT VARIABLES AFFECTING CIVIL AVIATION

ACCIDENTS AND INCIDENTS

Mohammad Kashef
The George Washington University

ABSTRACT

Flight safety has been an important topic for both academia and the industry. Aviation
experts and authorities, as well as commercial airline administrators, constantly seek to
improve flight safety. Researchers, on the other hand, have tried to model avionic fatalities
and suggest improvements or upgrades in flight systems to reduce risk. One approach has
been to use data from past accidents and incidents to capture and model the relationship
between the different factors involved in each event. However, some important factors are
not included in the databases maintained by entities such as the National Transportation
Safety Board. This study divides the factors involved into dependent variables (DVs) and
independent variables (IVs). IVs include flight factors—for instance, weather and pilot-
related data. DVs report the magnitude of the incident/accident, such as the number of
casualties. This research will improve existing databases—first, by adding variables, and
second, by using multivariate statistical analysis to assess the effect each group of IVs has
on correlations between flight factors and accident/incident-magnitude factors. Findings
demonstrate that pilot-related factors exert the most influence on the correlation between
the two categories. Our findings on the significance of factors or groups of factors will
assist researchers, policy makers, flight managers, and flight-crew schedulers in their
efforts to increase flight safety.

Keywords: flight accident/incident, causal analysis, multivariate analysis

Dr Mohammad Kashef is currently an Adjunct Professor at the George Washington

University teaching systems engineering and engineering management courses for
graduate and undergrad students. He has worked as Sr. Systems Engineer for National
Oceanic and Atmospheric Administration (NOAA) and has been a member of Joint Polar
Satellite System (JPSS) Integrated Product Team (IPT). Email: mkashef@gwu.edu, Phone:
+1 703-862-1938

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 24

1. INTRODUCTION

Air transportation is one of the fastest growing transportation modes, with an expected
growth rate of 5% to 6% over the next two decades (Netjasov & Janic, 2008). The
combination of the complexity of air-transportation systems and their respective
interconnectivity with other systems, such as air traffic control and navigation, makes their
management highly challenging. Air-transport management, which aims to improve flight
safety and reduce the associated costs, covers a broad range of disciplines, from risk
management to methods for flight-crew scheduling.

Despite major technological developments in the field, fatal accidents—often with high
numbers of casualties—occur with alarming frequency. Recent crashes, and in particular
Malaysian Airlines 370 (which disappeared on March 8, 2014, with 239 people on board),
Air Asia 8501 (which crashed into the Java Sea during bad weather, killing all 155
passengers), and Germanwings 9525 (which was deliberately crashed by the co-pilot,
killing all 150 people on board), have highlighted the critical importance of flight safety.

Aviation events are classified as either ‘incidents’ or ‘accidents’ (Nazeri, Barbara, Jong,
Donohue, & Sherry, 2008). In an aircraft incident, there are no fatalities, human injuries,
and/or substantial aircraft damage; nevertheless, flight safety is compromised. An accident
is one in which fatality, human injury, and/or substantial aircraft damage occurs.

Because of its severe consequences, aviation safety has become an important research
topic in the past decade (Orasanu, et al., 2001; Lee, 2006; Li & Baker, 2007; O’Connor,
Buttrey, O’Dea, & Kennedy, 2011; Cui & Ye, 2015), and it has been reviewed and studied
from a number of angles. Assessing and quantifying risk and safety in civil aviation has
been the focus of many studies, and possible approaches for improving the safety of
general aviation have been put forth (Janic, 2000; Li & Baker, 2007). In general, these
researches can be divided into three main groups, as shown in Figure 1. Some researchers
have studied aviation safety from a high-level managerial and administrative perspective
(Cacciabue, Cassani, Licata, Oddone, & Ottomaniello, 2015; Oster Jr., Strong, & Zorn,
2013; Tamasi & Demichela, 2011; Davison, Ciavarelli, Cohen, Fischer, & Slovic, 2001;
Netjasov & Janic, 2008). For instance, Netjasov & Janic (2008) describe four risk
categories: (1) risk to an individual, (2) statistical risk that an accident will occur, (3)
predicted risk, and (4) perceived risk. They also review different modelling methods of civil
aviation risk and safety and divide these into four groups: (1) causal, (2) collision risk, (3)
human-factor error, and (4) third-party risk.

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 25

 Figure 1 - Summary of Research on Civil Aviation Safety

The second group of research includes the application of risk assessment methods in
certain technical fields. Researchers in this group have investigated specific technical
domains of aviation risks, such as airport properties; airplane systems control; aviation
security screening; human factors, including pilot and air traffic controller; environmental
impacts; and others. Airport-runway properties and their effects on aviation safety have
been studied by researchers such as Waldron and Ford (2014), who investigated the airport
runway’s role in potential collisions and analysed how potential hazardous interactions can
vary among airports. In a related vein, Galle et al. (2010) have examined runway incursions
as a precursor to aviation accidents.

Another topic in this group is passenger security screening and how it affects aviation safety
risks. Nikolaev, Lee, and Jacobson (2012) have studied the problem of multistage,
sequential passenger screening with respect to passengers’ risk levels. Mook and
Scheinman (2011) have investigated risk-based screening systems to increase flight safety,
while Stewart and Mueller (2013) introduced a method for risk-reduction estimation in
commercial passenger airliners to prevent the aircraft from being hijacked.

Human error as a determining factor in aviation fatalities has also been studied in the
second group. Nelson (1997) states that more than 50% of accidents and incidents in
commercial aviation are caused by human error, and proposes a structured method to
identify and correct potential human errors in aviation operations. Shyur (2008) has

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 26

developed an analytical method to quantify aviation risks caused by human error, while
Naranji, Mazzuchi, and Sarkani (2015) use augmented cognition and automated systems
to reduce pilot error. Jin, Sun, and Kong (2010) examine the relationship between team
situation awareness (SA) and information sharing, and propose a method to reduce human
error. The authors also compare pilot SA and air traffic controller (ATC) requirements. Wei
et al. (2012) have studied the main factors that influence human error in the cockpit, and
developed a dynamic model for their prediction and evaluation. Human factors have also
been studied from another perspective, which is flight-crew scheduling and the airline
dispatcher’s role in flight management. For instance, Graves et al. (1993) developed a new
crew-scheduling system to reduce costs. The main concerns in such studies have been
reducing costs, minimising flight delays, and optimising flight routing (Graves, McBride,
Gershkoff, Anderson, & Mahidhara, 1993; Mercier & Soumis, 2007; Weide, Ryan, & Ehrgott,
2010; Nikulin & Drexl, 2010).

The third category includes studies that use mathematical and statistical models of civil
aviation risks. Since this category is the most relevant to our research, we will discuss these
in greater detail. Researchers have used a variety of mathematical tools to extract
meaningful patterns from aviation safety databases. Some of the newer techniques, such
as fuzzy logic, were applied by Lee (2006) to develop a quantitative model to assess
aviation safety risk factors. The factors included in the model are evaluated based on their
detectability, probability, criticality, etc. Other researchers have tried to capture patterns in
the occurrence of accidents using more rigorous methods. Wang and Gao (2013) analysed
the relationship between flight delays and aviation safety risk, and propose an approach
based on Bayesian networks to model safety risk assessment. Another Bayesian-based
model for avionic risk assessment was developed by Brooker (2011).

Causal methods can also be included in the third group; they are used to better determine
how factors that affect the level of risk can be employed to evaluate overall risk (Netjasov
& Janic, 2008). After each accident or incident, a causal report is prepared by related
agencies in which they identify causal factors (Luxhøj & Coit, 2006). Janic (2000) classifies
causal factors based on whether they are known or unknown and avoidable or unavoidable,
and further differentiates causal factors based on accident type—i.e., whether they can be
attributed to human error, mechanical failure, hazardous weather, sabotage, or military
operations.

Spouge (2004) further discusses the benefits of causal analysis, and argues that safety
managers and policy makers must understand the causes of accidents and evaluate the

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 27

benefits of different intervention policies before selecting measures for risk reduction.
Shyur, Keng, and Huang (2012) have developed an analytical model to analyse potential
aviation events using both accident and performance measures; they employ an extended
hazard-regression method to incorporate multiple safety performance indicators to assess
the probability of aviation events. Their model may not be suitable for estimating absolute
event probability, but it is valuable for understanding the structure of air events.

Common to these studies is the considerable emphasis placed on the use of different
approaches to study flight accidents and incidents. These, in turn, funnel into data and
prediction modelling. Underpinning these models are the data incorporated from aviation
safety databases maintained by the Federal Aviation Administration (FAA), the National
Transportation Safety Board (NTSB), and others, that have been used to model novel
approaches to assess risk, capture patterns, and construct prediction models. Modelling
the factors involved in aviation accidents/incidents has been at the core of these
researches, which have focused on managing flight risk and increasing flight safety. The
sheer range and diversity of these factors, however, significantly increases the difficulty of
determining how each factor contributes to an event. Christopher and Balamurugan (2013)
use data-mining approaches to predict aircraft accidents; they draw on the NTSB’s aviation
accident database, which does not include data on factors related to the pilot or weather.
Because these variables offer vital insight into the causes of fatal aircraft accidents and
improve data analysis, we have incorporated these factors in our database and will discuss
them in detail in the following sections.

Nazeri et al. (2008) used a method called ‘contrast-set mining of accidents and incidents’
to interpret the relationship between those two and propose a model for accident-risk
assessment. They found it difficult to identify a pattern in accidents, however, given the
rarity of their occurrence—an observation well documented by Janic (2000), who highlights
the difficulty in accurately locating, explaining, and managing overall aviation safety due to
the scarcity of events. In turn, the former research favors incidents as the predominant
tool in predicting the probability of an accident.

Though holistic in addressing all readily quantifiable data from either the FAA or NTSB
databases, other factors that may have a significant impact on the analysis of risk are not
included in these databases. Such factors are available, however, in NTSB Probable Cause
Reports (PCRs). Capturing these factors entails close review of individual PCRs and
translating relevant data points. Analyses that incorporate these factors would add
robustness to already rigorous prior research and allow the consideration of additional

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 28

factors. Nazeri et al. (2008) alludes to several such factors and notes, for example, the
importance of an event’s severity, phase of flight, and type of aircraft that, though
unavailable in public databases, would significantly enhance the value of the information
gained from the analysis.

Measuring how each factor affects an event—either individually or in combination—would

offer researchers and decision-makers a deeper understanding of aviation events and,
potentially, improve protocols and policies. It is worth mentioning that mathematical
explanations of factual observations in aviation safety are also of great value. For instance,
although the role of the pilot in flight safety seems obvious from an empirical point of view,
one can only study the effect of pilot contributions in combination with other factors by
using quantitative indices.

Differentiating and accentuating factors that have greater impact on events would save
time, money, and human resources—and, ultimately, increase flight safety and efficiency.
Therefore, investigating the relations between these factors—and specifically as dependent
and independent variables using multivariate correlation analysis—is the main focus of this
paper.

This study aims to examine how correlations between flight variables and incident/accident
variables are affected by different factors. This emphasis on correlative analysis is intended
to incorporate the aforementioned factors and demonstrate the approach’s ability to yield
highly specific results. Unlike researchers who have addressed the problem qualitatively,
such as Nazeri et al. (2008), our goal is to first enlarge the aviation safety database by
adding factors and values and then approach the problem quantitatively. This will not only
yield qualitative results, but will also enable us to apply our findings to more advanced
mathematical modelling that could be used by a variety of aviation personnel, such as flight
dispatchers and crew schedulers, to optimise flight risk. For example, a flight dispatcher
using the model could assess the risks imposed by weather on a specific flight against the
risks imposed by pilots (i.e., the combined risks of the pilot and co-pilot) and plan the flight
accordingly. The crew scheduler, in turn, could use the pilot variables to minimise risk by
selecting the optimal combination of pilot and co-pilot.

The paper is organised as follows: Section 2 discusses how the current study’s data were
obtained, and how the raw public database was improved to allow for subsequent analysis.
The section concludes by introducing dependent variables (DVs) and independent variables
(IVs). Section 3 introduces the multivariate statistical analysis used, and Section 4 presents

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 29

the results of our analytical method and discusses the significance of our findings. In
Section 5 we present our conclusions and discuss avenues for future research.

2. DATA

To obtain meaningful results, we first required a comprehensive and reliable database. The
second requirement was to define reasonable factors, including dependent and
independent variables, and the third requirement was a statistical tool capable of
measuring correlations between the variables. Careful selection of variables was crucial
for our analysis. Criteria for data selection and methods for data pre-processing, variable
selection, and grouping are described below. After building the database, a multivariate
statistical method will be introduced and applied to reveal correlations among variables and
identify the most influential.

Data Selection

The raw database for this research was obtained from the NTSB’s database, which contains
accident reports from 1962 to the present. Generally, a preliminary report is available online
shortly after an accident occurs. As the NTSB investigation progresses, more data are
added; upon completion of the investigation, the preliminary report is replaced by a final
description of the accident and its probable cause (NTSB, 2014).

For a database to be downloaded, one must specify certain information and submit a
relevant query. Preparing a database for retrieval often requires the provision of time
intervals, locations, and the type of aircraft involved. The raw database used in this
research was chosen from 10 different queries on the main NTSB repository; only accidents
with published PCRs were considered. Table 1 shows the details of the query selected for
the study, based on the data’s relevance, functionality, and feasibility; data from other
queries were either too cumbersome or too insignificant. The query selected includes 508
events, which comprise a sizable statistical population for data preparation.

Table 1. Selected Query Details

Query time interval 01/01/2003 to 12/31/2013

Location USA

Aircraft category Airplane

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 30

 Operation type Part 121: Air Carrier

Investigation type Accident/Incident

Report Status Probable Cause

In addition to the information provided in a downloadable spreadsheet, the PCR for each
event (accident or incident) is available as a PDF and is more detailed than the information
in the raw database.

The raw database was obtained and all corresponding PCRs downloaded. The database
consisted of rows and columns in which rows correspond to events and columns to
variables/factors. The database and PCR reports formed the basis for the process of data
preparation and database enhancement.

Data Preparation

As mentioned above, the raw database retrieved from the NTSB lacked information
pertinent to our study aims. We incorporated additional information as follows:
a) Grouping: Though public, the NTSB database is essentially intended for internal use;
therefore, significant effort is required to prepare the database to perform statistical
analysis. The first step was to group relevant factors into specific categories and reorder
the variables’ columns. For the purposes of this study, independent variables involving
accidents/incidents were categorised according to type. Pilot information was not included
in the original database, but because values were retrieved from PCRs in the next step and
added to the database, a category was created for pilot information. Independent variables
were divided into five categories:
 Flight information
 Weather information
 Airport information
 Aircraft information
 Pilot information

We selected three dependent variables, which concern the magnitude of the event:
 Event type (accident or incident)
 Severity of injuries/number of fatalities
 Level of damage to aircraft

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 31

b) New Variables: The study includes several critical variables, such as pilot information,
that are not provided in the raw NTSB database but are present, either explicitly or
implicitly, in the detailed Probable Cause Reports (PCRs). These variables were selected
based on advice from experts in the FAA and National Oceanic and Atmospheric
Administration (NOAA). Once the variables had been chosen, individual PCRs were carefully
examined to incorporate the new data into a more comprehensive database. Figure 2
shows the details of factors from the raw database and others that were collected from
narrative PCRs. Data shown in green are those used in the final analysis, which will be
discussed shortly.

Figure 2. Database Improvement using PCRs

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 32

c) Data from Additional Sources: Grouping and including new variables expanded the
database. In some instances, however, data for new variables—such as temperature, wind
speed, visibility, airspace type, and airport elevation—were missing from either the raw
databases or the PCRs. To acquire this information, we consulted sources other than the
NTSB, such as the NOAA database for weather information, the average daily temperature
(ADT) database of the University of Dayton, and the Federal Register for airport
information. These external sources filled critical gaps in the raw database. Database
improvement efforts are depicted in Figure 3. In some cases, the flight phase was not
explicitly stated in the report, but was implicit in the narrative. In such cases, we based
our judgment of the flight phase on the PCR’s narrative.

Figure 3. Database Improvement Using Additional Sources

d) Database Cleaning: Given the sheer number of events under consideration and the range
of variables, it was not possible to construct an exhaustive database. To ensure that the
data collected would be relevant, we removed factors that were irrelevant or insufficiently
significant (column cleaning). Likewise, events that were insufficiently significant or missing
too many variables were removed (row cleaning). These steps were performed only after
filling in as many gaps in the database as possible. The minimum acceptance threshold for
variables was 25%—i.e., variables that were missing values for more than 25% of events

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 33

were excluded. Percentages of missing values for each variable are shown in Figure 4; the
red line represents the 25% threshold. Some variables were removed because they were
almost uniformly constant—for example, the ‘Shoulder harness used?’ variable was either
‘yes’ or left blank in the PCR. There were also instances in which it was not possible to
quantify value—for example, Airport ID and Type of Airspace are not quantifiable. To reveal
their effects in the data analysis, however, they were included in the clustering phase,
which will be discussed later.

Figure 4. Percentage of Missing Values for Independent Variables

In addition to the above, the date of the event (in the form of MM/DD) and the local time
of occurrence (in the form of HH:MM) were normalised using the following formulas:
Date = (MM*30+DD)/365
Time = (HH+ (MM/60))/24

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 34

When it was necessary to convert qualitative data into quantitative data, we made logical
assumptions. For example, wind speeds that were reported as ‘calm’ were assigned a
numerical value of 0.5 mph. The database was now ready to perform statistical analyses,
and independent and dependent variables had been finalised. Table 2 shows the resulting
IVs and DVs, with information about type, range, and possible values for each variable.

Table 2. IVs and DVs for Statistical Analysis

Independent Variables (IVs)

Type and Possible Values Unit

1 Event date Normalised number between 0 andN/A

1

2 Event time Normalised number between 0 andN/A

1

3 Phase of Fight Standing, Taxi, Take Off, Climb, Cruise,

N/A

Descent, Approach, Landing

4 Temperature Continuous values Centigrade

5 Visibility Continuous values Statute Miles

6 Wind Speed Continuous values MPH

7 Number of Engines Discrete values N/A

8 Airframe Total Continuous values Hour

9 Age of pilot-in-command Discrete values Year

10 Pilot's Career Flight Time Continuous values Hour

11 Pilot's Specific Flight TimeContinuous

(with values Hour

accident/incident model)

Dependent Variables (DVs)

Type and Possible Values Unit

1 Event Type Binary: Accident (2) or Incident (1) N/A

2 Injury Severity Discrete values: Number of fatalities N/A

3 Level of aircraft damage None (0) , Minor (1), SubstantialN/A

(2),

Destroyed (3)

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 35

3. MULTIVARIABLE STATISTICAL ANALYSIS

To evaluate the effect of different IVs on the correlation between two sets of variables, a
multivariate statistical analysis tool was necessary. In multivariate statistics, multivariate
regression analysis is employed to investigate the relationship between a single DV and
multiple IVs (Hair et al. 2010). In cases in which both dependent and independent variables
are multivariate, the canonical correlation analysis (CCA) can be used to model the linear
relationship between multiple DVs and multiple IVs (Borga 2001, Hardoon et al. 2004).

CCA and its Application

Prior research has demonstrated the uses and value of the CCA method to predict multiple
DVs from multiple IVs (Bonner & Liu 2005, Singh et al. 2013, Singh et al. 2012). The aim
with CCA is to identify and quantify the interrelations between a p-dimensional variable X
and a q-dimensional variable Y (Dehon et al. 2000). The analysis looks for linear
combinations of the original variables, aTX and bTY, that have maximal correlation.

In mathematical terms, the CCA selects vectors α ∈ R and β ∈ R such that:

(α, β) = argmax , |Corr(a X, b Y)|

The selected univariate variables, U= . and = . , are referred to as canonical

variates. The number of pairs of canonical variates is equal to the minimum of p and q.
Each pair of canonical variates interprets the relationship in a given way. The CCA method
captures the highest correlation between linear combinations of IVs and linear
combinations of DVs. The most significant pairs are those with the highest correlations
(Nourzad & Pradhan, 2015). The single variables that represent X-values and Y-values,
respectively, are created using the formulas below:
U= . + . + ⋯+ .

V= . + . + ⋯+ .

We developed an approach to measure the correlation between DVs and IVs for flight
accidents/incidents using the CCA method. MATLAB statistical toolbox functions
(canoncorr) were used to run CCA. The first canonical correlation resulting from the
MATLAB function is the maximum correlation coefficient between U and V for all U and V
(Nourzad & Pradhan, 2015). The model’s effectiveness depends on the goodness of fit of
the captured linear relationships. The highest r-squared value (a measure of goodness of
fit) corresponds to the most effective model for capturing relationships between X-values
and Y-values. The main aim was to determine whether two sets of variables are related
and, if so, how different variables affect the r-squared values.

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 36

As stated in the previous section, we selected p=11 IVs and q=3 DVs (accident-magnitude
attributes) and used them to create canonical variates. The pairs with the highest r-squared
values have the strongest correlations. Figure 5 depicts our model, in which the r-squared
value will be measured and monitored depending on the change in the number of variables
employed.
Figure 5 - Research Model

Event Date
Flight
Local Time Information
Phase of Flight

Investigation
Temperature Type
Weather
Visibility Information

Wind Speed
Canonical Canonical
Variates r2 Variates Injury
Age Severity
for the X- for the Y-
Pilot's specific flight time
Pilot values values
Information

Pilot's career flight time

Aircraft
Number of Engines Damage
Aircraft
Information

Airframe Total Time

After database pre-processing, the first CCA run did not yield promising results. When all
IVs were included, the r-squared value was 0.36, which signifies a weak correlation. We
then performed clustering, which is a common approach in data analysis, to determine
whether better results could be achieved without losing the selected IVs. Clustering is
different from factors analysis; Cluster analysis tries to group cases/events that are more
similar to each other than to other types of cases whereas factors analysis attempts to
group features. Figure 6 is a generic illustration of how clustering can obtain stronger
results from multivariate analysis.

To select the best variable to cluster, four variables capable of being clustered were chosen:
Phase of Flight, Weather Condition, Flight Schedule, and Type of Clearance. Data clustering
was then performed on each variable, and the resulting r-squared values compared. As
shown in Figure 7, clustering based on Phase of Flight yielded the highest r-squared values.

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 37

 Figure 6 - Data Clustering

Figure 7 - Clustering Alternatives

As explained in Section 1, Nazeri et al. (2008) recommended that future studies include
the flight phase in which the accident/incident occurred. Together with results using other
variables (Figure 7), this led us to select Phase of Flight as the variable for clustering. Eight
flight phases were used as clusters for the database. To assess the effect of different
variables on the correlation between canonical variates, CCA was performed multiple times
on each cluster. Method details and results of the analysis are presented and discussed in
the next section.

4. RESULTS AND DISCUSSION

As mentioned earlier, the first CCA run on the entire database did not yield fruitful results,
since the r-squared value showed a weak correlation. Following clustering, the correlations
were strengthened significantly. Clustering was based on Phase of Flight IV, which lent
further relevance to nonnumerical values. The r-squared values for all eight phases, with
and without clustering, are shown in Figure 8.

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 38

 Figure 8 - Significance of Clustering

To investigate the effect of different variables on the r-squared values for each cluster, the
CCA statistical test was run six times with different variables. The first run included all IVs.
Successive runs were performed by excluding one group of IVs at a time while recording the
resulting changes. Consider, for example, the Cruise cluster, which includes all events that
occurred during that phase. The first run obtained 0.85 for the highest r-squared value
between canonical variates. The second run included all IVs except Weather Information. The
resulting r-squared value was 0.84, showing a minimal decrease in correlation. The third run
was performed including all IVs except Pilot Information. The resulting r-squared value was
0.56, showing a significant drop in correlation (34%). This supports the claim that the effect
of pilot-associated information is much more significant than weather information in the
investigation of correlations between different factors of flight events. Remaining runs were
performed in the same manner as the Cruise phase for the other seven Phases of Flight.
Detailed results are shown in Table 3.

Table 3 - r-squared Values for Different Run of CCA

Cruise Landing Climb Take-OffDescent Approach Taxi Standing All 8

1 All factors 0.85 0.56 0.92 0.74 0.88 0.88 0.72 0.65 0.36

2 Without 0.82 0.52 0.89 0.68 0.84 0.78 0.60 0.56 0.35

Flight Info

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 39

3 Without 0.84 0.47 0.81 0.67 0.86 0.83 0.65 0.52 0.33

Weather

Info

4 Without 0.56 0.47 0.91 0.60 0.81 0.69 0.47 0.62 0.26

pilot info

5 Without 0.83 0.56 0.88 0.74 0.83 0.80 0.70 0.63 0.32

Aircraft

Info

To gain a better understanding of the effect of different variables on goodness of fit, it is

necessary to calculate the level of drop in r-squared values when each group is excluded
from the analysis. Drops are calculated as percentages and shown in Table 4 and Figure 9.
As shown in Table 3, in five out of eight flight phases, pilot-associated data played the
most significant role in the correlation between DVs and IVs for accidents/incidents. This
phenomenon was observed by removing pilot-associated variables and monitoring the
changes in other variables. The highest drops are seen in the Taxi, Cruise, Approach, and
Take-off phases.
Table 4 - Drop in r-squared Values in %

Cruise Landing Climb Take-OffDescent Approach Taxi Standing All 8

Drop in r2 4% 7% 3% 8% 5% 11% 17% 14% 3%

for Flight

Info

Drop in r2 1% 16% 12% 9% 2% 6% 10% 20% 8%

for

Weather

Info

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 40

Drop in r2 34% 16% 1% 19% 8% 22% 35% 5% 28%

for Pilot

Info

Drop in r2 2% 0% 4% 0% 6% 9% 3% 3% 11%

for

Aircraft

Info

Our findings further demonstrate that even without clustering, pilot information has the
greatest effect of the IVs in all but two flight phases, Standing and Climb. The lower level
of correlation in the Standing phase can be attributed to the pilot’s low level of involvement;
it is reasonable that other factors, such as weather information or airport-related factors,
would be more influential, and this is corroborated by the results shown in Figure 9. In the
case of the Climb phase, the discrepancy in correlation may be attributed to the low number
of events recorded during this phase. The overall process of preparing the database,
performing multivariate statistical tests, and obtaining results is illustrated in Figure 10.

Figure 9 - Drop in r-squared Values for Different Flight Phases

35%
30%
25%
20% Drop in r-sq. for Flight Info
15% Drop in r-sq. for Weather Info
10% Drop in r-sq. for Pilot Info
5% Drop in r-sq. for Aircraft Info
0%

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 41

 Figure 10 - An Overview of Research Methodology

5. CONCLUSIONS AND FUTURE STUDIES

In this paper, CCA was used to analyse an enhanced aviation safety database to identify
the effects of different variables on correlations between flight factors and event factors.
The study’s focal point was to identify and assess relevant factors in aviation events. Prior
research with a similar aim has lacked a comprehensive database that incorporates not
only raw information from the NTSB, but, as with this study, additional data from sources
that are not immediately quantifiable (e.g., the NTSB’s PCRs). Database enhancement was
performed by studying all associated PCRs and retrieving new variables. The enhancement
process included grouping, introducing new variables, obtaining data from additional
sources, and database cleaning. Having said that, this research was limited to events
happened in USA and mentioned in the NTSB main database. The next step was to
determine whether the enhanced database would be suitable for CCA, with the goal of
discovering the most influential factor among the IVs considered. Initial results were not
promising, so a clustering method was proposed. Clustering based on Phase of Flight was
selected after comparing clustering options. CCA was run six times in each cluster with
different variables, based on the research model, to investigate the variables’ effects on r-
squared values between DVs and IVs.

Our findings statistically support the empirical observation that pilot-associated data,
including age, career flight time, and experience with the aircraft model involved in the
event, are the most effective factors in demonstrating a correlation between dependent

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 42

and independent variables of aviation events. The second, third, and fourth most significant
factors were variables associated with weather, flight time, and aircraft, respectively.

This research provides a framework for further inquiry and the construction of a predictive
model using the more comprehensive database we have made available. Such a predictive
model could be used by different stakeholders, such as risk managers, airline planners,
crew schedulers, and dispatchers, to minimise flight risk and improve flight safety. These
findings could be used to improve flight-crew scheduling and dispatching practices;
consideration of these factors when selecting pilots and co-pilots could also reduce flight
risk. Prior entering raw data in regular flight scheduling process, the above mentioned
predictive model can be used to assess the combination of those factors and the level of
risk they impose. This model can potentially tell schedulers that in specific weather
conditions, how assigning low experience pilot will increase the risk of flight. This model
can also be used to reduce the risks based on the known variables prior to flight. “Flight
variable assessment” based on this model can be added into existing flight scheduling
processes to measure the level of risks imposed by flight variable combination. For
example, a pilot with more experience and higher variable values could be paired with a
low-hours co-pilot with less experience to optimise flight risk and, possibly, lower cost.
Likewise, if weather factors based on our findings were included in the crew- scheduling
process, better results might be obtained. By evaluating the risks prior to flight, the
dispatcher or flight-crew scheduler could modify and reroute the flight, if necessary, based
on weather conditions and pilot variables.

CCA was applied in this research so it imposes its limitations and assumptions. Linear
relationship assumed for all variables in each set and also between sets. Applying none-
linear methods can improve results and contribute to findings of our study. Widening the
events selection criteria and including other countries aviation events, can potentially
improve the results. Our method is also adaptable for a wide range of research topics.
Other analytic methods, such as neural network analysis or fuzzy logic, could be used to
determine whether similar results can be obtained.

REFERENCES

 Bonner, A., & Liu, H. (2005). Canonical Correlation, an Approximation, and the

Prediction of Protein Abundance. Eighth Workshop on Mining Scientific and Engineering

Datasets (MSD’05), (pp. 29-38). Newport Beach.

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 43

 Brooker, P. (2011). Experts, Bayesian Belief Networks, rare events and aviation risk

estimates. Safety Science, 49, 1142-1155.

 Cacciabue, P. C., Cassani, M., Licata, V., Oddone, I., & Ottomaniello, A. (2015). A

Practical Approach To Assess Risk In Aviation Domains For Safety Management

Systems. Cogn Tech Work, 17, 249-267. doi:10.1007/s10111-014-0294-y

 Christopher, A., & Balamurugan, S. (2013). Data Mining Approaches for Aircraft

Accidents Prediction: An Empirical Study on Turkey Airline. IEEE International

Conference on Emerging Trends in Computing, Communication and Nanotechnology

(ICECCN 2013), (pp. 739-745).

 Cui, Q., & Ye, L. (2015). The Change Trend and Influencing Factors of Civil Aviation

Safety Efficiency: The Case of Chinese Airline Companies. Safety Science, 75, 56-63.

 Davison, J., Ciavarelli, A., Cohen, M., Fischer, U., & Slovic, P. (2001). The Many Faces

of Risk in Aviation Decision-Making, The Human Factors and Ergonomics Society 45th

Annual Meeting, 307 - 310.

 Dehon, C., Filzmoser, P., & Croux, C. (2000). Robust Methods for Canonical Correlation

Analysis. Data Analysis, Classification, and Related Methods, 321-326.

 Galle, K. M., Ale Jr., J. C., Hossain, M. M., Moliterno, M. J., Rowell, M. K., Revenko, N.

V., . . . Haimes, Y. Y. (2010). Risk-Based Airport Selection for Runway Safety

Assessments Through the Development and Application of Systems-Driven Prioritization

Methodologies. 2010 IEEE Systems and Information Engineering Design Symposium.

Charlottesville.

 Graves, G. W., McBride, R. D., Gershkoff, I., Anderson, D., & Mahidhara, D. (1993,

June). Flight Crew Scheduling. Management Science, 39(6), 736 - 745.

 Hardoon, D. R., Szedmak, S., & Shawe-Taylor, J. (2004). Canonical Correlation

Analysis: An Overview with Application to Learning Methods. Neural Computation, 16,

2639-2664.

 Janic, M. (2000). An Assessment of Risk and Safety in Civil Aviation. Journal of Air

Transport Management, 6, 43-50.

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 44

 Jin, H., Sun, R., & Kong, X. (2010). Enhance Team Situation Awareness by Sharing

Information to avoid Human Errors in Aviation. Third International Conference on

Knowledge Discovery and Data Mining.

 Lazarick, R. T. (1998). Airport Vulnerability Assessment - An Analytical Approach. SPIE

3575, Enforcement and Security Technologies.

 Lee, W.-K. (2006). Risk Assessment Modeling in Aviation Safety Management. Journal

of Air Transport Management, 12(5), 267-273.

 Li, G., & Baker, S. P. (2007). Crash Risk in General Aviation. Journal of American Medical

Association, 297(14), 1596-1598.

 Luxhøj, J. T., & Coit, D. W. (2006). Modeling Low Probability/High Consequence Events:

An Aviation Safety Risk Model. Proceedings of the 2006 Reliability and Maintainability

Symposium (RAMS). Newport Beach.

 Mercier, A., & Soumis, F. (2007). An integrated aircraft routing, crew scheduling and

flight retiming model. Computers & Operations Research, 34, 2251 - 2265.

 Mook, D., & Scheinman, E. D. (2011). Risk Based Screening and Explosive Detection at

the Passenger Screening Checkpoint. 2011 IEEE International Conference on

Technologies for Homeland Security (HST), (pp. 98-103).

 Naranji, E., Sarkani, S., & Mazzuchi, T. (2015, June). Reducing Human/Pilot Errors in

Aviation Using Augmented Cognition and Automation Systems in Aircraft Cockpit.

Transactions on Human-Computer Interaction, 7(2), 71 - 96.

 Nazeri, Z., Barbara, D., Jong, K. D., Donohue, G., & Sherry, L. (2008). Contrast-Set

Mining of Aircraft Accidents and Incidents. In Advances in Data Mining. Medical

Applications, E-Commerce, Marketing, and Theoretical Aspects (Vol. 5077, pp. 313-

322). Springer Berlin Heidelberg. doi:10.1007/978-3-540-70720-2_24

 Nelson, W. R. (1997). Structured Methods for Identifying and Correcting Potential

Human Errors in Aviation Operations. 1997 IEEE International Conference On Systems,

Man, And Cybernetics. 4, pp. 3132-3136. IEEE.

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 45

 Netjasov, F., & Janic, M. (2008). A review of research on risk and safety modelling in

civil aviation. Journal of Air Transport Management, 14, 213-220.

 Nikolaev, A. G., Lee, A. J., & Jacobson, S. H. (2012). Optimal Aviation Security

Screening Strategies With Dynamic Passenger Risk Updates. IEEE Transactions on

Intelligent Transportation Systems, 13(1), 203-212. doi:10.1109/TITS.2011.2167230

 Nikulin, Y., & Drexl, A. (2010). Theoretical aspects of multicriteria flight gate

scheduling: deterministic and fuzzy models. Journal of Scheduling, 13, 261 - 280.

doi:10.1007/s10951-009-0112-1

 Nourzad, S. H., & Pradhan, A. (2015). Computational Modeling of Networked

Infrastructures: A Macroscopic Multivariate Approach. Journal of computing in civil

engineering.

 O’Connor, P., Buttrey, S. E., O’Dea, A., & Kennedy, Q. (2011). An Assessment of Safety

Climate in U.S Naval Aviation. the Human Factors and Ergonomics Society 55th Annual

Meeting, (pp. 1740-1744). doi:10.1177/1071181311551361

 Orasanu, J., Davison, J., Ciavarelli, A., Cohen, M., Fischer, U., & Slovic, P. (2001). The

Many Faces of Risk in Aviation Decision Making. the Human Factors and Ergonomics

Society 45th Annual Meeting, (pp. 307-310).

 Oster Jr., C. V., Strong, J. S., & Zorn, C. K. (2013). Analyzing Aviation Safety: Problems,

Challenges, Opportunities. Research in Transportation Economics, 43, 148-164.

 Shyur, H.-J. (2008). A Quantitative Model for Aviation Safety Risk Assessment.

Computers & Industrial Engineering, 34-44.

 Shyur, H.-J., Keng, H., I, I.-K., & Huang, C.-L. (2012). Using Extended Hazard

Regression Model to Assess the Probability of Aviation Event. Applied Mathematics and

Computation, 218, 10647-10655.

 Singh, A., Acharya, N., Mohanty, U. C., & Mishra, G. (2013). Performance of Multi Model

Canonical Correlation Analysis (MMCCA) for Prediction of Indian Summer Monsoon

Rainfall Using GCMs Output. Comptes Rendus Geoscience, 345, 62-72.

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 46

 Singh, A., Kulkarni, M. A., Mohanty, U. C., Kar, S. C., Robertson, A. W., & Mishra, G.

(2012). Prediction of Indian Summer Monsoon Rainfall (ISMR) Using Canonical

Correlation Analysis of Global Circulation Model Products. Meteorological Applications,

19(2), 179-188. doi:10.1002/met.1333

 Spouge, J. (2004). A Demonstration Causal Model for Controlled Flight into Terrain.

London: Det Norske Veritas.

 Stewart, M. G., & Mueller, J. (2013). Aviation Security, Risk Assessment, and Risk

Aversion for Public Decisionmaking. Journal of Policy Analysis and Management, 32(3),

615-633. doi:10.1002/pam.21704

 Tamasi, G., & Demichela, M. (2011). Risk Assessment Techniques for Civil Aviation

Security. Reliability EngineeringandSystemSafety, 96, 892-899.

 Waldron, T. P., & Ford, A. T. (2014). Investigating the Causality of Potential Collisions

on the Airport Surface. 2014 IEEE/AIAA 33rd Digital Avionics Systems Conference

(DASC), (pp. 7C5-1-7C5-11). Colorado Springs. doi:10.1109/DASC.2014.6979520

 Wang, H., & Gao, J. (2013). Bayesian Network Assessment Method for Civil Aviation

Safety Based on Flight Delays. Mathematical Problems in Engineering. Retrieved from

http://dx.doi.org/10.1155/2013/594187

 Wei, Z., Zhuang, D., Wanyan, X., Wei, H., & Zhou, Y. (2012). Prediction and Analysis

of the Human Errors in the Aircraft Cockpit. 2012 5th International Conference on

BioMedical Engineering and Informatics (BMEI 2012). IEEE.

 Weide, O., Ryan, D., & Ehrgott, M. (2010). An iterative approach to robust and

integrated aircraft routing and crew scheduling. Computers & Operations Research, 37,

833 - 844.

Journal of Air Transport Studies, Volume 7, Issue 2, 2016 Page 47