Sunteți pe pagina 1din 7

Accred Qual Assur

DOI 10.1007/s00769-016-1245-5

GENERAL PAPER

An enhanced bootstrap method to detect possible fraudulent


behavior in testing facilities
Rosembergue P. de Souza1 Luiz F. R. C. Carmo1 Luci Pirmez1

Received: 12 March 2016 / Accepted: 12 November 2016


Springer-Verlag Berlin Heidelberg 2017

Abstract The testing process determines the specified Introduction


characteristics of an object. Testing is commonly used to
verify the quality, safety, reliability, efficiency, and envi- The word fraud describes the actions that are surrepti-
ronmental sustainability of products and services. In Brazil, tiously undertaken by one party to obtain an unjust
the National Institute of Metrology, Quality and Technol- advantage over another [1]. Undertaking fraudulent
ogy (Inmetro) is responsible for accrediting the facilities behavior when testing processes can result in products that
that undertake testing and calibration activities. Typical are dangerous to the ordinary citizen and to the environ-
accredited testing services include: water quality, clinical ment. Further, fraudulent behavior produces unfair
tests, emissions tests, and vehicle safety tests. Undertaking competition between concurrent facilities. In Brazil, the
fraudulent behavior when testing processes can result in National Institute of Metrology, Quality and Technology
products that are dangerous to the ordinary citizen and to (Inmetro) is responsible for accrediting the facilities that
the environment. Further, fraudulent behavior produces undertake testing and calibration activities.
unfair competition between concurrent organizations. This Inmetro evaluates these accredited organizations annu-
paper presents a method to detect possible fraudulent ally to verify their compliance with technical requirements.
behavior in accredited testing facilities. The proposed Inmetro also maintains a list of trained and qualified
technique combines the bootstrap method and the Demp- experts. Additionally, Inmetro undertakes some surprise
sterShafer theory of evidence, to investigate several visits to its accredited organizations. Those in-loco
fraudulent aspects at once. To validate the proposed inspections are to verify whether the accredited organiza-
method, we use a Brazilian real-world dataset with real tions are continually following the technical regulations.
cases of fraud. The proposed technique successfully detects However, the high inspection costs mean that it is not
the accredited testing facilities with fraudulent behavior. financially viable to inspect all of the accredited testing
facilities. To enhance the efficiency of such surveillance
Keywords Testing  Fraud  Bootstrap  activities, we propose a statistical method that combines
DempsterShafer theory the bootstrap method and DempsterShafer theory. The
objective is to identify the facilities with suspicious
behavior; that is, to identify the potential candidates for
Inmetro inspections. On identifying the suspicious facili-
ties, Inmetro could undertake detailed investigations on the
tests performed by these organizations, using the tech-
niques presented in [2, 3].
& Rosembergue P. de Souza Bootstrapping is suitable to identify suspicious testing
rosembergue.souza@ppgi.ufrj.br facilities, because it can be used to test hypotheses
1 involving comparisons between two or more groups. The
Av. Athos da Silveira Ramos, 274, Predio do CCMN -
Cidade Universitaria, Ilha do Fundao, technique does not require any distribution assumptions,
Rio de Janeiro 21941-916, Brazil and it is relatively easy to account for the specifics of each

123
Accred Qual Assur

situation. Further, bootstrapping allows for the use of alternative hypothesis, denoted by Ha, is the assertion
nonstandard test statistics [4]. Some studies use the that is contradictory to H0. The null hypothesis will be
bootstrap method as a tool in the fraud detection process. rejected in favor of the alternative hypothesis only if
Taylor et al. examine the suitability of existing methods the sample evidence suggests that H0 is false. If the
for fraud detection. They propose a new technique to sample evidence does not strongly contradict H0, then
identify fake questionnaire data, in both an informal we continue to believe in the plausibility of the null
graphical mode and as a bootstrap method. The proposed hypothesis. The two possible conclusions from a
technique exploits the underlying correlation structure of hypothesis testing analysis are (1) reject H0 and (2) fail
a questionnaire and the difficulty in fabricating such to reject H0 [10].
details [5]. Subsequently, considering that fraudulent data A one-sided hypothesis test is undertaken when the
cases have certain clustering structures, Wu and Carlsson critical area of the distribution falls under only one of the
discuss the potential for using statistical clustering distribution tails. The one-sided test decides whether a
methods in fraud detection. They identify and explore parameter is less than, greater than, or equal to some
angular, neighborhood, and repeated measurements clus- value. A two-sided hypothesis test is undertaken when the
tering. They propose corresponding test statistics and critical area of the distribution falls under both distribu-
carry out bootstrap tests [6]. Finally, Kirkwood et al. tion tails. The two-sided test decides whether a parameter
study the application of central statistical monitoring is either greater or less than some value. In testing a
methods, including bootstrapping, in data from real clin- hypothesis, the significance level a is the probability of
ical trials. In that case, bootstrapping is based on the sum rejecting the null hypothesis H0 when it is true. The
of the squared differences between the correlations in p value is the smallest level of significance that would
each site and the correlations in all of the data sites [7]. lead to rejection of the null hypothesis H0 with the given
While these proposed fraud detection methods are data. To make a proper decision about whether or not to
somewhat satisfactory, there is room for improvements. reject H0, one has to determine a cutoff probability for the
For example, such studies only consider one aspect of p value; this cutoff is the significance level a. If the
fraud behavior when using bootstrapping. p value is greater than or equal to a, then we fail to reject
We use the bootstrap method and DempsterShafer H0. Correspondingly, if the p value is less than a, then we
theory to merge several aspects of fraudulent behavior in reject H0 [11].
one degree of anomaly. DempsterShafer theory allows us In the bootstrap method, as in other hypothesis tests,
to combine evidence from different sources and yields a the first step is setting up the null hypothesis H0, the
degree of belief that takes into account all of the available alternative hypothesis Ha, and the significance level a.
evidence [8]. This method is particularly useful, because Then, using the significance level, the number M of
when data are falsified to pass certain statistical tests, resampling with replacement can be decided. M cannot be
these falsified data are likely to fail other tests [9]. less than 1000 for a test at the 5 % level of significance,
Combining multiple tests could make it more difficult to and M cannot be less than 5000 for a test at the 1 % level.
commit fraud. The data are then resampled through replacement, and the
The remainder of this paper is structured as follows: We statistic of interest S is calculated for the resampled data.
first present some basic concepts, followed by the general Finally, after repeating this procedure M times, the
model of the approach used in this work. We then check empirical distribution is built for S. The value s0 of S for
the proposed model with real-world data, and the final the observed data is compared with the distribution of
section concludes. S. For example, in a one-sided test to decide whether a
parameter is greater than some value, if s0 looks like a
typical value from the empirical distribution of S, then the
Background null hypotheses is accepted. Otherwise, if s0 is unusually
large, then the data are unlikely to have arisen if the null
Hypothesis testing with bootstrapping model is true, and it can be concluded that the alternative
hypothesis is more plausible. The significance level of s0
A hypothesis test decides which of two contradictory is the proportion or percentage of values that are as
claims about a statistical distribution is correct. A sta- extreme, or more extreme, than this value in the empirical
tistical hypothesis is a claim or assertion about (1) the distribution of S [4, 1215].
value of a single parameter, or (2) the values of several Traditionally, a bootstrap method is used to evaluate one
parameters, or (3) the form of an entire probability aspect at a time. To evaluate more than one aspect at the
distribution. The null hypothesis, denoted by H0, is the same time, a fusion techniquesuch as the Dempster
claim that is initially assumed to be true. The Shafer theoryis more appropriate.

123
Accred Qual Assur

DempsterShafer theory Enhanced bootstrap method

The DempsterShafer theory is a mathematical theory of We combine bootstrapping with DempsterShafer theory
evidence that it is based on two ideas. The first is the notion to propose an enhanced bootstrap method. This method
of obtaining degrees of belief for one question based on the combines several aspects of fraudulent behavior in one
subjective probabilities for a related question. The second degree of anomaly. We want to test whether some sites are
idea is Dempsters rule for combining the degrees of belief anomalous compared to others. Our hypotheses for the
when they are based on independent items of evidence [8]. proposed enhanced bootstrap method are:
The set of all of the possible states of the system con-
H0: The testing facility is normal
sidered is called the frame of discernment, which is
Ha: The testing facility is anomalous
represented by X. The set of all of the subsets of X is known
as the power set, which is denoted by the symbol 2X. A The statistic to discriminate between the null hypothesis
subset of those 2X sets may consist of a single possible state, and the alternative is the degree of anomaly S, calculated
or of a conjunction of possible states. Moreover, it is using the DempsterShafer theory of evidence. In the case
required that all of the possible states are unique, are not of anomaly detection, the possible states for the testing
overlapping, and are mutually exclusive. Pieces of evidence facility are anomalous A or normal (not anomalous)
could be symptoms or events that occurred or may occur N. Therefore, the frame of discernment for this context is
within a system. One piece of evidence is related to a single X = {A, N}. The power set is 2X = {;, A, N, V}, where
possible state, or to a set of possible states [16]. V represents the state of the subset {either A or N}, and ;
The function, called the basic belief assignment, is the represents the empty set. To obtain the degree of anomaly
degree of belief that a possible state or a set of possible of a site, we need to model the basic belief assignment
states is warranted, which is denoted by m: 2X ? [0, 1]. function for all of the evidence available that assign an
This function assigns an evidential weight to a subset in X. anomaly degree to that site. Then, we can calculate degree
Thus, analogous to the case of probabilities, each possible of anomaly S using B(A) = S.
event in X is assigned a value, ensuring that the sum of all The enhanced bootstrap procedure starts with calculat-
of the possible outcomes sums to unity. Within Dempster ing the degree of anomaly S for the selected site. Next, we
Shafer theory, these values denote the belief. A belief form a pseudo-site by sampling with replacement n sub-
measure is given by the function B: 2X ? [0, 1]. The belief jects from the entire study, where n is the number of
function can be interpreted as the total amount of justified elements in the dataset of the selected site. Then, we cal-
support given to a subset in X. Dempsters rule can be used culate the degree of anomaly S for the created pseudo-site.
to combine two independent sources of evidence, expres- We repeat this procedure of sampling with replacement and
sed by the following equation: anomaly degree calculation for M times, where M is
P defined according to the desired significance level. Sub-
C\DZ6; mC mD
mZ P ; 1 sequently, in the empirical distribution of S, we count the
1  C\D; mC mD
anomaly degree values that are as extreme, or more
with C; D; Z  X and the basic belief assignment functions extreme, than the anomaly degree of the selected site. If the
m(C) and m(D). In others words, the numerator represents counter is larger than aM, then we accept the null
the accumulated evidence for the sets C and D, which hypothesis H0 for the selected site. Otherwise, the alter-
supports the possible state Z, and the denominator sum native hypothesis Ha is more plausible for the selected site.
quantifies the amount of conflict between the two sets [16]. We perform the enhanced bootstrap procedure for all of the
sites being investigated. When the test rejects the null
hypothesis for a site, then this site is marked as suspicious.
Proposed procedure

This study proposes a technique that combines the boot- Illustrative case
strap method with DempsterShafer theory. The objective
is to detect accredited testing facilities that have possible Complaints about fraud in testing facilities
fraudulent behavior. The first step is to select the sites. We
then perform the enhanced bootstrap method on the In Brazil, Inmetro receives complaints and tip-offs about
selected sites. Then, when the sites are compared, we list fraudulent behavior in accredited testing facilities. Some of
all of the suspicious sites. This list can be used to identify this correspondence is from the customers and the
the sites that should be inspected to check for fraudulent employees of the relevant accredited organization. For
behavior. example:

123
Accred Qual Assur

Complaint 1: The testing facility exchanges a substan- and the misuse of a vehicle can also increase the likelihood of a
dard object with one with no defects to obtain legitimate vehicle failing the safety tests. In an accredited testing facility,
results. when falsifying the test results, the fraudster aims for a
Complaint 2: The testing facility edits the test reports determined range of safety limit values. Therefore, the vari-
with invented results. ability of the results is affected. The first symptom of potential
Complaint 3: The testing facility tampered with its fraud is in the variability of the test results. The lower this
equipment to falsify the test results. variability, the higher the suspicion of fraud.
Some of the complaints about fraud in accredited testing
When Inmetro receives complaints about fraudulent
facilities indicate that there may be some mechanism of
behavior, it undertakes a surprise visit to that testing
editing the test reports with invented results. Humans are
facility. There, Inmetro undertakes an in loco evaluation to
poor random number generators; hence, another way of
verify the facilitys compliance with the law.
checking for fraudulent data is looking for digit preference
[9]. Therefore, the second evidence of potential fraud is digit
Dataset
preferences in the test results, i.e., the tendency to use some
digits over others when falsifying test results. The higher the
The dataset used in this study is from six accredited vehicle
digit preference, the higher the suspicion of fraud.
testing facilities in Rio de Janeiro, Brazil. The data were
Using the variability and the digit preference as evi-
extracted from the sideslip tester measures from each
dence factors, we can now provide the basic belief
organization. A sideslip tester is used to verify the safety
assignment functions. In this case, the weight of the basic
conditions in vehicles. Two of these testing facilities were
belief assignment function can be interpreted as the degree
detected as fraudulent in unexpected visits performed by
of anomaly revealed by that evidence.
Inmetro. Table 1 outlines the dataset used in this study,
including quantitative sideslip measurements, and whether
the testing facility was detected as fraudulent. All of the Evidence 1: Variability
data were collected in 2014.
To measure the variability of the data, we use the standard
Modeling basic belief assignment function deviation. For simplicity, we use a linear function to model
our basic belief assignment function. This measures the
To use the DempsterShafer theory, we need to model the anomaly degree based on the standard deviation of the
basic belief assignment functions from the set of evidence. data. This basic belief assignment function has to map the
First, we must identify the evidence that can indicate values of the standard deviation r in the range [0, 1]. To
fraudulent behavior in accredited inspection bodies. Some complete our linear model, we need to determine the
of the patterns that could reveal fraud are: digit preference, domain range of the standard deviation. Where there are at
round number preference, too few or too many outliers, too least three numbers in the dataset, the standard deviation
little or too much variance, strange peaks, and data that are can never exceed 60 % of the range [17]. In this case, the
too skewed [9]. range of the sideslip tester is defined by -15 m/km to
In the case of vehicle safety tests, we expect a high vari- ?15 m/km. Considering the sideslip measures, a good
ability in the test results. Numerous factors affect the results estimation for the standard deviation domain is [0, 18].
across different vehicles, leading to variable results. For Equation 2 defines the basic belief assignment function
example, as vehicles age and accumulate mileage, they are to measure the anomaly degree based on the standard
more likely to fail the safety tests. Moreover, the vehicle deviation of the data.
model, the degree to which the owners maintain their vehicles, 9
m1 A w1 >
=
m1 N 0 2
Table 1 Sideslip measurements and detection of fraud in six >
;
accredited vehicle testing facilities in Rio de Janeiro, Brazil, in 2014 m1 V 1  m1 A
Site Sideslip measurements Fraudulent? where
1 137 No 18  r
w1 :
2 142 No 18
3 237 Yes
Evidence 2: Digit preference
4 117 No
5 130 No
Digit preference patterns can reveal possible fraud. We use
6 174 Yes
the Euclidean distance to compare the distribution of the

123
Accred Qual Assur

leading (first) digits used from each site with the distribution m1 A m2 A m1 A m2 V m1 V m2 A
from all of the other sites together. First, we calculate the B A
K
proportions of the leading digits in the selected site. This m1 N m2 N m1 N m2 V m1 V m2 N
calculation uses the ratio between the leading digit fre- B N
K
quency and the number of observations on that site m1 V m2 V
p0 ; p1 ; p2 ; . . .; p9 . We do this for the digits 09. Next, we B V
K
calculate the proportions of the leading digits from all of the
where
other sites together. This calculation uses the ratio between
the leading digit frequency and the number of observations K m1 A m2 A m1 A m2 V m1 V m2 A
in all of the remaining sites P0 ; P1 ; P2 ; . . .; P9 . We do this m1 N m2 N m1 N m2 V m1 V m2 N
for the digits 09. Finally, using these two vectors of pro- m1 V m2 V
portions, Eq. 3 calculates the digit preference distance dDP.
q Finally, consider combining multiple pieces of evidence
dDP p0  P0 2    p9  P9 2 3 from different properties of A. We can compute B(A) by
combining any evidence pair and then combine that result
With an index to compare the digit preference between with the remaining bits of evidence [18].
sites, we can now provide a basic belief assignment
function for this evidence. Remember, the higher the digit
preference, the higher the suspicion of fraud. Therefore, the Results
higher the digit preference distance, the higher the anomaly
degree of the site. This section empirically investigates the proposed proce-
For simplicity, we also use a linear function to model our dure. We use a dataset of six real-world sites that contain
basic belief assignment function. This measures the anomaly two real cases of fraud (sites 3 and 6). We use MATLAB
degree based on the digit preference distance of the data. in the analysis. We use a = 1 % level of significance, and
This function has to map the values of the digit preference M = 5000 iterations. Table 2 gives the results.
distance in the range [0, 1]. To complete the linear model, Table 2 shows that each of the six pseudo-sites created
we need to determine the domain range of the digit prefer- has standard deviation and digit preference distance values
ence distance. In an extreme situation, one site could have a that lie in the intervals of our basic belief assignment func-
digit preference for only one digit. The remaining sites could tion models. Therefore, the proposed basic belief assignment
have digit preferences for only the other leading digits. This functions models are valid to compare the testing facilities.
implies that thepmaximum
digit preference distance, as Figure 1 shows the empirical cumulative distribution
defined here, is 2. Then, a good estimation
p for the domain function for the anomaly degree of the six accredited vehicle
of the digit preference distance is [0, 2]. safety inspection bodies in the study. The results shown are
Equation 4 defines the basic belief assignment function after M = 5000 sampling with data replacement using the
to measure the anomaly degree based on the digit prefer- enhanced bootstrap method. The cumulative distribution
ence distance of the data. function returns the probability of the values in the sample
9 being less than (or equal to) some value in the sample [19].
m2 A w1 >
= For site 1, the anomaly degree is S1 = 0.833. The per-
m2 N 0 4
>
; centage of values that are as extreme or more extreme than
m2 V 1  m2 A the anomaly degree of site 1 in the empirical distribution is
greater than the chosen level of significance a = 1 %. In
where
this case, the p value is 100 %. Therefore, we fail to reject
dDP
w2 p :
2 Table 2 Standard deviation range and digit preference distance after
M = 5000 iterations
In each of the previous basic belief assignments, the
Site r range dDP range
basic belief assignments of the zero value for all
m(N) mean that those evidence give a zero degree of not 1 [2.074, 3.333] [0.017, 0.181]
anomalous behavior about the inspection activities [18]. 2 [2.173, 3.280] [0.017, 0.187]
3 [1.913, 3.124] [0.013, 0.127]
DempsterShafer rule of combination 4 [2.051, 3.334] [0.015, 0.213]
5 [2.064, 3.285] [0.017, 0.263]
To combine two independent sources of evidence, we can 6 [2.065, 3.190] [0.018, 0.374]
use the following equations:

123
Accred Qual Assur

Fig. 1 Empirical cumulative


distribution function and
anomaly degree for the six
accredited testing facilities after
M = 5000 iterations using the
enhanced bootstrap method,
where S1, S2, S3, S4, S5, and S6
are the observed anomaly
degrees for the sites 1, 2, 3, 4, 5,
and 6, respectively

the null hypothesis, and site 1 is not marked as suspicious. The results show that sites 3 and 6 are anomalous com-
Following, for site 2, the anomaly degree is S2 = 0.868. pared with the other sites. Although all of the sites have a
The percentage of values that are as extreme or more similar anomaly degree, the p values of sites 3 and 6
extreme than the anomaly degree of site 2 in the empirical indicate that those anomaly degrees rarely appear on their
distribution is greater than the chosen level of significance empirical distributions.
a = 1 %. In this case, the p value is 75.72 %. Therefore,
we fail to reject the null hypothesis, and site 2 is not
Conclusion
marked as suspicious. On the other hand, for site 3, the
anomaly degree is S3 = 0.926. The percentage of values
We present a technique to identify unusual test result pat-
that are as extreme or more extreme than the anomaly
terns in accredited testing facilities. This technique allows
degree of site 3 in the empirical distribution is less than the
us to map the accredited testing facilities with anomalous
chosen level of significance a = 1 %. In this case, the
behavior, considering several fraudulent aspects at once.
p value is 0.02 %. Therefore, we reject the null hypothesis,
To test the technique, we use a real dataset from
and site 3 is marked as suspicious. Next, for site 4, the
accredited testing facilities in Brazil. The proposed method
anomaly degree is S4 = 0.871. The percentage of values
successfully identifies the sites that presented fraudulent
that are as extreme or more extreme than the anomaly
data. Future studies could investigate methods to auto-
degree of site 4 in the empirical distribution is greater than
matically analyze the video records of the services
the chosen level of significance a = 1 %. In this case, the
performed by the testing facilities.
p value is 60.48 %. Therefore, we fail to reject the null
hypothesis, and site 4 is not marked as suspicious. Fol- Acknowledgements The authors would like to thank the National
lowing, for site 5, the anomaly degree is S5 = 0.879. The Institute of Metrology, Quality and Technology of Brazil for its
percentage of values that are as extreme or more extreme support.
than the anomaly degree of site 5 in the empirical distri-
bution is greater than the chosen level of significance
Compliance with ethical standards
a = 1 %. In this case, the p value is 30.88 %. Therefore,
we fail to reject the null hypothesis, and site 5 is not Conflict of interest The authors declare that they have no conflict of
marked as suspicious. Finally, for site 6, the anomaly interest.
degree is S6 = 0.917. Note that the percentage of values
that are as extreme or more extreme than the anomaly
degree of site 6 in the empirical distribution is less than the References
chosen level of significance a = 1 %. In this case, the
p value is 0.02 %. Therefore, we reject the null hypothesis, 1. Behdad M, Barone L, Bennamoun M, French T (2012) Nature-
inspired techniques in the context of fraud detection. IEEE Trans
and site 6 is marked as suspicious. Syst Man Cybern C 42:12731290
The proposed method successfully identifies the fraud- 2. Souza RP, Carmo LFRC, Pirmez L (2014) Deteccao de Dados
ulent sites, rejecting the null hypothesis for site 3 and site 6. Suspeitos de Fraude em Organismos de Inspecao Acreditados.

123
Accred Qual Assur

XIV Simposio Brasileiro em Seguranca da Informacao e de 11. Montgomery DC, Runger GC (2010) Applied statistics and
Sistemas Computacionais. http://www.sbseg2014.dcc.ufmg.br/ probability for engineers. Wiley, Massachusetts, USA. ISBN13:
files/anais.pdf. Accessed 6 Nov 2014 978-0-470-05304-1
3. de Souza RP, Carmo LFRC, Pirmez L (2016) A procedure to 12. Efron B (1979) Computers and the theory of statistics: thinking
detect suspected patterns of fraudulent behavior in vehicle the unthinkable. SIAM Rev 21:460480
emissions tests performed by an accredited inspection body. 13. Efron B, Tibshirani RJ (1993) An introduction to the bootstrap.
Accred Qual Assur 21:323. doi:10.1007/s00769-016-1231-y Monographs on statistics and applied probability, Vol. 57,
4. Manly BFJ (2006) Randomization, bootstrap and Monte Carlo Chapman & Hall, London, UK. ISBN-10: 0412042312
methods in biology, Chapman and Hall, New Zealand, USA 14. Wehrens R, Putter H, Buydens LMC (2000) The bootstrap: a
ISBN 978-0-412-36710-6 tutorial. Chemometr Intell Lab 54:3552
5. Taylor RN, McEntegart DJ, Stillman EC (2002) Statistical 15. Good P (1994) Permutation tests: a practical guide to resampling
techniques to detect fraud and other data irregularities in clinical methods for testing hypotheses. Springer, New York, USA. ISBN
questionnaire data. Drug Inf J 36:115125 978-1-4757-2348-9
6. Wu X, Carlsson M (2011) Detecting data fabrication in clinical 16. Kay RU (2007) Fundamentals of the DempsterShafer theory and
trials from cluster analysis perspective. Pharm Stat 10:257264 its applications to system safety and reliability modelling. Reliab
7. Kirkwood AA, Cox T, Hackshaw A (2013) Application of Theory Appl 3:173185
methods for central statistical monitoring in clinical trials. Clin 17. Croucher JS (2004) An upper bound on the value of the standard
Trials 10:783806 deviation. Teach Stat 26:5455
8. Ge Y, Xiong H, Liu C, Zhou Z (2011) A taxi driving fraud 18. Dong F, Shatz SM, Xu H (2009) Inference of online auction shills
detection system. In: IEEE 11th international conference on data using DempsterShafer theory. In: Sixth international conference
mining. doi:10.1109/ICDM.2011.18 on information technology, New Generations. doi:10.1109/ITNG.
9. Buyse M, George SL, Evans S, Geller NL, Ranstam J, Scherrer B, 2009.28
Lesaffre E, Murray G, Edler L, Hutton J, Colton T, Lachenbruch 19. Downey AB (2014) Think stats: probability and statistics for
P, Verma BL (1999) The role of biostatistics in the prevention, programmers. Green Tea Press, Massachusetts. ISBN-10:
detection and treatment of fraud in clinical trials. Stat Med 1449307116
18:34353451
10. Devore J (2012) Probability and statistics for engineering and the
sciences. Cengage Learning, Boston, USA. ISBN-10: 0-538-
73352-7

123

S-ar putea să vă placă și