Sensitivity Analysis in OLAP Databases
Emiel Caron and Hennie Daniels
Department of Management, Tilburg University, Warandelaan 2, Tilburg, The Netherlands
Keywords:
OLAP Databases, Business Analytics, Explanatory Analytics, Sensitivity Analysis, Decision-support Systems.
Abstract:
The theoretical underpinnings under which sensitivity analysis is valid in OLAP databases are dealt with in this
paper. Sensitivity analysis is considered to be the reverse of explanation generation in diagnostic reasoning.
Our exposition differentiates between sensitivity analysis in systems of purely drill-down equation and mixed
systems of equations with also business model equations. It is proven that there is an unique additive drill-
down measure defined on all cubes of the aggregation lattice. This proof is the basis for sensitivity analysis in
OLAP databases, where a change in some base cell in the lattice is propagated to all descendants in its upset.
For sensitivity analysis in mixed systems of equations a matrix notation is presented and the conditions for
solvability are discussed. Due to the fact that such systems are typically overdetermined in OLAP databases,
the implicit function theorem cannot be applied. Therefore, we proposed a method to reduce the number of
equations in the system and apply the implicit function theorem on a subsystem of the original system. We
conclude with an alternative method for what-if analysis in mixed systems of equations.
1 INTRODUCTION
The main goal of our research is “to extend the func-
tionality of multi-dimensional (or OLAP) business
databases with diagnostic capabilities to support man-
agerial decision-making” (Caron, 2013). In this pa-
per, the OLAP database is extended with function-
ality for sensitivity analysis. The purpose of the
methods and algorithms presented here, is to provide
OLAP databases with more powerful explanatory an-
alytics and reporting functions. In this paper, we de-
scribe how sensitivity analysis can be implemented in
a multi-dimensional database.
Sensitivity analysis in multi-dimensional
databases is related to the notion of comparative
statics in economics. Where the central issue is to
determine how changes in independent variables
affect dependent variables in an economic model.
Comparative statics is defined as the comparison of
two different equilibrium states solutions, before and
after a change in one of the independent variables,
keeping the other variables unchanged (Samuelson,
1941). It is one of the primary analytical methods
used in economics, where it is commonly used, for
example, in the study of changes in supply and de-
mand when analyzing a market. Instead of repeating
the phrase “keeping the other variables unchanged”,
economists use the more compact Latin equivalent
ceteris paribus (c.p.). The underlying model for
comparative statics is a set of equations that define
the vector of dependent variables y
1
,y
2
,...,y
m
as
functions of the vector of independent variables
x
1
,x
2
,...,x
n
, i.e.
y = f
l
(x),l = 1,2,...,m. (1)
On the one hand, this corresponds to a system of busi-
ness model equations between measures in an OLAP
database. Relations between measures are denoted by
y
i
(C) = f (x
i
(C)), (2)
where y and x = (x
1
,x
2
,...,x
n
) are measures on the
same cube C = [i
1
,i
2
,...,i
n
]. On the other , the ap-
plication of a specific aggregation function f , i.e.
SUM(), Count(), AVG(), etc., on the measure values
of each cube y(C) in L creates a system of drill-down
equations, given by
y
i
1
...i
q
...i
n
(C) = f (y
i
1
...(i
q
1)...i
n
(R
1
q
(C))). (3)
In Equation (2), the function f might be non-
linear, in Equation (3), the function f is linear for
SUM() and Count() aggregations, where R is a drill-
down operator in dimension q. In the latter situa-
tion we use the terms non-base variables and base
variables, for dependent and independent variables,
respectively. To implement sensitivity analysis in
OLAP, we define a new cube operator that supports
Caron, E. and Daniels, H.
Sensitivity Analysis in OLAP Databases.
DOI: 10.5220/0006791702210228
In Proceedings of the 20th International Conference on Enterprise Information Systems (ICEIS 2018), pages 221-228
ISBN: 978-989-758-298-1
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
221
the analyst in answering typical managerial what-if
questions, while navigating the cube. We distinguish
between two types of what-if questions:
Questions related to a system of drill-down equa-
tions. For example, “How is the profit in the year
2010 affected when the profit for a certain product
is changed with one percent in the first quarter in
The Netherlands, c.p.?”
Questions related to a system of business model
equations. For example, “How is the profit in the
year 2010 for a certain product affected when its
unit price is changed with one additional unit in
the sales model, c.p.?”
This paper is structured as follows. In Subsection
1.1 we discuss related work. In Section 2 we discuss
sensitivity analysis in systems that consist of purely
drill-down equations. In Section 3 we elaborate on
sensitivity analysis in systems that consist of purely
business model equations and mixed systems of equa-
tions. Finally, in Section 4 we draw some conclu-
sions.
1.1 Related Work
The variables, parameter values, and assumptions
of any business or economic model are subject to
change. Sensitivity analysis, generally defined, is
the investigation of these potential changes and their
impacts on conclusions to be drawn from the model
(e.g. (Baird, 1990)). There are many possible appli-
cations of sensitivity analysis, described here within
the categories of decision support, communication,
increased understanding or quantification of the sys-
tem, and model development (Pannell, 1997). There
is a very large literature on procedures and techniques
for sensitivity analysis (Clemson et al., 1995). Two
general classes of techniques for sensitivity analy-
sis are the implicit function theorem (Currier, 2000;
Heckman, 2000) and monotone comparative statics
(Milgrom and Shannon, 1994). These are methods
for characterizing whether an increase in a parame-
ter causes the dependent variable to increase or de-
crease. Historically the implicit function theorem was
used for this purpose and the implicit function theo-
rem not only tells you whether the dependent vari-
able increases or decreases but also the magnitude of
change. In contrast, monotone comparative statics
tells you only “up” or “down”, i.e., it gives an ordi-
nal rather than cardinal answer. In our research, we
focused solely on quantitative what-if analysis within
the multi-dimensional database.
To the best of our knowledge, (Balmin et al.,
2000) and (Lakshmanan et al., 2007) are the only pub-
lished research works that address sensitivity analysis
in OLAP databases in a significant way. In (Balmin
et al., 2000), the authors have developed the SESAME
system for the processing of hypothetical queries. For
this system query algebra operators are proposed that
are suitable for spreadsheet-style what-if computa-
tions. In the system hypothetical queries are mod-
eled as a list of hypothetical modifications on the data
in the fact table. A shortcoming of their approach is
that it lacks a good mathematical underpinning, to de-
cide whether a certain change is allowed in the model
or not, as opposed to our approach. In (Lakshmanan
et al., 2007), a different perspective is taken on what-
if analysis. They focus on what-if analysis related to
changes in dimensions and their hierarchical struc-
ture. However, our focus is on data-driven what-if
scenarios, as opposed to structural ones.
In many OLAP software products, sensitivity
analysis is not possible at the moment. If one wants
to do sensitivity analysis in these products one has
to copy the data to a reporting environment, for ex-
ample MS Excel, to compute manually the impact of
changes in certain cells of the data cube. An excep-
tion is the software product Clickview (Cliqview Cor-
poration, 2017), where a fixed change in a base vari-
able can be induced in a system of additive drill-down
measures, to determine its impact on non-base vari-
ables. The difference with our approach is that we can
induce variable changes in systems of additive and av-
erage drill-down measures and under certain condi-
tions in non-linear systems of business equations. For
this purpose we have designed a prototype application
for sensitivity analysis in MS Excel with Pivot tables,
with additional features implemented in Visual Basic.
2 SENSITIVITY ANALYSIS IN A
SYSTEM OF DRILL-DOWN
EQUATIONS
In this section we investigate the influence of a change
in a measure value of a cell in any OLAP cube, on a
higher level value of the same measure in the aggrega-
tion lattice. Or in formal notation, what is the effect
of changing y(c
0
) to y(c
0
) + δ on a dependent vari-
able y(c) in the upset of c
0
. To solve this consider the
lattice L
0
with top cube C
p
= [i
1
,i
2
,...,i
n
] and base
cube C
q
= [ j
1
, j
2
,..., j
n
]. Notice that L
0
is a sublat-
tice of L and L
0
= {↓ c} {↑ c
0
}. The values of the
measure y in the cube C
q
are denoted by x(c
0
i
), and
are called the base variables where i = 1,2,...,|C
q
|,
and the values of the measure y in {↑ C
q
} are de-
noted by y(c), and are called the non-base variables.
We distinguish between the original values of a mea-
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
222
sure without change x
r
(C
q
) and y
r
(C
p
), and the values
of the changed measure: x
a
(C
q
) and y
a
(C
p
), where
x
a
(C
q
) = x
r
(C
q
) except for one cell c
0
i
in the cube C
q
,
for which x
a
(c
0
i
) x
r
(c
0
i
) = δ.
The following theorem shows how the values of y
change in the lattice L
0
.
Theorem 1. There is an unique additive drill-
down measure y
a
(c) defined on all cube cells in the
sublattice L
0
such that:
y
a
(c) = y
r
(c) + β(c) · (x
a
(c
0
i
) x
r
(c
0
i
)), (4)
where:
β(c) = 1 if c {↑ c
0
i
}, and
β(c) = 0 if c / {↑ c
0
i
}.
Proof. To show that y
a
(c) is additive it is sufficient to
show that β(c) · (x
a
(c
0
i
) x
r
(c
0
i
)) is additive, because
the sum of additive measures is also additive and y
r
(c)
is additive by the consistency assumption. Hence, we
must show that:
β(c) =
q
β(R
1
q
(c)), (5)
where R
1
q
is the drill-down operation defined on a
cell c in the lattice L. Now there are two cases:
1. c {↑ c
0
i
}, i.e. c is an ancestor of c
0
i
. In that case c
0
i
is also a descendant of one of the cells in R
1
q
(c),
c
0
i
{↓ R
1
q
(c)}, which is a child of c in dimension
q. This property does not depend on dimension q.
So both sides of Equation (5) are equal to 1.
2. c / {↑ c
0
i
}, i.e. c is not an ancestor of c
0
i
. In that
case, c
0
i
is also not a descendant of one of the chil-
dren of c. Hence, both sides of Equation (5) are
zero.
Notice that the drill-down measure y
a
(c) is
unique. This follows from the general proposition that
every additive measure with given values on the base
cube is unique. This follows immediately from The-
orem on OLAP equations described in (Caron, 2013)
(page 40), and the fact that L
0
is a lattice of cubes.
In the case that c {↑ c
0
i
}, we can rewrite Equation
(4) as follows
y
a
(c) = y
r
(c) + inf(y
a
(c
0
),y
a
(c)). (6)
If y(c) is an additive drill-down measure then we use
inf(y
a;j
(c
0
),y
a;i
(c)) = y
a;j
(c
0
) y
r;j
(c
0
), (7)
for the computation of inf(y
a
(c
0
i
),y
a
(c)) in Equa-
tion (6) and if the variable x
r
(c
0
) is changed with
δ in sensitivity analysis then y
a
(c) is computed as
y
a
(c) = y
r
(c) + (x
a
(c
0
) x
r
(c
0
)). This result follows
immediately from the Theorem described in (Caron,
2013) (page 40).
Moreover, in the case that y
r
(c) is an average drill-
down measure we use specific influence measure for
the computation of inf(y
a
(c
0
i
),y
a
(c)) in Equation (6)
and if the variable x
r
(c
0
) is changed with δ in sen-
sitivity analysis then y
a
(c) is computed as y
a
(c) =
y
r
(c) +
1
|C
q
|
(x
a
(c
0
) x
r
(c
0
)), where C
q
is the context
cube under consideration. This result is not proven
here but the proof is similar to the proof of Theorem
described in (Caron, 2013) (page 40), with the differ-
ence that the RHS of the drill-down equation is di-
vided by the number of cells in the context cube.
Here we present a numeric example of a what-if
analysis in the cube C = Store × Products for the
measure sales, aggregated by the average function.
The data of the cube is depicted in Table 1. We want
to analyse a change δ in the cell (A,P
1
) on its upset
{↑ (A,P
1
)}. The reference value of the cell is given
by sales
r
(A,P
1
) = 1 and the actual value is given
by sales
a
(A,P
1
) = 1 + δ. By applying Equation (6)
we compute the effect of this change on {↑ (A,P
1
)};
these effects are given by,
sales
a
(All,P
1
) = sales
r
(All,P
1
) +
1
3
δ where |R
+1
Stores
(C)| = 3,
sales
a
(A,All) = sales
r
(A,All)+
1
4
δ where |R
+1
Products
(C)| = 4,
sales
a
(All,All) = sales
r
(All,All)+
1
12
δ where |C| = 12.
For example, here R
+1
Stores
(C), represent a roll-up (+1)
in the Stores’s dimension hierarchy, and |R
+1
Stores
(C)| =
3 represents the number of stores.
Table 1: Sensitivity analysis in the example cube Store ×
Products for the average drill-down measure sales. Here the
value of the cell (A, P
1
) is changed with δ and this change
is propagated in the cell’s upset.
AVG(sales) Stores
Products
A B C All
P
1
1 + δ 2 3 2 +
1
3
δ
P
2
4 5 6 5
P
3
7 8 9 8
P
4
10 11 12 11
All 5.5 +
1
4
δ 6.5 7.5 6.5 +
1
12
δ
The subsystem of drill-down equations that cor-
responds with {↑ c
0
} has an unique solution, after a
change in y(c
0
) with some δ, as a result of the theo-
rem (Caron, 2013) . However, the complete system of
equations becomes inconsistent because Equation (3)
does not hold in that case:
y
max
1
max
2
...max
n
(c) + δ(c
0
) 6=
c
n
R
max
n
n
...R
max
2
2
R
max
1
1
(c)
y
00...0
(c
n
).
In other words, when the change in what-if analysis
is not induced by a variable in the base cube, but by a
Sensitivity Analysis in OLAP Databases
223
(non-base) variable on some intermediate level in the
lattice L, the complete system of equations will be-
come inconsistent. For analysis restricted to {↑ c
0
}
this does not matter, however analysis in the com-
plete system is obviously not useful anymore. The
inconsistencies in the complete system of drill-down
equations, can be corrected by a straightforward pro-
cedure, that repairs the OLAP database (Caron and
Daniels, 2008).
3 SENSITIVITY ANALYSIS IN A
SYSTEM OF BUSINESS
EQUATIONS
In this section we discuss managerial what-if ques-
tions related to a system of business model equations
and a mixed system of drill-down and business model
equations.
Multiple related measures in the business model
and associated dimensions, result in a mixed, often
non-linear, system of equations. For example, con-
sider Table 2 with the equations of Figure 1. In 13
these equations are given in shorthand notation. The
system of equations in (13) are represented as a graph
in Figure 2. In this system we want to change an in-
dependent variable x
i
, e.g. x
4
(= Volume(2005.Q2))
and/or x
5
(= Unit Price(2005.Q2)), and study the im-
pact on its upset, in particular, the dependent root
variable y
1
(= Revenues(2005)). Notice that (13) is
overdetermined, because we have 4 independent vari-
ables and 5 equations.
In general, for a mixed system of equations
the equations are linear and non-linear, and
the system of equations is overdetermined.
Similarly, a system of solely drill-down equations is
also overdetermined in the case of multiple dimen-
sions. Equation (13) can be written as
f
l
(y,x) = 0. (8)
The linearization of (8) in a neighborhood of a solu-
tion (y
0
,x
0
) reads:
A
1
y + A
2
x = 0. (9)
The matrix A
1
is the l × m coefficient submatrix for
dependent variables and A
2
is the l ×n coefficient sub-
matrix for independent variables. Here the matrix of
the first derivatives of f with respect to y is represented
by A
1
= D
y
f(y,x) and the matrix of first derivatives of
f with respect to x is represented by A
2
= D
x
f(y,x).
With (9) we can examine the impact of a change in
one or more independent variables c.p., given by x,
on the dependent variables, given by y, where equa-
tion (8) has to be satisfied. In the next section, we in-
vestigate the conditions for consistency and solvabil-
ity of (9), which is a necessary condition for solvabil-
ity of (8).
3.1 Conditions for Solvability
A necessary condition for solvability in a system of
linear equations is the rank criterium. A system of
linear equations (9), of A
1
y + A
2
x = 0, is solvable if
and only if rank(A
1
| A
2
x) = rank(A
1
). The proof of
this theorem is, for example, given in (Schott, 1997).
In words, the rank criterium says that the vector A
2
x
must be in the column space (range) of A
1
for the sys-
tem to be solvable.
To investigate the solvability of (8), we assume
that
(y
0
,x
0
) = (y
0
1
,y
0
2
,...,y
0
m
,x
0
1
,x
0
2
,...,x
0
n
)
is a solution of (8). We substitute this solution in
the derivative matrices A
1
and A
2
to obtain the lin-
earized matrix [A
1
A
2
] at the solution (y
0
,x
0
). The
linearized system of equations A
1
y + A
2
x = 0 is
solvable if and only if rank(A
1
) = rank(A
1
| A
2
x).
Similarly, the linearized system of equations is solv-
able for an independent variable x
i
, if and only if,
rank(A
1
) = rank(A
1
|column x
i
from A
2
). A column
vector x
i
of the submatrix A
2
is represented by a
2
(i).
Accordingly, the rank criterium can be used to deter-
mine whether an independent variable x
i
qualifies for
what-if analysis in a system of business model equa-
tions. However, in the next section it is shown, that
this criterium is a necessary but not sufficient condi-
tion for the solvability of a non-linear system of equa-
tions.
When the submatrix A
1
is nonsingular then the so-
lution of A
1
y + A
2
x = 0 is unique and given by
y = A
1
1
A
2
x.
Notice that the rank criterium is a necessary but not
sufficient condition for the solvability of a non-linear
system of equations. Practically, this means that in
such models the number of equations must be equal
to the number of dependent variables to produce a
square submatrix A
1
(l = m).
Now suppose that we are given an overdetermined
system of equations as in (8) and a solution (y
0
,x
0
) to
this system such that all the equations are satisfied.
The first derivatives of the equations can be written
in matrix form as in (9). If the rank criterium for
consistency holds for a certain independent variable
x
i
, considered for what-if analysis, then the solution
f(y
0
,x
0
) = 0 is filled in Equation (9). Subsequently,
α
1
· eq. 1 + α
2
· eq. 2 + . . . + α
l
· eq. l = 0, (10)
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
224
holds if all the α
i
s exist. If the α
i
s exist we re-
move (l m) dependent equations from the system of
equations and derive a (m × m) submatrix A
1
. If the
remaining system of equations in A
1
is nonsingular
the implicit function theorem can be applied and the
α
i
s determined. In that case the removed equations
are satisfied too, because Equation (10) holds and the
general solution for x
i
can be determined.
3.2 What-if Analysis Example
In this example we want to change an independent
variable x
i
and study the impact on elements in its up-
set. The Jacobian of the system of equations in (13)
is given in (14). Observe that the vector
(y
0
x
0
) = (48 16 15 3.2|13 12 7 4 4 6 3 2 2.75 3 3.25),
is a solution to the system of equations. The Ja-
cobian at (x
0
,y
0
) is given in (15) The rank cri-
terium for solvability in this system is satisfied for
the variables x
4
(= Volume(2005.Q2)) and x
5
(= Unit
Price(2005.Q2)): rank(A
1
|a
2(4)
) = rank(A
1
) = 4 and
rank(A
1
|a
2(5)
) =rank(A
1
) = 4. It can easily be veri-
fied that the rank criterium is not satisfied for the other
independent variables. For example, for variable x
1
it can be concluded that rank(A
1
|a
2(1)
) > rank(A
1
).
Therefore, the only candidate independent variables
for what-if analysis in this example are x
4
and x
5
.
As we saw, the rank criterium is a necessary but
not sufficient condition for solvability. We cannot ap-
ply the implicit function theorem to verify solvability
here, because the submatrix A
1
is non-square (5 × 4).
But in this case we may eliminate one of the equations
because we can find α
i
such that:
α
1
· eq. 1 + α
2
· eq. 2 + α
3
· eq. 3+
α
4
· eq. 4 + α
5
· eq. 5 = 0.
(11)
These α
i
s are given by
α
1
α
2
α
3
α
4
α
5
=
1
1
1
0
y
3
.
Now we proceed as follows. In the system of equa-
tions in (13) all independent variables are replaced
by the solution (y
0
,x
0
) except the independent vari-
ables x
4
and x
5
, that are under consideration for what-
if analysis. From the original system of equations,
one dependent equation is removed and we derive a
reduced system of equations, where the matrix A
1
is
square. Removing eq. 2 yields
f (y
1
,y
2
,y
3
,y
4
,x
4
,x
5
) =
y
1
+ 32 + y
2
= 0
y
2
+ x
4
x
5
= 0
y
3
+ 11 + x
4
= 0
y
4
+ (32 + x
4
x
5
)/y
3
= 0.
(12)
(y
0
,x
0
) = (48,16,15,3.2,4,4) is a solution of (12).
The 4 × 4 derivative submatrix A
1
of f with respect to
y in (48, 16, 15, 3.2, 4, 4) is
D
y
f(48,16,15,3.2,4,4) =
1 1 0 0
0 1 0 0
0 0 1 0
0 0
48
225
1
= A
1
.
It can easily be verified that
A
1
A
1
1
=
1 1 0 0
0 1 0 0
0 0 1 0
0 0
48
225
1
1 1 0 0
0 1 0 0
0 0 1 0
0 0
48
225
1
= I
4
.
By the implicit function theorem we can find con-
tinuous differentiable functions ϕ
i
(x
4
,x
5
) : B R,
where B = B
r
(48,16,15,3.2,4,4), such that
y
1
= ϕ
1
(x
4
,x
5
)
y
2
= ϕ
2
(x
4
,x
5
)
y
3
= ϕ
3
(x
4
,x
5
)
y
4
= ϕ
4
(x
4
,x
5
),
is a solution of the system of equations (12). More-
over, also the removed equation y
1
+ y
3
y
4
= 0 (eq.
2) is satisfied because of (11). Computation gives:
y
1
= 32 + x
4
x
5
y
1
= (11 + x
4
)(
32+x
4
x
5
11+x
4
) = 32 + x
4
x
5
y
2
= x
4
x
5
y
3
= 11 + x
4
y
4
=
32+x
4
x
5
11+x
4
.
4 CONCLUSIONS
In this paper, we stated the theoretical underpinnings
under which sensitivity analysis is allowed in multi-
dimensional databases. We also discussed some the-
oretical issues and procedures related to sensitivity
analysis in OLAP databases.
For sensitivity analysis in systems of additive
drill-down measures we proved Theorem 1, and
Sensitivity Analysis in OLAP Databases
225
showed that there is an unique additive drill-down
measure y
a
(c) defined on all cubes of the aggregation
lattice. This theorem is the basis for sensitivity analy-
sis here, where a change in some base cell in the lat-
tice is propagated to all descendants in its upset. For
the average drill-down measure a similar expression
is determined. Moreover, sensitivity analysis might
cause the multi-dimensional database to become cor-
rupted, if the analysis is not carried out on cells in the
base cube. To overcome this problem we proposed a
correction procedure.
For sensitivity analysis in mixed systems of equa-
tions we introduced a matrix notation and we dis-
cussed the conditions for solvability. Because mixed
systems are typically overdetermined the implicit
function theorem cannot be applied. Therefore, we
proposed a method to reduce the number of equations
in the system and apply the implicit function theorem
on a subsystem.
REFERENCES
Baird, B. (1990). Managerial Decisions Under Uncer-
tainty: An Introduction to the Analysis of Decision
Making. John Wiley & Sons, Inc., New York.
Balmin, A., Papakonstantinou, Y., and Papadimitriou, T.
(2000). Optimization of hypothetical queries in an
olap environment. Data Engineering, International
Conference on, 0:311.
Caron, E. (2013). Explanation of Exceptional Values in
Multi-dimensional Business Databases. PhD thesis,
Erasmus University.
Caron, E. A. M. and Daniels, H. A. M. (2008). Ex-
tensions to the olap framework for business analy-
sis. In Cordeiro, J., Shishkov, B., Ranchordas, A.,
and Helfert, M., editors, ICSOFT (ISDM/ABF), pages
240–247. INSTICC Press.
Clemson, B., Yongming, T., Pyne, J., and Unal, R. (1995).
Efficient methods for sensitivity analysis. System Dy-
namics Review, 11(1):31–49.
Cliqview Corporation (2017). Cliqview.
Currier, K. (2000). Comparative Statics Analysis in Eco-
nomics. World Scientific Publishing Co, Singapore.
Heckman, J. (2000). Causal parameters and policy analysis
in economics: A twentieth century retrospective. The
Quarterly Journal of Economics, 115(1):45–97.
IBM Cognos Software (2017). Ibm cognos business intelli-
gence, powerplay.
Lakshmanan, L., Russakovsky, A., and Sashikanth, V.
(2007). What if olap queries with changing dimen-
sions perspectives are everything. In VLDB ’07: Pro-
ceedings of the 33th International Conference on Very
Large Data Bases, San Francisco, CA, USA. Morgan
Kaufmann Publishers Inc.
Milgrom, P. and Shannon, C. (1994). Monotone compara-
tive statics. Econometrica, 62(1):157–80.
Pannell, D. J. (1997). Sensitivity analysis of normative eco-
nomic models: theoretical framework and practical
strategies. Agricultural Economics, 16(2):139–152.
Samuelson, P. A. (1941). The stability of equilibrium:
Comparative statics and dynamics. Econometrica,
9(2):97–120.
Schott, J. (1997). Matrix Analysis for Statistics. Wiley, New
York.
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
226
APPENDIX
A star model representing a multi-dimensional financial database is shown in Figure 1 and is used as an illustra-
tive example in this paper. This database, called GoSales, contains the financial figures from a generic fictitious
company that sells sports equipment, obtained from the Cognos OLAP product PowerPlay (IBM Cognos Soft-
ware, 2017). Figure 1 depicts a central fact table and five dimensions tables. The central fact table represents the
Figure 1: Star model with five dimension tables and a central fact table representing the financial data set.
financial data set. It lists the measures of the data, like profit, revenues, costs, etc. The financial data set has ve
dimensions tables: Time (T), Product (P), Location (L), Customer (C), and Vendor (V), and all dimensions have
a 2-4 level hierarchy.
Table 2: Subsystem of business model and drill-down equations derived from a multi-dimensional financial database.
1. Rev.(2005) = Rev.(2005.Q1) + Rev.(2005.Q2) + Rev.(2005.Q3) + Rev.(2005.Q4)
2. Rev.(2005) = Vol.(2005) × Unit Pr.(2005)
3. Rev.(2005.Q2) = Vol.(2005.Q2) × Unit Pr.(2005.Q2)
4. Vol.(2005) = Vol.(2005.Q1) + Vol.(2005.Q2) + Vol.(2005.Q3) + Vol.(2005.Q4)
5. Unit Pr.(2005) = ((Vol.(*.Q1) × Unit Pr.(*.Q1)) + (Vol.(*.Q2) × Unit Pr.(*.Q2)) +
(Vol.(*.Q3) × Unit Pr.(*.Q3)) + (Vol.(*.Q4) × Unit Pr.(*.Q4))) / Unit Pr.(*)
Table 2 in shorthand notation
y
1
+ x
1
+ y
2
+ x
2
+ x
3
= 0
y
1
+ y
3
× y
4
= 0
y
2
+ x
4
× x
5
= 0
y
3
+ x
6
+ x
4
+ x
7
+ x
8
= 0
y
4
+ ((x
6
× x
9
) + (x
4
× x
5
) + (x
7
× x
10
) + (x
8
× x
11
))/y
3
= 0,
(13)
where y
i
with i = 1, 2, 3, 4 are the dependent variables and x
i
with i = 1, 2, . . . , 11 are the independent variables.
Sensitivity Analysis in OLAP Databases
227
Figure 2: Graph representation of the implicit system of equations.
A = [A
1
A
2
] =
1 1 0 0 1 1 1 0 0 0 0 0 0 0 0
1 0 y
4
y
3
0 0 0 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 x
5
x
4
0 0 0 0 0 0
0 0 1 0 0 0 0 1 0 1 1 1 0 0 0
0 0 )
1 0 0 0
x
5
y
3
x
4
y
3
x
9
y
3
x
10
y
3
x
11
y
3
x
6
y
3
x
7
y
3
x
8
y
3
.
)
=
x
6
x
9
+x
4
x
5
+x
7
x
10
+x
8
x
11
(y
3
)
2
(14)
A
0
= [A
1
A
2
] =
1 1 0 0 1 1 1 0 0 0 0 0 0 0 0
1 0 3.2 15 0 0 0 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 4 4 0 0 0 0 0 0
0 0 1 0 0 0 0 1 0 1 1 1 0 0 0
0 0
48
225
1 0 0 0
4
15
4
15
2.75
15
3
15
3.25
15
6
15
3
15
2
15
.
(15)
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
228