Efficiency of Banks in United Arab Emirates Masters in Corporate and Financial Risk Management School of Mathematical and Physical Sciences Sussex University - England Mohieb AbuZant 2015
Efficiency of Banks in United Arab Emirates
Masters in Corporate and Financial Risk
Management
School of Mathematical and Physical Sciences
Sussex University - England
Mohieb AbuZant
2015
ACKNOWLEDGEMENT
The MSc in Corporate and Financial Risk Management program at the University of Sussex has been
one of the most beneficial experiences throughout my life. Without the support of the below
mentioned people and many other people that I have not mentioned, this dissertation may not have
materialized and may not have been possible.
First of all, I would like to thank my parents. Without them, I will not be able to write this dissertation
now. They encouraged and supported me a lot during the time that I have stayed in the Sussex
University. My sister and brother were also very supportive while studying. Without their
encouragement I don’t think I can finish my degree successfully.
I would like to express my gratitude to my Course Convener and official supervisor Dr. Qi Tang for
his guidance, caring, patience and providing me with a good atmosphere for doing the dissertation.
I profoundly thank my unconditional advisor Dr. Malgorzata Sulimierska for her supervision and
guidance. Thanks for your help whenever I needed you.
I would like also to deeply thank all my tutors and classmates in my course which is MSc Corporate
and Financial Risk Management. They helped and provided me with a lot of knowledge during this
year.
Finally, my gratitude also goes to the School of Mathematical and Physical Sciences.
Abstract
The present thesis pursues a research into the hypothetical differentials in the efficiency of
UAE banking as driven by use of either Islamic or conventional financial models. The
foremost difficulty rests in defining efficiency as opposed to productivity or profit
maximization, along with its characteristically Islamic counterparts, deliverables, or targets. It
has been detected that return on equity (ROE) essentially captures all of the critical facets of
efficiency, when it comes to striking a careful balance between financial and operating
leverage versus excessive risk to be avoided. With the aid of prior formalization, it has been
demonstrated how a dummy based OLS model could parsimoniously outperform the more
advanced regression designs such as the GLS or logit. The panel based model lends strong
support to its prior cross-sectional reduction, with both pointing to but limited linear
significance of the channel whereby dividend or reserves policies affect ROE efficiency. This
could either hint at nonlinear mechanisms that might in turn conceal the inherent real-world
complexity of decision making along these lines or otherwise warrant further tests of Granger
causality. Somewhat surprisingly, both tests, along with the posterior CHI-squared and t-
statistics on individual fixed effects, point to their being no material performance gaps
between the two models. In risk or variance adjusted terms, however, Islamic banking appears
to outmatch conventional models, which net-of-uncertainty type deliverables constitute the
core of Islamic efficiency in the first place.
Key words: Islamic banking, gharar, ROE, efficiency, OLS, GLS, panel data, dummy
variables.
TABLE OF CONTENTS
Acknowledgements
Abstract
Introduction
Chapter 1 : Literature Review
Chapter 2 : Methodology
2.1 Background
2.2 Data
2.3 Research Objectives
2.4 Research Questions
2.5 Economical Assumptions
2.6 Analysis of Financial Ratios
2.7 Hypothesis testing & OLS Modeling
2.8 Generalized ANOVA
2.9 Afterthoughts on Triangulation & Bootstrapping
Chapter 3 : Modeling & Testing
3.1 Qualitative Analysis
3.2 Cross-Sectional Regression
3.3 Heterokedasticity, Autocorrelation & GLS
Chapter 4 : Auxiliary Modeling
4.1 Caveats & Extensions
4.2 An Auxiliary Formalization
4.3 An Augmented Model
4.4 Caveats on Data
4.5 Panel Estimates
Chapter 5 : Empirical Results
5.1 Prior Estimation
5.2 Two-Stage Inference on Fixed Effects
Conclusion
Introduction
The UAE has 56 banks in total. 28 of them are foreign banks, 23 are local banks & the 5 rest
banks are investment banks (Central Bank official website). The huge number of commercial
banks, the high density of the branches, the technological change, and the increasingly
competition in the banking industry, which is one of the major characteristics of the banking
industry in the United Arab Emirates have placed tremendous pressure to improve
performance. Therefore, it is interesting to investigate the efficiency of the UAE banks where
the greater the efficacy the higher the effectiveness and vice versa. (Spathis et al., 2002).
Among the Gulf Cooperation Council (GCC) countries, the United Arab Emirates is historically and
still is one of the strongest economies. In 2013 it was ranked second among the GCC countries in
terms of GDP per capita with a GDP per capita at current prices of USD 43,048. This figure shows an
impressive growth over the decade as in 2002 this figure was only USD 18,903. The UAE GDP per
capita is also second to Qatar which had an impressive GDP in 2013 per capita of USD 93,714 and is
followed by Kingdom of Saudi Arabia with a GDP per capita of USD 25,961 and all other GCC
countries falling within the range of USD 22-25,000. (IMF, 2012)
The Conventional Bank theories assume that the main way for conventional banks to earn
profit is the difference between interest rates. Where the bank purchase transaction deposits
from depositors at low interest rates & resell those funds to borrowers at higher interest rates.
On the other hand, Islamic Banks based on the Islamic rules doesn’t allow interest rates, Islam
does allow for a number of financial mechanisms which do allow for the bank to profit. The
Islamic banking principle differs from the basic commercial banking theory that it doesn’t
allow interest rates so it replaces it with theories based on risk sharing which depends on
trading rather than risk transferring which is usually used in conventional banks. They use
different concepts such as Profit sharing (Modaraba), Joint venture (Mosharakah)
Safekeeping (Wadea), Cost plus (Morabaha) and leasing (Ejar).
Chapter 1: Literature Review
Existing studies in this area show different results and opinions about the efficiency of both
Islamic and Conventional Banks.
Taufiq Hassan et al. (2009) who studied the efficiency of the Islamic Banks VS
Conventional Banks under international basis collected data for each year from BankScope
database over the period 1990-2005. They evaluated in several countries data taken from the
financial statements of 80 Banks from 21 Islamic countries. They used the Data envelopment
analysis non-parametric approach to come out with the results for their work. The results
showed that there was no significant main difference in the revenues, costs and profit between
the Islamic and conventional banks. The second argument shows that the results may change
if there will be a comparison between two different groups. One taken from less developed
countries and the other taken from more developed countries due to the different economic
environment for both. Hence, proof on the consistency of the results in their study is offered
which reports insignificant differences in cost, revenue and profit mean scores between
conventional and Islamic banks. The authors here will affect my work by the measures of
efficiency of Islamic and Conventional banks. But here they are using a cross country
equation for more than a country, while I will be focusing on measuring the efficiency in
banks for the United Arab Emirates only.
Hussein A. Hassan Al-Tamimi & Husni Charif (2011) who assessed the performance
factors in the UAE commercial banks by using multiple approaches taking into consideration
the effect of the bank size. collected the data from the annual reports of Emirates Banks
Association and then used the methodology of analyzing ratios such as: (ROA) Return on
Assets, (ROE) Return on Equity, (MARG) the ratio of net interest margin to current assets,
(LODE) the ratio of loans to deposits, (CALO) the ratio of current assets to loans , (TAEQ)
the ratio of total assets to equity and (EQTA) the ratio of equity to total assets . The results
show that large banks perform better than small banks. It also provides support to the
proposition that in an environment where banks are highly fragmented, mostly small, and are
not performing well, there is a need to consolidate the operations of these banks.
This will be will affecting my dissertation by measures of performance in the same country
which is the United Arab Emirates, I will be using some of the same ratios but for different
period.
Thorsten Beck et al (2012) Have studied the Islamic vs. Conventional banks efficiency and
stability between the years 1995-2009. Then who collected data from Bank scope. Which
provides us with information about listed & non-listed banks. To construct and compare
indicators of efficiency, asset quality, business orientation and stability of both the Islamic
banks and conventional banks. The results show some difference in the business orientation
between conventional banks and Islamic banks. Islamic banks were less cost-effective, had
higher asset quality and were capitalized much better. They also found a large cross-country
variation between them, better stock performance of listed Islamic banks during financial
crisis is also due to their higher capitalized and better asset quality.
In this dissertation, they focus on the periods before the financial crisis and during the
financial crisis. While I will be using data from year 2006 till 2014, which will study the
efficiency and effectiveness of Islamic vs Conventional banks before, during and after the
Global Financial Crisis of 2008.
Fatima S. Al Shamsi et al (2007) measured and explained the efficiencies of the United Arab
Emirates banking system, They analyzed the data using the DEA ( Data envelop analysis)
parametric and non-parametric approaches to measure efficiency . The data was collected
from Bank Scope by collecting data from balance sheets and income statements from most of
the commercial banks, operating in the UAE for the last 5 years. The three major inputs
targeted by this study are labor, capital and deposits and the two outputs are loans and
investments.
This dissertation measures the efficiencies of the United Arab Emirates for only 5 years,
while I will be extending this period to be more accurate about the results by using a different
methodology.
Hussein A. Hassan Al-Tamimi and Faris Mohammed Al-Mazrooei (2007) in their paper
entitled “Banks’ risk management: a comparison study of UAE national and foreign banks”
The writers created a modified questionnaire, and then divided it into two parts, the first part
covered six aspects: understanding risk and risk management, risk assessment & analysis,
identification of risk, Monitoring risk, The authors developed a modified questionnaire.
Divided into two parts. The; risk monitoring; Practices of risk management and the analysis
of credit risk. This part includes 43 closed-ended questions based on an interval scale. The
second part consists of two closed-ended questions based on an ordinal scale dealing with two
topics: methods of risk identification, and risks facing the sample banks, in their article they
answered a lot of questions such as:
RQ1. Do the UAE banks’ staff understand risk and risk management?
RQ2. Have the UAE banks clearly identified the potential risks relating to each of their
declared aims and objectives?
RQ3. Do the UAE banks efficiently asses and analyze risk in general?
And many other questions using the methodology of evaluating the reliability of the scales
using Cronbach’s alpha, which measures the consistency with which respondents answer
questions within a scale, then regression analysis and one-way ANOVA were run to test the
research hypotheses.
Their results show that a lot of risks were facing the UAE commercial & and foreign banks,
the most important types of risk facing the country were:
1- Foreign exchange risk.
2- Credit risk.
3- Operating risk.
This dissertation is really helpful and will be affecting my work since is studies and compares
the differences between local and foreign banks which are 23 local banks and 28 foreign
banks in the county but using a different methodology.
Bashir M. et al (1999) results show that the Islamic banks’ efficiency is much better than the
conventional banks. In their study they used a set of data between the years 1990-2005 then
used the (DEA) Data Envelopment analysis to test the revenue, comparative cost & profit
efficiency for both Islamic & Conventional banks.
In this dissertation, they compare between Islamic and conventional banks efficiency during a
long period using a different methodology as well.
S. KABLAN and O. YOUSFI (2011) who studied the efficiency of Islamic and Conventional
banking of 17 Islamic countries analyzed the Islamic banks efficiency between the years
2001-2008 and found that the banks were efficient at 78.9%. Their study displayed that Asia
had the highest score of efficiency with 84.64%. They found out also that the financial crisis
didn’t affect the Islamic banks as the results showed by the dummy variable. Also the
profitability and market power had a negative effect on the Islamic banking efficiency. In
their study they used the parametric approach and specify more the stochastic frontier analysis
(SFA). The objective function was the cost function. It allows to take into consideration banks
as financial companies and seek to evaluate their performance. Thereby minimize the costs
induced by the efficiency frontier. Then chose the Translog, as it best suits the multi-products
characteristic of banking technology, multiple inputs and outputs to be involved. When they
estimated the efficiency frontier they had to choose the inputs and outputs produced and used
by Islamic banks.
Chapter 2: Methodology
2.1 Background
The present study aims at comparing the efficiency-side performance of a set of banks
operating in the UAE jurisdiction as either Islamic or legacy-neutral entities. The two
alternate models could map into very different sets of implications likely to prove testable
along the regular econometric lines of ANOVA/ANCOVA inference versus ordinary-least
squares (OLS) modeling. The bulk of critical performance checks could be reduced to the
conventional investment ratios such as return on assets (ROA) or equity (ROE), which values
may swing around their historic averages while varying against the industry average or some
kind of alternative benchmark. Importantly, all of the relevant industry hurdles will be defined
from within the sample on hand, which is supposed to feature a reasonably complete set of
legacy trade-offs as shown in Table 1 below.
Table 1: Legacy Trade-Offs
Islamic Non-Islamic
UAE Non-UAE
It is important to know that some of the otherwise UAE headquartered banks could fall
under either legacy, Islamic or neutral, with the National Bank of Abu Dhabi being one
apparent case. Apart from the nearly corner outcomes (e.g. Abu Dhabi Islamic Bank=“UAE
and Islamic” or BNP Paribas=“non-UAE and non-Islamic”), of interest could be some interim
identities capturing players that are UAE yet non-Islamic or non-UAE and possibly outright
Western-headquartered yet characteristically Islamic-type.
In addition, one should distinguish between those faring as UAE nationals as opposed
to the operators falling under overseas jurisdictions yet working in UAE. For one thing, far
from all of the subsidiaries as headquartered in the UAE will have worked within their
homeland markets only. Conversely, many of the foreign banks will boast a strong or
emerging UAE presence amid this marketplace growing increasingly global. Although it
would be fully feasible technically as well as conceptually to add one extra dimension on to
the above identity matrix, visualizing the cubic or higher-dimensional construct might turn
out to be a rather involved enterprise. One way around the issue could be to view two
complementary matrices as stand-alone layers of analysis, with “UAE versus non-UAE” as
above referring to either the origin or the destination type jurisdiction. For now, this facet will
be assumed away for simplicity’s sake, even though in actuality, some further refinement will
be warranted to qualify the relevant scope as well as apparatus. The flipside or ‘silver lining’
is that the proposed set of identity trade-offs will be deployed on many a level for a variety of
specific tests and models.
When it comes to substantive underpinning of this taxonomy, any Islamic legacy
would involve a very distinct incentives structure likely to have affected the efficiency
outcomes in a systematic manner. Dwelling on these gaps could afford some prior analysis
beyond ad-hoc data mining with an eye toward rendering the candidate regressions or
ANOVA insights less of an atheoretic attempt.
Shari’a compliance, as long as it involves in-depth alignment rather than sporadic or
makeshift fatwa-based license, will target investment patterns that are truly productive as well
as equitably rewarding. Excessive risk is to be avoided, which also holds for ill-gotten
windfalls and unreasonable or speculative gains. In fact, this could formally be seen as
coming in line with generic optimization, in particular based on the Capital Asset Pricing
Model (CAPM). After all, even the risk seeking types are not supposed to buy into just any
risk or uncertainty extent, so long as less risky assets (or strategies as well as portfolios) are
conceivable as per the exact same payoff level. On the other hand, it is postulated under
strong-form market efficiency that there are limits to diversification or strategy building,
known as systemic or market risk, beyond which profitability will have to be compromised.
It is this grand trade-off between risk and expected return, or cost versus benefits, that
will be kept in mind throughout whenever judging on efficiency. In terms of Pareto
optimization or indeed CAPM, securing efficiency is about keeping the cost low as per any
given level of profitability. More specifically, a high ROE will implicitly refer to efficiency,
insofar as its embedded profitability versus leverage or gearing dimensions are distinct yet
inseparable. It will, inter alia, be demonstrated how efficiency implications may have to be
qualified for the Islamic representation, with excessive gearing at times referring to inherently
more cooperative securitization or scale sharing across banks or throughout the vertical value
chain.
2.2 Data
A total of 46 banks will constitute the cross-sectional sample drawing upon their
annual reports over the 2006 through 2014 time period. The respective number of
observations, suggesting a pooled or panel type dataset, affords 46*9=414 as per each metric
under study. In other words, regardless of the effective number of explanatory variables
acting to detract from the resultant degrees of freedom (DF), this value could be augmented
by multiplying the above number times that of the independent variables:
DF=46∗9∗N−N=413∗N
At that rate, any desired DF level can be ensured, albeit at the cost of mounting
complexity. Alternatively, as long as the structural shifts in the historical patterns, or fixed
and random effects, are less relevant, the full-blown scale could be collapsed to its cross-
sectional domain. For instance, average values could be obtained and plugged in place of the
time paths for each bank’s respective metric. Needless to say, though, the validity as well as
robustness of out-of-sample inference, be it in terms of posterior predictive power or prior
data quality with an eye on heteroscedasticity checks, cannot remain intact—which features
one other, meta-level efficiency trade-off. The good news is that a host of interim data tests,
such as serial or auto-correlation as well as endogeneity, can be waived. In any event, the
effective DF will be restricted subject to data availability, as the earlier reports or financial
statements might not be readily available on some of the banks.
Unless explicitly referred otherwise, the bulk of financial reports will be credited to a
standardized database implicitly restraining the sample to what amounts to a ‘population’
(UAE Central Bank 2015).
2.3 Research Objectives
The present thesis looks into whether banks that qualify as largely Islamic reveal
efficiency performance that is either superior quantitatively or downright distinct qualitatively
(substantively as well as structurally). In addition, approaches will be proposed to testing for
validity of Islamic self-identity beyond exogenous fatwas or claims to halal, adab, or mubah
compliance. Regardless of how far gharar is actually shunned or collective interests targeted
as well as maintained, a set of unbiased criteria could be invoked as part of an elegant yet
parsimonious optimization. Finally, one interim outcome would assess the feasibility of
mixed strategies or weak identities.
2.4 Research Questions
RQ1: Are efficiency gaps material between Islamic versus conventional banks that are
somehow associated with the UAE jurisdiction?
RQ2: Can these gaps be explained away in terms of the underlying legacy, with
qualitative analysis informing as well as qualifying the rigorous modeling?
2.5 Economical Assumptions
Among other things, the scope of efficiency checks can be reduced or contracted
bearing in mind the industry profile. For instance, gharar minimization could for banks boil
down to capital adequacy requirements as well as maturity gap management. However, the
former could be presumed as excessive in the Western legacy compared to Islamic setups,
given the nature of reserves or equity never acting as cushion against bank runs. It should
therefore come as little surprise if the respective equity proportions actually prove lower for
Islamic banks. For the same token, safe or minimalist gap management can be maintained for
Islamic banks inasmuch as they act largely as regular, non-financial type investment or
production networks—even if not engaged in the operations management directly.
Effectively, these areas of efficiency need not be tested in ways other than scale efficiency
assessment. This will in turn amount to the analysis of total as well as current asset
performance.
2.6 Analysis of Financial Ratios
The choice of ROA and ROE as representative of the relevant efficiency scope could
be motivated in a number of ways. Although these, first and foremost, capture the profitability
of assets versus capital, respectively, still they are so inherently intertwined that all of the
aforementioned facets of efficiency could be obtained as derivative parameters. Notably, both
ratios apply to real-sector and financial-type going concerns alike, thus amounting to a
transferable robustness bridge. Not least, the market stays largely unmoved as to whichever
asset types boast comparable performance that is balanced as well as sustainable—unless, of
course, the market is dominated by Islamic-profile investors exhibiting strong preferences for
assets showing minimum riba, gharar, or maysir.
On one level, the efficiency facet could be revealed via the Du Pont expansion of the
ROE structure building on the equity multiplier as the inverse of financial leverage (Brigham
& Ehrhardt 2011, pp.106-107):
ROE=Net IncomeEquity
=
Net IncomeTotal Assets
∗Total Assets
Equity=ROA∗EM
Remarkably, the equity multiplier reconciles asset versus capital performance while pointing
to genuine efficiency showing how the implied leverage, or the residual share of debt in the
capital structure, maps into superior profitability.
In fact, the exact same efficiency metric could further be unfolded by embarking on
one other Du Pont style representation, this time capturing asset performance in terms of scale
efficiency or the interplay of margins versus turnover:
ROE=
Net IncomeSales
∗Sales
Total Assets∗Total Assets
Equity=NM∗¿∗EM
Incidentally, the product of the net margin (NM), total asset turnover, and equity multiplier
captures it all. Whereas the former (drawing in the Islamic banking setup on salaf loans, ijara
leasing, and otherwise real-sector shirka stakes) captures the profitability upside, the turnover
metric points to scale efficiency as part of break-even point (BEP) analysis.
In light of the above, ROE based metrics could be thought of as the ultimate efficiency
check to be deployed as the dependent variable, while simultaneously capturing the ROA
core. Whichever instrument is chosen to represent it, one need not be log-separated, i.e.
modified via the logarithm of the underlying Du Pont product of ratios to appear as a linear
sum. For one thing, that would complicate the sign check, as any logarithm of values below
unity would yield a negative value thus modifying or blurring the estimates of the linear
coefficients. (Note that the turnover component of the log-product would be strongly positive
thus absorbing the other log’s negative contribution). Moreover, separating the dependent
variable rather than the independent or explanatory right-hand side would border on a vector
function suggesting very different estimation routines while ushering in unnecessary
computational complexity without matching it with an adequate increment in predictive
power.
2.7 Hypothesis Testing & OLS Modeling
The formal part will come in several sections, while at times presenting alternative or
mutually controlling devices in line with the triangulation agenda to be addressed at the end
of the chapter.
It would be straightforward to present the above rationale in terms of a simple OLS
model seeking to capture ROE performance in terms of size, or the book value of the total as
well as current assets. On the one hand, that would indirectly point to scale efficiency. On the
other hand, interpreting scale in Islamic versus conventional contexts would stumble into
multi-layered ambiguity or structural uncertainty. For one, Islamic agents may be perceived as
largely productivity and profit driven rather than efficiency concerned—which would suggest
scale efficiency is the last thing to be kept in mind at the outset. However, that would be a
very superficial conjecture, as partnership and vertical as well as horizontal sharing along the
musharaka lines could be consistent with either small or large assets, insofar as the former
can be reduced to the latter being aptly redesigned and participated in. The generic model
could look as follows:
ROE¿=α0+α 1∗TA¿+ε¿
This test spans a set of 46 banks’ ROE values over the time horizon of 2006-2014—and the
same goes for their total assets (TA). The applicable null hypothesis would posit
insignificance as the cut-off or weak criterion:
H 0 :α 0=α 1=0
The alternative hypothesis would constructively posit these values as non-zero at
whatever levels of significance chosen, with the free term or intercept plausibly referring to
either a minimum or the industry-average ROE, and the slope distributed much like the Du
Pont differential residual:
α 1∂ ROE∂ TA
= EM∗∂ ROA∂ TA
=−EM∗¿TA2 =−EM∗ROA
TA=−ROE
TA
Given the negative sign as well as the quadratic term in the denominator, the slope might well
turn out to be marginally non-positive, if at all significant based on the t-statistic. Put simply,
scale might act to detract from ROE based efficiency, as long as the intercept captures the
industry leader’s ROE as a benchmark. That said, the effect is second-order or decelerating,
with ROE itself implicitly acting as a first-order catalyst. On second thought, the differential
expansion could refer to the locally compensatory effect, whereas the ‘global’ or ‘levels’
counterpart is regular, or trivially positive.
Of course, a model such as this one can in no manner be deemed as either complete or
conclusive. Adding on extra explanatory variables might alter or offset this particular slope’s
sign, while rendering the potential inefficiency of the otherwise under-identified specification
less of an issue. In order to pin down the rest of the structural ambivalence, a different
modeling framework could be attempted building on dummy type instruments. The lefthand-
side dependent variable DROE will be represented by a status indicator or dummy variable
taking on a value of 1 if the bank’s ROE exceeds the industry or sample average, and zero
otherwise. The righthand-side or explanatory variables will all be dummy type as well. Better
yet, all of the identity dimensions will be separated and studied as either standalone channels
or as interactive patterns. The former version amounts to individual dummies, whereas the
latter to a product of these. For instance, the size dummy will take on a value of 1 in case the
asset size exceeds the sample average—and the same holds for the rest of individual dummies
capturing UAE identity, Islamic legacy, and their match as a product.
D¿ROE=β0+β1∗D¿
UAE+β2∗D¿ISL+β3∗( D¿¿¿UAE∗D¿
ISL)+ϵ ¿¿
In a complete setup like this, the residual should prove smaller, even though serial
correlations will have to be checked against, regardless. In addition, encoding the values as
indicator type variables should curb the bulk of multi-collinearity as well as spurious
covariance. That said, some of the channels, notably pertaining to couples such as,(β1 , β3),
and (β2 , β3) evidently featuring shared component factors, might be prone to formal multi-
collinearity. The flipside is that the alternative hypothesis could suggest a well-defined
structural inter-linkage across these propagation channels. By contrast, the null hypothesis
will again presume naïve prior insignificance:
H 0 : β0=β1=β2=β3=0
Outside the F-test revealing the model’s overall relevance, it is likely that only some of the
channels will prove [in]significant, thus reconciling the polar hypotheses.
2.8 Generalized ANOVA
Regardless of whether enough data is made available or technically fit to enable
BLUE (best linear unbiased estimate) efficiency in the OLS regression modeling, an
alternative test could be run irrespective of DF constraints. On the one hand, it could draw
upon the same ROE criterion as a matter of control test. Alternatively, the study could tap the
residual or supplementary layers of analysis. The generic approach could be based on the
CHI-square test in its Pearson form (Hayter 2012, pp. 469-470):
CHI 2 ∑i=1
46
∑t=2006
2014 ( x¿−e¿)2
e¿
The observations are checked against their respective expectation hurdles. The
subsequent section will dwell on how these could be defined in proportion terms as mutually
exclusive and collectively exhaustive series of identity matches. For now, suffice it to point to
the remarkable interim result reconciling the various and remote layers of analysis. Whereas
the entire set of tests could be done in regular ANOVA table terms, it happens that the
dummy variable approach as above and the CHI-squared inference amount to versions of
ANOVA or ANCOVA, too.
2.9 Afterthoughts on Triangulation & Bootstrapping
All of the attempted methods and approaches have been sought to allow for an
effective as well as efficient analytical framework. Bootstrapping could in this setup be
viewed as aligning a handful of straightforward approaches to whatever data available without
collapsing the analysis to sheer or theoretically unaided data mining. Triangulation has
targeted reliability and robustness in ways that posit all of the methods being proposed as
vehicles of mutual control and reinforcement. The dual methodological trade-off has to do
with how some of the methods have qualified or augmented others while revealing strains or
excesses of their own. Prospective research might look into how the more rewarding of these
could either be improved upon any further as standalone tools or re-arranged as systemic
frameworks building on well-defined concepts and criteria defining identity and performance.
Chapter 3: Modeling & Testing
3.1 Qualitative Analysis
The effective residual sample features some 27 banks out of the 46 as in the Central
Bank’s database. This implies a response or turnout rate of about 27/46=59%. In net or
adjusted terms, bearing in mind an 11% share of values missing in the panel of 27*9=243
observations, the effective response is secured at (27*9-270/(46*9)=59%*(1-0.11)=52%,
which amounts to the sample size based on the aggregate population. Although it would be
desirable to draw upon a larger time series, e.g. spanning a horizon of 1994 through 2014,
primary or open-source data constraints do affect the design being proposed, along with the
randomization issues. On the other hand, the ex-ante estimate of the share of missing values
could set the potential for improvement in the predictive power of whichever specifications
attempted. This issue will be addressed later in text, and for now suffice it to zoom in on the
qualitative shifts, if any, within as well as between the identity-based sub-samples along the
lines of the Table 1 taxonomy as suggested from the outset.
A total of 6 subsamples could be addressed based on the alternate as well as
complementary identities with an eye toward any possible performance gaps and subject to
varying sub-sample sizes. As Table 2 indicates, the number of non-Islamic banks dominates
as 20 versus 7. Out of these, it appears that all Islamic banks are associated with the UAE (7),
even though the latter pure or elemental identity is broader (18). A similar match is manifest
between banks that are non-UAE versus those neither UAE nor Islamic (9). In terms of
relative weights, these categories claim as follows: Islamic and UAE Islamic above one-
quarter each, non-UAE at large as well as non-Islamic outside of the UAE one-third, UAE
overall two-thirds, and non-Islamic on the whole just under three-quarters.
The analysis can be rendered even more meaningful based on the sub-sample or intra-
group average performance subject to the distributions. For instance, non-Islamic and non-
UAE entities have exhibited a slightly superior performance, with excess ROE ranging in
between 1.09% and 1.11% (13.16% and 13.6% for the respective categories as compared to
12.05% and 12.51% for the Islamic versus UAE benchmarks). However, the data for the
benchmarks have shown to be more homogeneous and possibly homoscedastic, as depicted in
the applicable standard deviations. The resultant ratios of average (in-sample or retrospective
expected) to SD, which is akin to the Sharp ratio based performance net of uncertainty,
demonstrates a systematically reversed pattern, with Islamic banking marked as strong leaders
(4.23 versus 2.43). Although the UAE identity is still marginally outmatched (2.57 versus
2.71), overall the Islamic UAE corner has outperformed its conjugate of those neither Islamic
nor UAE (4.23 versus 2.71).
The latter might be a first indication of the Islamic identity dominating as a contributor
to the explanatory power. This finding will be seconded in a naïve, cross-sectional regression
test drawing upon time averages or historical trends and equilibria rather than regular time
series. The reduced or collapsed version of the panel is warranted as a first estimate for lack
of temporal variability in the core explanatory variables such as Islamic, UAE, and combined
identity. An expanded panel version will then follow.
In terms of ROA based efficiency, the findings are even stronger and far less mixed, as
the ratio proves higher across all the elemental and interactive categories. For one, the bare-
bones ROA appears slightly higher for non-Islamic banks (1.73% versus 1.69%), which gap
reverses itself to a favorable .6% excess for UAE entities (1.92% versus 1.32%) and an
overall upside divide between the UAE Islamic banks versus those non-Islamic outside the
UAE (1.69% versus 1.32%). Evidently, the Islamic constituency again dominates amongst the
key explanatory factors, with the interactive coefficients largely mimicking the dominant
elementary ones.
3.2 Cross-Sectional Regression
This section will embark on some possibly non-manipulative data mining with an eye
on the constraints that more full-fledged panel testing is likely to confront at later stages.
Among other things, historical averages as point estimates will be used as a proxy for either
trend equilibria or actual time series. Whereas it might appear as a natural choice pertaining
to dummy identities which do not show enough variability anyway, imposing a similar
heuristic on ROE and ROA instruments, let alone actual values, should be taken with a grain
of salt. For the same token, although the intercept might capture the bulk of the higher-order
average, providing for one as part of the specification might usher in some awkward trade-
offs.
For one thing, even the sub-sample averages may have appeared largely convergent or
compressed around the gross (industry or sample) average. At the same time, assuming that
an intercept makes a difference would act to detract from the rest of the variables’ explanatory
or predictive power. The respective slopes or partial coefficients will be smaller, thus denying
any more active role to these propagation channels:
y=β0+∑i=1
m
β i∗xi+ε=0+∑i=1
m
β ' i∗x i+ε '
Not only will the latter specification effectively curb the naïve or exogenous component
amidst prior modeling incompleteness, the more degrees of freedom will amount to lower F
and t hurdles, thus going one further step toward higher modeling efficiency. The underlying
efficiency trade-off can rigorously be qualified by seeing whether and how the introduction of
extra variables measures up against a limited number of observations:
DF=N∗m− (m+1 )+1−∆ m=( N−1 )∗m−∆ m
Essentially, an increment in the predictive power should be weighed in against that in
the efficiency hurdle:
∆ R2
∆ F=
∂ R2
∂ m
∂∂ m
[( R2
1−R2 )∗N−m
m−1]
It can be shown that the ultimate sign of the trade-off, or local effect, will be determined by
that of the denominator, whose zero value turns the setup into an explosively inconclusive one
albeit well-defined structurally:
∂ R2
∂ m= N−1
( N−m )(m−1)∗R2∗(1−R2)
Given the oscillatory nature of the right-hand side with respect to the initial condition, or the
ongoing level of R-squared, along with the mixed impact arising from m in the denominator,
the knife-edge pattern should come as little surprise. In fact, the above relationship could be
used as a cut-off criterion. Based on a given N=27 and m=5 (as well as a delta step for m of
1), an increment of predictive power at the current level of, say, .25, should be at least
∆ R2= 1∗27−1(27−5 )(5−1)
∗.25∗(1−.25 )=.0554
In other words, phasing in another explanatory variable only pays if a resultant increment in
the explanatory power totals at least 5.54%, now hitting in excess of 30%. Conversely, if the
initial model featured m=6 independent variables, none of these could be omitted unless the
resultant loss of explanation or predictive power stops short of reaching 4.64%. Importantly, a
linear rule cannot apply, as the R-squared hurdle will have to be updated at each subsequent
stage. For instance, following the initial introduction of a variable given an R-squared of 25%,
its subsequent opportunity cost would be about 6.27% instead of 5.54%, and so on the setup
unwinds. With an initial m=4, the stake is even higher at 8%, which suggests how exacting
could an initially parsimonious model turn out to be. In contrast, it will be demonstrated later
on how mere omission of an intercept triggers an overwhelming increment in predictive
power far outweighing any complexity savings attainable otherwise.
This latter layer of data mining can be ensured empirically by trying out from amongst
a variety of alternate sub-specifications. Table 3 showcases standardized output as per
alternate specifications to be run on dummy counts as opposed to averages, and a zero-
intercept as distinguished from full-blown estimation. Notably, it is presumed that the squared
t-statistics do not converge to their F counterparts unless the sample is very large, in which
event both tend to the normal distribution. Otherwise, the t hurdle is taken for N-m=27-6=21
degrees of freedom in the intercept-active version at a weak alpha significance of .10 and 27-
5=22 at alpha=.01 as per the zero-intercept setup. For the F hurdle, the degrees of freedom
falling under nu1 and nu2 obtain as (5, 22) and (5, 23) respectively. These DF conventions are
in themselves divergent, which is reasonable for a multivariate setup. The null hypothesis
maintains insignificance, or statistical zero, for each channel being tested.
As the comparison of Tables 3A-B versus C-D suggests, only the no-intercept
specification suggests a significant model along with a high explanatory power. Both the
counts and averages based functional forms in the full-blown cases yield an R-squared of
26.38% amidst model insignificance, as depicted by a low F=1.97 which does not exceed the
2.13 hurdle even at a weak 10% significance level. By contrast, the zero-intercept scenario
fetches an enviable R-squared of 78.17% while marking the model as significant even at 1%,
with an F=20.59 clearly beating the 3.94 hurdle.
That said, overall significance need not translate into that of the individual
coefficients. As it happens, only the DROA dummy proves strongly significant, with its t=3.995
exceeding the 2.508 hurdle. As far as the rest are concerned, judgment shifts in between the
counts based versus averages denominated layouts, in that significance switches between the
Islamic versus interactive (UAE Islamic) dummy instruments. Interestingly, the prohibitively
large t or infinitesimal p-values appear in sharp contrast with the virtually zero coefficient
estimates, which implies that these channels’ mere presence is enough to make an explanatory
difference without necessarily affecting the predictive power or the actual forecast. Of course,
micronumerosity could be one other concern when it comes to testing for fine and intricate
mixed effects.
Peculiarly, the same holds for the nonzero-intercept cases, with the intercept versus
DROA strongly surpassing the t hurdle at 4.39 and 2.55, respectively, which may sustain at
more demanding significance levels in light of the low p-values. Whereas the low R-squared
in the intercept-laden specifications might lead one to presume that the model easily needs
major rethinking given the low overall and standalone significance, this is not the case bearing
in mind the alternate success stories. The latter building on a zero intercept, it may well be
that it was multicollinearity that accounted for the bulk of individually inefficient variables.
As was shown previously, this conjecture gains support in the prior qualitative
analysis, as the Islamic and UAE Islamic sub-samples showed a perfect match, which
convergence must have implied linear dependence between their underlying variable vectors.
Although the UAE identity is wider, the Islamic sub-domain proves dominant—as was shown
in the qualitative counts and the quantitative regression analysis alike. Somehow, the Islamic
identity shines through the provisionally intermittent or non-robust significance of the Islamic
and UAE Islamic instruments alike, with the overlap being apparent. Therefore, a natural
choice would be to omit either variable given their perfect collinearity, while lending some
extra significance to the rest. However, given that the UAE neither detracts from nor
reinforces the Islamic domain, it would be about as wise to go with the UAE Islamic
interactive coefficient only, while leaving the standalone or elemental identities out. The DROA
instrument should stay as well, as it bridges some important predictive gaps in reducing the
residual without showing any trivial correspondence with DROE.
When it comes to the (c+r) delta, its mediocre performance has not fallen short of
expectations. In fact, it will be shown that it is the lagged values of this differential, perhaps
by as long as two periods, that should have a material impact on ROE performance—be it an
Islamic setup or otherwise. For that matter, this value exhibited enough inter-temporal
variability for it to avoid being collapsed to a historic average. Therefore, this could be an
important layer of explanatory power yet to be addressed in a panel design. While at it,
regardless of how complete one might happen to be, the aforementioned 11% response or data
error slack could be invoked to update the ongoing R-squared of 78% to about a 89%
potential.
The more economical cross-sectional reduction could be re-specified as follows:
DiROE=β1∗Di
ISL∗DiUAE+β2∗D i
ROA+ϵ i
As Table 3E points out, the explanatory power remained nearly the same at 77.38%, whilst
the overall significance has soared better than twofold, as is evident in an F value of 42.76
rising above a hurdle of 5.57. However, the individual significance is now more mixed than
that. Whereas the DROE instrument has gained even more merit with a t=8.165 refuting the
null threshold of 2.485 at any alpha level, the prohibitively high p-value and low t for the
interactive dummy suggest that it is statistically zero for all practical purposes.
That said, it could still be used for explanatory purposes yet not as a matter of forecasting.
One way it might come in handy is by indicating that banks that are both UAE and Islamic
are only likely to under-perform should their ROA fall below average. The estimated equation
would look as follows:
E ( DROE )=.898∗DROA−.018∗D ISL∗DUAE
To illustrate the tentative predictive power, suppose the bank under study is UAE
Islamic. As long as its ROA exceeds the industry average, the odds are
about .898*1-.018*1*1=.88, or 88% that so too does its ROE. Now, of course the 12%
discount is unlikely to have been accounted by UAE Islamic identity—in fact, it should count
toward a 2% loss at the very worst, whereas a material residual was from the outset
potentially ascribed to data errors due to an 11% (or even 1-27/46=41%) non-response.
Moreover, banks that are either Islamic or UAE or both should boast more reliable
performance, which nets out as higher Sharpe-like efficiency. However, the present model
only addresses gross performance indicators—worse yet, as mapped into dummy instruments
which, by nature and design, cannot possibly claim anywhere near perfect predictive power.
Suppose, for ease of intuitive comparison, that the bank in question is either non-
Islamic or non-UAE, or neither. In the event its ROA is above average, odds are on the order
of 90% that so does its ROE. Either scenario illustrates consistency across alternate yet related
efficiency criteria. Now, presuming its ROA fell short, one can claim with near certainty that
ROE must have concurred. In fact, this suggests why the interactive coefficient might be
insignificant by the very nature of dummy normalization or additivity: After all, if it were
UAE Islamic, the expected status could prove slightly negative, which does not make much
economic sense other than seconding the same strong judgment. In other words, ROA
performance remains the core of analysis in this reduced, cross-sectional setup. As it happens,
the panel will not differ very critically.
3.3 Heteroskedasticity, Autocorrelation & GLS
Apart from the micronumerosity issue making room for outcomes as polar as excess
multi-collinearity versus spurious correlation, non-homoskedastic residuals may well render
the entire modeling as prone to structural shortcomings denying its OLS coefficients the
BLUE prerequisites. To begin with, not only do the two groups of banks in question—Islamic
versus regular entities—reveal some variance gap in the actual dependent variable and
possibly residuals. In fact, that would be an important piece of information in its own right,
provided it is statistically significant beyond divergent Sharp ratios. What is more,
performance may have shown some structural shift post-2008 in the aftermath of the crisis
back then.
On top of that, the very nature of the dummy based model being proposed, with both
the explanatory and the independent variables normalized as probability-like indicators, is
such that the resultant residuals are distributed binomially rather than showing a random
normality. Finally, the time series dimension of the panel yet to be addressed might exhibit
autocorrelation in the residuals, in which event some systemic pattern still remains untapped
while dimming the estimates’ efficiency. Among other things, heteroskedasticity is a cross-
sectional or static counterpart of temporal autocorrelation, insofar as various static error terms
might reveal non-zero covariances. On the other hand, the two alternative cases feature off-
diagonal covariances as either all zeros or otherwise, with diagonal variances either varying or
converging.
In fact, it is the dummy based design rather than particular specifications per se that
may either have smoothed the less significant gaps or exacerbated these. Interestingly, the
binomial or oscillatory distribution of the error terms may imply a seemingly chaotic pattern
in the first differences or derivatives looking as if the residuals were uncorrelated or more
generally independent in a difference based model such as the one under study.
One way of getting rid of non-constant variance, without having to look into the in-
sample nature of the added issues, could be to deploy Generalized Least Squares (GLS), with
OLS estimates being covariance-adjusted, which might waive the bulk of the above issues.
That said, it might be an option to stick with the OLS estimates for explanatory purposes, if it
were not for the regular and irreduced pattern in the error terms denying the model its
unbiased status. On the other hand, GLS lends or recoups the ‘best’ feature of BLUE, so that
interim or marginal improvements are not at stake anyway, when it comes to aligning
predictive and explanatory powers.
It is important to appreciate that the grand design or particular specifications need not
change, and nor does the GLS modification necessarily have to be re-run from scratch—as
applying a modifying matrix to whatever OLS coefficients should be enough. On the one
hand, based on the F-test, the cross-sectional model already proves a success, which should
carry over into the full-blown panel setup, so it is largely about the individual t-tests. On the
other hand, boosting the degrees of freedom or sample size by so many times would only
improve the resultant t-statistic by about the square root of that.
For instance, if the cross-section of banks were to be restored to the complete
population of 46 (which is not an option as the rest are either no longer out there or fail to
report their financials on a regular basis)—with observations being defined over twice as long
a horizon starting from 1994—the resultant increment in the t-statistic or contraction in the p-
value would at best be on the order of some 100%, or about 40% in case of either being
feasible. That could hardly remedy a .89 p-value for the non-lagged (c+r) differential or UAE
Islamic identity. Worse yet, since neither expansion is an option (with the actual response
being so objective as to posit the sample as the best proxy for the population), such data issues
should be assumed away with an eye on the binding data constraints. Not least, post-2008
structural shifts and learning effects may have turned out to be permanent yet less than
relevant in themselves as per the scope on hand, in which light delving into too distant a
history might never lend any major explanatory power to the model while ushering in an extra
data error of systemic, qualitative sort.
Again, there is no need to assume independence between the error terms and the
explanatory variables—much less with respect to Islamic or UAE identity, as the latter is,
among other things, what could come of interest given the scope of the present study. For that
matter, it should come as little surprise that the residuals’ distributions will vary depending on
the alternate designs or specifications. They may or may not show linear independence or
zero correlation with the explanatory variables, yet in any event the weaker prerequisite is
that the generally conditional or design-sensitive error expectation be zero ad hoc. The dual
outcome to be shunned is such that the errors supposedly independent by design ex ante turn
out to reveal heteroskedasticity ex post as a sample-specific feature at odds with either the
scope or the desired state.
The natural way to proceed in coming up with a GLS adjustment would be to presume
a knowledge of how exactly the variances differ from the constant OLS counterpart, and
adjust the variables accordingly. As it was proposed, the basic OLS estimates (slopes) would
not change, even though the intercept may end up affected:
VAR ( ε|X )=σ2∗G →~y= yG
=β0
G+∑
i
m ~β i∗xi
G+
ε i
G
~β i=∑~x∗~y
∑~x2
=
1G2∗∑ x∗y
1G2∗∑ x2
=β i
VAR (~βi )=VAR (~ε )1
G2∗∑k
N
x ik2=
1G2∗σ 2
1G2∗∑
k
N
xik2=VAR (β i
OLS)
In other words, a generic GLS adjustment does indeed yield a BLUE estimate whose variance
corresponds to that warranted by a qualified OLS procedure. In contrast, a naïve OLS design
overlooking the inefficiency issue would build on coefficients whose variances total a squared
factor of G times the regular OLS counterparts. Notably, in the particular setup being
attempted, the intercept will not be affected as it was put zero from the outset.
Now, of course, this is just the mnemonic scheme, whereas the actual G factor will
likely have to be estimated as a matrix in line with feasible GLS (or FGLS) rather than
assumed to have been known as a quasi-scalar function. Moreover, since the estimated G
operator will likely be nonlinear in the explanatory variables at times thus resulting in
nonlinear coefficient estimates, it has been observed that FGLS may not outcompete the OLS
in small samples despite asymptotic convergence, even though routinely it does (Verbeek
2000, p. 78).
To illustrate this in an alternate as well as generalized fashion, suppose there are only
2 core groups to define heterogeneity over: Islamic versus non-Islamic. Their implied error
variance-based GLS weights (omega) may well differ from either symmetry (50% each) or
the respective naïve OLS counterparts (alpha) that can be inferred based on an overall
coefficient (beta) versus standalone, sub-sample bA and bB. It can be shown that the GLS
versus OLS estimates would differ as follows:
βFGLS−βOLS= (ω−α )∗[bA−bB]
Since neither the weights nor the dummy based coefficients ever exceed 1, their
differences should be small—more so the product of these. In other words, the two alternate
coefficient estimates should prove very close unless heteroskedasticity is very pronounced. In
fact, this is not evident from the prior qualitative analysis. In the proposed model with 27
banks and 2 explanatory variables on hand, it can be demonstrated that even in the case of
perfectly equal group variances (as opposed to sample error variances drawing upon the
degrees of freedom), it is the divergent subsample or group sizes that might make the whole
difference. The resultant variance based weight could be about 58% while having little to do
with heteroskedasticity per se and thus blurring the inference rather than aiding it. By
contrast, heteroskedasticity induced weights may not map into a material gap between the
GLS versus OLS estimates. In addition, small sample and hence sub-sample sizes may render
the group estimates too inefficient ex ante, with posterior efficiency improvement thus largely
appearing to be a self-induced problem.
On the other hand, a two-group GLS might neither pay nor suffice, as the effective
number of would-be groups totals 8, i.e. 4 cross-sectional along the lines of the original
identity matrix and 2 temporal with respect to 2008 marking a potential structural or learning
landmark. One way of desolating the contingent component over and above the regular BLUE
variance as a meta-intercept would be to regress the squared OLS residuals on the exact same
model that was deployed as the OLS core yet to be refined as GLS:
log( y i−E( y iOLS))2≡ log ei
2=log σ2+β '1∗DiISL+β '2∗Di
UAE+…≡ β ' 0+E ( yiOLS )+ei '
However, the parsimony criterion serves as an embedded mechanism driving the
refinement choice. On the one hand, an OLS regression of the kind has already been run, and
is unlikely to prove more of a success in the residuals once again than it did in levels—much
less with an intercept being allowed for. Somewhat paradoxically, this test is bound to show
insignificance in all the variables except the BLUE variance intercept, which is why one could
safely presume no heteroskedasticity.
Based on all of the above, there is no need to run any GLS per se, as it is unlikely to
beat the OLS on hand. Extra complexity involved is hence not justified from the standpoint of
parsimony. It will be shown later on how a similar economical approach maintains a
minimalist design as robust, in that few if any extra variables appear to be warranted in a full-
fledged panel setting.
Chapter 4: Auxiliary Modeling
4.1 Caveats & Extensions
It has been proposed that any major structural gaps or shifts over time in performance
could largely be accounted for by either a differential impact of Islamic entrepreneurship and
management or UAE regulations and otherwise relevant institutions, or both—as implicitly
captured in the respective dummy type instruments. Notably, though, these can reasonably be
presumed to hold identically over time—which would deny the critical variability in the right-
hand side it takes for a dummy mediated model to explain that in the dependent variable. At
the very least, that could question the very panel’s relevance in the first place, even though
getting rid of one would deny an important layer of explanatory power. Alternatively, it might
be feasible technically to consider less stable, or non-stationary ‘effective’ identities, e.g.
based on a varying proportion of Islamic solutions in the total asset composition or capital
structure—or indeed define the dummies in terms of deviation relative to an industry average
which does change inter-temporally. However, that would be a rather artificial approach to
measuring Shari’a compliance as a contributor to ROE denominated performance.
Had it been for some institutional pillars that evolve with the passage of time without
varying across the entities, the cross-sectional stability of the sort could be depicted in the
fixed effects under a panel design. In actuality, the setup is reversed with material identities
only varying cross-sectionally yet not inter-temporally. Although it might be an option to just
reverse the dimensionality in the computation, still the SPSS design and any other software
would likely have been fine-tuned to a particular convention when it comes to specific
standardized tests building on a precise order or sequentiality of declaring the dimensions.
One way around the issue could be about supplying some extra explanatory variables
that show enough temporal variability to make it up in the initial specification. That said, a
few trade-offs will have to be considered. To begin with, one should stay wary of a
specification error or bias which might render the overall model more effective while denying
efficiency to the individual explanatory variables, or linear coefficients. That could be made
manifest in a higher R-squared along with a stronger F-test amidst weaker t-values. On the
other hand, these exact same metrics would in any event have to be traded off against a likely
increment in complexity in light of a more involved specification—as compared to the status
quo or incumbent model. At the end of the day, however, it would be rather strange to have
anticipated that any combination of Islamic and UAE identity, as well as their interaction or
downright denial, could possibly explain the entire short-run ROE difference, if any.
Consequently, a formal model will be proposed to further rationalize would-be ROE
performance while bridging some of the aforementioned testing gaps.
4.2 An Auxiliary Formalization
The premises of Islamic banking have little to do with the conventional ends as well as
means pertaining to effective monetary expansion as based on the issuing of loans and subject
to reserve constraints. For one thing, Islamic banking entities do sell a variety of lease- or
loan-like instruments that are more of an equity nature while building on non-interest payoffs
by and large. Legal reserves may well be there, yet the underlying rationale is not about bank
runs, or non-performing and ‘toxic’ assets per se. Some of the reserves may be intended as
cushion or hedging tools to offset partial losses, and in any event the discretionary layer far
outweighs the arbitrary reserves or fund requirements being imposed exogenously. Apart
from the structural gap extending beyond capital adequacy or weak Basel II compliance, the
very share of reserves could either be lower or higher depending on the business model or
asset structure of the Islamic bank, or possibly in line with the stage of the economy cycle, yet
definitely not as a mere constraint on loan issuance.
In a sense, Islamic banks are more characteristically contingent on the savings rate, or
the marginal propensity to save at the macro level, as well as by the retention ratio as its
loose, micro-level counterpart. In effective terms, their performance as well as expansion is
surprisingly akin to the Keynesian type investment multiplier building on the marginal
propensity to consume C as opposed to save (1-c), rather than a conventional money
multiplier as in Mishkin (2010, pp. 358-359), in terms of the banking system’s capacity to
create money or account for its velocity as in MV=PQ.
In fact, it should be straightforward to incorporate both, with the peripheral
mechanism of conventional banking reflected in terms of an exogenous parameter N, or bank
network size. What affects ROE is net income along with the ongoing equity. The latter may
largely be made up of retained or cumulative earnings as well as assets—or, at any rate, the
actual structure may either be unknown or arbitrarily reshuffled for reporting purposes.
However, the one pillar of relevance that has to do with ROE as the profitability of equity
pertains to growth in equity. By contrast, growth in assets is hardly the ultimate criterion of
profitability, or capital gain, because it can be secured in ways that deny intrinsic earning
power or organic expansion. Some of this phenomenon is evident in just how abruptly the
total assets may change when reporting on a holding as opposed to an entire group, or indeed
following ‘non-organic’ M&A expansion. It is fortunate that none of that has shown to alter
the core identity in the data on hand. More generally, ROE appears to be a rather
dimensionless metric whose scale-invariance qualifies it for meaningful inter-bank
comparison within as well as across any sub-samples in question.
4.3 An Augmented Model
The aforementioned formalization could serve as a critical add-on acting to augment
the core specification by possibly reducing the residual while lending some extra variability to
the explanatory part. The differential in ROE, while keeping this dependent variable
commensurate with the one in the core model, could be interpreted in a variety of ways. In
addition to generic sensitivity, it could refer to any gap as compared to the industry average.
For that matter, it could capture the gap in between the Islamic versus conventional business
models, insofar as it can be inferred from the sample structure. For instance, suppose the
effective sample (based on the actual response or turnout rate) features about 2/3 of the banks
qualifying as UAE (which makes the analysis largely UAE centered) and ¼ of them as
Islamic. This suggests that the industry average or sample mean would exhibit a bias toward
the non-Islamic and UAE legacy being incumbent. That could affect the actual (c+r)
difference as well as its composition. Although the generic response function would
structurally remain invariant, its actual values will likewise vary depending on the model.
Suppose an Islamic model shows, in theory, utter disregard for basic efficiency as
opposed to ultimate ROE efficacy, which might imply a low value of K ≡1−c−r. However,
it was pointed out that its zero asymptote might leads to a very high ROE (Figures 1A-1B).
Yet this oversimplification enables one to appreciate just how close the ROE gap comes to a
(c+r) differential anywhere around that level. In other words, any decrease in the basic
efficiency will compromise ROE efficacy by about the same percentage. Incidentally, the
response function amounts to an elasticity and not just a derivative. Therefore, depending on
the actual bank’s model or identity, the percentage response could either show overreaction to
or compression of whatever gap being determined by the sample composition (Figures 1A-
2B).
A wealth of efficiency implications could be drawn from the above conjecture. The
study of the basic efficiency differential could alone usher in a plethora of scenarios. For
instance, the individual impacts of legal versus discretionary reserves could be
compensatorily inverse or negatively related in case their combined ratio is constrained. For
the same token, they might net out to a rather mixed differential when it comes to industry
benchmarking. Whereas the idle or legal reserves could prove far lower in the Islamic setup,
the discretionary or productive counterpart might offset that.
For now, the whole of the ROE-augmented model to be addressed later in text could
be added to the core model as an interactive variable. Only in this case can linearity in the
coefficient be ensured, with invariance hardly attainable for any K. In other words, the
response function does vary, so that measuring it as part of the added coefficient of K might
not make a lot of sense in an OLS or dummy instrument LS (DILS) setup. The augmented
specification could look as follows:
D¿ROE=M +γ∗X ¿+u¿
In this specification, M captures the entire core, dummy-based model, and X refers to the
variable expansion yet to be covered. Although it might be an option to define the (c+r)
differential in dummy terms as well, the mixed structure within the same instrumental
variable could hardly prove productive beyond mechanical reduction in the residual. On the
other hand, paradoxically, some collinearity between this differential and Islamic identity
would be sought rather than shunned, which lends extra rationale behind keeping at least
some of the supposedly related variables as non-dummies so as to minimize the formal multi-
collinearity excess.
On second thought, it should be straightforward to motivate the entire setup in plain
and natural terms. To begin with, an add-on regression for delta ROE can indeed be run on a
(c+r) differential as an OLS regression. For one thing, if ever, the applicable coefficient
might prove linear in weakly differential or local rather than strong or levels denominated
terms:
∆ ROE=f (c , r )∗∆ (c+r )=
∂ ROE∂ X (c , r )
∗∂ X
∂(c+r )∗∆ (c+r )=β∗∆(c+r )
One may wonder if the implied and unknown function underlying the linear estimated
coefficient would likely be so well-behaved as to enable a pattern that is smooth over (c+r).
Remarkably, it is both ensured and irrelevant in the way of its actual representation. This is
easiest to show by noting that the above differential likely amounts to a Taylor expansion
around some initial conditions or benchmark like the industry average:
∆ ROE=ROE (c+r )−ROE (c0+r0 )≅ RO E ' ( c0+r 0 )∗[ (c+r )−( c0+r 0 )]
The implied mapping, be it a derivative or a coefficient, is inherently constant or fixed thus
rendering the transform linear, as it is defined at one point, e.g. industry average, baseline, or
initial conditions. In fact, it is for this reason that both the explanatory and the dependent add-
ons can be transformed arbitrarily around the linear stretching. As one possibility, both can be
represented in dummy terms, thus showing even more consistency with the core model:
At this point, however, one should come to realize the recursive endogeneity as
implied in a two-way and inherently auto-regressive interaction between (c+r) and ROE. On
the one hand, the present reserves and dividend policies should affect ROE performance in the
next period, e.g. as a compensatory cushion responding to an earnings drop. On the other
hand, ROE may in turn have an impact on these policies in the subsequent time period in
proportionate rather than compensatory terms, as a response to good rather than poor
earnings. The two-way setup suggests a double lag, i.e. a period over which the extra
variables significance is most pronounced.
However, including the second term would usher in linear dependence on the first one,
or multi-collinearity, by the very definition of a linear inter-linkage between the two as
conjectured in the lefthand-side scheme.
In fact, any attempts at further improving upon the model could prove about as
detailed as they too are superfluous. To begin with, the residual could further be reduced by
embarking on higher-order Taylor differentials based on the powers of (c+r) differences.
However, this series converges very quickly, because even if the sum is anomalously large
above 1 in levels, it is still regular and below unity in differences—and these die off
instantaneously starting with second-order terms. Therefore, this channel is unlikely to come
in handy, either.
4.4 Caveats on Data
The annual reports have not shown to be very accurate or reliable in a variety of ways.
For starters, the same terminology has tended to be used to denote very different layers of
income statement reporting, e.g. net operating revenues have at times been stated as operating
profit, as it were net of the operating costs. On the other hand, the Du Pont representation as
attempted has enabled meaningful control over the relevant interim ratios and the implied
base figures, given that the net margin can be discerned from ROE net of the equity
multiplier.
For that matter, accounting for the reserves has appeared rather manipulative. Not only
has the total value been at times adjusted in retrospect, it was not always possible to
distinguish between the regulatory or legal versus discretionary layers due to arbitrary
reshuffling. A similar challenge would confront anyone trying to infer the actual dividend
payout, which is why it had to be reconstructed based on a residual of net income net of
changes in reserves and retained earnings.
Whereas the systemic and likely cross-sectional component of discrepancy may have
had to do with the gap between reporting legacies (e.g. country-tailored GAAP as opposed to
uniform IFRS), idiosyncratic and possibly time-series type shifts largely stemmed from
reversals on impairments and provisions. Whilst the former captures minor counterparts of
bank runs, the latter could take on forms as diverse as, cushion against non-performing loans,
hedging vehicles, and translation or reporting-specific disparities—let alone currency risk
mitigation in a cross-border setup. Pension and insurance type commitments could be one
other layer of importance showing how provisions may pertain to gap management and fixed-
income portfolio immunization beyond derivatives hedging per se.
The re-allocation setup becomes even more entwined in light of provisions such as
‘dividend reserves’ (ABN Amro 2014, p. 184), which lend some extra rationale behind a
proposed setup in which retention and reserve ratio enter ROE symmetrically—even if totally
unrelated ex ante or are defined against very different bases, e.g. net income versus
cumulative or retained earnings which are trivially related in terms of flows versus stock and
hence compatible anyway. On the other hand, this dividend reserve may have been tied in to a
dividend payout target (ABN Amro 2011, p. 113), which essentially amounts to a C input in
ROE being posited as inseparable from the r impact. In passing note an important distinction
between the two corner dividend policies. For instance, whereas the non-Islamic, non-UAE
legacy may fare on a 40% share of net income (ABN Amro 2011), its Islamic UAE
counterparts declare an intended 40% on any paid-in capital (ADCB 2014, p. 9).
Moreover, although dividend payout versus reservation rate could be seen as related to
working capital versus fixed assets potentially, their complementarity can neither be
overlooked nor fully appreciated as technology-accommodative c/r ratio. At the end of the
day, processes should not vary materially across banks within the same legacy, so that it is the
latter dimension that makes the difference time and time again.
Importantly, the bulk of these may first have been accounted for as special income
statement items, followed by posterior transfer into either the retained earnings or offsetting
reserves falling under the respective balance sheet categories. In any event, the single most
important challenge has been about the meta-translation or reconciliation uncertainty as
affecting the data quality in ways that can hardly outmatch bank-specific translation and
allocation efforts.
The more serious divide rests with the very gap in conventions whereby sukuk, as an
effective supplement or extension of regular dividends or at any rate difficult formally to
distinguish from preferred payouts, is routinely reported as a tax deductible expense item in
the income statement, i.e. counting toward pre-tax allocation rather than after-tax distribution
(ADIB 2014).
Although manipulable items cannot possibly provide a careful account of the ultimate
variability in the key performance indicators (KPI), what might be possible to detect, among
other things, is the group-systemic or legacy-specific patterns of deviation or reporting errors
that could prove statistically significant in their own right. This is another way of saying that,
say, Islamic banks might tend to under-report their reserves in somewhat distinct ways that
are still of relevance to ROE. In a sense, this higher-order dimension of efficiency pertains to
a layer of transaction or deliberation costs, which exhibits an overlap or interaction between
macro-level institutional characteristics versus bank-specific discretion, rather than confining
performance to micro-level operations or internal environment alone.
It remains to be seen whether institutional convergence, or indeed reporting
harmonization, can either be expedited by hedging against the implied translation mismatch,
or whether that could be a major facet of market efficiency accounting for some of the
potential ROE slack yet to be seized. Put simply, it is unclear whether harmonization can
implicitly be fostered ex post by making hedging or reconciliatory choices, or if these should
be made ex ante as a matter of hedging against a smaller expected ROE or its higher standard
deviation. One way or the other, the attempted study could shed some light on whether the
two alternate legacies have been more or less apt at addressing the aforementioned efficiency
trade-offs.
4.5 Panel Estimates
The typical choice facing any panel setup is between fixed versus random effect
approaches. In a sense, the latter could be treated as an extension of the former, with both the
regular intercept and the error terms rethought in major ways. Whereas in the fixed-effects
case, the intercept pertains to a fixed part that varies cross-sectionally, in the random effects a
stable core is added on along with residuals that could thus be treated as largely the temporal
shocks or short-run propagations.
Although it should be reasonable to discard random effects at the outset due to the
questionable explanatory impact of a fixed-core intercept as before, a fixed effects version
could well be afforded as well as warranted, in that a higher R-squared might be salvaged
along with the now-more significant and efficient estimates. The latter is secured with an eye
on the added time dimension as a natural extension of the cross-sectional sample, whose
increase by a factor of T would bring in a significance boost of about the square root of that.
In other words, considering 9 time periods might secure a triple increase as opposed to an
adequate reduction in the t-statistics and p-values respectively.
Although the presence of fixed effects might not be desirable to define and measure
conditional probabilities such as those building on a dummy setup, it could later on be
possible to see whether the averaged group intercepts have shown to vary. On second thought,
that might put into question the deployment of identity dummies in the explanatory vector
which do not vary temporally anyway. Should there be a good reason to expect that these
variables remain insignificant the regular way, it might be worthwhile to include them
implicitly in the fixed-effect free terms so as to measure any of the 4 cross-sectional identity
impacts later on. At that rate, the model should prove even more economical as well as panel
fit—more so given that the other 2 temporal dimensions or transition shifts are not
immediately relevant to the attempted scope anyway.
The parsimony trade-off might not be that straightforward, bearing in mind that even if
all but one or (m-1) explanatory variables are removed, another N fixed-effect host is
introduced, which results in updated degrees of freedom applying to t-tests.
The candidate specification could look as follows:
D¿ROE=α i+ β1∗D¿
ROA+β2∗D i, t−1c+r +ϵ t
H 0 : β1=0=β2
In fact, there is no need to explicitly restrict the individual effects to zeros as well. After all,
they ‘explain’ very little by themselves, should the rest of the coefficients turn out
insignificant.
It is remarkable that a model represented in deviations from the temporal means,
which is the equivalent keeping the fixed effect terms irrelevant, resembles the cross-section
counterpart building on differentials with respect to industry averages. In fact, this lends even
more support to the original decision of assuming a zero intercept in that cross-sectional
reduction.
Since temporal shifts are less relevant to the intended scope, and the cross-sectional
heteroskedasticity has been discussed previously, an OLS panel estimation procedure can be
attempted. For that matter, no endogeneity issues are envisaged, as these have been eliminated
along with the would-be extra multicollinearity between the candidate lagged regressors.
Finally, employing a fixed effects model is all the more reasonable given that the dummy
identities as implied in the alpha individual effects are plausibly correlated with the rest of the
explanatory variables—or so the underlying theory predicts.
As per the time series dimension of the panel design, the dummy based specification
secures ‘stationarity’ by the very nature of binary or limited variables that represent
transforms of first differences or gaps around the static means. On second thought, ‘co-
integration’ issues are less relevant, in that auto-regression has been avoided as a matter of
sterilizing multicollinearity while boosting the explanatory rather than predictive power
around the proposed regressors.
Interestingly, the fixed effects design appears to fit into the micronumerosity issue
thus making it a forte rather than an efficiency constraint. As proposed before, the rationale is
that a total of N extra individual-effect variables are introduced as per a cross-sectional
sample size of N, which suggests it is conceivable for small samples only—unless the T
dimension features very many periods and not just a few extra years. Incidentally, the
individual effects cannot be estimated consistently in any event, so it is no wonder if the issue
of less efficient dummies carries over into the panel setup. The silver lining, though, could be
that the cross-sectional reduction was a reasonably good and representative starting point
along many lines.
A conditional maximum likelihood (CML) as one way around the issue could be seen
as an analogue of the aforementioned conditional variance decomposition, which both are
likely to depend on exactly the dummy variables being approximated by individual effects yet
to be bypassed in turn—and possibly with as meager a chance. Much like in that case
showing effective GLS irrelevance, the basic linear setup need not be altered this time around
either. For the same token, the alternative designs such as logit, tobit, or probit will not be
contemplated, as less relevant theoretically or less fit for comparative purposes against the
cross-sectional reduction or at odds with fixed effects. Worse yet, the ‘initial conditions’ issue
pertains to the inherent difficulty to argue that the initial probabilistic value is fully exogenous
rather than path-dependent or arbitrarily chosen in an otherwise cross-sectionally determined
setup.
Finally, the panel can be presumed balanced for all practical purposes, as there was a
near complete overlap between the banks available and those reporting their financials
throughout.
Chapter 5: Panel Findings
5.1 Prior Estimation
The panel data test has been run on SPSS/PASW 17.0, with ‘mixed modeling’
corresponding to a panel, ‘subject’ to cross-sectional dimension for the purpose of fixed
effects estimation, and the entire setup regressed under the restricted estimates maximum
likelihood (REML) mode. The common or random effects intercept has been waived, whereas
the individual fixed effects have been estimated for a total of 27 banks. The missing values
have totaled 32 for DROE and DROA and 52 for D(c+r) as lagged by one period. The estimated
coefficients along with their significance parameters have been provided in the output below.
Table 4: Panel or Mixed Modeling Output (SPSS/PASW 17.0)
Coefficient Estimate t-stat p-value Correlation DISL DUAE
(1) DROA=1 .5908 5.165 .0000 (1,4)=-.651,
(1,5)=-.623,
(1,6)=-.633,
(1,7)=-.681,
(1,8)=-.669,
(1,9)=-.766,
(1,10)=-.611
(2) Bank=AAIB .6062 3.734 .0000 (2,4)=.420,
(2,5)=.418,
(2,6)=.411
0 0
(3) Bank=ABK .4806 3.209 .0020 (3,4)=.438,
(3,5)=.437,
(3,6)=.442,
(3,7)=.429,
(3,9)=.443=(3,10)
0 1
(4) Bank=ABN .4054 3.055 .0030 (4,5)=.512,
(4,6)=.510,
(4,7)=.505,
(4,9)=.508
0 0
(5) Bank=ADCB .4788 3.643 .0000 (5, 4)=.512,
(5,6)=.507
1 1
(6) Bank=ADIB .4781 3.642 .0000 (6, 4)=.510,
(6,5)=.507,
1 1
(6,10)=.502
(7) Bank=Baroda .3780 2.704 .0080 (7,4)=.505,
(7,9)=.528
0 0
(8) Bank=BLOM .5279 3.348 .0010 (8,9)=.515 0 0
(9) Bank=BNP .4826 3.361 .0010 (9,4)=.508,
(9,7)=.528,
(9,8)=.515
0 0
(10) Bank=FGB .4258 3.259 .0010 (10,6)=.502 0 1
Pseudo R2 .87
Average bank
FE
2/9 4/9
The expanded model shows a clear improvement on the cross-sectional reduction,
even though the core of it has shown to be transferable and robust. For instance, based on the
ROA dummy as an invariably dominant factor now enjoying an F at 83.47, that appears to
measure up against the expected significance boost as an outcome of temporal sample
expansion, which effectively totaled [27*9-(32+32+52)]/27=4.7—exactly by how much the F
statistic has increased, with the t-counterpart going up by about the square root of that, i.e.
5.17 up from 2.49 for DROA. The pseudo-R2 can now be estimated at about 87%, which is a
clear improvement on the original 78% albeit again in line with the 11% expectation on data
error slack. Notably, all of the individual bank fixed effects that proved highly significant at
most levels have seen strongly convergent estimates anywhere in between .40 and .60 (with
mode around .47) and t-statistics (3.055 to 3.73 with a mode at 3.64) alike. Interestingly, all of
the Akaike related criteria have likewise been compressed around 84 to 88 values.
Although for the most part these banks are non-Islamic, the proportion of UAE players
is about twice as large as that of Islamic legacy. Spectacularly, this smaller sample showing
utmost significance is representative of the overall structure. The lagged reserves and
dividend differential proved insignificant again, yet for the most part its inverse relationship
with the ROE and ROA is strongly apparent throughout, with missing values not denying the
regularity. That said, this pattern is clearly far from linear, which renders GLS and related
estimates nearly pointless as per this regressor.
The estimates have turned out to be remarkably similar at about .48 for the few Islamic
as well as these and UAE banks for which they proved very significant, even though for UAE
banks as such these are more disparate (.48 versus .43). For that matter, the correlations across
the banks’ individual effects have been rather compressed around .4 to .5, even though DROA
induces some excessive correlations with these varying anywhere within the -.6 to -.8 range.
It should come as little surprise that there is a compensatory trade-off between the explanatory
terms, if only insofar as these have to be restricted to 1.
The non-Islamic banks are far more scattered when it comes to the unconditional
expectation component of enjoying an excessive ROE regardless of ROA gap—even though
the average for the two groups is about the same at nearly 50%. In fact, that seconds the cross-
sectional qualitative reduction showing superior Sharpe based, or risk adjusted performance
as boasted by Islamic banks. The Islamic bank, and one of UAE Islamic legacy for that
matter, will enjoy the 48% sure bet on seeing an excess ROE, with the contingent component
adding up to near certainty in the event of there being an excess ROA.
In fact, the bulk of these macro-outcomes could have been discerned from visual
inspection of the data distributions. Figure C1 showcases all of the core inter-relationships
within as well as across datasets. Apparently, ROE and ROA gap dummies are distributed
much alike, with sparse areas or zero values dominant (as is evident from Figure C2).
Although the (c+r) gap dummy appears more of a normally distributed set, one is led to
accede to the conjecture explaining why it lacks the linear efficiency. Whereas ROA and ROE
dummies appear monotonously convergent, that other dummy is either dually or orthogonally
co-distributed with them. Put simply, evident is sheer lack of linear correspondence, even
though that is not to deny the conjectured inverse relationship as predicted theoretically ex
ante.
5.2 Two-Stage Inference on Fixed Effects
It should now be readily apparent how the fixed effects could be utilized in tracing
through the revealed yet latent identity impacts. The identity matrix as in Table 1 can now be
filled in with respect to each sub-domain, to arrive at the 4 group averages to run the CHI-
squared and t-tests on.
Table 5A: Identity Matrix Made Operational
ISLAMIC NON-ISLAMIC
UA
E
ADCB
ADIB
ABK
FGB
Non
-UA
E n/a
AAIB
ABN
Baroda
BLOM
BNP
Table 5B: Fixed Effects Averages
ISL*UA
E
ISL*nUA
E nISL*UAE
nISL*nUA
E ISL nISL UAE nUAE
.47845 0 .4532 .48002 .47845 .47236 .46583 .48002
The above table lends ultimate support to the original qualitative finding whereby
there is a perfect match between Islamic and UAE Islamic models, and very little difference
in unadjusted terms between Islamic versus non-Islamic performances. In fact, it may appear
ironic that the two polar extremes, ‘UAE Islamic’ versus ‘non-UAE non-Islamic’ can hardly
be distinguished between. It is unlikely that a naïve CHI-squared test against the common
average ever shows material inter-group gaps. Althought it may appear that the non-UAE
Islamic is an outlier in that it is not represented, still that would suggest a zero gap between
each [missing] individual parameter and the [non-existent] subgroup average by definition. In
order to avoid the missing value or selection bias, a CHI-squared will be run based on sub-
sample averages rather than a general sample average.
It is for this reason that one may want to run a t-test making use of asymmetry such as
drastically differential sub-sample sizes which might usher in statistically significant chasms
even amidst the otherwise indistinguishable sub-sample averages.
The CHI-squared across the 4 identity subgroups barely totals a 1.075, with the bulk of
it stemming from the non-UAE Islamic outlier that alone brings in as much as 1. Even so, it is
impossible to reject the null hypothesis of there being no material differences with respect to
any fine or aggregated identity gaps.
By contrast, a t-test will be run for Islamic versus non-Islamic and UAE versus non-
UAE or weakly dual identities. Notably, for two subsamples of unequal size, the resultant t
statistic could be inferred as (Hayter 2012, p. 402),
t=μ ISL−μnonISL
√ σ ISL2
nISL+
σ nonISL2
N−nISL
Largely due to the size asymmetries for N=9 degrees of freedom, the tests for Islamic
versus UAE identity gaps yields 1.716 and 2.02 respectively, which barely suffices to reject
the null hypothesis of similarity on margin at about 6% and 2% significance levels,
respectively. In other words, resorting to rigorous tools and tests did not secure any major
qualifying concerns on the initial, naïve analysis. In fact, the latter has been supported
consistently.
Conclusion
The present study aims at looking into the key and intertwined drivers of efficiency for
UAE banks, in particular what pertains to approaches and mechanisms distinguishing between
Islamic versus conventional finance. It has been shown that, for all practical as well as formal
purposes, efficiency as risk-adjusted versus leveraged slack picking could be represented in
terms of ROE and its underlying Du Pont decomposition touching on areas as diverse as
profitability margin, financial leverage, and scale efficiency or operating leverage.
A set of related modeling tools has been proposed to arrive at similar findings as
stemming from very distant sets of assumptions. In particular, the inter-relationship between
reserves and dividend policies versus ROE performance has been treated from a variety of
standpoints. Lagged responses have been modeled in ways that are economical as well as
autocorrelation sparing. For that matter, the very design of the grand approach and particular
models has economized on extra dimensions without compromising explanatory rigor and
predictive validity.
Remarkably, the bulk of the selection analysis was done based on design-driving
theoretical considerations as well as study-accommodating specifications that delineate
structure as a robust choice beyond sample-specific tests or manipulative data mining. The
empirical part is closely related to a formal modeling rationale in ways that secure parsimony
and bootstrapping with an eye toward hard data constraints.
The less formal qualitative analyses as well as technical scrutiny unequivocally point
to very minor distinctions between the alternate models in unadjusted terms. On second
thought, in terms of Sharpe-like performance net of excessive risks, which is particularly in
line with the Islamic discipline, Islamic and UAE identity has lent itself with superior
performance. Whereas dividends and reserves have posited a controversial agenda in its own
right, as evident in the initial theoretical discussion, still the reportedly insignificant linear
impact amidst an apparently inverse inter-linkage with respect to ROE and ROA alike may
hint at non-linear response patterns yet to be addressed in future research. On second thought,
it remains to be seen whether the perceived non-linearity hides the utter complexity of
underpinning decisions that cannot otherwise be reduced to any elegant or generalized
mechanisms or rationales. In fact, this agenda reaches far beyond residual minimization or
predictive pragmatics.
One further direction could be attempted, with the Islamic identity dummy acting as a
dependent variable running on the rest as explanatory regressors rather than the other way
around. In a sense, this dual setup would appear to be complementary with respect to the one
undertaken in this study, and largely descriptive or positive in determining what an Islamic
identity is about rather than normative in judging on what an optimum identity should be.
References
Taufiq Hassan Shamsher Mohamad Mohammed Khaled I. Bader, (2009),"Efficiency of
conventional versus Islamic banks: evidence from the Middle East", International Journal of
Islamic and Middle Eastern Finance and Management, Vol. 2 Iss 1 pp. (46 – 65)
Hussein A. Hassan Al-Tamimi Husni Charif, (2011),"Multiple approaches in performance
assessment of UAE commercial banks", International Journal of Islamic and Middle Eastern
Finance and Management, Vol. 4 Iss 1 pp. (74 – 82)
Thorsten Beck, Asli Demirguc-Kunt, Ourda Merrouche (2013), “Islamic vs Conventional
banking: Business model, Efficiency and Stability”, Journal of Banking and Finance, Volume
27 (433-447)
Fatima S. Al Shamsi , Hassan Y. Aly & Mohamed Y. El-Bassiouni (2009) Measuring and
explaining the efficiencies of the United Arab Emirates banking system, Applied Economics,
41:27, (3505-3519)
Hussein A. Hassan Al-Tamimi Faris Mohammed Al-Mazrooei, (2007),"Banks' risk
management: a comparison study of UAE national and foreign banks", The Journal of Risk
Finance, Vol. 8 Iss 4 pp. 394 409
Bashir, M. (1999), “Risk and Profitability Measures in Islamic Banks: The case of two
Sudanese banks, Islamic Economic Studies, 6(2), (1–24).
ABN Amro 2014, Annual report, viewed 8 July 2015, < http://www.abnamro.com >
ABN Amro 2011, Annual report, viewed 8 July 2015, < http://www.abnamro.com >
Abu Dhabi Exchange 2014, Abu Dhabi Islamic Bank annual report, viewed 9 July 2015, <
http://adx.ae >
Brigham, EF & Ehrhardt, MC 2011, Financial management: theory and practice, 13th edn,
Cengage, New York
Hayter, A 2012, Probability and statistics for engineers and scientists, 4th edn, Cengage, New
York
Mishkin, F 2010, The economics of money, banking, and financial markets, 9th edn, Pearson,
New York
UAE Central Bank 2015, Database of banks, viewed 7 July 2015, <
http://www.centralbank.ae/en/index.php?
option=com_content&view=article&id=149&Itemid=97# >
Verbeek, M 2000, A guide to modern econometrics, Wiley, New York
Appendix
Appendix A
Table 2A: ROE Based Qualitative Cross-Sectional Comparison (By Author 2015)
ROE
average D^ISL
D^non-
ISL D^UAE
D^non-
UAE D^ISL&UAE
D^non-ISL &
non-UAE
AAIB 15.25% 0 1 0 1 0 1
ABK 9.08% 0 1 1 0 0 0
ABN 17.15% 0 1 0 1 0 1
ADCB 11.46% 1 0 1 0 1 0
ADIB 12.45% 1 0 1 0 1 0
ARBIFT 4.10% 0 1 1 0 0 0
Baroda 16.46% 0 1 0 1 0 1
BLOM 15.25% 0 1 0 1 0 1
BNP 8.35% 0 1 0 1 0 1
CBD 15.09% 1 0 1 0 1 0
CBI 7.78% 0 1 1 0 0 0
DIB 14.45% 1 0 1 0 1 0
FGB 15.90% 0 1 1 0 0 0
HBL 18.17% 0 1 1 0 0 0
HSBC ME 17.68% 0 1 0 1 0 1
Janata 2.40% 0 1 0 1 0 1
Mashreq 12.96% 1 0 1 0 1 0
NBAD 14.41% 0 1 1 0 0 0
NBB 15.90% 0 1 0 1 0 1
NBF 9.59% 0 1 1 0 0 0
NBO 13.97% 0 1 0 1 0 1
NBQ 10.62% 0 1 1 0 0 0
NoorI 11.56% 1 0 1 0 1 0
RAK 24.63% 0 1 1 0 0 0
Sharjah 8.70% 0 1 1 0 0 0
SharjahI 6.39% 1 0 1 0 1 0
UAB 17.85% 0 1 1 0 0 0
count (sub- 7 20 18 9 7 9
sample size)
weight .259 .741 .667 .333 .259 .333
average 12.05% 13.16% 12.51% 13.60% 12.05% 13.60%
SD .0285 .0542 0.0488 .0502 .0285 .0502
ratio 4.231 2.429 2.565 2.711 4.231 2.711
Table 2B: ROA Based Qualitative Cross-Sectional Comparison (By Author 2015)
ROA
average D^ISL
D^non
-ISL D^UAE
D^non-
UAE D^ISL&UAE
D^non-ISL &
non-UAE
AAIB 1.63% 0 1 0 1 0 1
ABK 1.35% 0 1 1 0 0 0
ABN .57% 0 1 0 1 0 1
ADCB 1.41% 1 0 1 0 1 0
ADIB 1.49% 1 0 1 0 1 0
ARBIFT 1.07% 0 1 1 0 0 0
Baroda .95% 0 1 0 1 0 1
BLOM 2.45% 0 1 0 1 0 1
BNP .34% 0 1 0 1 0 1
CBD 2.43% 1 0 1 0 1 0
CBI 1.14% 0 1 1 0 0 0
DIB 1.78% 1 0 1 0 1 0
FGB 2.64% 0 1 1 0 0 0
HBL 1.65% 0 1 1 0 0 0
HSBC ME 1.52% 0 1 0 1 0 1
Janata .55% 0 1 0 1 0 1
Mashreq 1.91% 1 0 1 0 1 0
NBAD 1.59% 0 1 1 0 0 0
NBB 1.95% 0 1 0 1 0 1
NBF 1.32% 0 1 1 0 0 0
NBO 1.91% 0 1 0 1 0 1
NBQ 2.63% 0 1 1 0 0 0
NoorI 1.29% 1 0 1 0 1 0
RAK 4.33% 0 1 1 0 0 0
Sharjah 1.84% 0 1 1 0 0 0
SharjahI 1.50% 1 0 1 0 1 0
UAB 3.19% 0 1 1 0 0 0
count (sub-
sample size) 7 20 18 9 7 9
weight .259 .741 .667 .333 .259 .333
average 1.69% 1.73% 1.92% 1.32% 1.69% 1.32%
SD .0039 .0096 .0084 .0074 .0039 .0074
ratio 4.293 1.798 2.287 1.776 4.293 1.776
Table 3A: Cross-Sectional, Counts Based Regression
Regression output
R-squared .263796018
Modified R-squared .084486203
Standard error
2.19517418
4
Observations 27
df SS MS F
F
hurdle
Regression 5 37.98662662 7.597325323
1.97075557
3 2.13
Residual 22 106.0133734 4.818789699
Total 27 144
Coefficients SE t-stat p-value t hurdle
Intercept
4.79104355
5 1.090346987 4.394054013 .000230413 1.323
D^ISL -.404150999 1.069688266 -.377821289 .709184079
D^UAE -1.35090712 1.098274581 -1.230026756 .231678664
D*D 0 0 65535
D^(c+r) -.175414787 .158482535 -1.106839857 .280319295
D^ROA .452262114 .177582207 2.546776068 .018382884
Table 3B: Cross-Sectional, Averages Based Regression
Regression output
R-square .263796018
Modified R-
square .084486203
Standard error .243908243
Observations 27
df SS MS F
F
hurdle
Regression 5 .468970699 .09379414 1.970755573 2.13
Residual 22 1.308807079 .059491231
Total 27 1.777777778
Coefficients SE t-stat p-value t hurdle
Intercept .532338173 .121149665 4.394054013 .000230413 1.323
D^ISL 0 0 65535
D^UAE -.150100791 .122030509 -1.230026756 .231678664
D*D -.044905667 .118854252 -.377821289 .709184079
D^(c+r) -.175414787 .158482535 -1.106839857 .280319295
D^ROA .452262114 .177582207 2.546776068 .018382884
Table 3C: Cross-Sectional, Counts Based Regression—No Intercept
Regression output
R-squared .781739944
Modified R-squared .66631472
Standard error
2.94185072
4
Observations 27
df SS MS F F
hurdle
Regression 5 712.9468293 142.5893659
20.5947197
7 3.94
Residual 23 199.0531707 8.654485681
Total 28 912
Coefficients SE t-stat p-value t hurdle
Intercept 0 n/a n/a n/a 2.508
D^ISL 0 0 65535
D^UAE -.423604612 1.444419067 -.293269884 .77194571
D*D -.109862719 1.430724316 -.076788182 .939456127
D^(c+r) .166112689 .185093086 .897454857 .378774885
D^ROA .831178367 .20804276 3.995228517 .000569247
Table 3D: Cross-Sectional, Averages Based Regression—No Intercept
Regression output
R-squared .781739944
Modified R-squared .66631472
Standard error .326872303
Observations 27
df SS MS F
F
hurdle
Regression 5 8.801812708 1.760362542
20.5947197
7 3.94
Residual 23 2.457446551 .106845502
Total 28 11.25925926
Coefficients SE t-stat p-value t hurdle
Intercept 0 n/a n/a n/a 2.508
D^ISL
-.01220696
9 .158969368 -.076788182 .939456127
D^UAE
-.04706717
9 .160491007 -.293269884 .77194571
D*D 0 0 65535
D^(c+r) .166112689 .185093086 .897454857 .378774885
D^ROA .831178367 .20804276 3.995228517 .000569247
Table 3E: Cross-Sectional Reduced, Averages Based Regression—No Intercept
Regression output
R-squared .773815129
Modified R-squared .724767734
Standard error .319166045
Observations 27
Analysis of
variance
df SS MS F
F
hurdle
Regression 2 8.712585153 4.356292576
42.7645273
2 5.57
Residual 25 2.546674107 .101866964
Total 27 11.25925926
Coefficients SE t-stat p-value t hurdle
Intercept 0 n/a n/a n/a 2.485
D*D -.017646649 .137657237 -.128192669 .899022167
D^ROA .897677338 .109935941 8.165458305 1.61456E-08
Appendix B: Simulated ROE Distribution
Figure 1A: A Plot for ROE(c) (By Author 2015, from MS Excels)
5% 10%
15%
20%
25%
30%
35%
40%
45%
50%
55%
60%
65%
70%
75%
80%
85%
90%
0%5%
10%15%20%25%30%35%40%45%
ROE
ROE
Figure 1B: A Plot for ∂ ROE
∂ c (By Author 2015, from MS Excels)
c 5% 10%
15%
20%
25%
30%
35%
40%
45%
50%
55%
60%
65%
70%
75%
80%
85%
90%
00.020.040.060.080.1
0.120.140.160.180.2
ROE'
ROE'
Figure 2A: A Plot for ROE (c, r)=ROE(c+r) (By Author 2015)
5%10%15%20%25%30%35%40%45%50%55%60%65%70%75%80%85%90%
-200%
-150%
-100%
-50%
0%
50%
100%
5%30% 55%80%
Figure 2B: A Plot for ∂ ROE
∂ r=∂ ROE
∂ c (By Author 2015)
5%10%
15%
20%
25%
30%
35%
40%
45%
50%
55%
60%
65%
70%
75%
80%
85%
90%
-15
-10
-5
0
5
10
15
5%20%35%50%65%80%
Appendix C: Graphic Output on Panel
Figure C1: Visualized Joint Distributions in the Data (By Author 2015, from SPSS PASW
17.0)
Figure C2: Zooming in on ROE Distribution (By Author 2015, from SPSS PASW 17.0)