Professional Documents
Culture Documents
Introduction:
The introduction of telecommunication industry has been pretty late in our country compared to that of the neighboring countries. This is very much attributable to the political unrest in early 1990s. Though the first telecom company, Citycell, was introduced ages ago, the device has not become so pervasive until 1997, the year when the biggest telecom of the country, GrameenPhone (GP) hit the market with its GSM technology. Since then there was no looking back. The industry grew at such an incredible rate in just a decade that anyone could hardly imagine. Now there are a number of players battling so hard for their respective market share and the consumers as well as the economy benefitted tremendously from this fierce competition. The following sections elaborate on almost every aspect of the industry, ranging from how it become so big, what fueled the growth to what are the risks and opportunities that it currently renders for the players.
1.1 Objectives:
a) Broad Objective: The broad objective of the research is: To analyze the relationship between the advertisement expenditure and the number of subscription in the telecom industry of Bangladesh. b) Specific Objectives: Specific objectives of this research are: o To give a brief overview of the telecom industry. o To demonstrate the ad spends and changes in the number of subscribers over time for six major mobile operators of Bangladesh. o To find out the correlation between the advertisements spend and number of subscription.
1.2 Methodology:
Research Scope: In this research paper the effectiveness of advertisement expenditure and the change in the number of subscribers for different operators in Bangladesh is analyzed and evaluated. The correlation between these two variables is focused on.
Page 1 of 24
Research Design: The research is exploratory in nature as there the problem has not been clearly defined. Type of Data: The underlying research on the Effectiveness of Advertisement on Number of Subscribers in the Telecom Industry of Bangladesh is based on secondary data source and is analyzed to get an insight of the actual situation. Methods of Data Collection: The data is collected from the Software Media Express and AdEdge by Sirius Communication, and different articles, journals, and web sites.
1.3 Expected Result:
There is a common concept that if any company increases their advertisements spend their number of subscribers will surely increase. In this research this assumption has been verified. This research is to focus on the correlation between advertisement spend and the increase of the number of subscribers over time. This paper will be useful to those who want to work on this area.
1.4 Limitations:
Unavailable data: As the idea of people meter software was conceived by the end of 2006, the related data on advertisement spend of different mobile operators of Bangladesh before 2007 were unavailable.
Limited access: Data used in this research are monitoring data taken from the software, as the researcher did not have access to the real data from the companies.
Lack of experience and time: A research as broad as this needs proper experience in the researchers side where as a longer period of time is needed to.
Page 2 of 24
currently nearly 2.6 million EDGE/GPRS users in the Grameenphone network. Today, Grameenphone is the leading telecommunications service provider in Bangladesh with more than 36 million subscribers as of December 2011.
Robi - Robi Axiata Limited is a joint venture company between Axiata Group
Berhad, Malaysia & NTT DOCOMO INC, Japan. It was formerly known as Telekom Malaysia International (Bangladesh) which commenced operations in Bangladesh in Page 4 of 24
1997 with the brand name AKTEL. On 28th March 2010, the service name was rebranded as Robi & the company came to be known as Robi Axiata Limited. Robi is truly a people-oriented brand of Bangladesh. Robi, the people's champion, is there for the people of Bangladesh, where they want & the way they want. Having the local tradition at its core, Robi marches ahead with innovation & creativity. To ensure leading-edge technology, Robi draws from the international expertise of Axiata & NTT DOCOMO INC. It supports 2G voice, CAMEL Phase II & III & GPRS/EDGE service with high speed internet connectivity. Its GSM service is based on a robust network architecture & cutting edge technology such as Intelligent Network (IN), which provides peace-of-mind solutions in terms of voice clarity, extensive nationwide network coverage & multiple global partners for international roaming. It has the widest International Roaming coverage in Bangladesh connecting 600 operators across more than 200 countries. Its customer centric solution includes value added services (VAS), quality customer care, easy access call centers, digital network security & flexible tariff rates.
operation outside of India. The company launched its mobile services in Sri Lanka in January 2009 on a state-of-the-art 3.5G network. On July 19, 2007, the company crossed the 1 million customers mark in the first 70 days of operation. Airtel Bangladesh had 5.045 million subscribers as of June 2011.
Page 6 of 24
3.
The data has two main variables Ad Spending (A), and Number of Subscribers (S). In the numbers, both are measured in Million units. By number of subscribers the customers who are using the Sim card of a Mobile Phone Operator in a given time are concentrated on.
The monthly data spans from May 2007 to 2011 August, i.e. 52 months. There are six mobile phone operators under consideration Grameen Phone, Banglalink, Airtel, Citycell, Robi and Teletalk. The August 2009 entry is missing for Teletalk, and appropriate missing value corrections were taken during the data analysis.
Number of Subscribers is the variable of interest, also called the response variable or dependent variable in regression terminology. Ad Spending is the explanatory variable, also known as the predictor variable in regression. The investigation will be on whether Number of Subscribers is in any way related to Ad Spending. Intuitively, it is thought that increase in ad spending will ideally bring in more subscribers.
When the Number of Subscribers is simply plotted across time in the next plot, it is seen that for all the operators, subscriber base increases in time, which is only natural to expect, given the steady growth of population and the spread of technology. This automatically tells a keen observer that time could be another explanatory variable for this data analysis. And so, Time (T), measure in month units, is also included in the data. Overall, the trend for Number of Subscribers looks to be a linear growth with Time. So, the principle of parsimony tells that while performing regression, there is no worth in including terms of T higher than first order, e.g. T2, T3, or, for example, using other models like exponential.
Page 7 of 24
Next, there is the other variable Ad Spending. Unlike Number of Subscribers, which is pretty smooth over time, this has heavy fluctuations, and the variable itself is distributed quite haphazardly Since linear regression works best with, and often assumes, an underlying Normal (Gaussian) distribution, a Quantile-Quantile Plot is made, separately for each operator. The Q-Q plots, and also the histograms, show that this variable is quite positively skewed (i.e. more mass at the smaller values, and a long, heavy tail).
Page 8 of 24
So, in order to standardize it, a log transform on the sample values is applied. The QQ plot with the transformed values show that the data distribution is more normal now, with only a few extreme values deviating from the diagonal line.
Page 9 of 24
From now on, whenever the variable Ad Spending is referred to, it would mean the log-transformed values, not the original ones. A plot of this variable across time for the various operators is shown next.
Page 10 of 24
It is seen that the nature of Ad Spending is not smooth over time. The high level of fluctuations also tells right away that, the effect of Ad Spending on the Number of Subscribers, if any, would not be very high, since the other one is considerably smooth. One can also easily spot some patterns, for example, that the trend in spending by Citycell actually decreased over time, while the trend for Robi increased at a decently high rate.
Page 11 of 24
3.2 Subscriber Takeover A behavior of interest would be subscriber takeover whether a telecom operator can take away a part of the subscriber base of another operator, probably through aggressive advertising or by offering bonus features etc. Also if a potential new customer of one operator joins another instead, that will be reflected in the reduced growth rate of the first operator. While the first-hand data of subscribers actually switching operators is not available, an idea is generated by focusing on the behavior by comparing the Number of Subscribers data across different operators.
But first, the linear trend over time is to be removed from these numbers. Otherwise, it is the strongest component of the data, and any other effects are masked. The trend is removed via a linear regression of Number of Subscribers against Time, and the residuals are taken for our purpose. The residuals are shown in the next plot.
Page 12 of 24
<
Now, looking at the residual subscriber numbers, it can be observed how they
compare to each other. For this, the pairwise correlations for all the operators are looked at. In the next plot, the correlation matrix for the six operators is represented. The color indicates the direction of the association black if the two operators are positively correlated, and grey if negatively correlated. The size of the marker denotes the strength of the association the larger the magnitude (absolute value) of the correlation, the larger the marker. There are no diagonal entries, obviously, because every variable is perfectly correlated with itself, so it is of no interest to plot.
Page 13 of 24
It is seen that Grameen Phone, Banglalink and Airtel have similar subscriber patterns, while that of Citycell is rather the opposite of them. Robi and Teletalk do not show any markedly strong relation with any operators.
3.3 Regression Studies: In the initial analysis, some basic adjustments were made to the data via transformations, & also got insights into what kind of models would be suitable for the subsequent regression analysis. It was also ensured that the variables are continuous numbers (not categorical or binary etc.) & are not oddly behaved, which allows to perform simple Linear Regression using Ordinary Least Squares. To recapitulate, the response variable is Number of Subscribers (S), & explanatory variables are Time (T) & Ad Spending (A). A is already being log-transformed. It is also seen that S is significantly (p-value almost equal to 0) dependent on T, linearly, for every mobile phone operator. Next, an interesting question is whether the other explanatory variable Ad Spending (A) also depends on Time (T). Just like Number of Subscribers, a linear regression is done. The regression coefficient (b), P-values, and significance levels of those values are shown in the next table for each operator. Operator Grameen Phone Banglalink Airtel Citycell Robi Teletalk Coeff. (b) 0.008 0.032 -0.023 -0.021 0.044 -0.004 P-value 0.025 0 0.111 0 0 0.709 ***** ***** Significance * *****
The notation for significance is as follows if the coefficient is significant at 5% level, i.e. P-value < 0.05, is denoted by *; if it is significant at 0.1% level, i.e. P-value < 0.001, is denoted by *****; if the coefficient is not significant at all, it is left blank. When using simple linear regression model where response variable is Number of Subscribers (S) & Ad Spending (A), the result is:
Page 14 of 24
For Grameen Phone, the RSquare is 0.0256, which means that the correlation between the two variables is very low to bring a
spend is increased. Although the value of the regression coefficient, 2.32, seems large and positive (a positive coefficient means that increased ad spending would lead to increasing subscriber numbers), it is actually not statistically significant at all (P-value 0.257), which means that ad spending and subscriber number is not associated in fact, and as a result, increased ad spending would not affect the number of subscribers.
For Banglalink, the R-Square is 0.2676, which means that the correlation between the two variables is moderate, and there will be change to some extent number in if the ad subscriber spend is
increased. The value of the regression coefficient, 3.009, seems large and positive (a positive coefficient means that increased ad spending would lead to increasing subscriber numbers), and here it is actually statistically significant (P-value 0), which means that ad spending and subscriber number is really associated. Therefore, increased ad spending can actually have an effect on the number of subscribers.
For Airtel, the R-Square is 0.000047, which means that the correlation between the two variables subscriber number and ad spend is too low. And the value of the
Page 15 of 24
regression
coefficient,
0.0047, is also very small (it is positive, but that is of no real consequence), and in the plot the regression line looks almost indeed, horizontal. the And
regression
coefficient it not statistically significant at all (P-value 0.962), which means that ad spending and subscriber number is not associated. So, increased ad spending would not affect the number of subscribers.
For Citycell, the R-Square is 0.1855, which means that the correlation between the two variables is moderate, and there will be change to some extent number in if the ad subscriber spend is
0.196, is also moderate, and noticeably, negative (a negative coefficient means that increased ad spending would actually lead to fewer subscriber numbers). The value is actually statistically significant (P-value 0.001), which means that ad spending and subscriber number is indeed negatively associated, and as a result, increased ad spending can have a negative effect on the number of subscribers.
For Robi, the R-Square is 0.3622, which means that the correlation between the two variables is moderate, & there will be change to some extent in the subscriber number if ad spend is increased. The value of the regression coefficient, 1.43, seems large and positive (a positive coefficient means that increased ad spending would lead to
Page 16 of 24
increasing
subscriber
value 0), which means that ad spending & subscriber number is in fact associated. Hence, increased ad spending can actually have a positive effect on the number of subscribers.
For Teletalk, the R-Square is 0.0399, which means that the correlation between the two variables is quite low, and there might not be in a the
significant
change
0.0249, is also low. It is negative (a negative coefficient means that increased ad spending would actually lead to fewer subscriber numbers), but with such a small magnitude, the sign is probably not important. As one can expect, the regression coefficient it not statistically significant either (P-value 0.162), which means that ad spending and subscriber number is not associated. So, increased ad spending would not actually affect the number of subscribers.
As it appears, not all operators ad spend significantly depend on time. Referring back to the plot & discussion in the first section (section 3.1), it can be seen that how the patterns of different operators differ. And checking association is important because, if both T & A as explanatory variables are kept in the regression model, this means that these variables are significantly correlated themselves. And therefore, the
Page 17 of 24
regression results will be very different when T is kept in the model versus when it is not. From the objective, the initial regression model was to regress Number of Subscribers on Ad Spending. In notations, the model could be concisely expressed as S~A or via the more standard linear model notation, S = c + b A Where c is baseline constant, and b is the regression coefficient. To recall, a P-value is the probability of the regression coefficient b being statistically significant (i.e. significantly different from 0) under the null hypothesis, which is that S is not associated with A.
But seeing the trend of S over time, the model has been extended to include Time as well. So, the bigger model can be written either way as S~A+T S = c + bA + dT Where d is the regression coefficient corresponding to the variable T. It has been seen that d is always significant, since Number of Subscribers strongly depends on Time.
But the more important factor is that, since in some cases T is associated with A itself, that indicates that the behavior of b under the two different models could be very different. And that brings to the tricky territory of Causality, which postulates about the possible underlying mechanisms that lead to a data. It is demonstrated with two different hypothetical situations: (i) Consider the simple situation when, if the operator increases spending towards
advertisement, that directly has an effect on the public perception, and as a result the
Page 18 of 24
number of subscribers increase. This is the case where A is indeed a causal factor behind S. And of course, time T keeps on increasing in the background.
(ii)
As explained in the first section (section 3.1), it is only natural to expect that
subscriber size would increase over time, given various factors such as the steady growth of population and the spread of technology. There are also similar factors, such as increasing cost, which can lead to required expense for a fixed amount of service increase over time that is, S also depends on T. However, it is assumed that S doesnt actually depend on A, that is, increased ad spending does not really bring in extra subscribers. In here, T is the real causal factor behind both S and A, and the similar pattern that will be seen between those two is not because A is controlling S.
Can it be really distinguishable between the two situations from our data? Not really, and that is the disadvantage of doing such exploratory analyses compared to a properly designed experiment; but that is something nearly impossible in most situations, including here.
In the first case, the direct model S ~ A is the proper one, while in the second case, where T is a causal factor, S ~ A + T would be the correct model to consider. Since it is not known which is the real one, the best bet is to fit both. Therefore, in the next table, regression results are shown coefficient, P-value and significance for both the models, for each operator.
A side note is that, if it is recalled that the residuals (R) from the previous section (section 3.2) which were obtained by fitting S on T, it can be realized that fitting the second model S ~ A + T is the same as working with the residuals via the first model, which would now become R ~ A. That is because, S = c + bA + dT is equivalent to R = S dT = c + bA In the following table, the regression R2 obtained in the first graph, i.e. fitting S ~ A, is compared with the other model, S ~ A + T. These are the first two columns of the table. It is to be noted that the R2 in the second model is much higher, which justifies
Page 19 of 24
the addition of the Time variable in the regression, and is consistent with the earlier finding that S is strongly correlated with T. Also, in the last column, the R2 corresponding to fitting R ~ A is presented, which shows the goodness of fit of the b coefficient in the S ~ A + T model. It can be seen that their behavior is quite different in the two situations. It will be shown later that the b coefficients themselves are also quite different.
Now the regression results for both the models are plotted. The first graph presents the data points (via a scatterplot) and the regression line from the model S ~ A. The second graph involves the model S ~ A + T, and plots the data points residuals (R) versus A as well as the fitted regression line.
Page 20 of 24
Page 21 of 24
Now in the following table, the regression results for both the models are presented.
S~A Operator Grameen Phone Banglalink Airtel Citycell Robi Teletalk
Coeff. (b)
Signif. ***** * *
The point to notice from the above table is that the operators who had significant A ~ T association in the previous section (section 3.2), had their findings drastically switched between the two models: Grameen Phone is only significant in the second model, whereas Banglalink, Citycell and Robi show much stronger association in the first model. The interpretation is open to scope, including the decision on which model to apply in reality.
(i) In the first case of Grameen Phone, it can be said that taking away the common factor T has helped bring out the true relationship between S and A in in the data, and a strong linear relationship between R and A can be seen from the plot. (ii) Similar, statistically significant associations, between R and A can be seen in case of Airtel and Teletalk as well, albeit less strong than Grameen Phone. (iii) On the other hand, Citycell and Robi have strong associations in the original data, but that goes away when the effect of time is removed. (iv) The case of Banglalink is rather interesting while a strong positive association in the original data is seen, a remaining significantly negative association between R and A is also seen when Time is adjusted for. In this case, the second model might actually be the true one, with the real causal effect of A on S coming out after adjusting for the common covariate T.
,
Page 22 of 24
Another side note: Since the relationship between S and A is not that straightforward as that of between S and T, we tried fitting higher order regression models to the data, e.g. R ~ A + A2. But it turns out that the higher order polynomial models do not provide any better fit, and so they were not included in the analysis presented.
The highlight of the above regression findings would be that, most of the significant regression coefficients (4 out of 7) are in fact negative! That is quite contrary to the intuition that the subscriber base should increase if more money is spent on advertisement. So, irrespective of which model is true, it might be more prudent not to go overboard with ad spending but concentrate on other factors such as better call quality, better coverage, lower call rates, etc.
Page 23 of 24
Recommendations: As the assumption that the more money spent on advertisement increases the number of subscriber too is proved to be wrong, it might be more prudent not to go overboard with ad spending.
This might even be an indication that the cell phone market is being saturated, so, providers must introduce more value added services, schemes along with better network coverage, better signal quality for the current subscribers.
Teletalk, Citycell must come up with new innovation effectively use their advertisement.
4.2 Conclusion: This research proved one assumption of a regular marketer to be incorrect. It was a common assumption that the more presence in the media electronic or other, the more the number of subscribers will increase. So it is learnt from this research that spending a huge amount is poorly designed marketing campaigns may not make the number of subscribers go up but it can even be negative to some extent.
The mobile telecommunication of Bangladesh is very competitive & the rivals are facing huge competition in the market every day. To compete with the changing customers need and want they need to give more new tariff plans, value added services, better network coverage, and better signal quality and provide interesting schemes to retain subscribers to switch to a different operator. The mobile companies are changing their strategies day to day to survive in this ever changing market to stay in the positive tide and make the profit they desire.
Page 24 of 24