4. Data
4.2 Firm data
4.2.4 Potential bias
The dataset used for the estimation contains all available bank connections of firms that satisfy the requirements. If firms have more than one bank connection, then they appear more than once in the dataset. It is interesting to examine if it has an effect on the firms’ switching costs. A simple variable that contains the number of bank connections each firm has in the dataset is therefore included in the estimation.
The last variable that is included in the dataset is the year of establishment. It is not a continuous variable and there are too many years to include each year as a dummy. Therefore the seven groups used in Figure 5 are used. The groups are included in the estimation as six dummy variables, with the group of firms established before 1950 as the default group. Each dummy thus represents the
switching costs of the firms that were established in the time period, compared to firms that were established before 1950.
year a firm were established, but rather on other characteristics of the firms. In the following size of firms will be examined, as it should be highly correlated with the other characteristics. The size of the firm is thus used a proxy of the firms’ demand. The relevancy of the examination depends on how well the size of the firm explain the demand of the firms.
The supply of credit is based on the firms’ ability to fulfill their credit commitments. This is mostly depending on future factors and therefore hard to measure. Usually a subjective assessment by each bank decides the terms of the contract. It is not possible to observe the individual contracts, so a formal examination of this is not possible. The subjective assessment is however partly based on financial figures, so it is likely to be correlated to the observed figures in the estimation.
If the newly established firms have different preferences than the existing firms, observed
differences in the market shares can be caused by consumer heterogeneity and not the lock-‐in effect.
The preferences of consumers are likely to depend on the size of the consumer, it is therefore important that the two distributions of markets shares that are compared are based on a
homogenous distribution of firms. The first examination looks at the distribution of both the median balance and median capital of each group depending on year the firms were established. The median is presented instead of the average, because the switching cost estimates are based on the absolute number of firms that each bank serve in each period. Thus each firm has identical weights,
independent of their characteristics, when estimating the switching costs4. The amount of firms in each category is also added to the graph, to illustrate the weight each group has in the estimations.
The financial figures approximate the size of the firm. If the sizes of firms are independent of the year the firms were established, the median financial figures would be the same for each group of firms.
Below are the two aforementioned graphs.
4 The averages are similar but influed by the large firms, most notably in the groups consisting of the oldest firms.
Figure 6: Median capital of firms distributed according to year of establishment.
Figure 7: Median balance of firms distributed according to year of establishment.
It is clear that firms are not homogenously distributed according to size over the year they were established. Firms established a longer time ago are generally larger than newly established firms, which can be seen from the larger median capital and median balance. The median capital of firms, illustrated by Figure 6, is largest for the two groups containing the oldest firms. It is constant for the group of firms established in 1960-‐1989, but considerably smaller for the group of firms established in 2011 and after. The median balance of firms, illustrated by Figure 7, is continuously decreasing from the group of firms established the longest time ago, to the group of firms established in 2011 and after. These graphs are not surprising, as firms are likely to either grow or go out of business.
This can unfortunately potentially bias the observations, if the sizes of the firms are also correlated with the choice of bank. As mentioned before firms are not weighted according to their size in the switching cost estimations, but all count equally towards the market shares. In Figure 7 and Figure 8 the number of firms in each group are also graphed. The number of firms displays how much each group contribute to the estimations. The graph of the number of firms is in line with the tendency for
firms to either grow or go out of business. While the medians indicate that the estimation is
potentially biased, the number of firms in each group reduces this effect. The groups with the largest medians that potentially bias the estimation are very small compared to the groups with medians most similar to those of the new firms. The groups most similar to the newest firms are the largest groups and because firms are not weighted, this suggest that while the estimation does have
potential to be biased, the issue is not as severe as the medians suggest without taking the size of the groups into account.
One way to reduce the potential bias that arise because of firm heterogeneity, could be to use another group of firms that consists of firms that has been established after 1990, or some other year, and compare that group to the group of new firms. The same method for estimating the
consumer switching costs could then be applied to the two distributions of market shares. As the two groups of firms are more similar than the total population of firms, it would reduce the potential bias.
The group is not more correct than the original but will contain firms that are more similar to the newly established firms and still have a significant number of observations. The disadvantage of this approach is that many firms that could potentially have large switching costs, due to a longer lasting relationship with a bank, will be excluded. The alternative group definition yields very similar results, but with lower statistical power – as would be expected. The results are not reported independently.
It is evident that the size of the firms are correlated with the year the firms were established. It will bias the estimations if the sizes of the firms have an affect on the firms’ choice of bank. If all banks serve consumers of the same size, the market share differences will not be biased as a consequence of size differces of firms. To examine if the size of the firms are correlated with the choice of bank, the median capital and median balance are graphed for each bank. The interesting aspect is the
difference of the medians across banks. If some banks banks serve customers with low medians, ie small firms, the market share of new firms will be to high compared to the situation with
homogenous firms, and likewise will banks that serve firms with high medians have a high market share among the existing firms. This means that the switching cost estimates of banks that serve small firms will be biased such that their switching cost estimates are too low, and similarly will the switching cost estimations of banks that serve large firms will be too high. Below are the median
capital and median balance of firms that are customers at the ten largest banks graphed on separate graphs. The ten largest banks accounts for the majority of the market, so for a better overview the smallest banks are left out. Graphs including all banks can be found in appendix A-‐1.
Figure 8: Median capital of firms served by the ten largest banks.
Figure 8 illustrates that there are differences in the amount of capital that the customers of the banks have. Bank number three and bank number five to ten serve customers with the same median
capital, while other banks serve customers with a higher median capital. The fourth largest bank serves customers with much larger capital than the other banks, while the two largest banks are also quite a bit above the other banks. These banks, especially bank number four, will have a tendency to have a too high market share among existing customers, compared to new customers, as they
generally serve larger customers. This will bias their switching costs upwards. The amount of firms these banks serve accounts for a large part of the market, which suggest that the switching costs of those banks could be upwards biased. Generally however it seems that there is not much difference
between the sizes of customers that most of the banks serve. So while Figure 6 and Figure 7 suggests that there are significant size differences between the groups, Figure 8 suggests that there is not a general tendency for banks to serve customers of different size.
Figure 9: Median balance of firms served by the ten largest banks.
The median balance of firms served by each bank has more variance than the median capital. The four largest banks serve firms with comparable balance, but there seems to be a general tendency for smaller banks to serve firms with a smaller balance. This problem is enhanced as the large banks serve most of the market, so a large part of the market is affected by potentially biased switching costs. Figure 7 shows that the existing firms are generally larger and Figure 9 shows that the larger banks generally serve larger customers. As before this has the consequence that the switching costs of the larger banks are potentially upwards biased.
The potential bias ascending because of heterogeneity of firms cannot be controlled for in the
estimations conducted in this thesis, and is therefore an unavoidable downside to using this method to estimate consumer switching costs. The potential bias is of less importance when doing the regression of firm characteristics on switching costs, as the switching cost estimates are all
potentially upwards biased, but the concern of the estimation, is only the sign and significance of the parameters. The sign of the parameters will not change if the switching cost estimates are increased, but their individual significance may be overestimated. Chen and Hitt estimate consumer switching costs with and without controlling for consumer heterogeneity and find no substantial differences between the two estimations (Chen & Hitt 2002). This suggests that their estimation of consumer switching costs is not very sensitive to consumer heterogeneity. If the estimation in this thesis yields the same traits, the estimates should be practically unbiased.