Strategic Groups in the French Franchising Sector

Abstract

We provide a picture of the French franchising sector, based on the strategic group approach. We use a recent 4-year panel dataset from the French Federation of Franchising, for the period 2010–2013, and sophisticated statistical and supervised learning models. Five main strategic groups of franchisors are distinguished in the French system, characterized by specific strategies and performance outcomes. We first survey the literature dealing with strategic groups and then conduct a multidimensional statistical analysis (Principal Components Analysis and Ascending Hierarchical Clustering), highlighting three factorial axes and five clusters. We test the stability of network behaviors with a classification model. Finally, we observe and comment on the differences in strategic group performances.

Appendices

Appendix 1: Principal Component Analysis

Table 7 presents the explained variance of the resulting factor analysis. Usually, the number of factors is given by the number of eigenvalues higher than 1. Although the number of eigenvalues that are higher than 1 is 4, we make the choice to keep only three of them, considering the gap between the third and the fourth eigenvalue. In other words, since the amount of variation explained drops after the third principal component, we keep the three first components which explain more than 52% of the total variance.

Table 7 Explained variance of resulting factor analysis

To detect clusters, we use three factors in the PCA analysis, we overcome the potential correlation between variables using these factors, and the resulting PCA factors are uncorrelated. The analysis of factor correlations presented in Table 8 highlights three main factorial axes. The first is related to the financial obstacles for the franchisees. It associates higher financial barriers for the franchisees (higher entrance fees and initial investment, high surface area and estimated turnover) with a lower franchise rate (and vice versa). Higher financial barriers to franchising for franchisees are associated with a longer contract term. Hence, this factor also relates to the length of the payback period for the franchisee. This axis can be labeled “(financial) easiness and quickness of payback period for the franchisee.”

Table 8 Correlation between the principal components and the variables

The second factor associates high age, high experience before franchising, large size, and low entrance fees and financial investment for franchisees with a low rate of franchising. Hence, this axis opposes two types of networks. The first group consists of networks that are well endowed with human and managerial capital (high age, experience, and size) and also with financial capital (large size and low investment and entrance fees from the franchisees). These networks do not make much use of franchising. On the other end, this axis presents networks with a low endowment of human and managerial capital (low age, experience, and size) and requiring substantial financial resources from their franchisees (high initial investment and entrance fees). These networks have a high franchising rate. Thus, this axis can be labeled “franchising as a means to overcome resources scarcity.”

The third relevant factorial axis is more about the value-added rate of the activity and may encompass the sectorial affiliation of the network. On one extremity are highlighted networks with large outlets and a high predicted turnover, in addition to a low royalty rate, entrance fees, and experience before franchising the first unit. This picture is consistent with low value-added activities. On the other extremity are networks with higher royalty rates, entrance fees, and experience before franchising. The typical outlet as described by the franchisor is smaller. This picture is consistent with networks having higher value-added activities. This factorial axis is related to the value added and the complexity of network activities. Thus, it is relevant to assume that this axis integrates sectorial effects (e.g., services vs. retail) as it refers to the size (turnover and surface area) of the outlets. We label this factorial axis “value-added rate.”

Appendix 2: Predictive Modeling

First, we split the data into training (2/3), to evaluate the model, and test set (1/3). The test set, i.e., the part of data never used in the training process, will be only used to validate the model.

We apply random sampling using stratified random methods (Kruskal and Mosteller 2004) and obtain a balanced training set and test set taking into account the cluster distributions in the original sample (Table 9). Then we normalize the variables since they have heterogeneous measures.

Table 9 % of stratified random sample

We use ten repeats of tenfold cross-validation (Picard and Cook 1984). This method aims at providing a nonbiased estimation of model errors. The results from the folds are averaged to produce a single estimation of accuracy. Accuracy is calculated for each model and represents the proportion of the total number of correct predictions. The confusion matrix shows (Table 10) the number of correct and incorrect predictions made by the classification model, compared to the actual outcomes in the data. In our case the matrix is 5 × 5, where 5 is the number of target values (clusters). The following table displays a 5 × 5 confusion matrix for five clusters. So we estimate the rate error using ten repeats of tenfold cross-validation.

Table 10 Theoretical confusion matrix and accuracy formula

Table 11 compares the results of four classification models (support vector machines, random forest, logistic regression, and conditional inference tree). The best model is support vector machine, with more than 90% of correct predictions.

Table 11 Model evaluation