‘Types’ of private family firms: an exploratory conceptual and empirical analysis

Abstract

Family firms that can leverage entrepreneurial experience and knowledge can shape local economic development. Practitioners concerned with fostering enterprise sustainability need to be aware that family firms cite contrasting goals, resource profiles and requirements. Family firms are not a homogeneous entity. The ‘targeting’ of support to ‘types’ of family firms could enable practitioners to satisfy their wealth creation and social inclusion objectives. To stimulate increased critical reflection, insights from agency and stewardship theories were drawn upon to illustrate six conceptualized ‘types’ of private firms based on company ownership and management structures as well as company objectives. Cross-sectional survey evidence was gathered from key informants in family firms in the UK. An agglomerative hierarchical QUICK CLUSTER analysis identified seven empirical ‘types’ of family firms. Four out of the six conceptualized ‘types’ were validated by the exploratory empirical taxonomy. Implications for policy-makers and practitioners as well as researchers are discussed.

Keywords:

• family firms
• enterprise sustainability
• types
• agency and stewardship theory
• cluster analysis.

Acknowledgements

The authors would like to thank the Leverhulme Trust and the BDO Stoy Hayward Centre for Family Businesses for their financial support. Our views do not necessarily coincide with those of the sponsors. Views expressed here have benefited considerably from the comments from two anonymous referees and Bengt Johannisson.

Notes

Notes

1. There is a lack of consensus surrounding the theoretical and operational definition of a family firm (Handler 1989, Litz 1995, Chua et al. 1999). Definitions have been based upon three key issues: majority share ownership (Cromie et al. 1995); whether members of an ‘emotional kinship group’ perceive it to be a family business (Gasson et al. 1988, Ram and Holliday 1993); and management by members of a single dominant family group (Daily and Dollinger 1992). In addition, some researchers have considered multiple conditions (see Westhead and Cowling 1998). Chua et al. (1999: 24) have warned, ‘companies with the same level of family involvement in ownership and management may or may not consider themselves family businesses and, more importantly, may or may not behave as family businesses’. Some have suggested that each private firm should be monitored with regard to the degree of ‘familiness’ (i.e. the extent of family influence) (Astrachan et al. 2002, Habbershon et al. 2003, Klein et al. 2005).

2. The stereotypical family firm is closely-held, family owned and managed with little outside influence, and the firm's objectives are entangled with family objectives.

3. With reference to intentions of being a family firm and the dispersion of ownership and management, Litz (1995) has presented a conceptual model of family firm definitions. The purpose of this study was not to explore the validity of Litz's conceptual model nor Sharma's (2004) two by two conceptual matrix. This study also did not seek to present or validate a general purpose classification that distinguishes family firms from non-family firms or between different types of family firms (Sharma 2002, 2004). Shaw and Wheeler (1985: 255) have asserted that ‘there is no need for concern that classification schemes of the same phenomenon may not yield identical results’.

4. Johannisson and Huse (2000) presented an ideological framework to explain the selection process of outside directors by small family firms. They suggest that organizations enable owners to achieve their goals (or dreams) and organizations are arenas for emotions and politics. This study is complementary to that of Johannisson and Huse (2000). We did not seek to explore the validity of their conceptual model, which was not designed to specifically examine different ‘types’ of private family firms.

5. The aim is to classify cases (i.e. firms) into groups (or clusters) comprising similar individuals, and thereby to separate dissimilar individuals into different groups. Johnston (1980: 219) suggested that a ‘classification should be based on some underlying theory about the nature of a group, even if not about the probable groups in the data set to be explored’. Further, Shaw and Wheeler (1985: 255) have asserted that ‘In all cases the classification should be geared towards the specific needs of the study and should only be undertaken within some framework’. The analyst needs to be selective and it is rarely desirable to consider a very large number of themes and variables.

6. Owners of family firms, who grow their firms in order to employ family members in key managerial positions, may do this at the expense of firm profitability (Singell and Thornton 1997). The favouring of family members may lead to more able non-family managers seeking employment outside the family firm and may also retard firm development.

7. The problem of missing data was considered from the outset of the research design. The random and stratified sampling procedure was selected because it is a probability sampling method. Hair et al. (1995: 44) have asserted that ‘The process of generalizing to the population is really an attempt to overcome the missing data of observations in the sample. The researcher makes this missing data ignorable by using probability sampling to select respondents. Probability sampling allows the analyst to specify that the missing data process leading to the omitted observations is random and that the missing data can be accounted for as sampling error in the statistical procedures’.

8. Content validity was considered and the structured questionnaire was tested during a pilot exercise. An early version was revised in line with comments and feedback from family firm practitioners and academics. To source potential problems and address the problem of face validity, 30 limited liability companies located on the Isle of Wight were contacted to pilot test the revised questionnaire. No major problems were detected.

9. To reduce measurement error (Hair et al. 1995), the questionnaire was sent to key informants (Kumar et al. 1993) who should have had sufficient knowledge to answer all presented questions.

10. Strong positive correlations were detected between the information gathered by the questionnaire and the archival data provided by Dun and Bradstreet relating to business age and employment size (i.e. Pearson correlation coefficients of 0.84 and 0.89, respectively, both significant at the 0.001 level (one-tailed test)). We inferred that the data collected from the key informant was reliable.

11. Chi-square and Student's t test analyses were conducted to detect response bias. With regard to industry; location of the business by standard region as well as ‘assisted’ area location; age; employment size and sales revenue of the company, no statistically significant response differences were detected between the 427 valid respondents and 460 valid non-respondents. This evidence did not eliminate the concern relating to non-response bias, but it did indicate some representativeness.

12. The randomness of the missing data relating to the 272 firms was monitored. Guided by Hair et al. (1995: 53) a missing at random (MAR) randomness test was conducted. Each of the 17 ‘raw’ variables was converted into dichotomous variables. With reference to each variable, valid values were allocated a value of one and missing data allocated a value of zero. A correlation matrix was computed, and only four significant correlations at the 0.05 level (1-tailed tests) were detected. It was concluded that no single missing data process was significantly affecting a substantial number of variables.

13. There are two basic factor analysis methods: namely, common factor analysis and principal components analysis (PCA) (see Davies 1984, Shaw and Wheeler 1985: 278–279, Hair et al. 1995: 375–377 for a detailed summary of both methods). The PCA method is selected if the aim is the identification of uncorrelated linear combinations of the original set of variables (Norusis 1990), and where the minimum number of components are required to account for the maximum portion of the variance represented in the original set of variables (Hair et al. 1995). A closed system is assumed where the statistical variation in the variables is explained by the variables themselves. PCA assumes high correlations between all variables, with high common variances and low unique variances. A component is a linear combination of the variables introduced in to the PCA. Components represent the ‘underlying dimensions (or constructs) that summarize or account for the original set of observed variables’ (Hair et al. 1995: 365). The common factor analysis method is selected if the objective of the analysis is to search for some underlying variable structure. Analysts select the latter method when high levels of measurement error are detected.

14. The degree of generalizability of the results from the PCA to the population was considered. A key aspect of generalizability is the stability of the results from the PCA model. Component stability is shaped by the size of the sample and the number of cases (i.e. firms) per variable. General rules are to explore at least 100 cases and a minimum of 5 cases for each variable (Hair et al. 1995: 373). Information from 237 firms relating to 17 variables was explored in the final model, representing approximately 14 cases per variable. The reported analysis, therefore, did not suffer from the ‘overfitting’ data issue problem. We acknowledge that further confirmatory analysis is required, using larger databases to explore the validity of the components identified by the PCA.

15. The assumptions of PCA are discussed in detail elsewhere (Davies 1984, Shaw and Wheeler 1985, Hair et al. 1995). The conceptual assumption underlying PCA is that the selected variables are an appropriate set of variables relating to theory and/or a conceptual model. PCA assumes that some underling structure exists in the set of selected variables (Hair et al. 1995: 375). The correlation matrix relating to the 17 ‘raw’ variables was inspected, and several correlation coefficients greater than 0.3 were noted. It was inferred that the data matrix had sufficient correlations to justify the application of a PCA. The Bartlett test of sphericity is a statistical test for the presence of correlations among the variables. Table 1 shows that the test statistic was significant at the 0.0001 level. We inferred that this assumption had been satisfied, and ‘the correlation matrix had significant correlations among at least some of the variables’ (Hair et al. 1995: 374). The Kaiser-Meyer-Olkin measure of sampling adequacy is an additional measure to quantify the degree of intercorrelations among the selected variables, and the appropriateness of the PCA. ‘The index ranges from zero to one, reaching one when each variable is perfectly predicted without error by the other variables’ (Hair et al. 1995: 374). The score of 0.7 was ‘meritorious’ and highly acceptable.

16. A goal of PCA is to identify components that are substantively meaningful. The rotation phase transforms the initial matrix into one that is easier to interpret. A varimax rotation method seeks to minimize the number of variables that have high loadings on a component, in order to enhance the interpretability of the components (Norusis 1990).

17. Each component loading was individually significant with regard to a sample size of 237 firms (Hair et al. 1995: 385, table 7.3), and convergent validity was apparent with regard to all constructs (Bagozzi and Yi 1991). Also, the constructs appeared to exhibit discriminant validity. Each variable loaded significantly on only one of the six components.

18. On the downside, 38% of the total variance in the data was not accounted for by the six components. The issue of loss of information is acknowledged (see section 3.3.1 and table 3).

19. Cluster analysis groups objects (i.e. cases/firms), while PCA groups variables. The QUICK CLUSTER procedure can be used to cluster relatively large numbers of firms efficiently without requiring substantial computer resources. Further, the procedure produces only one solution for the number of clusters requested (Norusis 1990).

20. A key decision relates to the selection criterion, that is to say, the initial choice of variables determines the characteristics that can be used to identify sub-groups (Hambrick 1984, Norusis 1990). Insights from agency and stewardship theories provided a platform to identify conceptual ‘types’ of family firms (figure 1). The selected cluster solution (or classification) solely relates to insights from these theoretical perspectives. An exploratory cluster analysis was conducted in order to identify an objective classification of objects (i.e. family firms). The identification of a general classification of family firm ‘types’ was not the objective of this study. Further, the purpose of this study was not to explore the wider applicability of classifications presented elsewhere. A confirmatory cluster analysis was, therefore, not conducted.

21. Cluster analysis, unlike PCA, is not a statistical inference technique, and does not have the same stringent assumptions. The representativeness of the sample and multicollinearity assumptions, however, should be considered. As indicated in section 3.1, a representative sample of respondents responded to the survey, and tests did not indicate any problem of response bias. Information from 272 family firms was gathered, of which 237 firms provided complete data for the 17 ‘raw’ variables analysed within the final PCA model. No marked differences were detected between the demographic profiles of the 237 firms explored with the PCA and cluster analysis and the profiles of the total 272 responding firms. This evidence did not eliminate the concern relating to the representativeness of the sample analysed, but it did indicate some representativeness. Also, the multicollinearity issue was considered. The cluster analysis explored the 237 firms by 6 components matrix produced by the final PCA model. Each of the 6 variables was an independent and orthogonal dimension (or construct). Consequently, the cluster analysis was not distorted by variables significantly associated with one another.

22. Cluster analysis groups objects (e.g. firms) based on the characteristics they possess. Objects are grouped into clusters with other similar objects with regard to some predetermined selection criteria (i.e. variables). The selected grouping solution (i.e. the selected number of clusters) should exhibit high internal (within-cluster) homogeneity and high external (between-cluster) heterogeneity (Hair et al. 1995). The agglomerative hierarchical procedure begins with each of the 237 firms in a separate cluster. In subsequent steps, objects (i.e. clusters/firms) that are closest together in squared Euclidean distance (i.e. a measure of similarity between two objects) were combined to build a new aggregate cluster. The centroid method calculates the distance between the two clusters as the distance between their means for all of the variables (Norusis 1990).

23. Cluster analysis is sensitive to the inclusion of irrelevant variables (or undifferentiated variables). Each of the 17 selected ‘raw’ variables was found to be distinctive in one or more of the seven clusters (table 3). The selected cluster solution did not appear to be contaminated by the inclusion of irrelevant variables.

24. Cluster 2 had only three members. It was viewed as a valid structural component in the sample (Hair et al. 1995) and was not deleted as an unrepresentative outlier. Firms in cluster 2 exhibited the profiles of ‘large open’ firms, which were conceptualized in figure 1.

25. There is considerable debate surrounding whether policy intervention is warranted to alleviate barriers to entrepreneurial behaviour and business development (Holtz-Eakin 2000). To justify intervention in a market economy it is necessary to identify precisely where the market failure exists, and whether it is possible to rectify that market failure through intervention. The costs of the intervention have to be carefully assessed and the benefits estimated (Storey 1994). A decision widely perceived as ‘correct’ in the current time period may lead to an undesirable outcome in the future (Ferguson and Ferguson 1994). Advocates of a free enterprises economy system caution against interference with market forces (Wright et al. 2007). Perfectly competitive markets are something of a myth (Bridge et al. 2003), and neo-classical economic theory can be viewed as being an inappropriate basis for public policy prescriptions (Ferguson and Ferguson 1994). Barriers that discriminate and prevent a level playing field create market imperfections or market failure (Bridge et al. 2003) that can constrain firm development (DTI 2004). Inevitably, owners of smaller private firms concerned with uncertainty and risk will face attitudinal, resource (i.e. information, technology, finance, legitimacy, marketing, ‘family agenda’, etc.) operational and strategic barriers to business development. The case for intervention to address barriers can be supported from a public choice theoretical perspective.