Worker inflow, outflow, and churning

Abstract

Linked employer–employee data from Finnish business sector is used in analysing worker turnover. The data set is an unbalanced panel with over 219 000 observations in 1991–97. The churning (excess worker turnover), worker inflow, and worker outflow rates are explained by plant and employee characteristics. The probabilities of observing non-zero churning, inflow, and outflow rates increase with plant size. The magnitudes of the non-zero churning and inflow rates depend positively on size, but the magnitude of outflow rate depends negatively on size. High-wage plants have low turnover; plants with large within-plant variation in wages have high turnover. Average tenure of employees has a negative impact on turnover. High plant employment growth increases churning and separation but reduces hiring in the next year. Also controlled are average age and education of employees, shares of women and homeowners among employees, foreign ownership, ownership changes, and regional unemployment.

Acknowledgements

This research has been supported by the Academy of Finland (project 50950). When this research was carried out, Mika Maliranta was working for Statistics Finland. Earlier versions of this paper have been presented at EALE Conference, Jyväskylä, and Conference on Microeconomic Analyses of Labour Reallocation at Upjohn Institute of Employment Research, Kalamazoo, Michigan. The data set used in the paper can be accessed in the Research Laboratory of the Business Structures unit of Statistics Finland.

Notes

In economics, Parsons (1977, 1986), Farber (1999), and Davis and Haltiwanger (1999) survey the turnover literature from different angles. For the analysis of turnover in the human resource management literature, see Cotton and Tutle (1986) and Griffeth et al. (2000).

We analyse the cyclical behaviour of the gross job and worker flows at the aggregate level and in main industries in a separate paper (Ilmakunnas and Maliranta, 2003).

A measure that equals CF/2 is called replacement rate by Albæk and Sorensen (1998) and excess turnover by Barth and Dale-Olsen (1999).

The idea of this triangle is adopted from Burgess et al. (2000a). In Ilmakunnas and Maliranta (2000) we present the distribution of Finnish manufacturing plants in the triangle.

Previously, Oi (1962) used the minimum of separations and hirings as a measure of replacement hiring, using aggregate (industry-level) data. However, the result discussed in the text needs not hold at the aggregate level (or for a group of plants) where there is simultaneous job creation and destruction. Instead, at the aggregate level it holds that CF + EJR = 2*min(WIF,WOF). Hence, Oi's measure includes, besides genuine replacement hiring or churning, also excess job reallocation across firms in the industries.

For example in manufacturing, the exiting plants accounted for 16% and the entering plants for 10% of the total number of plants during the recession years (Ilmakunnas and Maliranta, 2000). We have interpreted as an exited plant such a plant that was in the register in year t − 1, but no longer in year t. Besides true exit, there may be other reasons why a plant disappears from the registers, but plant data is much less problematic in this sense than firm data that may be influenced, for example, by mergers.

In wage-setting models the distinction between quits and firing need not be essential. The efficient turnover hypothesis argues that in principle quits and layoffs are equivalent from the point of view of both parties. When the firm rejects a wage demand, the worker quits, and when a worker rejects a wage cut, he is laid off (McLaughlin, 1991).

The turnover of personnel may coincide with changes in the technology of the firm, i.e. investment in new technology may cause exit of some old workers and entry of new ones that have skills appropriate for the new equipment (Bellman and Boeri, 1998; Maliranta, 2000).

If quits and layoffs could be identified, the behaviour of the plants could be modelled with hiring and layoff equations, as in Hassink and Broersma (2003), but even then some quits may be initiated by the employer.

We have examined hiring from unemployment and separations to unemployment in a separate paper (Ilmakunnas and Maliranta, 2004).

In principle, observed hiring is not equal to desired hiring, since the plants may have open vacancies that they have not yet been able to fill. (The process of filling vacancies has been examined in several papers, e.g. Gorter et al., 1996; van Ommeren and Russo, 1997). This may underestimate the desired inflow. The nature of vacancies, and hence the characteristics of the plant and its employees, may also affect the success of filling the vacancies (van Ommeren and Russo, 1997). In practice, however, vacancy durations are relatively short compared to our data that is based on end-of-year comparisons. Most outflows during the year would have already led to a replacement hiring. Any open vacancies at the end of the year would have been posted only a few weeks before. Further, it is unlikely that many vacancies would be posted just before the Christmas and New Year season.

The two-year lag arises because we use lagged NET(−1) as an explanatory variable. NET(−1) in turn is calculated by using the average employment in years t − 1 and t − 2 as the denominator.

A justification for using the logit transformation is that if at the level of individuals the quit or hiring decisions are based on logit models, the plant-level flow rates can be regarded as grouped data. This leads to a logistic transformation of plant-level flows and heteroscedasticity in the error term (Greene, 2003, pp. 687–8). When the flow rate is in the interval [0,2], it could be transformed to the form log((X + c)/(2 − X + c)), where X is a flow rate and c is a small constant. However, this does not solve the problem of high concentration of zero values. The transformation just shifts the peak to the negative value log((c/(2 + c)).

The same issue of corner solution would arise also in the ordinary Tobit model. There is no censoring, as there cannot be negative censored flow rates.

In principle, the models could be estimated as a system, but this would unnecessarily complicate the analysis. See Hassink and Broersma (2003) for an example of joint estimation of Tobit models for hiring and layoffs.

It should be noted that the composition of the workforce in a plant in terms of average age, education etc. is partly determined by the matching process. For example, if young workers have a high tendency to quit, the average age of workers increases. Lagging the variables also reduces problems of this kind of simultaneity of turnover with the workforce characteristics.

This is also evidenced by the fact that if our proxy for quits, WOF–WOFU is regressed on the same variables, the coefficients of the wage groups decline when we go from group 2 to group 5 (results not reported in the tables).