The competitive environment hypothesis revisited: non-linearity, nonstationarity and profit persistence

Abstract

Much empirical literature dealing with the competitive environment hypothesis tends to find nonstationary behaviour and very high persistence in time series of company profits. Profit time series is modelled using a simple threshold autoregressive model that allows for nonstationary behaviour over subsamples. Using a new dataset consisting of profits for more than 150 US companies over a time period of 50 years, statistical evidence is presented that the high persistence observed in profits when using linear autoregressive models is often due to the misspecification of the data generating process.

Acknowledgements

The authors thank John Cable, Robert Kunst, Dennis C. Mueller, Burcin Yurtoglu and an anonymous referee for very helpful comments on earlier drafts of this work.

Notes

¹ See for example Chang (2002), Levin et al. (2002) and Im et al. (2003).

² The empirical literature on profit persistence uses two different but interrelated definitions of persistence of profits. The persistence measure related to long-run deviations from normal profits is given by the unconditional expectation of the AR(1) process, as defined above. Short run persistence (which corresponds to the context in which ‘persistence’ is usually used in time-series analysis), on the other hand, is given by the size of the parameter λ _i . This will be referred to in the latter sense. The cases where λ _i  = 1 cannot be rejected will be concentrated on, implying perfect persistence in the short run and the impossibility of using the unconditional expectation as a measure of long-run persistence.

³ Concerning estimates for other economies, Goddard and Wilson (1999) summarize the persistence estimates for eight countries (seven advanced economies plus India) for varying periods between 1950 and 1985 and tend find persistence estimates above 0.4.

⁴ A simple Jarque–Bera test gives evidence against Gaussianity in the distribution of for the linear estimates, while the null of normal distribution cannot be rejected at any sensible significance level for the estimates including bands of inaction.