In this paper we consider a particular form of cointegration, called hidden cointegration, which arises between positive and/or negative components of a time series. Hidden cointegration is especially interesting to model asymmetric behaviours, but it requires specific estimation and testing procedures. In order to detect the existence of hidden cointegration we propose a bootstrap version of the two stage Engle and Granger procedure (originally thought for linear cointegration). We also present some Monte Carlo evidence and an application to real data.