I use the spivreg routine from STATA and would like to compare the estimates obtained with the usual IV procedure and those obtained through the three steps procedure underlying spivreg. My concern is that in the related STATA Journal paper mentions that when running spivreg, the interpretation of the beta coefficient is not like in the linear regression context. Although I can see why intuitively, I do not understand what this statement implies in terms of causality. In other words, is the causal relationship between the regressors and the dependent variable still valid (assuming that the instrument I have used is a valid and relevant one and that the weighting matrix for spatial correlation is appropriately chosen)? If this is the case, does the beta coefficient still represent a marginal effect of the regressors on the dependent variable?