The first one is to use the Pseudo Poisson maximum likelihood (PPML) estimation proposed by Santos
Silva and Tenreyo (2006), the implicit assumption of which is that zeros are the result of rounding errors and hence missing observations are wrongly recorded at zero