This paper presents: (1) statistical justification and calibration of model B4 using laboratory creep data and long-term bridge deflection data, and (2) statistical comparisons of various types with the existing creep prediction models of engineering societies. The comparisons include the 1995 RILEM Recommendation (Model B3), fib Model Code 1999, Model Code 2010, ACI Committee-209 Model, and the 2000 Canadian Model by Gardner and Lockman. The statistics and comparisons rely on a separately presented combined database of laboratory tests and multi-decade bridge deflection measurements, which has been developed at Northwestern University (NU). The laboratory data assembled in the NU database more than double the size of the previous RILEM database. The collected bridge data include multi-decade deflections of 69 large-span prestressed bridge spans, most of them excessive. The multi-decade bridge data are the only available and a significant source for long-term calibration because only 5 % of laboratory creep tests in the database had durations > 6 years, and only 3 % are > 12 years. Joint optimization of the laboratory and bridge data is conducted. Improved equations are obtained to predict the basic parameters of the compliance function for creep from the environmental conditions and concrete composition parameters, including the water-cement and aggregate-cement ratios, cement content and type, and admixture content. Comparisons with measured individual compliance curves are included as an essential check to validate the form of the compliance function.