The existence of a correlation between observed radio spectral index and redshift has long been used as a method for selecting high-redshift radio galaxy candidates. We use nine highly spectroscopically complete radio samples, selected at different frequencies and flux limits, to determine the efficiency of this method and compare consistently observed correlations between spectral index (α), luminosity (P), linear size (D) and redshift (z) in our samples. We observe a weak correlation between z and α which remains even when Malmquist bias is removed. The strength of the z–α correlation is dependent on both the k-correction and sample selection frequency, in addition to the frequency at which α is measured, and consistent results for both high- and low-frequency-selected samples are only seen if analysis is restricted to just extended radio galaxies. This fits with the popular interpretation that the spectra steepen with z because the radio lobes work against a denser intergalactic medium environment as z increases, out to z∼ 2–3. However, we also note that the majority of sources known at z > 4 are very compact and often display a negatively curved or peaked spectrum, indicative of youth or merger activity, and therefore the low-frequency radio spectrum as a whole should be determined; this is something for which the new LOw Frequency ARray will be crucial. We quantify both the efficiency and the completeness of various techniques used to select high-z radio candidates. A steep-spectrum cut applied to low-frequency-selected samples can more than double the fraction of high-z sources, but at a cost of excluding over half of the high-z sources present in the original sample. An angular size cut is an almost as equally effective radio-based method as a steep-spectrum cut for maximizing the high-z content of large radio samples, and works for both high- and low-frequency-selected samples. In multiwavelength data, selection first of infrared-faint radio sources remains by far the most efficient method of selecting high-z sources. We present a simple method for selecting high-z radio sources, based purely on combining their observed radio properties of α and angular size, with the addition of the K-band magnitude if available.