BACKGROUND: Mixed, polyclonal Mycobacterium tuberculosis infection occurs in natural populations. Developing an effective method for detecting such cases is important in measuring the success of treatment and reconstruction of transmission between patients. Using whole genome sequence (WGS) data, we assess two methods for detecting mixed infection: (i) a combination of the number of heterozygous sites and the proportion of heterozygous sites to total SNPs, and (ii) Bayesian model-based clustering of allele frequencies from sequencing reads at heterozygous sites. RESULTS: In silico and in vitro artificially mixed and known pure M. tuberculosis samples were analysed to determine the specificity and sensitivity of each method. We found that both approaches were effective in distinguishing between pure strains and mixed infection where there was relatively high (> 10%) proportion of a minor strain in the mixture. A large dataset of clinical isolates (n = 1963) from the Karonga Prevention Study in Northern Malawi was tested to examine correlations with patient characteristics and outcomes with mixed infection. The frequency of mixed infection in the population was found to be around 10%, with an association with year of diagnosis, but no association with age, sex, HIV status or previous tuberculosis. CONCLUSIONS: Mixed Mycobacterium tuberculosis infection was identified in silico using whole genome sequence data. The methods presented here can be applied to population-wide analyses of tuberculosis to estimate the frequency of mixed infection, and to identify individual cases of mixed infections. These cases are important when considering the evolution and transmission of the disease, and in patient treatment.