Roger Pielke Jr suggests that doping prevalence can be estimated by drug testing of athletes (Nature 517, 529; 2015). I contend that this method is flawed: as the autobiographies of some athletes attest, regular dopers have a track record of avoiding testing positive.

To estimate doping's true prevalence, two procedures that circumvent inherent weaknesses in simple counts of positive test results are useful.

First, Bayesian inference methods can be used to compare two distributions of biological parameters affected by doping: one distribution among sampled athletes and the other in a suitable reference population, allowing the number of manipulated samples to be estimated (P.-E. Sottas et al. Clin. Chem. 57, 762–769; 2011). A major advantage of this population-level analysis is that it can recognize abnormalities even when dopers' values remain within the 'normal' range.

Second, specially tailored questionnaires allow athletes to give honest answers to doping questions under cover of anonymity (W. Pitsch et al. Eur. J. Sport. Soc. 4, 89–102; 2007). The questionnaires are based on the randomized response technique and are used by social scientists to study illegal and deviant behaviours.

A combination of these approaches estimates that 14–39% of elite athletes have intentionally doped (see O. de Hon et al. Sports Med. 45, 57–69; 2015). This contrasts markedly with the 2% of samples designated as suspect in the World Anti-Doping Agency's published statistics.