In the three Italian centres involved in the European Community Respiratory Health Survey (ECRHS), prevalence of asthma-like symptoms was assessed through a mailback questionnaire. Since the nonresponse rate was not negligible, ranging 10-18%, we investigated whether nonresponse bias affected the results and, if so, whether the bias could be eliminated from the final estimates of prevalence. A screening questionnaire was sent by mail to 7,000 randomly selected subjects 20-44 yrs of age, and nonresponders were contacted again by phone. Additional information was collected on a subsample of the respondents through a clinical interview. A logistic regression analysis showed that, except for one symptom (awakening for coughing), symptom prevalence significantly decreased from the first to the subsequent contact, when controlling for age, sex, centre and season of interview. The decrease in symptom prevalence was largely independent of smoking habits and socioeconomic status, and was seemingly caused by a symptom-related self-selection. When correcting results according to a linear regression model, observed estimates appeared to be slightly overestimated, by 4-10%. A simulation with the Italian data showed that the bias increased steeply at nonresponse rate higher than 30%, a situation quite common in asthma surveys. In conclusion, nonresponse bias affects the results of ECRHS in Italy, slightly inflating prevalence estimates. To make reliable comparisons on international data in the presence of different nonresponse rates, a correction of the observed prevalence seems necessary.