1%) Perifosine and our outcome of interest was not rare (37.8% nonresponse to follow-up). However, our estimates could still be biased if nonrespondents do indeed differ systematically from respondents, which is as yet unproven. Specifically, we found that, on average, nonrespondents at baseline were 3.6 years younger than respondents. This difference is somewhat more pronounced than the age effect for late response to follow-up, but still small. Given that bias is a systematic deviation from the truth and not a random deviation,1 nonrespondents can only bias estimates if at least one relevant characteristic systematically divides respondents and nonrespondents. Until such a characteristic is identified, there is insufficient evidence that nonrespondents bias estimates, and it remains reasonable to assume that nonresponse is random.

This does not mean, however, that outcome estimates will necessarily be the same for nonrespondents and respondents. For example, a study found that, despite equally distributed smoking habits, respiratory symptoms, and lung function, the outcome��hospital admissions during follow-up��was twice as high in nonrespondents (9.9%) as in respondents (5.0%). The authors concluded that estimates were biased by nonresponse.34 We computed exact 95% CIs for the estimates of this study using STATA 11 for Windows and found that the estimate for the whole sample (5.6%, n = 1070) was within the 95% CIs for both the nonrespondents (5.5%�C16.0%, n = 142) and the respondents (3.6%�C6.6%, n = 928). If estimates vary randomly, ie, not because of bias, the true value will stay within the 95% CI in 95% of cases.

35 We conclude that the difference between nonrespondents and respondents in that study can be explained by random variance alone. We nevertheless agree with the principal conclusion of the authors, that equal distribution of baseline characteristics is not sufficient to exclude nonrespondent bias.34 Research must continue to move forward and analyze more than common baseline characteristics. Meta-analytic methods might be useful in distinguishing random differences from biases. In an email survey of 2127 clinicians, nonrespondents received as many as 5 email reminders and, if necessary, a sixth by fax. The outcome was the prescribing of contraindicated medications. Subgroups were defined by number of reminders needed.

The estimate for the total group was within the 95% CIs of all 7 subgroups. An I2 statistic of zero indicated that there was no inconsistency among groups, other than random differences.36 To summarize, our findings show Carfilzomib that avoidance coping and negative affectivity are unlikely to differ among respondents and nonrespondents to a questionnaire survey. In addition, our survey of the literature findings revealed no decisive factors underlying nonresponse. Further study of this topic is important because nonresponse is very frequent.