The construction of statistical models of perioperative risk and long-term postoperative survival is a useful activity. It facilitates fair, assessment of surgical outcomes and provides insight into the association between certain clinical features and outcome. It provides quantitative estimates of risk or long-term survival. There are, however, a number of limitations to the use of such models in informing decisions concerning the selection of patients for lung resection. In essence, the limitations described in this article are those of work-up bias come full circle. Concerning the use of scoring systems in selecting patients for resection, one should remember the advice of the wise Gene Blackstone: caveat emptor. The findings of model-building exercises, if based on surgical databases, can only augment, and not replace, clinical judgment. When models suggest that certain patient groups do well, the prior selection of these patients should be borne in mind. When models of perioperative risk or long-term survival suggest that certain patient groups, despite being carefully selected by clinical teams, do badly, this information should be heeded. That said, moves to deny informed patients lung resection on the basis of estimates of risk or "poor" survival should be considered carefully. For example the British Thoracic Society Guidelines on Surgery for lung cancer state that mortality following resection should not be in excess of 4% for lobectomy. It is not exactly clear what is intended by publishing that statement. It represents some form of audit standard but clearly fails if one thinks in terms of the individual patient. A patient who has a curable cancer and who faces a life expectancy likely to be under 2 years without surgery might well accept a greater than 1 in 25 chance of perioperative death. If 25 patients in a room were facing that prospect, all 25 might reasonably hope to be among the 24 expected survivors and opt for surgery.