Shepard D S, Neutra R
Am J Public Health. 1977 Aug;67(8):743-50. doi: 10.2105/ajph.67.8.743.
Samples of outpatient visits often must be used to identify users of a health facility with a given chronic condition. Such samples can lead to biases, however, because patients with more frequent visits are overrepresented. These biases can be avoided by a weighting procedure in which each sampled visit is weighted inversely to the number of clinic visits made by that patient during the sample period. This procedure proved critical in estimating the number and characteristics of hypertensive patients seen in the medical clinic of a teaching hospital. The unweighted estimate of the number of hypertensives was 7,373 patients, more than three times the weighted estimate of 2,250. Similarly,, the number of visits per year by these patients would be overestimated by almost 50 per cent without weighting. The estimated proportion of hypertensives still under treatment after 18 months was 68 per cent without weighting, compared to 51 per cent with weighting. Thus biases from failure to weight may be substantial. Analogous biases and solutions apply to other sampling problems in health services research.
通常必须使用门诊就诊样本，来识别患有特定慢性病的医疗机构使用者。然而，这样的样本可能会导致偏差，因为就诊更频繁的患者在样本中占比过高。通过加权程序可以避免这些偏差，在该程序中，每个抽样就诊的权重与该患者在样本期间的门诊就诊次数成反比。这一程序在估计教学医院门诊部高血压患者的数量和特征时被证明至关重要。高血压患者数量的未加权估计为7373例，是加权估计2250例的三倍多。同样，若不进行加权，这些患者每年的就诊次数将被高估近50%。未加权时，估计18个月后仍在接受治疗的高血压患者比例为68%，加权后为51%。因此，未加权导致的偏差可能很大。类似的偏差和解决方案适用于卫生服务研究中的其他抽样问题。
