In order to determine the research to evaluate the impact of publication bias, the following search terms are used in Medine( 1993 1 to 2003 1): publication bias, citation bias, language bias, regional bias, reference bias and repeated publication. At the same time, the literature about publication bias in Cochrane systematic review was searched. The retrieval is limited to those published articles whose main research is biased, and the acceptance of publication depends on whether the bias is found. There are 265 references. Among them, we selected 148 articles for comprehensive investigation, and obtained another 26 papers from the bibliography. We excluded 148 papers without original data, and the remaining 26 papers were included in this analysis.

We use a funnel chart to evaluate biased reports. In a funnel diagram, the magnitude of the action is the abscissa, and the measure of its accuracy, such as the sample size, is the ordinate. With the increase of sample size, the random change of effect decreases. In this way, if there is no publication bias, the data obtained from various studies will be symmetrically distributed in an inverted funnel shape on the chart. On the contrary, the asymmetric inverted funnel model indicates that there is bias in the research sample.

In the coordinates, we mark the coordinates of the action size corresponding to the sample size (see figure 1). The size of the effect is the ratio of the probability that the positive result is published to the probability that the negative result is published. Scientists convert reported relative risks into ratios, but do not convert risk ratios. A horizontal line indicates that there is no publication bias. Of these 26 studies, 23 positive results will be published first.

It is reported that the average ratio is 2.3, which means that positive reports can be published first. Make a regression diagonal with the report function as the dependent variable. There is no significant difference between slope and 0 (P=0. 13), which shows that the asymmetry of data is not statistically significant.