The Appropriate Selection of a Representative Sample Population: The Case of Surveillance, Epidemiology, and End Results and Gastrointestinal Stromal Tumor Free

In Response: We read with great interest the comments of Cheung et al. and understand the nature of their criticisms. Their primary concern centers on the exclusion of patients who may have had gastrointestinal stromal tumor (GIST) but were potentially misclassified within the Surveillance, Epidemiology, and End Results (SEER) database, especially before the year 2000. The implication is that some sort of selection bias has occurred. The nature and the source of the purported selection bias, however, are not made clear. Having carefully considered their comments, we reaffirm that what Cheung et al. call a “major design flaw” was actually a carefully considered, hypothesis-based exclusion of cases in light of the realities and limitations of the SEER database.

With respect to the issue of the appropriate classification of GIST, we recognize that GISTs were misclassified to a large degree in the 1990s, and that only with further understanding of the disease did appropriate diagnostic classification occur. In their comments, Dr. Koniaris's group cites two studies they themselves have published on the subject (1, 2). In performing these two studies, they included 4 histology codes—8890, 8935, 8936, 9560—and suggest that we do the same. For the purpose of clarity, the code 8890 refers to leimyomas and leimyosarcomas, 8935 to stromal sarcomas not otherwise specified, 8936 to gastrointestinal stromal sarcomas, and 9560 to neurilemmoma. Our study focused on patients with histology code 8936, and only those with metastatic disease (3).

It should be noted that histology codes do not specify or exclude any particular disease site. Therefore, a large proportion of tumors with histology codes 8890, 8935, and 9560 in the SEER database are actually from nongastrointestinal sites. Table 1 presents a breakdown of cases with the above histologies within the SEER database from 1995 to 2004. Also shown in the table is a breakdown of patients from gastrointestinal sites alone, which excludes a significant number of retroperitoneal sarcomas as well gynecologic, head and neck, and airway tumors. Because our study focused on metastatic disease alone, we have presented a similar breakdown of patients with metastatic tumors of gastrointestinal origin (Table 1). These cases are further stratified by year of diagnosis in Table 2.

As evident from the tables, what Cheung et al. consider a major design flaw is the exclusion of far fewer patients than implied in their comments. The question remains whether these additional patients should be included in the study population. In general, the result of excluding misclassified patients is a reduction in sample size, not necessarily an introduction of bias. The goal of our study, however, was not to capture every last patient with GIST but to capture a homogenous population representative of true GIST; and the exclusion of non-GIST sarcomas, which comprise a proportion of these additional histologies, enhances this homogeneity.

Cheung et al. suggest that we include these additional histologies to ensure that two “equal” GIST populations are being examined. The suggestion, although unclear from the comments, seems to be to include the additional histologies only in the preimatinib group, and not in the postimatinib group. On further examination of the data, patients with gastrointestinal tumors coded as 8890 and 9560 represent a subset of patients with worse survival than those coded as GIST (data not shown). Therefore, treating the data in this fashion would introduce significant selection bias, in fact making the two comparison groups far less equal. If the additional histologies are to be included, the statistically valid approach would be to add them to both comparison groups, although, as stated above, doing so is unnecessary. Nevertheless, we have done the suggested analysis and present it below. Figure 1 presents 4 survival curves, with both the preimatinib and postimatinib curves presented in their original fashion and with the addition of the suggested histologies. The result of the addition is a decrease in median survival of 1 month for each group, with no change whatsoever in the final interpretation. With the above in mind, we reaffirm that the homogeneity of our original sample population is indeed a strength of our study and not a design flaw in any way.