To the Editor In their article on how stromal proportion (stroma) and p53 relate to survival from resected adenocarcinoma of the lung, Demarchi et al. emphasize how important it is to control for other, more traditional prognostic variables such as T and N stages (1). In this regard I was surprised to see that the P values they obtained for N stage were as high as 0.01 in Tables 5 and 6. In fact, the P values for N stage were higher than those for T stage. Because they provided the full data on their patients in their Table 1 but did not indicate how they coded the N variable, I decided to perform additional Cox model analyses on their published data. My further analyses strengthen Demarchi et al.'s conclusions, but they also emphasize the importance of coding of prognostic variables for optimal fitting of a survival model to the data. Finally, these new analyses raise questions about the optimal use of stroma and p53 for prognosis of adenocarcinoma of the lung.

TABLE 1 Cox Model Reanalysis of Survival in Demarchi et al.'s Dataa
Full size table

Because Demarchi et al.'s major results used the endpoint of survival time, this is also the endpoint I analyzed, and like the authors I included female gender as a variable in all the analyses. First, I examined the optimal use of T and N staging variables in their data. Using dummy variables (2) and T stage of 1 as the baseline, I discovered that T stage 3 was significantly related to survival time in their patients but that T stages 2 and 4 were not. This preliminary result and the observation that just two patients comprised stage T4 implied that the most useful way to code T stage was into two categories, namely T1&2 versus T3&4, and this is how the authors seemed to code T stage in their Table 2 and Figure 2. Once again, using dummy variables I performed a similar analysis for the N stages and discovered that the optimal coding of N stages was also as two categories: N0&1 versus N2&3. Coding N stage this way yielded results that differ from the authors'. Whereas they obtained a P value of 0.003 for N stage in Table 4, mine was 0.000068. Whereas they obtained a P value of 0.01 for N stage in Table 5, mine was 0.00055. Whereas they obtained a P value of 0.01 for N stage in Table 6, mine was 0.000097. More importantly, coding N stage this way yielded lower P values for the associations between survival time and stroma as well as survival and staining for p53. I obtained Cox model P values for these two variables of 0.004 and 0.00025—in contrast to the authors' reported P values of 0.03 and 0.002. In this analysis I coded stroma and p53 as continuous variables. Although these new results do not change the authors' conclusion that there was a significant association between survival time and stroma as well as p53, the lower P values suggest a model that fits the data better, and they demonstrate that the association between these two variables and survival was stronger than the authors reported.

In addition, I used univariable survival plots and Cox models to explore whether stroma and p53 were best used as continuous variables or in cutpoints. The survival plots suggested that tumors with stroma greater than 15% had similar survival times regardless of the exact percentage, but those with stroma less than 15% clearly survived a shorter time. Using this single cutpoint of 15% for stroma produced a P value of 0.001 for stroma in contrast to the P value of 0.004 when stroma was used as a continuous variable, and the likelihood ratio for the model rose from 40.9 to 42.5. Thus, the optimal way to use stroma may be with a cutpoint, and if this conclusion is verified with additional studies, pathologists might avoid the painstaking technique of morphometry for this variable in favor of a simple binary observation of percentage stroma less than 15% versus greater than 15%. This is at least a hypothesis that deserves further study. By contrast, I was unable to demonstrate a successful cutpoint for p53 in Demarchi et al.'s data, so that the authors' morphometric technique for immunohistochemical staining may be the optimal way to record p53. In my hands, then, the optimal survival model for the authors' data is given in the enclosed Table 1.

Optimizing the way the model fits the data can be important if one plans to use the results. For example, the model of Table 1 can be used to form a hazard score (hs) as follows:

with the variables in this equation defined as in Table 1. Such an hs could be used to predict who is likely to suffer an early death after surgery for adenocarcinoma, because the Cox model results imply that those with higher values of hs will suffer earlier deaths. For example, Figure 1 demonstrates the relationship between the above hs and the probability of dying within 2 years of surgery in authors' study patients.

  • Robin T Vollmer, M.D.

  • Durham VA Medical Center Durham, North Carolina

FIGURE 1
figure 1

Plot of the observed probability of dying within 2 years of surgical resection of adenocarcinoma in Demarchi et al.'s study patients versus the hazard score (hs) given by the equation in the text. The line gives the relationship, and it was obtained through the lowness function of S-PLUS (MathSoft, Inc., Seattle, Washington).

Full size image

REFERENCES

1. Demarchi LMMF, Reis MM, Palomina SAP, Farhat C, Takagaki TY, Beyruti R, et al. Prognostic values of stromal proportion and PCNA, Ki-67, and p53 proteins in patients with resected adenocarcinoma of the lung. Mod Pathol 2000; 13:511–520.

2. Vollmer RT. Multivariate statistical analysis for pathology. Part II: failure time analysis. Am J Clin Pathol 1996; 106: 522–534.

In reply: We are sending by e-mail our response about the useful comments made by Dr. Vollmer in our article “Prognostic values of stromal proportion and PCNA, Ki-67 and p53 proteins with resected adenocarcinoma of the lung.”

The main purpose of publishing a paper in a journal of international circulation and prestige such as Modern Pathology is to evoke constructive discussion and to benefit from the input provided by the scientific community. This point could not have been better achieved after reading the comments made by Dr. Vollmer on our study. Dr. Vollmer has a well-established reputation in the area of statistical modelling in pathology and spent a great deal of effort in analyzing our data and further improving our analysis. Basically, the analysis made by Dr. Vollmer differed from ours in the way he established the cutpoints for the explanatory variables of our model, as well as in the way he used fewer functions to depict the survival curves of our patients. The contributions made are positive and go in the same direction pointed out in our article. The main reason our group always presents data from our studies on individual basis is that this procedure allows other people to re-analyze our models. We made it several times, but this was the first time we received so much positive and constructive input.

  • Vera Capelozzi

  • Paulo Saldiva