Making the most of your prognostic factors: presenting a more accurate survival model for breast cancer patients

• Published in: Breast cancer research and treatment Vol. 22; no. 3; p. 251
• Main Authors: Knorr, K L, Hilsenbeck, S G, Wenger, C R, Pounds, G, Oldaker, T, Vendely, P, Pandian, M R, Harrington, D, Clark, G M
• Format: Journal Article
• Language: English
• Published: Netherlands 01.01.1992
• Subjects: Breast Neoplasms - mortality; Breast Neoplasms - pathology; Female; Humans; Lymphatic Metastasis; Models, Statistical; Prognosis; Proportional Hazards Models; Regression Analysis; Survival Rate
• ISSN: 0167-6806
• DOI: 10.1007/BF01840838

Determining an appropriate level of adjuvant therapy is one of the most difficult facets of treating breast cancer patients. Although the myriad of prognostic factors aid in this decision, often they give conflicting reports of a patient's prognosis. What we need is a survival model which can properly utilize the information contained in these factors and give an accurate, reliable account of the patient's probability of recurrence. We also need a method of evaluating these models' predictive ability instead of simply measuring goodness-of-fit, as is currently done. Often, prognostic factors are broken into two categories such as positive or negative. But this dichotomization may hide valuable prognostic information. We investigated whether continuous representations of factors, including standard transformations--logarithmic, square root, categorical, and smoothers--might more accurately estimate the underlying relationship between each factor and survival. We chose the logistic regression model, a special case of the commonly used Cox model, to test our hypothesis. The model containing continuous transformed factors fit the data more closely than the model containing the traditional dichotomized factors. In order to appropriately evaluate these models, we introduce three predictive validity statistics--the Calibration score, the Overall Calibration score, and the Brier score--designed to assess the model's accuracy and reliability. These standardized scores showed the transformed factors predicted three year survival accurately and reliably. The scores can also be used to assess models or compare across studies.