History of Survival Analysis at Mayo Clinic

Early publication of survival results

The analysis of survival or other time to event data has played a key role in medical research done at Mayo Clinic since the clinic's earliest days. In 1926, Gordon B. New published an article titled "End Results of the Treatment of Malignant Tumor of the Antrum" in Proceedings of the Weekly Staff Meetings of the Mayo Clinic.

Methods research

Researchers at Mayo have also played a pivotal role in the development of survival analysis methodology and software. All were developed because of analysis needs.

Joseph Berkson came to Mayo as a Macy Foundation Fellow in 1931 and was named chairman of the new division of Biometry and Medical Statistics in 1934. That same year he published the first of several papers on survival analysis methods in Mayo Clinic Proceedings involving the appropriate construction of a life table to describe the survival experience of a group of patients following an operation. An important concept introduced through these papers is the need to account for censoring when estimating survival rates. Several medical manuscripts were subsequently published using this method, and the methodological work culminated in two papers with Robert R. Gage, another member of the department.

Lillian (Lila) Elveback joined the department in 1965 and added important practical and theoretical justification to the methods. Her guidance on how to lay out the tabular results and plots of a survival computation guided the early software in the department and is still visible in the output of the R survival package and Mayo SAS macros. The underlying calculations have been almost entirely superseded by the Kaplan-Meier method. However, this approach was not feasible for any but the smallest datasets before the advent of modern computers.

Analysis of clinical data has continued to spur research in survival analysis. Methods for testing survival curves were contributed by Peter C. O'Brien; Thomas R. Fleming, Judith R. O'Fallon and David P. Harrington; and Daniel J. Schaid, H. Samuel (Sam) Wieand and Terry M. Therneau.

Methods and software for the comparison of observed survival for a cohort to what would be expected in the population at large, useful for the assessment of a surgical cure, were developed by Kenneth P. Offord, Erik J. Bergstralh and others and later extended by Therneau (multiple HSR technical reports).

Diagnostic methods for survival models (e.g., functional form exploration) were explored by Therneau, Patricia M. Grambsch and Fleming, and by Cynthia S. Crowson, Elizabeth (Beth) J. Atkinson and Therneau.

The addition of random effect terms to survival models has been explored by Daniel J. (Dan) Sargent and Therneau, Grambsch and V. Shane Pankratz.

Recent work by Therneau has focused on multistate models, type III tests and adjusted survival curves. Vignettes, available with the R package, explore these topics.

Software: SAS

In 1978 SAS added the ability to create user-written procedures; in 1980 they released part of the base SAS product along with the SAS User Written Procedures Guide documenting these additions.

The Mayo written COXREGR procedure was presented at the January 1979 SAS Users Group International (SUGI) meeting and was part of this first release. COXREGR was the first Cox model procedure available for SAS. In later years the SURVDIFF, SURVFIT and PERSONYRS procedures were added.

As these procedures were criticized in the early 1990s, multiple local SAS macros were written to replace the functionality. These macros have been shared with other institutions via the department internet website. Most of this functionality is now available with the standard SAS procedures.

Software: R

In 1985 some members of the department began to do work with the S software package from Bell Labs, which later grew into the Splus and later the R statistical systems. The survival package, written by Therneau, began as an internal project that was later released for Splus (statlib 1987), became a part of Splus and is now one of the standard components of R. Over 300 other contributed R packages currently depend on features of this base package.

Selected publications