Saturday, 15 January 2005 - 4:00 PM

This presentation is part of: Methodological Issues in Randomized Trials

Is Survival the Fittest? A Post-hoc Evaluation of Event History Estimations in an Experimental Design

Purpose: Survival analysis has become one of the most common statistical techniques for analyzing longitudinal data in the social sciences, largely due to its ability to produce unbiased estimates in the presence of censored or incomplete data. Event history techniques have also been used to analyze data from experimental designs in other fields (e.g., medicine), where differential lengths of follow-up among subjects is common. However, the dearth of experimental studies in the social sciences has meant that the application of event history methods in such designs has been limited. An exception that used an event history model in an experimental design tested whether an intervention would decrease the time between termination of parental rights (TPR) and adoption was published in 2002 in Social Work Research (Festinger & Pratt). Although the use of event history methods was appropriate for the question at hand, the intervention was so successful that a disproportionate number of cases in the control group were right censored. Theoretically, this should have little bearing on estimates. A follow-up case-record review of this sample offers an opportunity to investigate the utility of event history techniques in experimental social research by ascertaining whether original estimates that used incomplete or censored data turned out to be reasonably correct. The comparative analysis also provides an occasion to review two major assumptions of event history procedures in the context of social work research employing experimental designs.

Methods: A follow-up case record review of subjects (N=175) from this experimental study revisits original survival estimates and, using confidence intervals, compares them to estimates generated using nearly complete data.

Results: Results indicate that Kaplan-Meier median estimates of survival time and parameter estimates using Cox Proportional Hazards Regression were relatively accurate despite substantial differential censoring between the treatment and control groups. Estimated mean averages, however, tended to be inaccurate in the presence of substantial censoring.

Implications: If explanatory factors are sufficiently accounted for during modeling and the assumption of independent censoring of observations is met (i.e., the censoring itself is unrelated to the probability that an individual will experience the event), standard forms of survival analysis provide unbiased estimates in experimental studies involving subjects with varying lengths of follow-up. This appears to be the case even when censoring is differential between the treatment and control groups.

See more of Methodological Issues in Randomized Trials
See more of Oral and Poster

See more of Celebrating a Decade of SSWR (January 13 - 16, 2005)