Affiliation: Gillings School of Global Public Health, Department of Biostatistics
Use of a generalized linear mixed model with a binary outcome and logit link function is proposed to generate trajectories of the probability of use of novel medical procedures. It is hypothesized that the shape of these adoption trajectories vary by institution and region and are influenced by patient, institutional, and geographic factors. The example of the adoption of sentinel lymph node biopsy in the treatment of early stage breast cancer is used to demonstrate the models utility and improvement over those typically used in registry and claims based research. Surveillance, Epidemiology, and End Results (SEER)-Medicare data from 1999 to 2007 was used as the basis for these model based trajectories. Fixed effects included patient, institution, and regional variables including a cubic polynomial of time for each region. Random effects were at the institution level and included a cubic polynomial of time. Results indicated a better fit of the multilevel model with a polynomial of time in comparison to standard models and that patient, institutional, and geographic factors influence the shape of the adoption trajectory of this novel medical procedure. Additionally, an evidence-based medical implementation index (EMII) was developed and tested using sentinel node biopsy adoption trends. Data were analyzed in aggregate and at the institution level. A single summary metric, based upon the area under the curve, was derived to quantify the pattern of adoption ranging from 0-100, with higher scores reflecting earlier adoption. The EMII was compared between SEER regions and between institutions. Differences in adoption patterns were found for SEER regions and institutions (p<0.001 for each effect). For SEER regions (n=15) the SLNB EMII range was 33 (New Mexico) to 66 (Seattle). For all institutions: n = 720, range = 4 - 87, mean = 46, S.D. = 20, bell shaped distribution. Finally, four estimation techniques for the random effects parameters were compared to maximum likelihood using quadrature based estimates, two types of pseudo-likelihood (PL), and jack-knifed estimates based on these. The estimates were compared via D-, A-, and E-efficiency. Results indicated that even with jackknifing PL estimates of the variance and thier confidence intervals were biased.