OBJECTIVES: Since its establishment, NICE has become increasingly explicit about the way it uses evidence on cost effectiveness in decision-making - and, more recently, about the other factors it considers.  This, together with other ways in which decision making has evolved, suggests a number of testable hypotheses.  We propose and empirically test alternative ways that NICE decision-making might be modelled, building on and extending Devlin and Parkin (2004) and Dakin et al (2006). The large number of NICE decisions now observable facilitates the use of more sophisticated modelling techniques. METHODS: NICE’s decisions are characterised as binary choices: yes or no to a technology in a specifically defined patient group or indication. NICE Guidance often contains multiple such decisions. The probability of NICE recommending a technology is modelled as depending on evidence on effectiveness and cost-effectiveness; characteristics of the patients, disease or treatment; and contextual factors. Data were obtained from HTAinSite (www.htainsite.com) to November 2009.  RESULTS: Initial results, drawing on data for 262 decisions, suggest cost effectiveness alone explains the vast majority of NICE’s decisions, correctly classifying 85%, with high sensitivity and specificity. The estimated threshold, around £40k, is higher than NICE's stated threshold (20k – £30k) but similar to that estimated by Devlin and Parkin (2004). Results across alternative model specifications showed that almost none of the other variables exert a statistically significant effect on decisions, with two exceptions. First, technologies for the treatment of cancer have a significantly higher probability of being accepted, ceteris paribus, implying a willingness to pay an additional >£10k per QALY gained by cancer patients.  Second, analysis of the sub-set of decisions made after NICE’s second ‘social value judgement’ document suggest an increased probability of rejection.   CONCLUSIONS: This is work in progress; further results will be available to report from additional data extraction and modeling.