01 January 2018
Substance abuse is a serious issue in both modern and traditional societies. Besides health complications such as depression, cancer and HIV, social complications such as loss of concentration, loss of job, and legal problems are among the numerous hazards substance use disorder imposes on societies. Understanding the causes of substance abuse and preventing its negative effects continues to be the focus of much research. Substance use behaviors, symptoms and signs are usually measured in form of ordinal data, which are often modeled under threshold models in Structural Equation Modeling (SEM). In this dissertation, we have developed a general nonlinear optimizer for the software package OpenMx, which is a SEM package in widespread use in the fields of psychology and genetics. The optimizer solves nonlinearly constrained optimization problems using a Sequential Quadratic Programming (SQP) algorithm. We have tested the performance of our optimizer on ordinal data and compared the results with two other optimizers (implementing SQP algorithm) available in the OpenMx package. While all three optimizers reach the same minimum, our new optimizer is faster than the other two. We then applied OpenMx with our optimization engine to a very large population-based drug abuse dataset, collected in Sweden from over one million pairs, to investigate the effects of genetic and environmental factors on liability to drug use. Finally, we investigated the reasons behind better performance of our optimizer by profiling all three optimizers as well as analyzing their memory consumption. We found that objective function evaluation is the most expensive task for all three optimizers, and that our optimizer needs fewer number of calls to this function to find the minimum. In terms of memory consumption, the optimizers use the same amount of memory.
Page generated in 0.0373 seconds