Programming an IRT, 2Pl Model from Scratch (So You Don’t Have to) I recently had a client that was looking to build a 2PL model from scratch. The 2PL describes the process through which someone gets a correct or incorrect response on a test. More specifically, the probability of person j providing a positive answer to item i is given by: \[Pr(Y_{ij} = 1| \theta) = exp(a_i(\theta_j - b_i)/(1 + exp(a_i(\theta_j - b_i))\] \[\theta \sim N(0, 1)\]

Thoughts Warning: Massive Spoilers to Follow You Were Never Really Here is the best meditation on Post-Traumatic Stress Disorder that I have seen at the movies. Unlike most movies that deal with Post-Traumatic Stress Disorder, Post-Traumatic Stress Disorder is never mentioned. It’s not even alluded to. You Were Never Really Here doesn’t have a lot to say about Post-Traumatic Stress Disorder, but that’s because the movie makes the viewer feel the symtpoms of Post Traumatic Stress Disorder.

Introduction I recently finished reading Strangers in Their Own Land, in which Arlie Hochschild profiles Tea Party supporters and activists in Louisianna. She mostly writes about their opposition to government restriction and their embrace of family, community, and G-d. While I was reading, I wondered, what policies does the White Working Class want? Or is their policy preference defined entirely in opposition to government doing anything. To answer this question, I looked at a dataset from DataForProgress, in which they asked respondents about their preference for a variety of different policies.

For my dissertation, I’m going to apply causal inference techniques to three separate datasets to examine the impact of family income on school readiness skills, while examining supportive parenting and cognitive stimulation as mediators. For one dataset, the Secondary Education Childcare and Youth Development (SECCYD) dataset, I plan to use a longitundal fixed effects approach. Unfortunately, this approach makes examining supportive parenting as a mediator difficult because the items that assess supportive parenting change at different assessment points.

In a recent post I pitted listedwise deletion against Full Information Maximum Likelihood (FIML) to see which outperformed which in an instrumental variables analysis (listwise deletion won). However, a big caveat of that analysis was that I didn’t use FIML to generate predicted values of x because lavaan can’t produce predicted values with incomplete data. So, this post is the same analysis, but with multiple imputation instead of FIML so that we can generated predicted values of x with the incomplete data.