CAT Simulation Sample Clauses
CAT Simulation. We used a post hoc CAT simulation to assess how efficient and precise a CAT version of the Anxiety item bank may be in clinical and general population subjects, and how accurate this CAT version may be compared with a full item bank administration. Previous studies have shown that post hoc CAT simulations are useful for this purpose as the results tend to be very similar to that of a real CAT administration (Kocalevent et al., 2009). Below, we provide the details on the CAT simulation settings and the assessment of efficiency, precision, and accuracy. The CAT simulation was performed using the R package mirtCAT (Version 0.5; ▇▇▇▇▇▇▇▇, 2015). A CAT administration/simulation consists of four basic building blocks: a starting item, a method for estimating θ, an item selection procedure, and a stopping rule. The administration/simulation starts by presenting a first item. After a response is given, the software estimates θ and calculates the corresponding measurement precision (standard error [SE]). It then evaluates whether the obtained results meet the stopping rule. If not, a new item is selected and the procedure is repeated until the stopping rule is met, or all items have been presented. As starting item, the CAT simulation used the item with the highest Fisher’s information (▇▇▇▇▇▇▇▇▇ & Reise, 2000; ▇▇▇▇▇▇ et al., 2000) at the average value of the latent trait in the general population (θ = 0). This item was I felt tense, which was coded as EDANX54 (Emotional Distress – ANXiety item bank, item 54) in the original US PROMIS item bank (▇▇▇.▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇.▇▇▇). To estimate θ, we could choose from two methods: maximum likelihood (ML) and Bayesian estimation (▇▇▇▇▇▇▇▇▇ & ▇▇▇▇▇, 2000). Bayesian estimation is often chosen because it uses an a priori population distribution of the latent variable. This property ensures that θ can be estimated for all response patterns. A drawback of Bayesian estimation, however, is that the estimation of θ is also influenced by the a priori distribution; it pulls θ estimates toward the center of the population distribution, which may result in bias (▇▇▇▇▇ et al., 2017; ▇▇▇▇▇, 2016). ML, by contrast, does not use an a priori distribution, and is therefore not able to estimate θ for response patterns that exclusively comprise extreme responses. It is, however, a more stable estimator considering possible bias. ML can also result in bias, but generally to a lesser extent compared with Bayesian estimation, especially using CAT (▇...
CAT Simulation. We simulated a separate CAT on each of the three MASQ scales (PA, NA and SA) from the item responses that were obtained from the patients. The item responses were selected for each patient from all the item responses in the corresponding scale and were evaluated as if they were collected adaptively. Basically, the CAT simulation started with the same item for every individual and then estimated the latent scale score and measurement precision using both item response and item properties. From here, either a new item was selected according to the item properties and the estimated latent trait level, or the simulation stopped when the prespecified value of measurement precision was obtained. The selection of new items, and the estimation of latent trait score and measurement precision using all collected item scores so far, continued until this prespecified measurement precision was reached, or when all items were used; items were used only once. To apply this procedure, we made several decisions regarding (a) the IRT model that estimates the item parameters, (b) the methods for selecting new items and (c) estimating patients’ latent scale scores (θ), and (d) the starting level and (e) stopping rule for the CAT. A program (▇▇▇▇▇ et al., 2011; ▇▇▇▇▇ et al., 2012) was written in the statistical environment R (R Core Team, 2014) to implement these decisions into three separate CAT simulations. Below, we will present the details concerning the decisions rules. First, as an appropriate IRT model for estimating item parameters, we used ▇▇▇▇▇▇▇▇’▇ (1969) GRM for polytomous items. The GRM is often the preferred IRT model, because it is easier to illustrate to test users than other models, and the item parameters are easy to interpret with regard to responder behavior (▇▇▇▇▇▇, ▇▇▇▇▇▇▇▇▇, & ▇▇▇▇▇▇, 2015; ▇▇▇▇▇ et al., 2011). These advantages are especially desirable when CAT is implemented on a large scale, as is mostly the case in clinical measures, because clinicians should generally understand how CAT works. The GRM model uses two types of parameters. The discrimination parameter a specifies to what extent persons with similar scores on the latent trait can be differentiated by the item. Furthermore, the GRM uses the location parameters b (the number of location parameters for an item is equal to the number of response categories minus one) which specifies the θ location on which a patient is expected to choose from a lower to a higher item response. We fitted the ...