This May I attended and presented at the annual APS Convention in San Francisco. This was the first U.S. organized convention/conference I attended, and it was quite the experience. I have gotten used to being only one of very few Dutch people the last two years, but during APS, it sometimes felt as if I was back in the Netherlands. The Dutch were especially present in the Methodology corner of the convention, which made me feel a strange sense of pride (examples include, but are not limited to Eric-Jan Wagenmakers, Eiko Fried, Riet van Bork, Sacha Epskamp, Julia Haaf, Johnny van Doorn, Casper Albers, Tim van der Zee, Bobby-Lee Houtkoop, and honorary Dutchman Oisín Ryan)!
Of course, I wasn’t just there to fawn over the other amazing researchers who were there. I was there to present my own research! This time, I presented a poster which focused on the results of my first simulation study completed in graduate school (see below).
The focus of my research is on improving estimation of models of change. For this specific study, I looked at the second-order latent growth model (this model incorporates a measurement model for the observed measures at each time point) and whether some estimation methods were better than others when measurement non-invariance is present. Measurement non-invariance happens when participants in your study start to interpret and answer the questions in your study differently as time progresses. This can happen when participants grow older (e.g., a three-year-old will probably interpret a question differently than an eight-year-old), or when they go through an event that alters their attitudes substantially (e.g., when a soldier experiences trauma while deployed).
The results show that the appropriate type of estimation depends on a lot of characteristics of your data and study design. For example, if you have continuous observed variables (e.g., participants answer on a scale from 0 to 100), a relatively large sample, and you know, based on theory, where non-invariance occurs, you can use traditional frequentist methods and specify a partially invariant model. However, if none of these things are true for your study and data, it might be safer to switch to Bayesian estimation.
I’m laying the final hand on the manuscript for this study to submit it for publication. I’m curious to see what peer feedback I’m going to receive, as this is the first study I mostly designed and executed myself. Of course with support from my advisor, Sarah Depaoli!