Joanna Sleigh joins SPSM, with Mike Conway, to discuss the issues related to social media data access and ethics, 4/16/17, 9pCT.
Conway recommends we check out this paper (which includes SPSM alum Glen Coppersmith as co-author).
We’ll be discussing ideas from SPSM alum.
From Dr. Jonathan Singer:
- “Traditional research has long recognized that response rate can be a strength or limitation to the research. If I sent a survey to 2000 people and only 200 responded I would have to say that the response rate was 10% and that this was a limitation of the research. If you know something about your sample then you can say something about the people who did not respond. This can help interpret results. Social media doesn’t provide a sample size, so response rate is impossible to establish.
- “If we want to target select groups on SoMe we can do snowball samples (e.g. send out a link to a group of people and ask them to share it with similar folks). Alternately if we pay Facebook (for example) to reach out to a demographic then our sample will be limited to the FB users who fit the marketing profile, which can further bias our sample.
- “These comments suggest the following questions:
- What are the strengths and limitations of using publicly accessible SoMe data for suicide prevention research?
- If we want to recruit a more nuanced sample are traditional survey / data collection methods sufficient or do we need new methodologies?
- If we recruit participants using FB, Amazon’s Mechanical Turk, or other for-profit sources how might that change our findings?”
From Dr. Philip Resnik:
“Ethical issues in data analytics in general and NLP in particular are becoming a hot topic. Worth giving a shout out to the first workshop in Ethics in NLP, http://ethicsinnlp.org/, coming up April 4, among whose organizers is Meg Mitchell, who was one of the three co-founders of the ACL workshops in Computational Linguistics and Clinical Psychology. Glen and two colleagues have a paper there on ethical protocols for social media health research.
“In terms of design issues, one of the biggest nuts to crack is what we data-driven folks call “ground truth” — without at least some of it you can’t evaluate your predictive models, and without decent quantities of it you can’t train them. I’ve railed for years about HIPAA having been designed with a huge blind spot when it comes to research uses, as you know; I’ve also given up on there being a way to solve that problem in HIPAA — right now my money is on data donation as the right solution. That doesn’t solve every problem but it’s a big start.
“Also related to ground truth, one of the challenges I’ve discovered recently has to do with good definitions of “positive” when we’re talking about assessing suicide risk. As you will recall, I was rather shocked a couple of months ago to discover that I have stumbled into doing original research on assessment of suicidality from language samples, having thought I was simply going to be applying existing work that’s already been well established.”
Watch us LIVE here:
Suicide Prevention Social Media #SPSM 