Thread by @olson_km

#JSM2018 The brilliant Susan Murphy is this year’s Fisher Lecture award recipient!

#JSM2018 Murphy Lab does sequential experimentation in improving health. Some for companies.

#JSM2018 Murphy Experimentation and continual optimization is key. How do we use learning as an experiment is put into the field to improve outcomes for individuals? Mobile interventions are key here. Intervention may be either a push intervention or pull intervention

#JSM2018 Murphy Pull intervention requires you to be aware that you need help. Push can go without you doing anything but can have negative impact (people delete app)

#JSM2018 Murphy Sense2Stop is designed as a smoking cessation program. Use stress reduction method to buffer real-like stressors. Should device notify you to go to exercises?

#JSM2018 Murphy Participant wears sensors that measures physiological responses. Has machine learning algorithm determining when you are stressed. Is a reminder effective? Does it vary by context? Thus, stratified micro-randomized trials

#JSM2018 Murphy Looking at time intervals of every minute for 10
Days straight. !!! Data comes in at different levels throughout the day (sensor, weather, etc.)

#JSM2018 Murphy Contains an indicator of whether appropriate to try to provide a treatment (no treatment while driving, for example).Only randomize if there is more than one treatment available at the time. If randomized intro treatment, then remind to access mindfulness exercise

#JSM2018 Murphy Proximal response is what you did within an hour of the treatment. (Can pick other time frames)

#JSM2018 Murphy Randomization means we can assess causal effects of the reminder and whether varies by context. Need to stratify so that you can actually have time points where you are actually stressed and not stressed.

#JSM2018 Murphy On average, for every 1 minute stressed, participants had 6 minutes not stressed

#JSM2018 Murphy Need to give a budget for how often it’s appropriate to try to provide treatment. Constrained this experiment to about 1.5 times per day for times stressed and not stressed

#JSM2018 Murphy I just made up an optimization criterion of 1.5 times per day on average for each stressed/not stressed time. Have uniform distribution of pinging participants across all times of day (not just in morning)

#JSM2018 Need to forecast expected number of times you will be stressed during today. Probability of getting treatment depends on desire number of treatments per day, how many treatments received, and anticipated number of future stressed times

#JSM2018 Murphy Schematic for the study

#JSM2018 Murphy What about the causal treatment effect? Call it causal excursion effect. Coming from a potential outcomes framework.

#JSM2018 Murphy Following Rubin’s potential outcomes. All of the treatments that occur could affect your outcome to all time up to time up to and including 59 minutes later from the treatment.

#JSM2018 Murphy Have a collection of all the Ys that could occur on your forehead. Researchers see a subset of these.

#JSM2018 Murphy All excursions have a time dimension on them - time of all treatments up to now and getting the treatment vs same collection of treatments and not getting the treatment

#JSM2018 Murphy Of course, individual level causal effect is not estimable. So need to look at averages. Causal excursion effect at time t beginning in strata x

#JSM2018 Murphy Now we need to place our bets on a particular hypothesis and one particular question. Want to contrast two treatments now and for the subsequent hour. So primary test - is there a signal going on here at all? Test of main effect

#JSM2018 Murphy Concerned about diminishing returns to treatment. So alternative is a quadratic decreasing effect and a linear decreasing effect.

#JSM2018 Murphy Can develop models that are very similar to GEE - familiar to many practicing statisticians. Use weighted and centered least squares.

#JSM2018 Murphy Can use contrast coding - adds robustness to test. Use a projection of excursion effect through time.

#JSM2018 Murphy Then use simulation based sample size calculator. Low dimensional alternative hypotheses increase power

#JSM2018 Murphy Treatment design and experimental design intertwined here. Treatment design is an algorithm! Developing part of the treatment, not just evaluating the treatment.

#JSM2018 Murphy has a music cue! This should be our mantra, Murphy urges.

#JSM2018 Murphy We are the ones who enrich science through experimentation.

#JSM2018 Murphy This requires lots of collaboration

#JSM2018 And this is exactly why Susan Murphy is a MacArthur Genius!