Experiments designed to help the participants Maximilian Kasy - - PowerPoint PPT Presentation

experiments designed to help the participants
SMART_READER_LITE
LIVE PREVIEW

Experiments designed to help the participants Maximilian Kasy - - PowerPoint PPT Presentation

Experiments designed to help the participants Maximilian Kasy 7/10/2019 Why experiments? Are your programs effective in helping refugees? How to find out? Possibility 1: Compare the outcomes of those who got the programs to others who


slide-1
SLIDE 1

Experiments designed to help the participants

Maximilian Kasy 7/10/2019

slide-2
SLIDE 2

Why experiments?

◮ Are your programs effective in helping refugees?

◮ How to find out?

◮ Possibility 1: Compare the outcomes of those who got the programs to others who didn’t.

◮ Problem: These groups might be different for other reasons.

◮ Think about a doctor prescribing a medical treatment.

◮ Then the patients who got the treatment might die more often. ◮ But only because they were more sick to begin with! ◮ “Selection problem.”

slide-3
SLIDE 3

The standard way of doing experiments

◮ Possibility 2: Randomized experiment.

◮ Create groups that are ex-ante similar, by randomly assigning participants to groups. ◮ To compare apples with apples.

◮ Conventionally:

◮ Divide the sample equally between treatments. ◮ Wait until experiment is done. ◮ Then compare average outcomes. ◮ Use statistical tests to see whether there was any effect.

slide-4
SLIDE 4

Drawbacks of conventional experiments

◮ This approach gets the causal effects right. ◮ And it gets precise estimates for every policy. ◮ But we need to wait a long time until we learn something. ◮ And we might not do the best we can for our participants. ◮ Think again of a medical experiment:

◮ Suppose in the first few months, everybody who got the new treatment died. ◮ Then you better stop the experiment!!!

slide-5
SLIDE 5

Preliminary estimates for our experiment

  • cash

information psychological control 0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20

success rate treatment treatment ●

  • cash

information psychological control

◮ We already have suggestive evidence that the psychological treatment performs better.

slide-6
SLIDE 6

A different objective: Helping participants

◮ The standard approach is optimal when you want to get precise estimates of policy effects. ◮ But we want to instead help participants as much as possible. ◮ Cf. Immanuel Kant: “Act in such a way that you treat humanity, whether in your own person or in the person of any other, never merely as a means to an end, but always at the same time as an end.” ◮ This requires using the information we already have, when deciding which policy to assign people to. ◮ But we also want to continue learning, to do better in the future.

slide-7
SLIDE 7

The exploitation / exploration tradeoff

◮ Possibility 1: Assign each participant to the policy we currently think is best.

◮ Good for the current participant. ◮ Problem: We might stop learning, getting stuck with a sub-optimal policy.

◮ Possibility 2: Assign participants to each policy with fixed probability over time.

◮ Good for learning policy effects. ◮ But not optimal for current participants.

◮ Possibility 3: Optimal strategies shift to better performing policies over time. ◮ For instance Thompson sampling:

◮ Assign each treatment with probability ◮ equal to the current probability that it is optimal.

slide-8
SLIDE 8

Assignment probabilities in our experiment

Start of adaptive assignment Outage Ramadan

  • 0.0

0.2 0.4 0.6 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

week of the experiment assignment probability treatment

  • cash

information psychological control

◮ As we learn that the psychological treatment does better, more participants are assigned to this treatment.

slide-9
SLIDE 9

Assignment frequencies in our experiment

Start of adaptive assignment Outage Ramadan

  • 10

20 30 40 50 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

week of the experiment number of observations treatment

  • cash

information psychological control

slide-10
SLIDE 10

Targeting

◮ Not every policy is good for everybody. ◮ Some things work better

◮ for those with more or less work experience, ◮ for those with more ore less education, ◮ for women or men.

◮ We can do better than just going with “one size fits all.” ◮ Try to get each group what works best for them.

slide-11
SLIDE 11

Combining information

◮ Problem: For each group and policy, we might only have very few observations. ◮ This means averages are unreliable estimates. ◮ Solution: Combining information between groups. ◮ Estimate effect on a group by combining

◮ their own average outcomes, ◮ and the average outcomes for everybody else.

◮ Bayesian hierarchical models do this optimally.

slide-12
SLIDE 12

Effect heterogeneity in our experiment

  • Jor,F,>= HS,ever emp

Jor,F,>= HS,never emp Jor,F,< HS,ever emp Jor,F,< HS,never emp Jor,M,>= HS,ever emp Jor,M,>= HS,never emp Jor,M,< HS,ever emp Jor,M,< HS,never emp Syr,F,>= HS,ever emp Syr,F,>= HS,never emp Syr,F,< HS,ever emp Syr,F,< HS,never emp Syr,M,>= HS,ever emp Syr,M,>= HS,never emp Syr,M,< HS,ever emp Syr,M,< HS,never emp 0.0 0.1 0.2 0.3 0.4

success probability

  • cash

information psychological control

slide-13
SLIDE 13

THANK YOU

For all your work in making this experiment happen!