Peer Participation in the Medicare Shared Savings Program

class: center, middle, inverse, title-slide

.title[
# Peer Participation in the Medicare Shared Savings Program
]
.subtitle[
## Presenter: Youcheng Hou, UNC
]
.author[
### Discussant: Ian McCarthy, Emory University and NBER
]
.date[
### ASHEcon Annual Meeting, June 28, 2022
]

---

<style type="text/css">
.remark-slide-content {
    font-size: 30px;
    padding: 1em 2em 1em 2em;    
}
.remark-code, .remark-inline-code { 
    font-size: 20px;
}
</style>

# Physician networks and voluntary participation
- Lots of work measuring static physician networks
- Less work studying the effects of such networks
  - Even less work on the formation and evolution of those networks (more on this later)
- Very little work on voluntary program participation in health care
  - Look into "Selection and Causal Effects under Voluntary Programs" by Gupta et al.

---
# This paper
*Main Question*: Does my peers' participation in a shared savings program affect my own participation?

<br>
*Approach*: LPM regression of my participation against a weighted average of my lagged peers' participation, instrumented using lagged peers-of-peers

<br>
A ton of impressive data work!

---
class: inverse, center, middle
name: questions

# Some thoughts and questions

---
# Some descriptive work
- Would love to see some graphs describing the networks, especially the instrument (peer-of-peers)
- Specifically showing extent of peers-of-peers that are not my own peers

.center[
![:scale 50%](images/pcp-ref-graph.png)
]

---
# Reflection
- Are lagged values necessary?
- Already using leave-one-out peer average
- Ahn and Trogdon (2017) use contemporaneous (leave-one-out) peer average with lagged peer-of-peer average as instrument

---
# Weighting
- Results much larger when not weighting peer averages by patient counts
- Could instead be a reflection of measurement error in the peer group instrument (Angrist, 2014)

---
# Mechanisms
- HHI highly correlated with peer group structure
- Is recruitment effect identified separately from "social learning"?
- Recruitment is a two-way street...more likely to recruit if program is running successfully
- Suggests that, as I learn about my peers' performance, recruitment may be harder or easier
  - Is there some way to leverage "successful" versus "unsuccessful" ACOs here?

---
# Network formation
- Is my peer-of-peer exogenous?
  - If the existence of an edge is endogenous, then the absence of an edge should be endogenous too
- Analysis allows for time varying peer groups, but what assumptions are needed on the evolution of the network
- Given the need for sufficiently large ACOs, practices must be forward-thinking
  - Could you exploit system-level information here? Plausibly exogenous shock to ACO participation