Trust is good, but is control better?
Rethinking supervision for learning health systems
HTSR | University of Twente
September 24, 2025
Control versus trust
Control
Assures (minimum) standards; reduces variance, may prevent fraudulent behavior,
but: can create gaming, paperwork load, and fear → less experimentation & learning
Trust
Psychological safety → enhances learning: experiment, evaluatate, reflect and adapt
When control backfires
Allocation of capacity in a risk-based approach
Defensive medicine
Wider evidence base for defensive medicine
- Williams, Williams, and Williams (2021) shows how physicians across specialties practice defensively, leading to unnecessary diagnostics, procedures, and referrals.
- It increases costs, can reduce quality, and undermines the doctor–patient relationship and shared-decision making.
Case: C-sections in the U.S.
- Mushinski, Zahran, and Frazier (2022) show that litigation risk raised C-section rates for low-risk mothers (where discretion exists), but not for high-risk cases.
- Mushinski, Zahran, and Frazier (2022) show that litigation risk raised C-section rates for low-risk mothers (where discretion exists), but not for high-risk cases.
Trust is needed to facilitate learning
The challenge
Fragmentation
Experimenting in Amsterdam-Noord
Learning healthcare systems
Trust is essential for learning health systems
Control ensures minimum standards, but may create fear and rigidity
Our challenges (aging, rising costs, workforce shortages) require experimentation
Experimentation needs trust: psychological safety to test, fail, and adapt
Coordination and shared learning across systems can turn local insights into international progress
Appendix
How the inspection budget is updated
Let total inspections per period be \(B\)
allocation shares \(a_A^{(t)}\) and \(a_B^{(t)}=1-a_A^{(t)}\).
\[ I_A^{(t)} = \mathrm{round}\!\big(B\,a_A^{(t)}\big), \quad I_B^{(t)} = B - I_A^{(t)}. \]
Detections (true event probabilities \(p_A, p_B\)):
\[ D_A^{(t)} \sim \mathrm{Binomial}\!\big(I_A^{(t)}, p_A\big), \quad D_B^{(t)} \sim \mathrm{Binomial}\!\big(I_B^{(t)}, p_B\big). \]
Naïve observed “risk” (detections per capita) with smoothing \(\varepsilon>0\):
\[ r_A^{(t)}=\frac{D_A^{(t)}}{N_A}, \quad r_B^{(t)}=\frac{D_B^{(t)}}{N_B}. \]
Target share from observed risks:
\[ \tilde a_A^{(t+1)}=\frac{r_A^{(t)}+\varepsilon}{r_A^{(t)}+r_B^{(t)}+2\varepsilon}, \quad \tilde a_B^{(t+1)} = 1-\tilde a_A^{(t+1)}. \]
Update with inertia \(\lambda\in(0,1]\):
\[ a_A^{(t+1)}=(1-\lambda)\,a_A^{(t)}+\lambda\,\tilde a_A^{(t+1)}, \qquad a_B^{(t+1)}=1-a_A^{(t+1)}. \]
References
39th EPSO Conference in Dublin-Ireland