prior-elicitation

Name: prior-elicitation
Author: pymc-labs

28installs10forks

How to install prior-elicitation?

pymc-labs/python-analytics-skills/prior-elicitation

$npx -y skills add pymc-labs/python-analytics-skills --skill prior-elicitation

Installs into the current project.

Summary

Load when the user is choosing priors, running prior predictive checks, calling find_constrained_prior, using PreliZ, or otherwise eliciting domain knowledge into a Bayesian model. Covers weakly informative priors, constrained priors, sensitivity analysis, and elicitation workflo

Files · 1

View on GitHub

SKILL.md

1	# Prior Elicitation for PyMC Models
2
3	## Decision Flowchart: Choosing a Prior Strategy
4
5	```
6	Do you have domain expertise or expert access?
7	├── YES: Can the expert quantify beliefs precisely?
8	│ ├── YES → Expert elicitation (SHELF protocol, PreliZ roulette/quartile)
9	│ └── NO → Constrained priors (find_constrained_prior, PreliZ maxent)
10	└── NO: Do you know the plausible scale of the parameter?
11	├── YES → Weakly informative priors (Normal, HalfNormal, Student-t)
12	└── NO → Use prior predictive checks to calibrate
13	└── Generate predictions → Do they cover plausible outcomes?
14	├── YES → Prior is acceptable
15	└── NO → Tighten or widen prior, repeat
16	```
17
18	### When to Use Each Strategy
19
20	\| Strategy \| Use When \| Example \|
21	\|---\|---\|---\|
22	\| Weakly informative \| You know rough scale but not shape \| `pm.Normal("beta", 0, 10)` for standardized predictors \|
23	\| Constrained \| "95% sure the value is between A and B" \| `pm.find_constrained_prior(pm.Normal, ...)` \|
24	\| MaxEnt \| You have bounds and want least-informative prior \| `preliz.maxent(preliz.Normal(), -1, 1, 0.94)` \|
25	\| Expert-elicited \| Domain expert provides quantiles or probabilities \| SHELF protocol with PreliZ \|
26	\| Hierarchical \| Group-level parameters with partial pooling \| `pm.Normal("mu_group", mu=mu_hyper, sigma=sigma_hyper)` \|
27
28	## pm.find_constrained_prior
29
30	Finds distribution parameters such that a specified mass falls within given bounds.
31
32	```python
33	import pymc as pm
34
35	# "I'm 95% sure the effect is between -2 and 2"
36	params = pm.find_constrained_prior(
37	pm.Normal,
38	lower=-2, upper=2,
39	mass=0.95,
40	init_guess={"mu": 0, "sigma": 1},
41	)
42	# Returns: {"mu": 0.0, "sigma": 1.02}
43
44	# Use directly in model
45	with pm.Model() as model:
46	beta = pm.Normal("beta", **params)
47	```
48
49	### Common Patterns
50
51	```python
52	# Positive parameter, 90% between 0.1 and 10
53	params = pm.find_constrained_prior(
54	pm.LogNormal,
55	lower=0.1, upper=10,
56	mass=0.90,
57	init_guess={"mu": 0, "sigma": 1},
58	)
59
60	# Rate parameter, 80% between 0.01 and 0.5
61	params = pm.find_constrained_prior(
62	pm.Gamma,
63	lower=0.01, upper=0.5,
64	mass=0.80,
65	init_guess={"alpha": 2, "mu": 0.1},
66	)
67
68	# Bounded parameter (probability), 95% between 0.2 and 0.8
69	params = pm.find_constrained_prior(
70	pm.Beta,
71	lower=0.2, upper=0.8,
72	mass=0.95,
73	init_guess={"alpha": 5, "beta": 5},
74	)
75	```
76
77	See `references/find_constrained_prior.md` for full API details.
78
79	## PreliZ Integration
80
81	PreliZ is a library for prior elicitation. It provides tools to translate domain knowledge into probability distributions.
82
83	### Maximum Entropy (maxent)
84
85	Find the least-informative distribution consistent with constraints:
86
87	```python
88	import preliz as pz
89
90	# Least-informative Normal with 94% mass in [-1, 1]
91	dist = pz.maxent(pz.Normal(), -1, 1, 0.94)
92
93	# Least-informative HalfNormal with 94% mass below 5
94	dist = pz.maxent(pz.HalfNormal(), 0, 5, 0.94)
95
96	# Use result in PyMC
97	with pm.Model():
98	sigma = pm.HalfNormal("sigma", sigma=dist.sigma)
99	```
100
101	### Roulette (chip-and-bin elicitation)
102
103	Interactive method where an expert allocates "chips" to bins:
104
105	```python
106	# Define bins and expert-allocated chip counts
107	pz.roulette(x_min=0, x_max=100, nrows=10)
108	# Opens interactive widget — expert distributes chips across bins
109	# Returns fitted distribution
110	```
111
112	### Predictive Finder (predictive_elicitation)
113
114	Elicit priors by reasoning about observable predictions:
115
116	```python
117	def my_model(x, beta0, beta1, sigma):
118	mu = beta0 + beta1 * x
119	return pz.Normal(mu=mu, sigma=sigma)
120
121	# Expert specifies: "when x=5, y is typically between 10 and 30"
122	pz.predictive_finder(my_model, target=pz.Normal())
123	```
124
125	### Quartile Method
126
127	Specify distribution via expert-provided quartiles:
128
129	```python
130	# Expert says: "median is 50, Q1 is 30, Q3 is 70"
131	dist = pz.quartile(pz.Normal(), q1=30, q2=50, q3=70)
132	```
133
134	See `references/preliz.md` for comprehensive PreliZ reference.
135
136	## Prior Predictive Interpretation Checklist
137
138	Always run prior predictive checks before sampling:
139
140	```python
141	with model:
142	prior_pred = pm.sample_prior_predictive(draws=500, random_seed=42)
143
144	# prior_pred is a DataTree (ArviZ 1.1)
145	prior_samples = prior_pred["prior"].ds
146	prior_predictive = prior_pred["prior_predictive"].ds
147	```
148
149	### What to Check
150
151	1. Range of predictions: Do simulated outcomes cover the plausible data range?
152	- Plot: `az.plot_ppc_dist(prior_pred, group="prior_predictive", kind="ecdf")`
153	2. Impossible values: Are any simulated outcomes physically imposs

People who liked this also viewed

6 picks

Type

prior-elicitation

28installs10forks

View on GitHub

How to install prior-elicitation?

pymc-labs/python-analytics-skills/prior-elicitation

$npx -y skills add pymc-labs/python-analytics-skills --skill prior-elicitation

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Prior Elicitation for PyMC Models
2
3	## Decision Flowchart: Choosing a Prior Strategy
4
5	```
6	Do you have domain expertise or expert access?
7	├── YES: Can the expert quantify beliefs precisely?
8	│ ├── YES → Expert elicitation (SHELF protocol, PreliZ roulette/quartile)
9	│ └── NO → Constrained priors (find_constrained_prior, PreliZ maxent)
10	└── NO: Do you know the plausible scale of the parameter?
11	├── YES → Weakly informative priors (Normal, HalfNormal, Student-t)
12	└── NO → Use prior predictive checks to calibrate
13	└── Generate predictions → Do they cover plausible outcomes?
14	├── YES → Prior is acceptable
15	└── NO → Tighten or widen prior, repeat
16	```
17
18	### When to Use Each Strategy
19
20	\| Strategy \| Use When \| Example \|
21	\|---\|---\|---\|
22	\| Weakly informative \| You know rough scale but not shape \| `pm.Normal("beta", 0, 10)` for standardized predictors \|
23	\| Constrained \| "95% sure the value is between A and B" \| `pm.find_constrained_prior(pm.Normal, ...)` \|
24	\| MaxEnt \| You have bounds and want least-informative prior \| `preliz.maxent(preliz.Normal(), -1, 1, 0.94)` \|
25	\| Expert-elicited \| Domain expert provides quantiles or probabilities \| SHELF protocol with PreliZ \|
26	\| Hierarchical \| Group-level parameters with partial pooling \| `pm.Normal("mu_group", mu=mu_hyper, sigma=sigma_hyper)` \|
27
28	## pm.find_constrained_prior
29
30	Finds distribution parameters such that a specified mass falls within given bounds.
31
32	```python
33	import pymc as pm
34
35	# "I'm 95% sure the effect is between -2 and 2"
36	params = pm.find_constrained_prior(
37	pm.Normal,
38	lower=-2, upper=2,
39	mass=0.95,
40	init_guess={"mu": 0, "sigma": 1},
41	)
42	# Returns: {"mu": 0.0, "sigma": 1.02}
43
44	# Use directly in model
45	with pm.Model() as model:
46	beta = pm.Normal("beta", **params)
47	```
48
49	### Common Patterns
50
51	```python
52	# Positive parameter, 90% between 0.1 and 10
53	params = pm.find_constrained_prior(
54	pm.LogNormal,
55	lower=0.1, upper=10,
56	mass=0.90,
57	init_guess={"mu": 0, "sigma": 1},
58	)
59
60	# Rate parameter, 80% between 0.01 and 0.5
61	params = pm.find_constrained_prior(
62	pm.Gamma,
63	lower=0.01, upper=0.5,
64	mass=0.80,
65	init_guess={"alpha": 2, "mu": 0.1},
66	)
67
68	# Bounded parameter (probability), 95% between 0.2 and 0.8
69	params = pm.find_constrained_prior(
70	pm.Beta,
71	lower=0.2, upper=0.8,
72	mass=0.95,
73	init_guess={"alpha": 5, "beta": 5},
74	)
75	```
76
77	See `references/find_constrained_prior.md` for full API details.
78
79	## PreliZ Integration
80
81	PreliZ is a library for prior elicitation. It provides tools to translate domain knowledge into probability distributions.
82
83	### Maximum Entropy (maxent)
84
85	Find the least-informative distribution consistent with constraints:
86
87	```python
88	import preliz as pz
89
90	# Least-informative Normal with 94% mass in [-1, 1]
91	dist = pz.maxent(pz.Normal(), -1, 1, 0.94)
92
93	# Least-informative HalfNormal with 94% mass below 5
94	dist = pz.maxent(pz.HalfNormal(), 0, 5, 0.94)
95
96	# Use result in PyMC
97	with pm.Model():
98	sigma = pm.HalfNormal("sigma", sigma=dist.sigma)
99	```
100
101	### Roulette (chip-and-bin elicitation)
102
103	Interactive method where an expert allocates "chips" to bins:
104
105	```python
106	# Define bins and expert-allocated chip counts
107	pz.roulette(x_min=0, x_max=100, nrows=10)
108	# Opens interactive widget — expert distributes chips across bins
109	# Returns fitted distribution
110	```
111
112	### Predictive Finder (predictive_elicitation)
113
114	Elicit priors by reasoning about observable predictions:
115
116	```python
117	def my_model(x, beta0, beta1, sigma):
118	mu = beta0 + beta1 * x
119	return pz.Normal(mu=mu, sigma=sigma)
120
121	# Expert specifies: "when x=5, y is typically between 10 and 30"
122	pz.predictive_finder(my_model, target=pz.Normal())
123	```
124
125	### Quartile Method
126
127	Specify distribution via expert-provided quartiles:
128
129	```python
130	# Expert says: "median is 50, Q1 is 30, Q3 is 70"
131	dist = pz.quartile(pz.Normal(), q1=30, q2=50, q3=70)
132	```
133
134	See `references/preliz.md` for comprehensive PreliZ reference.
135
136	## Prior Predictive Interpretation Checklist
137
138	Always run prior predictive checks before sampling:
139
140	```python
141	with model:
142	prior_pred = pm.sample_prior_predictive(draws=500, random_seed=42)
143
144	# prior_pred is a DataTree (ArviZ 1.1)
145	prior_samples = prior_pred["prior"].ds
146	prior_predictive = prior_pred["prior_predictive"].ds
147	```
148
149	### What to Check
150
151	1. Range of predictions: Do simulated outcomes cover the plausible data range?
152	- Plot: `az.plot_ppc_dist(prior_pred, group="prior_predictive", kind="ecdf")`
153	2. Impossible values: Are any simulated outcomes physically imposs

People who liked this also viewed

6 picks

Type