This section describes the default prior distributions for the Meridian
model. All prior distributions are specified by the `prior_distribution`

argument, which accepts a
`PriorDistribution`

object. Each parameter has its own argument in the `PriorDistribution`

constructor, and the joint prior distribution assumes that all of the priors are
independent.

Distributions can be specified as either a vector (such as,
`tfp.distributions.Normal([1, 2, 3], [1, 1, 2])`

) or as a scalar (such as
`tfp.distributions.Normal(1, 2)`

). All scalar distributions are broadcast to the
length of the parameter vector that they represent.

`knot_values`

**Parameter:** \(b_k\)

**Default Prior:** `Normal(0, 5)`

**Rationale:**

- Uninformative prior dictating how much time can have an effect.
- Uninformative because you want the flexibility to allow time to have a strong effect.
- Can be learned from the data given multiple geos per time period, and also multiple time periods per knot when the number of knots is low.

`tau_g_excl_baseline`

**Parameter:** \(\tau_g\)

**Default Prior:** `Normal(0, 5)`

**Rationale:**

- Uninformative prior dictating geo-differences.
- Uninformative because you want the flexibility to allow geo to have a strong effect.
- Can be learned from the data given multiple time periods per geo.

`roi_m`

and `roi_rf`

**Parameter:** \(ROI_m,ROI_n^{(rf)}\)

**Default Prior:** `LogNormal(0.2, 0.9)`

**Rationale:**

- This prior says that
*a priori*the mean ROI across channels are 1.83, 50% of ROIs are greater than 1.22, 80% are between 0.5 and 6.0, 95% are between 0.25 and 9.0, and 99% are less than 10.0. - By default, each channel is assigned the same ROI prior.

`beta_m`

and `beta_rf`

**Parameter:** \(\beta_m,\beta_n^{(rf)}\)

**Default Prior:** `HalfNormal(5)`

**Rationale:**

- Uninformative prior on hierarchical mean of geo-level media effects
(
`beta_gm; beta_grf`

) for impression and reach and frequency media channels respectively. - Uninformative because the interpretation of
`beta_m`

can vary widely given transformations, scaling, and kind of media execution.

`eta_m`

and `eta_rf`

**Parameter:** \(\eta_m,\eta_n^{(rf)}\)

**Default Prior:** `HalfNormal(1)`

**Rationale:**

Moderate regularization encourages pooling across geos. This leads to lower variance estimates at the cost of increased bias, and allows the model to use the data more efficiently.

`gamma_c`

**Parameter:** \(\gamma_c\)

**Default Prior:** `Normal(0, 5)`

**Rationale:**

Uninformative because of the wide range of control variables you can possibly see.

`xi_c`

**Parameter:** \(\xi_c\)

**Default Prior:** `HalfNormal(5)`

**Rationale:**

- Uninformative to allow a wide range of geo variation in control variable effects.
- By default, pooling is weaker for control effects than for media effects because control effects are simple linear effects (without the complexity of Hill and Adstock transformations).

`alpha_m`

and `alpha_rf`

**Parameter:** \(\alpha_m,\alpha_n^{(rf)}\)

**Default Prior:** `Uniform(0, 1)`

**Rationale:** Uninformative to allow data to inform the decay rate.

`ec_m`

**Parameter:** \(ec_m\)

**Default Prior:** `TruncatedNormal(0.8, 0.8, 0.1, 10)`

. This is the conditional
distribution \(X|0.1 < X < 10\), where \(X \sim N(0.8,0.8)\).

**Rationale:**

- The data is scaled such that when \(ec=1\), the half-saturation happens at the median of the non-zero media units per capita across geos and time. \(ec=k\) means that the half-saturation happens at \(X\) times the median value.
- This prior has mean near one, which is a reasonable a priori assumption of where the half-saturation happens.
- The truncation is done to keep the parameter within a reasonable range for parameter identifiability.
- If a channel is way under-saturated (\(ec > 10\)) or way over-saturated
(\(ec < 0.1\)), the data does not really contain information about the
half-saturation point anyway. In such cases, the
`ec_m`

parameter determines the shape of the response curve, but shouldn't be interpreted as an accurate estimate of half-saturation.

`ec_rf`

**Parameter:** \(ec_n^{(rf)}\)

**Default Prior:** `LogNormal(0.7, 0.4) + 1`

```
# Tensorflow Probability Syntax
tfp.distributions.TransformedDistribution(
tfp.distributions.LogNormal(0.7, 0.4),
tfp.bijectors.Shift(0.1)
)
```

**Rationale:**

- Moderately informative to prevent non-identification with
`slope_rf`

. - Set in conjunction with the
`slope_rf`

prior so that the prior distribution for optimal frequency has a mean of 2.1 and 90% CI of`[1.0, 4.4]`

. This is considered to be a reasonable range of optimal frequency.

`slope_m`

**Parameter:** \(slope_m\)

**Default Prior:** `Deterministic(1)`

**Rationale:**

- Difficult to learn because of identifiability reasons.
`Deterministic (1)`

means it is restricted to concave Hill curves.- The budget optimization algorithm produces a global optimum when Hill curves are concave. Changing this prior can lead to non-concave Hill curves and budget optimization can no longer produce a global optimum.

`slope_rf`

**Parameter:** \(slope_n^{(rf)}\)

**Default Prior:** `LogNormal(0.7, 0.4)`

**Rationale:**

- Moderately informative to prevent non-identification with
`ec_rf`

. - Set in conjunction with the
`ec_rf`

prior so that the prior distribution for optimal frequency has a mean of 2.1 and 90% CI of`[1, 4.4]`

, a reasonable range of optimal frequency.

`sigma`

**Parameter:** \(\sigma_g\)

**Default Prior:** `HalfNormal(5)`

**Rationale:**

Uninformative because residual variance varies widely by advertiser.