The Model-Parameter-Value Model
The Model-Parameter-Value model categorizes claims on three levels: a claim is about the model, about a parameter (including its type) or about a value.
The claim:
Two people who disagree about a conclusion usually disagree on one of three levels: the model, a parameter, or a value. When they argue on different levels, no argument lands.
The three levels
| Level | Question | Light as example |
|---|---|---|
| Model | Which things are related, and how? | Perceived brightness depends on the light that reaches the eye and on the sensitivity of the viewer. |
| Parameter | Which axes does the model have, and of what type? | Luminance: one continuous axis from dark to bright. |
| Value | Where on the axis? | This screen is too bright. |
“Parameter” is used as in physics: one axis of the model’s parameter space.
Two people can share the model and the parameter and still disagree about the value. Neither is wrong about the physics.
The levels stack
A value needs a parameter. A parameter needs a model. A disagreement on a lower level cannot be settled on a higher one.
One question locates it:
Do we disagree about the model, a parameter, or a value?
Type: the disagreement inside the parameter level
A parameter has a type: binary, ordered or continuous, and one axis or two. The type is usually assumed and rarely stated.
Are good and bad the two ends of one axis, or two separate axes, so that something can be both or neither? Psychology argued this for decades about positive and negative feeling. The dispute narrowed to how the axes had been defined.
Until the type is agreed, values cannot be compared.
Two styles of building a model
| Compressive | Exhaustive | |
|---|---|---|
| Starts with | few parameters | many parameters, listed up front |
| Changes the model when | a parameter is added on sufficient reason, or because the discussion needs it | a parameter is removed when it is shown not to matter |
| Parameters | broad: one parameter holds many effects | narrow: one parameter per effect |
| Move in a discussion | reduce: “that is part of X” | add: “what about Y?” |
| Typical argument | the relevant parameters point this way | this case breaks it |
| Assumptions hide in | what is left out | the values that are filled in |
| Risk | the omitted parameter was the one that mattered | the list has no end and nothing gets weighed |
| Check | flip test: how large must an omitted parameter be to reverse the conclusion? | removal test: which parameters can be deleted without changing the conclusion? |
Compressive does not mean compression at all cost. A parameter is added when there is sufficient reason for it, or when the discussion needs it.
The two styles argue on different levels:
- An exhaustive objection is a model-level move: the model lacks parameter Y.
- The compressive reply is a value-level move: Y is too small to matter.
- An objection names a parameter. It becomes an argument when it carries a value large enough to change the conclusion.
- An edge case is a parameter that matters only at extreme values.
Both styles look for the dominant parameter. The exhaustive style succeeds when it finds one parameter strong enough to be the bottleneck. At that point its model is as small as the compressive one.
Example: one case for each style
A prediction starts from a trend. The compressive style extrapolates it. The exhaustive style lists what could stop it.
| Case | Compressive | Exhaustive | Outcome |
|---|---|---|---|
| Chip density | doubles about every two years, so it keeps doubling | limits: heat, lithography, cost, physics | It kept doubling for about fifty years. No listed limit was the bottleneck in that range. |
| Height of a child | 6 cm per year, so above 3 metres at age 40 | limits: the growth plates close | Growth stops. One limit was the bottleneck. |
One question in both styles: how tall is a child at age 40? Compressive: one broad parameter, extrapolated. A second parameter is added (dashed) when there is sufficient reason. Exhaustive: twelve narrow parameters, listed up front. One of them is the bottleneck.
A prediction cannot be tested when it is discussed
- A prediction with many unknowns cannot be tested at the time of the discussion. This holds for both styles.
- The argument is therefore not yet about truth. It is about which model to use: a model-level disagreement.
- The two checks prove nothing. They bound the claim.
- A prediction becomes testable later if it carries a date and a threshold. An objection is testable now if it carries a value.
Related concepts
- Structural versus parametric uncertainty (modelling): uncertainty about the form of a model is separate from uncertainty about its inputs. This matches the model and value levels.
- Opening stage (pragma-dialectics): the parties in a discussion agree on starting points before they argue.
- Double crux: find the one belief that would change both conclusions.
- One-factor-at-a-time versus global sensitivity analysis: a related split in simulation work: varying parameters one by one misses their interactions.
- Scale types (Stevens, 1946): nominal, ordinal, interval, ratio.
- Levins (1966): no model maximizes generality, realism and precision at once.
Limitations
MPV is itself a compressive model: three broad parameters. It has that style’s risk: it leaves things out. It does not describe how the levels interact. A change of type, for example, can change the model.