Dietrich and List on reasons
Let's return to my recent explorations into the formal structure of reasons. One important approach that I haven't talked about yet is that of Dietrich and List, described in Dietrich and List (2013a), Dietrich and List (2013b), and Dietrich and List (2016).
In essence, Dietrich and List (D&L) propose a rebranded version of multi-attribute utility theory, with an added element of flexibility. I'll briefly explain the theory, then we'll look into what it might say about reasons.
Theoretical presentations of decision theory often take as primitive a utility assignment or preference relation over certain "options" or "outcomes" or "alternatives". We assume, for example, that you prefer a cup of coffee to a cup of peppermint tea. Intuitively, such a preference isn't a basic fact about your psychology. There will be a reason why you prefer the coffee. Perhaps you prefer the taste. Or perhaps you crave the caffeine. In general, if you prefer one alternative to another then this is usually because the alternatives have certain properties that you care about, and you like the properties of the preferred alternative more than the properties of the other alternative.
In more applied areas of decision theory (such as the field of decision analysis), there is a rich body of work that fills in this gap. A straightforward and standard approach is multi-attribute utility theory. Here we assume that the ultimate objects of preference or utility are combinations of properties (or "attributes"). An alternative with relevant properties A, B and C is represented by the list (A,B,C).
D&L propose a similar model, with the same motivation: to explain "where preferences come from" (which is the title of Dietrich and List (2013b)). An agent's preferences over concrete alternatives, they suggest, are determined from an underlying preference order over property combinations. What's new and distinctive about their proposal is that the relevant properties are not held fixed. Instead, every decision context selects a set of properties by which the alternatives are evaluated. As a consequence, we can't simply represent an alternative as lists of properties: if two situations both involve a cup of coffee as an alternative, then the taste of coffee may be selected in the first but not in the second.
D&L's motivation for this added flexibility is to explain certain shifts in preferences. The idea is that the properties we attend to vary from situation to situation. D&L suggest that this might lie behind some apparently irrational behaviour like framing effects.
This is an interesting idea. I'm not entirely convinced. To explain framing effects, the "properties" would have to be individuated hyperintensionally, and then an outcome may no longer determine a list of properties, even if we know which properties are selected. D&L argue that their model can also explain other phenomena like menu-dependence, but here we don't need the added flexibility.
Anyway, all of this looks fine to me – mainly because I'm already a fan of multi-attribute utility theory.
What I want to comment on is D&L's idea that their model provides a model of reasons. In any context, they suggest, the selected properties are the reasons for the agent's preferences and thereby for their choice.
Is this plausible? Are all selected properties reasons for choice? Conversely, are all reasons for choice selected properties?
To assess these questions, we need to know more about what it means for a property to be selected, and about which properties are eligible to be selected. This, in turn, depends on how we construe the "alternatives".
A common idea in economics is that the "options" or "outcomes" or "alternatives" are superficial, coarse-grained, context-independent things like a cup of coffee or a biscuit. Candidate properties of such an alternative might be contains caffeine, tastes like coffee, etc. Real people don't just care about these sorts of properties. They also care about (say) whether taking the biscuit would be polite or impolite. Politeness isn't an intrinsic property of the biscuit. To model realistic preferences and reasons, we therefore need a more fine-grained individuation of alternatives.
Let's assume, then, that the "alternatives" are finely individuated, so that they at least specify everything that matters to the agent. In Dietrich and List (2016) and Dietrich and List (2017), D&L suggest that these fine-grained alternatives can be factored into coarse-grained "options" and a "context" that is the same for each alternative in a given choice situation. We can then distinguish three kinds of properties that might be selected: option properties that are intrinsic to the options, context properties that only depend on the context, and relational properties that depend on both the relevant option and the context (see Dietrich and List (2016, 184ff.)).
With sufficiently fine-grained alternatives, there's at least a chance that an agent's reasons in a choice situation might map onto the selected properties. (We won't miss politeness and other such reasons.) But what does it mean that a property is "selected"?
Suppose you prefer the cup of coffee to the cup of peppermint tea because you prefer the taste of coffee. If the coffee is served in red cups and the tea in blue cups, why can't we say that the "selected" properties include the colour of the cup? Why can't we say that this is the only selected property, even though it is not a reason for you at all? We'd still predict the correct preference over alternatives.
D&L don't give an informative account of what makes a property selected. But here's what they could say. (Perhaps it's what they are saying, I'm not sure.)
(Def) A property is selected (in a situation) iff it is a reason for the agent's preferences over the alternatives (in the situation).
The 'iff' is supposed to be the 'iff' of definition. The idea is to define the selected properties in terms of reasons. We therefore can't go on and analyse reasons in terms of selected properties. Conceptually, reasons are basic.
One might think that (Def) trivially ensures that every reason for an agent's preference or choice can be found among the selected properties. But this is false.
Take a different example. It is raining, and you walk around with an umbrella. Why do you do this? What reason do you have for carrying an umbrella? Well, one good reason is that it is raining. If this were a selected property in your choice situation, it would be a "context property": it's not affected by your choice. But it's probably not a selected property.
Why not? By (Def), a property is selected iff it is a reason for your preference over alternatives. The alternative you have chosen is not to carry an umbrella. This might be an "option", but it's not fine-grained enough to be a credible alternative. Recall that an alternative settles at least everything that you care about. The hypothesis that you carry an umbrella settles almost nothing that you plausibly care about. Here's a slightly more fine-grained and plausible individuation of the relevant alternatives:
(A1) It is raining. You carry an umbrella. Your arm position is a little inconvenient. You remain dry and warm. …
(A2) It is raining. You don't carry an umbrella. You have the hands free. You get wet and cold. …
You prefer (A1) to (A2). Why? What reason (or reasons) do you have for this preference? We want to explain the preference by certain properties instantiated by (A1) and (A2). There are properties of (A1) whose combination you like better than the corresponding properties of (A2). What are these properties?
You see the point. The main thing that makes (A1) better is that you remain dry and warm. If we want to take into account all your reasons, we could also include ("select") the convenience property, which is a reason against (A1). But that it's raining isn't a plausible reason. I certainly wouldn't want this to figure in a multi-attribute utility analysis (unless you have a basic desire for rain).
The point generalises. When we talk about reasons for a preference or choice, we often assume a coarse-grained individuation of the alternatives. We wonder what reasons you have for carrying an umbrella, for buying coffee, against meeting Jones. In this context, all sorts of things show up as reasons that shouldn't figure as "selected properties" in D&L's model, or as attributes in a multi-attribute utility analysis. The reasons we cite are often not properties of the "options" or "outcomes", but of the "state". A reason for buying coffee might be that doing so is likely to keep you awake during a meeting. This reason is a partial description of a "dependency hypothesis", in the sense of Lewis (1981): it says how the things you really care about depend on your choice.
The upshot is that we need to distinguish two kinds of reasons, or two roles for reasons.
One role is internal to the theory of value or desire. A system of value should say what kinds of things ultimately matter and how they matter (compare Dietrich and List (2017) – although they don't see this as part of axiology). Perhaps it ultimately matters to you that you are warm. This is one of your basic desires. If so, it is a (pro tanto) reason for you to prefer possible situations in which you are warm over possible situations in which you are cold. Your desire for warmth partly explains where your preferences over finely individuated alternatives come from.
Once we have a system of value and some coarsely individuated options, we may ask whether a particular option is likely to promote your values. Anything that suggests that it does is a reason in favour of the option. That it's raining suggests that carrying an umbrella promotes your desire to stay warm. So it's a reason in favour of carrying an umbrella. This is the kind of reason that Sher (2019) has tried to model. D&L only model the other kind of reason.
What's the connection between these two kinds of reasons? Perhaps it is this. In both cases, a reason is something that makes a difference to (expected) value: a relevant object that has R is either or better or worse than an otherwise similar object that lacks R.
(These considerations reveal a possible weakness in D&L's account. D&L suggest that some human behaviour can be explained by assuming that we attend to different properties in different choice situations, so that different properties are "selected" in these situations. It's not clear, however, what attention has to do with selection, if selection is defined in terms of reasons.)