[From Bill Powers (2009.11.18.2335 MDT)]
Martin Taylor 2009.11.18.23.54
The problem is that after the
first time a person has bid on the object, he has established a mental
price for that object. We are asking the question of whether a prior
suggestion of a price will affect the price the person
sets.
I’m not assuming that people work that way. In my model, the (composite)
person establishes a reference price that remains the same throughout.
What the person changes is the bid price, with those changes being like
an attempt to bring the price down to the reference price. The higher the
asked price is relative to the reference level, the harder the person
tries to lower the price, but that increased effort is caused by an
increased error signal when the asked price goes up. The bid price
increases, but not as much as it would if the buyer didn’t resist what he
sees as the seller’s excessive asked price.
This would be easier to understand if we included more than just this one
control system. If you ask why the person doesn’t just pay the asking
price whatever it is, I think the reason is simply that the person has a
limited budget and needs to spend money on other things, too. So spending
the money on the item disturbs other control processes like buying food,
keeping warm, running a car, and so on. If it weren’t for the need to
spend the money on other things, the person would just pay the asking
price, the way we do when we shop in stores.
It often happens that we want or need something, but find that the price
is more than we wanted to spend on it. This is like one round of
bargaining. The asked price for milk is $3.50 per gallon instead of the
$2.85 we anticipated paying, so we have to decide whether to buy it
anyway – that is, to offer more than our reference price for it. Maybe
our gain for keeping the price down is so high that the price we’re
willing to pay is only $3.00 when the asked price is $3.50. In that case
we don’t buy the milk, but perhaps go shopping at a different store, or
decide to buy skimmed milk which is closer to $2.85 per gallon, though we
don’t want it as much.
I’m challenging the assumption that subtle effects of perceptions on
other perceptions are enough to cause significant changes in behavior.
I’m saying there is another explanation for the observed apparent
effects in the Arielly experiment. The control-system model fits the data
well enough so my challenge isn’t thrown out automatically, which leaves
us to ask how well another theory can plausibly explain the same
observations. I don’t know if we could come up with a plausible
control-system model in all such cases, but it seems to work quite well
in this case.
Of course we could just fit y = ax + b to the data, but that doesn’t
suggest any explanation for why that should be observed. In fact, that is
the observation that needs to be explained. Aside from the control-system
model, what other explanation is there, beside “That’s just the way
it happens”?
We want that prior
suggestion to be just the random number, but if the person already has a
price in mind, it’s unlikely that the prior suggestion will be only the
random number. It’s quite likely that the price bid the last time will
have a much stronger influence than the random number, perhaps to the
extent of acting as a reference value for the price to be
bid.
I don’t think prior bids have any influence on the behavior, nor do the
social security numbers nor would the random numbers. Of course I could
be wrong, but you’d have to show me a model indicating the mechanism from
which these influences could be deduced, as we deduce the observed
effects from the control system model. Otherwise just saying
“influences” doesn’t tell us anything (certainly nothing
quantitative), and all we really have is curve-fitting.
The first time you do it,
the random number will very probably have a much bigger influence than on
subsequent occasions.
What model are you assuming that makes that seem probable? If you’re just
telling me you have a hunch that this will occur, you’re not making any
quantitative prediction or describing a method for constructing one. Any
amount of influence from the minimum observable to the maximum measurable
would make your statement true: an effect of one millionth of a cent
change in the bid price is clearly much larger than an effect of one
nanocent, so you can hardly go wrong with such a prediction. But how
about a real test?
Even if you scramble a
very large number of objects, you will not be able to guarantee that the
person has forgotten what price he set for that object on a previous
occasion.
So your model says that memory has something to do with it. Maybe it does
– when you work out the model enough so we can determine parameters and
fit it to the observations, we can compare it with the control system
model and see which does better. If your prediction about what will
happen under my proposed experiment turns out to be correct, then we will
find my model doesn’t predict individual behavior as well as it does for
the virtual population control system (if at all), and we will know more
than we know now. But I don’t have to guarantee anything to justify doing
the experiment as I propose. You’re just saying that under your model, if
the person remembers his previous bid, that will influence his next bid.
I am saying that his next bid will be determined by the quantitative
relationships in the control model I proposed, not by memory of the
previous bid. So one of us is wrong. What’s your justification for saying
I’m the one?
To ask someone to ignore the
previous bids is likely to work about as well as a judge asking a jury to
ignore some prejudicial information the prosecution
“inadvertently” let slip about the defendant.
Perhaps so. But you’ll have to spell out the model you’re using which
explains how memory of previous bids affects future bids, and show how it
predicts the data in the experiment I propose. If your model is the
better one, it will predict behavior that is different from the behavior
my model would predict and yours will be quantitatively more like the
observed behavior.
I think we’re at the point here where only quantitative models applied to
real data can resolve any issues.
Best,
Bill P.