Good science takes time and usually raises more questions
than it answers. This is no exception.
When #thedress first came out in February 2015, vision
scientists had plenty of ideas why some people might be
seeing it differently than others, but no one knew for
sure. Now we have some evidence as to what might be going on.
The illumination source in the original image of
the dress is unclear. It is unclear whether the image
was taken in daylight or artificial light, and if the light
comes from above or behind. If things are unclear, people
assume that it was illuminated with the light that they
have seen more often in the past. In general, the human
visual system has to take the color of the illumination
into account when determining the color of objects. This
is called color constancy.
That’s why a sweater looks largely the same inside a
house and outside, even though the wavelengths hitting the
retina are very different (due to the different
illumination). So if someone assumes blue light, they
will mentally subtract that and see the image as yellow.
If someone assumes yellow light, they will mentally
subtract it and see blue. The sky is blue, so if someone
assumes daylight, they will see the dress as gold.
Artificial incandescent light is relatively long-wavelength
(appearing yellow-ish), so if someone assumes that, they
will see it as blue. People
who get up in the morning see more daylight
lifetime and tend to see the dress as white and gold,
up later and stay up late see more artificial light in
their lifetime and tend to see the dress as
black and blue.
This is a flashy result. Which should be concerning because
scientific publishing seems to have traded off rigor with appeal in
the past. However, I really do not believe that this
was the case here. In terms of scientific standards, the paper
has the following features:
*High power: > 13,000 participants
*Conservative p-value: Voluntarily adopted p
< 0.01 as a reasonable significance threshold to guard
against multiple comparison issues.
*Internal replication prior to publication:
This led to a publication delay of over a year, but it is
important to be sure.
*No excluding of participants or flexible
stopping: Everyone who had taken the survey by the
time of lodging the paper for review at the journal was
*#CitizenScience: As this effect holds up “in
the wild”, it is reasonable to assume that it doesn’t fall
apart outside of carefully controlled laboratory conditions.
*Open science: Shortly (once I put the
infrastructure in place), data and analysis code will be made
openly available for download. Also, the paper was published –
on purpose – in an open-access journal.
Good science takes time and usually raises more questions than
it answers. This is no exception. If you want to help us out,
take this brief 5-minute survey.
The more data we have, the more useful the data we already have
This post also appeared at Pascal’s website Pascal’s Pensées