4 ms·

I remain skeptical. The effect size here is really small. They presented subjects with an image of a rectangle sitting on a table with a human face either look

I remain skeptical. The effect size here is really small. They presented subjects with an image of a rectangle sitting on a table with a human face either looking at it or not looking at it, from either the left or the right, and then asked them to work out what angle the rectangle would need to be tilted at to be perfectly balanced on one point. The idea being that if people model vision as a force beam emitted from the eyes, then people will assume that that force needs to be countered by leaning the rectangle slightly towards the eyes. The difference between a face looking at it or not looking at it was less than 1 degree of rotation[0]. They also don't say how many pixels tall the rectangle was. For all we know it might only be 1 pixel different in the position at the top. [0] https://www.pnas.org/content/pnas/116/1/328/F2.large.jpg https://www.pnas.org/content/pnas/116/1/328/F2.large.jpg

gus_massa 8y ago
If the images in the article are at scale, the height of the rectangle is 1/8 of the width of the image, so if we assume that it is as width as posible with a screen of ~1200 pixels, then the height of the rectangle is ~200 pixels. Also, 1 degree = pi/180 ~=1/60 radians. The difference they are measuring is approximately 1/2 degree. So with a height of 200 the change of 1/2 degree is approximately 2 pixels. [In case it is not obvious, I'm making a lot of assumptions and approximation.] The text explain the size of the groups: 175 (online), 25, 15, 15. [I'm not sure if this numbers are counted before or after eliminating the subjects that didn't understand the task.] Looking at the graph my objection is that the three results where the "face" was looking at the rectangle are too consistent. They are .64±.23, .67±.30 and .63±.26. The difference in the three averages is only .04, but the standard deviation is like .25 in each case. And with 15 subjects you don't reduce the deviation of the average too much, so these values look fishy. The other five results look more natural. The standard deviation is something like .20 or .30, and they are scattered randomly in a range of a similar width. Are they reporting all the experiments they made, or they made a lot of experiments and cherrypicked four of them? Also, in the graphic "The asterisk indicates D significantly greater than 0, P < 0.05." not that the two measurements have a difference that is statistically significantly. In particular in the third graph the error bars are too close to each other.

gus_massa 8y ago

If the images in the article are at scale, the height of the rectangle is 1/8 of the width of the image, so if we assume that it is as width as posible with a screen of ~1200 pixels, then the height of the rectangle is ~200 pixels. Also, 1 degree = pi/180 ~=1/60 radians. The difference they are measuring is approximately 1/2 degree. So with a height of 200 the change of 1/2 degree is approximately 2 pixels. [In case it is not obvious, I'm making a lot of assumptions and approximation.] The text explain the size of the groups: 175 (online), 25, 15, 15. [I'm not sure if this numbers are counted before or after eliminating the subjects that didn't understand the task.] Looking at the graph my objection is that the three results where the "face" was looking at the rectangle are too consistent. They are .64±.23, .67±.30 and .63±.26. The difference in the three averages is only .04, but the standard deviation is like .25 in each case. And with 15 subjects you don't reduce the deviation of the average too much, so these values look fishy. The other five results look more natural. The standard deviation is something like .20 or .30, and they are scattered randomly in a range of a similar width. Are they reporting all the experiments they made, or they made a lot of experiments and cherrypicked four of them? Also, in the graphic "The asterisk indicates D significantly greater than 0, P < 0.05." not that the two measurements have a difference that is statistically significantly. In particular in the third graph the error bars are too close to each other.