Science reports that recordings of 400 university students found men and women both say about 16,000 words per day: We have developed a method for recording natural language using the electronically activated recorder (EAR) … participants wear the EAR for several days during their waking hours. The device is programmed to record for 30 s every 12.5 min. All captured words spoken by the participant are transcribed. … The data suggest that women spoke on average 16,215 (SD = 7301) words and men 15,669 (SD = 8633) words over an assumed period of, on average, 17 waking hours … the difference does not meet conventional thresholds for statistical significance.
Here are the contract specific rules as stated on Intrade. --------------------For this contract to settle (expire) at 100 ($10.00) the following conditions must be met:
1. The study must be first reported on or after July 10th, 20072. The study must find that women talk exactly 10% or more then men over a typical whole day3. The study must include more then just university students, but need not include university students
This contract will settle (expire) at 0 ($0.00) if the next study finds that women do not talk exactly 10% or more then men.-------------------
I suspect that the point 2. above has been misstated, or at least sounds ambigous. More precise statements could be that -... women talk exactly 10% or more in excess of men ... or,... women talk exactly 110% as much as men or more ...
You should also add restrictions as to what sample size, and/or level of statistical significance would be considered acceptable for settling this contract.
Mark, how many words people say per minute of conversation and how many words they say per day sound like very different questions to me - the difference of course being: how many minutes a day do people converse?
The large amount of within-group variation, and the smallness of the between-group differences in comparison, was the whole point of the research.
For graphs of the whole distribution of word counts from a more representative sample of Americans (all ages, regions and educational levels), see here.
Those are word counts from brief (10-minute) telephone conversations -- but many other studies, of many other sorts of interactions, from a number of different countries, all show essentially the same results.
I'd hazard that this 'recording' had a tendency to suppress some people, and encourage others to 'perform', even at 1 in 25 for any given statement to be recorded. I, for one, would feel weird about going straight back to my room to watch TV or study all afternoon, but I'd also be self-conscious about talking. If men were more likely to 'perform' while women more likely to be suppressed, this would skew the samples. I don't know how to adjust for this observation effect, if it exists, short of secretly bugging a sample, which strikes me as unlikely to get past the human subjects ethics committee.
Thanks Chi. Now that is pretty interesting! And it sure tells a very different story than the abstract. Why focus so much on the average and not the wide distribution of data points? Sometimes I don't get academics ;-).
Brad, they do just that at the last paragraph of page one, and the rest of page 2 in the Supplemental data (that you can't see, heh) at http://www.sciencemag.org/c...
Summarized, the results are:* Both men and women's word counts were positively skewed, but of the six groups of subjects, all mens' subgroups had higher skews than the womens'.* Men and women were equally represented (50/50, although total participation was 47/53) in the top 17% of word users from each subgroup.* The top 3 word users drawn from the total pool were men
So this makes the hypothesis that: a few chatty women serve as outliers we all remember, unlikely.
Visually, in the histogram for word usage of men vs. women, it looks like the women's peak is slightly higher (mode ~16000) than the men's (mode ~10000), with more men in the high tail (reaching ~48000/42000).
Above, I'm talking about the usefulness of a conclusion like "women talk only slightly more than men on average". I'd like to see the actual distributions, not statistics about them. Each data point is obviously bounded on the low end by 0, but not necessarily by 2x the average on the high end. The Science article is behind a paid firewall. I assume it has the pictures I want unless the study was striving to be politically correct blather. The men have a flatter distribution. I'd bet its a little heavier on one end too. If the peculiarities of the distributions are left out of the popular narrative, it seems to me something is lost. Isolate and compare the top 50%, 25%, 10% of both groups. How do things look?
If the point was just to debunk the 20K v. 7K figure, well, ok, duh. Experience is gonna tell you there would be lots of variance in a result that showed that too.
I see no problem giving exact data values in an academic paper. They didn't say their estimate was that precise, just their data. The large standard deviation shows that individuals vary widely in talkativeness. I don't see how that undermines their basic result.
Great point Felix. Numbers quoted to that level of precision given the SD's seem to be ridiculous.
And I agree that ideas futures for forthcoming studies are a natural good idea. Perhaps such articles could have a standard link to the relevant ideas future, the way slate articles have standard links to fray commentary.
I'm not a statistician but presume "SD" means standard deviation. Aren't those pretty big SD numbers? Real flat curves?
Heck, maybe this is just another instance of words doing a bad job of describing a graph. Instead of 16215 (not, mind you, 16214) and 15669, it might be, "mostly between 9000 and 23000" and "mostly between 7000 and 24000". Heck, that even supports the validity of the numbers as it sounds more plausibly expectable, given that males usually have flatter bell curves than females. But, that's probably 'cause of diet. (Absolute SD values would not alone imply bell curve pancakeness, right?)
Here are the contract specific rules as stated on Intrade. --------------------For this contract to settle (expire) at 100 ($10.00) the following conditions must be met:
1. The study must be first reported on or after July 10th, 20072. The study must find that women talk exactly 10% or more then men over a typical whole day3. The study must include more then just university students, but need not include university students
This contract will settle (expire) at 0 ($0.00) if the next study finds that women do not talk exactly 10% or more then men.-------------------
I suspect that the point 2. above has been misstated, or at least sounds ambigous. More precise statements could be that -... women talk exactly 10% or more in excess of men ... or,... women talk exactly 110% as much as men or more ...
You should also add restrictions as to what sample size, and/or level of statistical significance would be considered acceptable for settling this contract.
Mark, how many words people say per minute of conversation and how many words they say per day sound like very different questions to me - the difference of course being: how many minutes a day do people converse?
The large amount of within-group variation, and the smallness of the between-group differences in comparison, was the whole point of the research.
For graphs of the whole distribution of word counts from a more representative sample of Americans (all ages, regions and educational levels), see here.
Those are word counts from brief (10-minute) telephone conversations -- but many other studies, of many other sorts of interactions, from a number of different countries, all show essentially the same results.
You can now bet on this topic at InTrade!
I'd hazard that this 'recording' had a tendency to suppress some people, and encourage others to 'perform', even at 1 in 25 for any given statement to be recorded. I, for one, would feel weird about going straight back to my room to watch TV or study all afternoon, but I'd also be self-conscious about talking. If men were more likely to 'perform' while women more likely to be suppressed, this would skew the samples. I don't know how to adjust for this observation effect, if it exists, short of secretly bugging a sample, which strikes me as unlikely to get past the human subjects ethics committee.
Also interesting is that men had a larger variance than women.
Thanks Chi. Now that is pretty interesting! And it sure tells a very different story than the abstract. Why focus so much on the average and not the wide distribution of data points? Sometimes I don't get academics ;-).
Do swear words count?
Brad, they do just that at the last paragraph of page one, and the rest of page 2 in the Supplemental data (that you can't see, heh) at http://www.sciencemag.org/c...
Summarized, the results are:* Both men and women's word counts were positively skewed, but of the six groups of subjects, all mens' subgroups had higher skews than the womens'.* Men and women were equally represented (50/50, although total participation was 47/53) in the top 17% of word users from each subgroup.* The top 3 word users drawn from the total pool were men
So this makes the hypothesis that: a few chatty women serve as outliers we all remember, unlikely.
Visually, in the histogram for word usage of men vs. women, it looks like the women's peak is slightly higher (mode ~16000) than the men's (mode ~10000), with more men in the high tail (reaching ~48000/42000).
Above, I'm talking about the usefulness of a conclusion like "women talk only slightly more than men on average". I'd like to see the actual distributions, not statistics about them. Each data point is obviously bounded on the low end by 0, but not necessarily by 2x the average on the high end. The Science article is behind a paid firewall. I assume it has the pictures I want unless the study was striving to be politically correct blather. The men have a flatter distribution. I'd bet its a little heavier on one end too. If the peculiarities of the distributions are left out of the popular narrative, it seems to me something is lost. Isolate and compare the top 50%, 25%, 10% of both groups. How do things look?
If the point was just to debunk the 20K v. 7K figure, well, ok, duh. Experience is gonna tell you there would be lots of variance in a result that showed that too.
I see no problem giving exact data values in an academic paper. They didn't say their estimate was that precise, just their data. The large standard deviation shows that individuals vary widely in talkativeness. I don't see how that undermines their basic result.
Yeah, this result looks slightly more useful than noting that the average human being has one breast and one testicle (SD = 1).
Great point Felix. Numbers quoted to that level of precision given the SD's seem to be ridiculous.
And I agree that ideas futures for forthcoming studies are a natural good idea. Perhaps such articles could have a standard link to the relevant ideas future, the way slate articles have standard links to fray commentary.
I'm not a statistician but presume "SD" means standard deviation. Aren't those pretty big SD numbers? Real flat curves?
Heck, maybe this is just another instance of words doing a bad job of describing a graph. Instead of 16215 (not, mind you, 16214) and 15669, it might be, "mostly between 9000 and 23000" and "mostly between 7000 and 24000". Heck, that even supports the validity of the numbers as it sounds more plausibly expectable, given that males usually have flatter bell curves than females. But, that's probably 'cause of diet. (Absolute SD values would not alone imply bell curve pancakeness, right?)