Hmm that's a lot of scenarios, I'll have to find the time! But I'm interested in stress-testing my model as well + vetting it for extensibility with the multiple scenarios. I'll get back to you soon.
Insights copied from my writeup:For an individual, the riskiness of going outside scales linearly with four factors: the number of people they encounter (N_p_b), the number of surfaces they touch (N_s), the number of other people that have touched those surfaces since the surfaces were last sanitized (N_p_s), and the probability pᵢ that any given person outside is contagious.
Surgical masks offer you the same protection as 3x increased distance. T-shirt masks? Same as 2x distance.
6 feet of distancing keeps you pretty darn safe. BUT any amount less than that is much, much riskier!
Cool, only I went to some length to reduce the chance of getting infected by pizza delivery - something that I can't do to the same degree with food shopping (though I tried). Now I have no clue what my total prob is :-o
I modified the spreadsheet decreasing the probability of a random person being infected from 100% to 5% to see how it affect correlation.Correlation dropped to ~ 0.6.
This shows that I was wrong when assumed that people don't think about the worst-case when filling polls. The next time I should do math BEFORE drawing conclusions.
Besides my calculations are based on a huge number of assumptions I am not sure to be true. So, I'd better trust poll results than my calculations when comparing risks of different activities.
I think the spreadsheet still may be useful as it gives intuition on how pizza index maps to actual probabilities.
If working in an office with an infected person is not a significant risk, then the current work-from-home policies seem like a bad idea.
I'd want the entire office to have been tested, not just asymptomatic and untested, and that pattern-matches to low confidence hedging to avoid admitting if I was wrong... But given the actual asymptomatic numbers in the data, and also that office workers are less likely to be immune compromised, and also some vague thoughts about low viral load infections, I think that it might not be surprising if a small office had five infections and only one person became symptomatic.
You may be right. We could check that by investigating if infected people tend to cluster by office.
If your hypothesis is correct, we should find it rare if only one person in an office is ever infected, since symptoms don't appear until at least four doublings.
I added correlation calculations to the sheet.Correlation is 0.7351706336
I skipped some activities because I didn't know how to classify them in terms of the number of touches, bodily contacts and time spent in proximity.E.G. I don't know whether to count Pizza delivery as a bodily contact with the cook (you eat stuff he touched).
I also supposed that each activity except kissing is done with protective measures.
So, what does this suggest about the risk of being the person who delivers pizza? If someone delivers N pizzas in a week, is their total risk from pizza delivery 1, N, log(N)+1, e^(N-1), or something else entirely? Given that the delivery driver has that risk this week, is the risk of getting a delivery next week much higher than this week?
I see that as "In conditions that crowded, each infected person infected an average of one person per day"
An office is approximately half as many hours per day, and in my limited experience in cruises and office work I spent about as much interaction-with-person per hour in each. (If I was going to roll fort saves vs disease in a model, I would roll once per interaction, not once per day if any interaction). Thus, with those numbers, I estimate that if there is an N infected people in a workplace, two days later there will be 2N infected people there, until the conditions for exponential growth are no longer true (which simplify to "everyone is infected or no longer coming to the office").
That estimate is based on a homogeneous population of spherical vectors, so it won't actually hold up in reality, but I lack any reasonable basis to predict which direction I've erred.
On the Diamond Princess, the original infected person was on the ship for five days. Quarantine went into effect 11 days after that.
In those 11 days, about 20% of the passengers and crew became infected (half of those symptomatic). that's about 700 out of 3711.
Not sure what that tells us, though.
1 person to 700 people is 9 doublings in 11 days. Call it one doubling a day. That means for every infected person on the ship, you have a 1/3700 chance of becoming infected that day. With 700 people infected, you have ... well, if it's all independent, you have a 17 percent chance of becoming infected in the next day.
Does that sound reasonable? How can we apply it to other situations?
Is it a problem that the risk of getting/eating a pizza is zero so everything else is multiplied by zero? Or has there been a case of someone getting infected that way? Most of these are undefined because they don't say how close you get to other people and how many, which is by far the most important factor.
The Diamond Princess data suggest that about half of infections are symptomatic.That's the only case I know of where (almost) every person in an interesting cohort was tested.
I'd love to know the probability of getting infected, even for a single specific situation.
My gut says if the second was that high, we'd see a lot more transmission. The Loblaws/Shoppers Drug Mart chain in Canada seems to report every time an employee gets sick (they sanitize that store overnight). So far, I've seen only four reports from them. And there are a lot of employees working a lot of hours, without masks, touching merchandise already on shelves.
Even one fixed reference point would help my gut calibrate to different situations.
Hmm that's a lot of scenarios, I'll have to find the time! But I'm interested in stress-testing my model as well + vetting it for extensibility with the multiple scenarios. I'll get back to you soon.
Can you use your estimator on the risks I rated, so we can see how strongly our estimations correlate?
Hi Robin, I just published a direct physics-based risk estimator here: https://medium.com/swlh/so-...
Insights copied from my writeup:For an individual, the riskiness of going outside scales linearly with four factors: the number of people they encounter (N_p_b), the number of surfaces they touch (N_s), the number of other people that have touched those surfaces since the surfaces were last sanitized (N_p_s), and the probability pᵢ that any given person outside is contagious.
Surgical masks offer you the same protection as 3x increased distance. T-shirt masks? Same as 2x distance.
6 feet of distancing keeps you pretty darn safe. BUT any amount less than that is much, much riskier!
Cool, only I went to some length to reduce the chance of getting infected by pizza delivery - something that I can't do to the same degree with food shopping (though I tried). Now I have no clue what my total prob is :-o
Convert pizza index to probability of getting infected
No doubt.
I modified the spreadsheet decreasing the probability of a random person being infected from 100% to 5% to see how it affect correlation.Correlation dropped to ~ 0.6.
This shows that I was wrong when assumed that people don't think about the worst-case when filling polls. The next time I should do math BEFORE drawing conclusions.
Besides my calculations are based on a huge number of assumptions I am not sure to be true. So, I'd better trust poll results than my calculations when comparing risks of different activities.
I think the spreadsheet still may be useful as it gives intuition on how pizza index maps to actual probabilities.
Thank you
If working in an office with an infected person is not a significant risk, then the current work-from-home policies seem like a bad idea.
I'd want the entire office to have been tested, not just asymptomatic and untested, and that pattern-matches to low confidence hedging to avoid admitting if I was wrong... But given the actual asymptomatic numbers in the data, and also that office workers are less likely to be immune compromised, and also some vague thoughts about low viral load infections, I think that it might not be surprising if a small office had five infections and only one person became symptomatic.
You may be right. We could check that by investigating if infected people tend to cluster by office.
If your hypothesis is correct, we should find it rare if only one person in an office is ever infected, since symptoms don't appear until at least four doublings.
Then according to your analysis, my poll respondents' response contained a lot of info. :)
I added correlation calculations to the sheet.Correlation is 0.7351706336
I skipped some activities because I didn't know how to classify them in terms of the number of touches, bodily contacts and time spent in proximity.E.G. I don't know whether to count Pizza delivery as a bodily contact with the cook (you eat stuff he touched).
I also supposed that each activity except kissing is done with protective measures.
So, what does this suggest about the risk of being the person who delivers pizza? If someone delivers N pizzas in a week, is their total risk from pizza delivery 1, N, log(N)+1, e^(N-1), or something else entirely? Given that the delivery driver has that risk this week, is the risk of getting a delivery next week much higher than this week?
I see that as "In conditions that crowded, each infected person infected an average of one person per day"
An office is approximately half as many hours per day, and in my limited experience in cruises and office work I spent about as much interaction-with-person per hour in each. (If I was going to roll fort saves vs disease in a model, I would roll once per interaction, not once per day if any interaction). Thus, with those numbers, I estimate that if there is an N infected people in a workplace, two days later there will be 2N infected people there, until the conditions for exponential growth are no longer true (which simplify to "everyone is infected or no longer coming to the office").
That estimate is based on a homogeneous population of spherical vectors, so it won't actually hold up in reality, but I lack any reasonable basis to predict which direction I've erred.
On the Diamond Princess, the original infected person was on the ship for five days. Quarantine went into effect 11 days after that.
In those 11 days, about 20% of the passengers and crew became infected (half of those symptomatic). that's about 700 out of 3711.
Not sure what that tells us, though.
1 person to 700 people is 9 doublings in 11 days. Call it one doubling a day. That means for every infected person on the ship, you have a 1/3700 chance of becoming infected that day. With 700 people infected, you have ... well, if it's all independent, you have a 17 percent chance of becoming infected in the next day.
Does that sound reasonable? How can we apply it to other situations?
Is it a problem that the risk of getting/eating a pizza is zero so everything else is multiplied by zero? Or has there been a case of someone getting infected that way? Most of these are undefined because they don't say how close you get to other people and how many, which is by far the most important factor.
The Diamond Princess data suggest that about half of infections are symptomatic.That's the only case I know of where (almost) every person in an interesting cohort was tested.
I'd love to know the probability of getting infected, even for a single specific situation.
My gut says if the second was that high, we'd see a lot more transmission. The Loblaws/Shoppers Drug Mart chain in Canada seems to report every time an employee gets sick (they sanitize that store overnight). So far, I've seen only four reports from them. And there are a lot of employees working a lot of hours, without masks, touching merchandise already on shelves.
Even one fixed reference point would help my gut calibrate to different situations.