5 ms·

We all know it is super-easy to cheat in science, but not much can be done about it. Short of requiring replication for every result published which isn't feasi

We all know it is super-easy to cheat in science, but not much can be done about it. Short of requiring replication for every result published which isn't feasible if the study was costly to produce. And the problem isn't confined to medicine either. How many studies in hpc of the type "we setup this benchmark and our novel algorithm/implementation won!" aren't also faked or flawed?

soperj 3y ago
If the study is costly to produce, then it's even more important to replicate it, otherwise you'd wasting a large amount of money on a study with no real sense of whether it is flawed or not.
- mcmoor 3y ago
  I've read that the real reason why researches generated by war crimes are worthless, is because we can't replicate it. Using a result that its validity can't be checked is just a way to disaster. And this is researches with the highest cost (human lives) and we still can't use it. Lots of "interesting" psychology experiments in the days of yore turns out to have lots of damaging confounding variables and we can't just redo the experiment because, well, we shouldn't.
gmd63 3y ago
Plenty can be done about it. We can start by increasing the standards for experimental proof past words on a piece of paper. Stream or record your experiment. If a 12 year old has the ability to stream his gaming life on twitch, scientists should be able to record what they are doing in more detail. You could start a new journal with higher level of prestige that only publishes experiments that adhere to more modern methods of proof.
- gustavus 3y ago
  This seems overly simplistic, I mean part of the reason so many studies are so difficult is because they can happen over the period of months or years. This isn't like your High School chemistry experiment. These studies can be massive longitudinal studies that take years to conclude and work on. If the answers were easy I'm sure someone would've implemented it already, it turns out those kind of things are hard. I'd recommend checking out the following article. https://slatestarcodex.com/2014/04/28/the-control-group-is-out-of-control/ https://slatestarcodex.com/2014/04/28/the-control-group-is-o...
  - gmd63 3y ago
    I'm interested more in why it specifically would be hard to just show people what you are doing (or did) instead of telling them a story at the end of months or years. Critical details and oversights can be lost in translation between reality and LaTeX that could be easily pointed out by third party observers.
    - MockObject 3y ago
      Maybe experiments can take thousands of hours.
    - mattkrause 3y ago
      Who is going to watch thousands[0] of hours of video, most of which will show a scientist fiddling around with setscrews or hunting down a loose ground wire? And what scientist is going to be happy with constant surveillance that's intended to be shared (so no picking your nose, losing your temper, or making mistakes)? [0] Literally, thousands. I just finished a paper about a project that we started three years ago. This was the main project for two people, plus a tech: 2000 hrs/yr * 3 * 80% effort = 4800 hours of footage.
      - gmd63 3y ago
        That does sound like a lot of footage, but you wouldn't watch every second just as you likely haven't read every word of every scientific paper you've skimmed or referenced. The footage is there to support inquiries into how you did your experiment, just as you would skim a youtube video about repairing an appliance to get to the part you need to see. And yeah, it would have to be a new level of prestige backing this way of documenting your methods. Obviously people prefer being trusted, but p-hacking and replication problems have proven that we can't have nice things.
- sfink 3y ago
  > We can start by increasing the standards for experimental proof past words on a piece of paper. That is overly cynical. We are already well past that standard. We require words on a piece of paper that were written by somebody from a recognizable institution, or by an AI operated by somebody from a recognizable institution.
michaelrpeskin 3y ago
The "setup this benchmark" happens in medicine all of the time, and it's super insidious and no one sees it. It's hard to tease out. For example, say you have a new (patentable) drug that you're trying to get approved and replace an old (generic, cheap) drug. You need to prove that the new drug is at least as safe as the old one and/or more effective than the old one. Which sounds reasonable. Let's say that the old drug is super safe and that the new one may not be. What you can do is set up your RCT so that you surreptitiously underdose the control arm so the safe drug looks less effective. And then you can say that your new drug is more effective than the old one. Peer review doesn't notice this because you can easily hide it in the narrative of the methods section. I've seen it a couple of times. So you can easily have "gold standard RCTs" that get the result you want just by subtle changes in the study design.
noslenwerdna 3y ago
Require experimental analysis to be pre-registered? That is how it's done in particle physics, and it works well.
- mike_hearn 3y ago
  It helps a lot but rests on the assumption that the people with whom you're pre-registering will check that you did the same thing you claimed you'd do. And also, that they won't tolerate gamed pre-registrations e.g. pre-registrations so vague that they admit many different post-hoc analyses. Sometimes this is true. One example is when the people verifying the pre-registration work for a regulator. The introduction of pre-registration seems to be at least partly behind the collapse in pharma productivity: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0132382 https://journals.plos.org/plosone/article?id=10.1371/journal... 17 of 30 studies (57%) published prior to 2000 showed a significant benefit of intervention on the primary outcome in comparison to only 2 among the 25 (8%) trials published after 2000 (χ2=12.2,df= 1, p=0.0005). There has been no change in the proportion of trials that compared treatment to placebo versus active comparator. Industry co-sponsorship was unrelated to the probability of reporting a significant benefit. Pre-registration in clinicaltrials.gov was strongly associated with the trend toward null findings. But when journals do this the results seem to be worse. They aren't strongly incentivized to improve anything, so you can get studies that claim to be pre-registered but do something different to what the pre-registration said they'd do.

soperj 3y ago

If the study is costly to produce, then it's even more important to replicate it, otherwise you'd wasting a large amount of money on a study with no real sense of whether it is flawed or not.

mcmoor 3y ago
I've read that the real reason why researches generated by war crimes are worthless, is because we can't replicate it. Using a result that its validity can't be checked is just a way to disaster. And this is researches with the highest cost (human lives) and we still can't use it. Lots of "interesting" psychology experiments in the days of yore turns out to have lots of damaging confounding variables and we can't just redo the experiment because, well, we shouldn't.

gmd63 3y ago

Plenty can be done about it. We can start by increasing the standards for experimental proof past words on a piece of paper. Stream or record your experiment. If a 12 year old has the ability to stream his gaming life on twitch, scientists should be able to record what they are doing in more detail. You could start a new journal with higher level of prestige that only publishes experiments that adhere to more modern methods of proof.

gustavus 3y ago
This seems overly simplistic, I mean part of the reason so many studies are so difficult is because they can happen over the period of months or years. This isn't like your High School chemistry experiment. These studies can be massive longitudinal studies that take years to conclude and work on. If the answers were easy I'm sure someone would've implemented it already, it turns out those kind of things are hard. I'd recommend checking out the following article. https://slatestarcodex.com/2014/04/28/the-control-group-is-out-of-control/ https://slatestarcodex.com/2014/04/28/the-control-group-is-o...
- gmd63 3y ago
  I'm interested more in why it specifically would be hard to just show people what you are doing (or did) instead of telling them a story at the end of months or years. Critical details and oversights can be lost in translation between reality and LaTeX that could be easily pointed out by third party observers.
  - MockObject 3y ago
    Maybe experiments can take thousands of hours.
  - mattkrause 3y ago
    Who is going to watch thousands[0] of hours of video, most of which will show a scientist fiddling around with setscrews or hunting down a loose ground wire? And what scientist is going to be happy with constant surveillance that's intended to be shared (so no picking your nose, losing your temper, or making mistakes)? [0] Literally, thousands. I just finished a paper about a project that we started three years ago. This was the main project for two people, plus a tech: 2000 hrs/yr * 3 * 80% effort = 4800 hours of footage.
    - gmd63 3y ago
      That does sound like a lot of footage, but you wouldn't watch every second just as you likely haven't read every word of every scientific paper you've skimmed or referenced. The footage is there to support inquiries into how you did your experiment, just as you would skim a youtube video about repairing an appliance to get to the part you need to see. And yeah, it would have to be a new level of prestige backing this way of documenting your methods. Obviously people prefer being trusted, but p-hacking and replication problems have proven that we can't have nice things.
sfink 3y ago
> We can start by increasing the standards for experimental proof past words on a piece of paper. That is overly cynical. We are already well past that standard. We require words on a piece of paper that were written by somebody from a recognizable institution, or by an AI operated by somebody from a recognizable institution.

michaelrpeskin 3y ago

The "setup this benchmark" happens in medicine all of the time, and it's super insidious and no one sees it. It's hard to tease out. For example, say you have a new (patentable) drug that you're trying to get approved and replace an old (generic, cheap) drug. You need to prove that the new drug is at least as safe as the old one and/or more effective than the old one. Which sounds reasonable. Let's say that the old drug is super safe and that the new one may not be. What you can do is set up your RCT so that you surreptitiously underdose the control arm so the safe drug looks less effective. And then you can say that your new drug is more effective than the old one. Peer review doesn't notice this because you can easily hide it in the narrative of the methods section. I've seen it a couple of times. So you can easily have "gold standard RCTs" that get the result you want just by subtle changes in the study design.

noslenwerdna 3y ago

Require experimental analysis to be pre-registered? That is how it's done in particle physics, and it works well.

mike_hearn 3y ago
It helps a lot but rests on the assumption that the people with whom you're pre-registering will check that you did the same thing you claimed you'd do. And also, that they won't tolerate gamed pre-registrations e.g. pre-registrations so vague that they admit many different post-hoc analyses. Sometimes this is true. One example is when the people verifying the pre-registration work for a regulator. The introduction of pre-registration seems to be at least partly behind the collapse in pharma productivity: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0132382 https://journals.plos.org/plosone/article?id=10.1371/journal... 17 of 30 studies (57%) published prior to 2000 showed a significant benefit of intervention on the primary outcome in comparison to only 2 among the 25 (8%) trials published after 2000 (χ2=12.2,df= 1, p=0.0005). There has been no change in the proportion of trials that compared treatment to placebo versus active comparator. Industry co-sponsorship was unrelated to the probability of reporting a significant benefit. Pre-registration in clinicaltrials.gov was strongly associated with the trend toward null findings. But when journals do this the results seem to be worse. They aren't strongly incentivized to improve anything, so you can get studies that claim to be pre-registered but do something different to what the pre-registration said they'd do.