4 ms·

Ditto on the "130 papers in 7 months." I am not familiar with the field, but I assume the process would look like this: * Read and understand paper * Find and

Ditto on the "130 papers in 7 months." I am not familiar with the field, but I assume the process would look like this: * Read and understand paper * Find and download appropriate input data * Code paper model and validate (he said he wrote his own code) I can see myself being able to do this for ONE paper in maybe a week. He claims he was doing 1-2 of these per day. Wow. So either there is some exaggeration on his part, or he is a total wizard in his field.

sgt101 7y ago
I think you are a quick study; typically it takes me a week to figure out the detail of what's being done in a paper. Getting the data and testing would take longer. Charitably he/she has a framework with all the data required sitting ready to go and is just writing wrappers to the models. But even downloading frameworks from Github and getting them working takes a couple of days - for me. For example I've been playing with the Graph-network code from deepmind for a few weeks - I had to learn how the graphs were represented, how to build them and access them and how the models were made and put together. Just working that out was a solid three day job. Now I can build things and test out what's going on in the examples and get a feel for the framework, probably (if there was a problem) I would be in a reasonable position to say "this doesn't work like they think it does" (it does, but no surprise) but unless you've done that leg work I think you can't really. I think a proper replication effort is really 1 man month of expert time - or really you're just throwing stones.
BorisVSchmid 7y ago
Depending on the complexity of the model it would take me at least a month for a single paper. What makes it fully unbelievable to me is the claim of detecting p-value hacking in many of these 130 papers while doing 3 papers every 2 days. To make that claim for a single paper I would 1. have to be able to reproduce their p-value, and 2. spend enough time with the model to understand how/what assumptions were unfairly tweaked to get to that p-value. Just running your own implementation of a model on your own dataset and getting an insignificant or different p-value is not enough. You might just have implemented the model wrongly.

sgt101 7y ago

I think you are a quick study; typically it takes me a week to figure out the detail of what's being done in a paper. Getting the data and testing would take longer. Charitably he/she has a framework with all the data required sitting ready to go and is just writing wrappers to the models. But even downloading frameworks from Github and getting them working takes a couple of days - for me. For example I've been playing with the Graph-network code from deepmind for a few weeks - I had to learn how the graphs were represented, how to build them and access them and how the models were made and put together. Just working that out was a solid three day job. Now I can build things and test out what's going on in the examples and get a feel for the framework, probably (if there was a problem) I would be in a reasonable position to say "this doesn't work like they think it does" (it does, but no surprise) but unless you've done that leg work I think you can't really. I think a proper replication effort is really 1 man month of expert time - or really you're just throwing stones.

BorisVSchmid 7y ago

Depending on the complexity of the model it would take me at least a month for a single paper. What makes it fully unbelievable to me is the claim of detecting p-value hacking in many of these 130 papers while doing 3 papers every 2 days. To make that claim for a single paper I would 1. have to be able to reproduce their p-value, and 2. spend enough time with the model to understand how/what assumptions were unfairly tweaked to get to that p-value. Just running your own implementation of a model on your own dataset and getting an insignificant or different p-value is not enough. You might just have implemented the model wrongly.