Replication, reproduction, confirmation. What is the optimal mix?
Issues surrounding the relationship of Open Research and replication seems to be the meme of the week. Abhishek Tiwari provided notes on a debate describing concerns about how open research could damage replication and Sabine Hossenfelder explored the same issue in a blog post. The concern fundamentally is that by providing more of the details of our research we may actually be damaging the research effort by reducing the motivation to reproduce published findings or worse, as Sabine suggests, encouraging group think and a lack of creative questioning.
I have to admit that even at a naive level I find this argument peculiar. There is no question that in aiming to reproduce or confirm experimental findings it may be helpful to carry out that process in isolation, or with some portion of the available information withheld. This can obviously increase the quality and power of the confirmation, making it more general. Indeed the question of how and when to do this most effectively is very interesting and bears some thought. The optimization of these descisions in specific cases will be important part of improving research quality. What I find peculiar is the apparent belief in many quarters (but not necessarily Sabine who I think has a much more sophisticated view) that this optimization is best encouraged by not bothering to make information available. We can always choose not to access information if it is available but if it is not we cannot choose to look at it. Indeed to allow the optimization of the confirmation process it is crucial that we could have access to the information if we so decided.
But I think there is a deeper problem than the optimization issue. I think that the argument also involves two category errors. Firstly we need to distinguish between different types of confirmation. There is pure mechanical replication, perhaps just to improve statistical power or to re-use a technique for a different experiment. In this case you want as much detail as possible about how the original process was carried out because there is no point in changing the details. The whole point of the exercise is to keep things as similar as possible. I would suggest the use of the term “reproduction” to mean a slightly different case. Here the process or experiment “looks the same” and is intended to produce the same result but the details of the process are not tightly controlled. The purpose of the exercise is to determine how robust the result or output is to modified conditions. Here withholding, or not using, some information could be very useful. Finally there is the process of actually doing something quite different with the intention of testing an idea or a claim from a different direction with an entirely different experiment or process. I would refer to this as “confirmation”. The concerns of those arguing against providing detailed information lie primarily with confirmation, but the data and process sharing we are talking about relates more to replication and reproduction. The main efficiency gains lie in simply re-using shared process to get down a scientific path more rapidly rather than situations where the process itself is the subject of the scientific investigation.
The second category error is somewhat related in as much as the concerns around “group-think” refer to claims and ideas whereas the objects we are trying to encourage sharing when we talk about open research are more likely to be tools and data. Again, it seems peculiar to argue that the process of thinking independently about research claims is aided by reducing the amount of data available. There is a more subtle argument that Sabine is making and possibly Harry Collins would make a similar one, that the expression of tools and data may be inseparable from the ideas that drove their creation and collection. I would still argue however that it is better to actively choose to omit information from creative or critical thinking rather than be forced to work in the dark. I agree that we may need to think carefully about how we can effectively do this and I think that would be an interesting discussion to have with people like Harry.
But the argument that we shouldn’t share because it makes life “too easy” seems dangerous to me. Taking that argument to its extreme we should remove the methods section from papers altogether. In many cases it feels like we already have and I have to say that in day to day research that certainly doesn’t feel helpful.
Sabine also makes a good point, that Michael Nielsen also has from time to time, that these discussions are very focussed on experimental and specifically hypothesis driven research. It bears some thinking about but I don’t really know enough about theoretical research to have anything useful to add. But it is the reason that some of the language in this post may seem a bit tortured.