The criteria people typically use to distinguish the more reliable ‘hard sciences’ from the less reliable ‘soft sciences’ do not match our understanding of which is which. Instead, I think that the difference lies in the history of the fields. For hard sciences, the science precedes the community. For soft sciences, the community precedes the science. Hard sciences have norms of empiricism. Soft sciences try to import rules of empiricism. If you want to harden a science, it is more important to grow the norms than to establish the rules. One way to accomplish this is to require that early career scientists replicate at least one major result in their field as part of their education.
Prerequisites: Awareness of the difference between hard and soft sciences.
Originally Written: March 2021.
Confidence Level: A bit more than speculation.
The sciences are often divided into hard sciences and soft sciences.
Hard sciences include astronomy, physics, chemistry, and geology. Soft sciences include medicine, psychology, economics, and anything political or social. Mathematics and computer science are either hard sciences or in separate categories.
What distinguishes these categories?
Some common criteria include:
- Complexity. Soft sciences tend to deal with complex systems with many interacting parts. It is easier to isolate and test a single variable in the hard sciences.
- Observation vs Experiment. Hard sciences tend to deal with subjects that can be experimented on. Soft sciences rely more on observation because experiments are either impractical or unethical.
- Mathematical rigor. Hard sciences rely more extensively on math and can more easily use rigorous proofs. This is often associated with have stricter standards for discoveries.
- Age. Soft sciences are typically younger than hard sciences.
I do not think that any of these criteria actually match how we use ‘hard science’ and ‘soft science’. Some examples will help demonstrate the problems with these distinguishing criteria.
Astronomy is a hard science. It is ancient. Even the oldest problem in astronomy, the motion of the planets, can be chaotic. Modern problems, like the dynamics of a galaxy or the internal structure of a neutron star, are extremely complex. Astronomy exclusively relies on observations and these observations are limited. Almost everything we know about space comes from starlight. Only recently have other modes of observation, like neutrinos or gravitational waves, become available. Astronomy often uses sophisticated mathematics, but not precise mathematics. Making a prediction that is only five times too big is often good for astronomy, and a prediction that is only fifty times too big is sometimes acceptable. Based on these criteria, astronomy sounds kind of squishy.
Medicine is just as ancient as astronomy. Experiments are possible, but often unethical. People have done many unethical experiments anyway and we still have their data. While the human body is extremely complex, many of the most important discoveries, like vaccination or germ theory or DNA, are quite simple. Medicine usually lacks mathematical sophistication, except for some subfields like protein folding, but their statistics are highly sophisticated. Few physics experiments would consider adding a statistician to their team even if none of the physicists have training in statistics; this is common in medical research. And yet, medicine is a soft science. People talk about scientific medicine as a goal, while scientific astronomy is just astronomy.
Geology is a moderately old hard science. It seems like it has lower complexity, although that might just be due to my ignorance. Some subfields, like volcanology or paleontology, are completely observational, while other subfields have some experiments. Mathematical sophistication is optional and usually low.
Economics is a moderately old soft science. Like geology, it has mostly observations. Experiments are only possible in some subfields. Economics has a high degree of mathematical sophistication, and not just in statistics. There are some branches of mathematics, like stochastic differential equations, where most of the relevant applications are in economics and not hard sciences like physics and computer science.
There are also young hard sciences like epidemiology or meteorology. These both have high complexity and high mathematical sophistication. They both rely mostly on observation with limited opportunity for experiment.
Psychology is a young soft science. It has high complexity. The statistical methods tend to be sophisticated, although other mathematical sophistication is optional. Experimentation is usually straightforward, but may have questionable ethics.
The common criteria I listed above to distinguish between hard and soft sciences do not match how we use those words.
So is there a real difference? Or is the only thing that makes a science ‘hard’ or ‘soft’ its reputation?
I think that there is a real difference. There are some criteria that do show the difference, like how often graphs are used or how often textbooks need to get updated. These criteria are consequences of the underlying distinction, but they do indicate that a real distinction is there.
Hard sciences and soft sciences have different histories.
Hard sciences begin with individual empiricists or mathematicians who begin developing new techniques to investigate new problems. Physics has Galileo Galilei and Isaac Newton. Epidemiology has John Snow and Ronald Ross. If they are successful, an intellectual community forms around them. If they are unsuccessful, they are forgotten, or maybe someone in the next generation is able to make their ideas work. The science precedes the community.
Soft sciences begin as an intellectual community which then tries to import empiricism. Medicine is an ancient tradition involving the Hippocratic Oath and the sign of Moses’s staff. Doctors practiced medicine long before they practiced science. Economics and psychology both began as philosophical movements that then tried to improve themselves by copying the scientific method. The community precedes the science.
This history explains some further differences between hard and soft sciences.
Soft sciences often have internal critics of the scientific method: doctors who claim that clinical experience is worth more than research or political theorists who argue that methods from the humanities are more important. While the hard sciences do (and should) have significant internal criticism, it assumes that the field could not exist without its empirical or mathematical content. The most philosophical subfield of physics is quantum interpretation, and here, it is believed that in order to understand quantum mechanics you have to work through the calculations yourself.
Soft sciences usually have precise standards for what counts as a discovery. Hard sciences usually do not. The easiest way to see if something is a soft science is to check if their journals require $p < 0.05$. It is not the case that soft sciences require $p < 0.05$ and hard sciences require $p < 0.01$. Journals in the hard sciences typically do not have a specific $p$ value. The one exception I am aware of is experimental particle physics, where discoveries can be declared when the evidence reaches $5 \sigma$ (about $p <$ one in a million). In other branches of physics, whether there is enough evidence is determined on a case-by-case basis. Hard sciences can rely on empirical norms which already exist in the community. Since soft sciences are trying to import empiricism, they have to establish precise rules.
Hard sciences have a reputation for being more reliable than soft sciences, so people often try to convert soft sciences into hard sciences. I believe that some of the techniques used are ineffective if not counterproductive.
Often, the strategy to harden a science is to impose more and stricter rules. This cannot work because hard sciences have fewer rules (but more norms) than soft sciences.
Any precise rule can be exploited. If you are willing to publish only $1/20$ or your work, you can consistently get $p < 0.05$ even if none of your results are valid. This ‘$p$-hacking’ is a major contributing factor to the replication crisis. Changing the standard to $p < 0.01$ would mean that it takes more work to exploit the system, but also would exclude many valid results. The larger sample sizes needed would further shift research towards larger institutions which would still be capable of $p$-hacking. Naively, we would expect that using a standard of $p < 0.05$ would mean that $5\%$ of our results are invalid and a standard of $p < 0.01$ would mean that $1\%$ of our results are invalid. Currently, much more than $5\%$ of the results in the soft sciences do not replicate and we should not expect this to improve by a factor of five if we adopted the stricter standard.
So how can you turn soft sciences into hard sciences?
One method that has worked repeatedly is invasion. Some people from one field of science start working on problems traditionally associated with a different field. If they succeed in solving at least a few important problems, then a new subfield with the norms of the invading field develops. Biochemistry is the result of chemists invading medicine and parts of neuroscience are the result of computer scientists invading psychology. Astronomy and physics have been continually invading each other for as long as science has existed.
Occasionally, the invaders find themselves more capable at almost every problem and conquer the field. The most dramatic example is the conquest of theoretical chemistry by physics. After the development of quantum mechanics, some physicists decided to use it to investigate some atomic, molecular, and optical (AMO) problems traditionally associated with chemistry. They were wildly successful. Theoretical chemists were forced to learn quantum mechanics just to keep up with their own field. Today, theoretical chemistry is called physical chemistry. Physical chemists need to learn much more physics than AMO physicists need to learn chemistry.
Invasion by other fields is sometimes successful and is frequently attempted. For some soft sciences, it has been extremely difficult for hard scientists to make much progress.
To harden these soft sciences, we need to remember that the goal is not better rules. The goal is to embed empirical or mathematical norms within the community so we can relax the rules. The rules should be a temporary aid to help our community learn to be scientific.
As part of this effort, I propose that early career scientists should have to replicate some of the classic results in their field. More replication attempts is one of the main suggestions for how to address the replication crisis. My contribution here is to suggest that this should be a required part of curricula in the soft sciences.
This is done in physics. As part of my undergraduate degree, I was required to take a class called Advanced Lab. The lab had equipment to replicate ten of the classic experiments in physics. I had to do four of them over the course of the semester.
Whenever a physicist begins working with a new experiment or new software, you are first expected to repeat a few earlier runs. If you cannot get the software or experiment to produce the same results as before, then any new results should be suspect. If this happens often, the experiment or software is suspect, instead of the researcher.
Soft sciences should have an Advanced Lab, either as part of their undergraduate or graduate curriculum. Every student should be required to select at least one major result in their field and to replicate it as closely as possible. This could be a semester long class with equipment specifically for certain experiments like my physics lab or it could be more open ended.
Requiring early career scientists to replicate major results is good for two reasons.
First, replication by independent scientists is a good way to make sure that the results are valid. Many proposals for increased replication try to get major results replicated a few times. My proposal would involve replicating major results hundreds if not thousands of times.
Second, this is good training for early career scientists and can help to develop good empirical norms. They get some experience with how the great achievements occurred and can later imitate them in their own work. Designing a good experiment for the first time is much easier if you are following a good example than if you are doing entirely original research.
In the hard sciences, empiricism is a collection of organically grown norms. In the soft sciences, empiricism is a collection of imposed rules. To harden science, we should focus on how to grow better norms, not how to impose better rules.