Although hypothesis tests have been the basis of modern science since the middle of the 20th century, they have been plagued by misconceptions from the outset; this has led to what has been described as a crisis in science in the last few years: some journals have gone so far as to ban p -value s outright. 6 This is not because of any flaw in the concept of a p -value, but because of a lack of understanding of what they mean.
Possibly the most pervasive misunderstanding is the belief that the p- value is the chance that the null hypothesis is true, or that the p- value represents the frequency with which you will be wrong if you reject the null hypothesis (i.e. claim to have found a difference). This interpretation has frequently made it into the literature, and is a very easy trap to fall into when discussing hypothesis tests. To avoid this, it is important to remember that the p- value is telling us something about our sample , not about the null hypothesis. Put in simple terms, we would like to know the probability that the null hypothesis is true, given our data. The p- value tells us the probability of getting these data if the null hypothesis were true, which is not the same thing. This fallacy is referred to as ‘flipping the conditional’; the probability of an outcome under certain conditions is not the same as the probability of those conditions given that the outcome has happened.
A useful example is to imagine a magic trick in which you select a card from a normal deck of 52 cards, and the performer reveals your chosen card in a surprising manner. If the performer were relying purely on chance, this would only happen on average once in every 52 attempts. On the basis of this, we conclude that it is unlikely that the magician is simply relying on chance. Although simple, we have just performed an entire hypothesis test. We have declared a null hypothesis (the performer was relying on chance); we have even calculated a p -value (1 in 52, ≈0.02); and on the basis of this low p- value we have rejected our null hypothesis. We would, however, be wrong to suggest that there is a probability of 0.02 that the performer is relying on chance—that is not what our figure of 0.02 is telling us.
To explore this further we can create two populations, and watch what happens when we use simulation to take repeated samples to compare these populations. Computers allow us to do this repeatedly, and to see what p- value s are generated (see Supplementary online material). 7 Fig 1 illustrates the results of 100,000 simulated t -tests, generated in two set of circumstances. In Fig 1 a , we have a situation in which there is a difference between the two populations. The p- value s cluster below the 0.05 cut-off, although there is a small proportion with p >0.05. Interestingly, the proportion of comparisons where p <0.05 is 0.8 or 80%, which is the power of the study (the sample size was specifically calculated to give a power of 80%).
The p- value s generated when 100,000 t -tests are used to compare two samples taken from defined populations. ( a ) The populations have a difference and the p- value s are mostly significant. ( b ) The samples were taken from the same population (i.e. the null hypothesis is true) and the p- value s are distributed uniformly.
Figure 1 b depicts the situation where repeated samples are taken from the same parent population (i.e. the null hypothesis is true). Somewhat surprisingly, all p- value s occur with equal frequency, with p <0.05 occurring exactly 5% of the time. Thus, when the null hypothesis is true, a type I error will occur with a frequency equal to the alpha significance cut-off.
Figure 1 highlights the underlying problem: when presented with a p -value <0.05, is it possible with no further information, to determine whether you are looking at something from Fig 1 a or Fig 1 b ?
Finally, it cannot be stressed enough that although hypothesis testing identifies whether or not a difference is likely, it is up to us as clinicians to decide whether or not a statistically significant difference is also significant clinically.
As mentioned above, some have suggested moving away from p -values, but it is not entirely clear what we should use instead. Some sources have advocated focussing more on effect size; however, without a measure of significance we have merely returned to our original problem: how do we know that our difference is not just a result of sampling variation?
One solution is to use Bayesian statistics. Up until very recently, these techniques have been considered both too difficult and not sufficiently rigorous. However, recent advances in computing have led to the development of Bayesian equivalents of a number of standard hypothesis tests. 8 These generate a ‘Bayes Factor’ (BF), which tells us how more (or less) likely the alternative hypothesis is after our experiment. A BF of 1.0 indicates that the likelihood of the alternate hypothesis has not changed. A BF of 10 indicates that the alternate hypothesis is 10 times more likely than we originally thought. A number of classifications for BF exist; greater than 10 can be considered ‘strong evidence’, while BF greater than 100 can be classed as ‘decisive’.
Figures such as the BF can be quoted in conjunction with the traditional p- value, but it remains to be seen whether they will become mainstream.
The author declares that they have no conflict of interest.
The associated MCQs (to support CME/CPD activity) will be accessible at www.bjaed.org/cme/home by subscribers to BJA Education .
Jason Walker FRCA FRSS BSc (Hons) Math Stat is a consultant anaesthetist at Ysbyty Gwynedd Hospital, Bangor, Wales, and an honorary senior lecturer at Bangor University. He is vice chair of his local research ethics committee, and an examiner for the Primary FRCA.
Matrix codes: 1A03, 2A04, 3J03
Supplementary data to this article can be found online at https://doi.org/10.1016/j.bjae.2019.03.006 .
The following is the Supplementary data to this article:
While hypothesis testing is a highly formalized activity, hypothesis generation remains largely informal. We propose a systematic procedure to generate novel hypotheses about human behavior, which uses the capacity of machine learning algorithms to notice patterns people might not. We illustrate the procedure with a concrete application: judge decisions about who to jail. We begin with a striking fact: The defendant’s face alone matters greatly for the judge’s jailing decision. In fact, an algorithm given only the pixels in the defendant’s mugshot accounts for up to half of the predictable variation. We develop a procedure that allows human subjects to interact with this black-box algorithm to produce hypotheses about what in the face influences judge decisions. The procedure generates hypotheses that are both interpretable and novel: They are not explained by demographics (e.g. race) or existing psychology research; nor are they already known (even if tacitly) to people or even experts. Though these results are specific, our procedure is general. It provides a way to produce novel, interpretable hypotheses from any high-dimensional dataset (e.g. cell phones, satellites, online behavior, news headlines, corporate filings, and high-frequency time series). A central tenet of our paper is that hypothesis generation is in and of itself a valuable activity, and hope this encourages future work in this largely “pre-scientific” stage of science.
This is a revised version of Chicago Booth working paper 22-15 “Algorithmic Behavioral Science: Machine Learning as a Tool for Scientific Discovery.” We gratefully acknowledge support from the Alfred P. Sloan Foundation, Emmanuel Roman, and the Center for Applied Artificial Intelligence at the University of Chicago. For valuable comments we thank Andrei Shliefer, Larry Katz and five anonymous referees, as well as Marianne Bertrand, Jesse Bruhn, Steven Durlauf, Joel Ferguson, Emma Harrington, Supreet Kaur, Matteo Magnaricotte, Dev Patel, Betsy Levy Paluck, Roberto Rocha, Evan Rose, Suproteem Sarkar, Josh Schwartzstein, Nick Swanson, Nadav Tadelis, Richard Thaler, Alex Todorov, Jenny Wang and Heather Yang, as well as seminar participants at Bocconi, Brown, Columbia, ETH Zurich, Harvard, MIT, Stanford, the University of California Berkeley, the University of Chicago, the University of Pennsylvania, the 2022 Behavioral Economics Annual Meetings and the 2022 NBER summer institute. For invaluable assistance with the data and analysis we thank Cecilia Cook, Logan Crowl, Arshia Elyaderani, and especially Jonas Knecht and James Ross. This research was reviewed by the University of Chicago Social and Behavioral Sciences Institutional Review Board (IRB20-0917) and deemed exempt because the project relies on secondary analysis of public data sources. All opinions and any errors are of course our own. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.
MARC RIS BibTeΧ
Download Citation Data
Jens Ludwig & Sendhil Mullainathan, 2024. " Machine Learning as a Tool for Hypothesis Generation, " The Quarterly Journal of Economics, vol 139(2), pages 751-827.
Conferences, more from nber.
In addition to working papers , the NBER disseminates affiliates’ latest findings through a range of free periodicals — the NBER Reporter , the NBER Digest , the Bulletin on Retirement and Disability , the Bulletin on Health , and the Bulletin on Entrepreneurship — as well as online conference reports , video lectures , and interviews .
COMMENTS
generation represent two distinct research objectives. In hypothesis testing research, the researcher specifies one or more a priori hypotheses, based on existing theory and/or data, and then puts these hypotheses to an empirical test with a new set of data. In hypothesis generating research, the researcher explores a set of data searching
Formulating Hypotheses for Different Study Designs. Generating a testable working hypothesis is the first step towards conducting original research. Such research may prove or disprove the proposed hypothesis. Case reports, case series, online surveys and other observational studies, clinical trials, and narrative reviews help to generate ...
The hypothesis is a tentative prediction of the nature and direction of relationships between sets of data, phrased as a declarative statement. ... Studies that seek to answer descriptive research questions do not test hypotheses, but they can be used for hypothesis generation. Those hypotheses would then be tested in subsequent studies.
1. An Overview. Broadly, preclinical research can be classified into two distinct categories depending on the aim and purpose of the experiment, namely, "hypothesis generating" (exploratory) and "hypothesis testing" (confirmatory) research (Fig. 1).Hypothesis generating studies are often scientifically-informed, curiosity and intuition-driven explorations which may generate testable ...
Hypothesis generation is an early and critical step in any hypothesis-driven clinical research project. Because it is not yet a well-understood cognitive process, the need to improve the process goes unrecognized. Without an impactful hypothesis, the significance of any research project can be questionable, regardless of the rigor or diligence applied in other steps of the study, e.g., study ...
Table of contents. Step 1: State your null and alternate hypothesis. Step 2: Collect data. Step 3: Perform a statistical test. Step 4: Decide whether to reject or fail to reject your null hypothesis. Step 5: Present your findings. Other interesting articles. Frequently asked questions about hypothesis testing.
Background: Scientific hypothesis generation is a critical step in scientific research that determines the direction and impact of any investigation. Despite its vital role, we have limited ... After formulating a scientific hypothesis, researchers design studies to test the scientific hypothesis to determine the answer to research questions 2,4.
Using data to generate potential discoveries and using data to subject those discoveries to tests are distinct processes. This distinction is known as exploratory (or hypothesis-generating) research and confirmatory (or hypothesis-testing) research. In the daily practice of doing research, it is easy to confuse which one is being done.
Hypothesis-generating (Qualitative hypothesis-generating research) - Qualitative research uses inductive reasoning. - This involves data collection from study participants or the literature regarding a phenomenon of interest, using the collected data to develop a formal hypothesis, and using the formal hypothesis as a framework for testing the ...
Hypothesis Generation vs. Hypothesis Testing. This is a very common mistake data science beginners make. Hypothesis generation is a process beginning with an educated guess whereas hypothesis testing is a process to conclude that the educated guess is true/false or the relationship between the variables is statistically significant or not.
testing 50 null hypotheses, which would have required a cor-rected significance level of .05/50 = 0.001. In a confirmatory study, it is mandatory to show that the ... Hypothesis-generating studies are much more common than confirmatory, because the latter are logistically more complex, more laborious, more time-consuming, more expensive, and
Hypothesis Testing vs Hypothesis Generation . In the world of Data Science, there are two parts to consider when putting together a hypothesis. Hypothesis Testing is when the team builds a strong hypothesis based on the available dataset. This will help direct the team and plan accordingly throughout the data science project.
The purpose of our hypothesis-generating study is to test the feasibility of using MPS to generate clinical hypotheses, and to approach the return of results as an experimental manipulation. Issues to consider in both designs include: volume and nature of the potential results, primary versus secondary results, return of individual results ...
Abstract. While hypothesis testing is a highly formalized activity, hypothesis generation remains largely informal. We propose a systematic procedure to generate novel hypotheses about human behavior, which uses the capacity of machine learning algorithms to notice patterns people might not.
The paradigm of hypothesis-generating research does not replace or undermine hypothesis-testing modes of research; instead, it complements them and has facilitated discoveries that may not have been possible with hypothesis-testing research. The hypothesis-generating mode of research has been primarily practiced in basic science but has ...
The paradigm of hypothesis-generating research does not replace or undermine hypothesis-testing modes of research; instead, it complements them and has facilitated discoveries that may not have been possible with hypothesis-testing research. The hypothesis-generating mode of research has been primarily practiced in basic science but has ...
Abstract To correct a common imbalance in methodology courses, focusing almost entirely on hypothesis-testing issues to the neglect of hypothesis-generating issues which are at least as important, 49 creative heuristics are described, divided into 5 categories and 14 subcategories. Each of these heuristics has often been used to generate hypotheses in psychological research, and each is ...
The replication crisis may be partly explained by scientists' overconfidence in the replicability of their results. It has been argued that one source of this overconfidence is the false portrayal of exploratory hypothesis tests as confirmatory hypothesis tests (Nosek et al., Citation 2018; Wagenmakers et al., Citation 2012). Exploratory hypothesis tests involve unplanned tests of post hoc ...
Instead, companies should focus organizational energy on hypothesis generation and testing. Hypotheses force individuals to articulate in advance why they believe a given course of action will succeed. A failure then exposes an incorrect hypothesis — which can more reliably convert into organizational learning. What Exactly Is a Hypothesis?
Presenting an outcome from a hypothesis-generating study as if it had been produced in a confirmatory study is misleading and represents methodological ignorance or scientific misconduct. Hypothesis-generating studies differ methodologically from confirmatory studies. A generated hypothesis must be confirmed in a new study.
A hypothesis test is a procedure used in statistics to assess whether a particular viewpoint is likely to be true. They follow a strict protocol, and they generate a 'p-value', on the basis of which a decision is made about the truth of the hypothesis under investigation.All of the routine statistical 'tests' used in research—t-tests, χ 2 tests, Mann-Whitney tests, etc.—are all ...
While hypothesis testing is a highly formalized activity, hypothesis generation remains largely informal. We propose a systematic procedure to generate novel hypotheses about human behavior, which uses the capacity of machine learning algorithms to notice patterns people might not. We illustrate the procedure with a concrete application: judge ...