We scientists are pretty smart. We pose hypotheses and consequently confirm them in a series of logically connected experiments. Desired results follow in quick succession; our certainty grows with every step. Almost unfailingly the results have statistical significance, sometimes to the 5 % level, sometimes the p-value also has a whole string of zeros. Some of our experiments are independent of each other, some are dependent, for they use the same material, e.g. for molecular biology and histology. Now we turn tired but happy to the job of illustrating and writing up our results. Not only had we had a hand in the initial hypothesis, now confirmed. No, our luck was all the lovelier when we saw that the chain of significant p-values remained unbroken. That is comparable to the purchase of several lottery tickets which one after the other turn out to be a winner. If we then manage to convince the reviewer, our work will be printed just as it is.
Am I exaggerating? A casual glance through the leading journals (Nature, Cell, Science, etc.) will attest that the overwhelming majority of original articles published in them follow this pattern. The linearity of the pattern, unclouded by aborts or dead ends, typically starting with the phrase “First we…” and then followed by many repetitions of “Next we…”. A further indication is the almost complete absence of any non-significant results. Where you occasionally find an “n.s.”, it usually belongs there; if a significance had been reached in those, it would have jeopardized the hypothesis. As in a group that is not supposed to differ from a control in which the same gene, for example, had been manipulated with differing experimental strategies.
The naïve observer would have to come to the conclusion that the authors of these studies are not only incredibly smart but also implausibly lucky. He could even consider them braggards or swindlers. After a few years in science, however, we know that there is something quite different behind it all. We are telling each other stories. The long years spent in the laboratory working on the story were quite different. Many things went wrong, some were ambiguous, or the results didn’t match the hypothesis. Strategies were changed. The hypothesis was revised. And so forth. So the “smooth running” story was developed and told ex-post. So actually, it really is a “story”.
But is that a problem? We all know that it doesn’t work the way it’s been told. And besides, we are for good reason not interested in all the problems we encountered and the wrong paths we followed in our scientific explorations. They don’t make good reading and would swamp us with useless information. On the other hand, however, telling stories opens the flood gates to a number of bad habits, such as “outcome switching” and the selective use of results obtained. This has been compared to firing a shot at a wooden wall and then drawing a bullseye around the bullet hole. With the hole in the middle. Right in the heart! That way you can “prove” any old hypothesis! And we don’t learn anything about results that didn’t make it into the story but would lead us to other hypotheses and new discoveries.
So allow us to ask at this point: Just how did it happen that reporting on scientific discoveries has almost completely detached itself from the processes in the laboratory upon which it is based? Is it a product of our preference for stories that are smooth and spectacular? Our academic rewards system, which especially rewards them when they are published in journals with a high impact factor? Surprisingly, no. The rhetoric of a linear, uninterrupted and impeccable, logical, timely chain, necessarily pressing forward towards proof of the initial hypothesis, is several hundred years old. Towards the end of the 17th century experiments were hardly ever published, but rather presented to a public persons who would thus be their witnesses. The expansion and internationalization of the “scientific community”, which was carried out first by privatizing gentlemen and then more and more by “professionals”, made broadly accessible publication necessary. These publications developed under the patronage of the scientific societies founded at this time. In the lead was the Royal Society in England with is “Proceedings”, still being published today. In as far as the experiments were now without “witnesses”, and addressed to a very mixed and yet little specialized public, the readers had to be interested in the object of the study and convinced by its results. The rest is in the true sense of the word “history”. The dissociation of the actual logic and practice of a study from its representation in the corresponding scientific publication in favor of a “story” is today’s standard – not only in biomedicine.
A long tradition; we have become accustomed to it, and publication in the journals is accepted only in this manner. So everything is just fine, right?
I don’t think so. For one thing, because substantially more studies are being published and they contain substantially more information in the form of sub-studies, and these are essentially more complex in their methods and conceptions. This means that the “degree of freedom” has massively increased, which makes it possible for authors by selecting “desired results” to “substantiate” every thinkable hypothesis. Furthermore, it has become normal today to mix together in single study the generation of hypotheses by exploration and their confirmation. And this supposedly affords the reader a clear overview. But many of the experiments carried out did not make it into the publication? And why not? Was the hypothesis “un-biased” by means of explorative experiments and then confirmed in subsequent independent experiments? Was the hypothesis formulated unambiguously for the confirmation, was the number of required cases determined and bias excluded as far as possible? I.e. were the experiments randomized and blinded?
But how could the risk that we and our readers be led astray by the selective use of results for the proposes of “story-telling” be minimized? How can we make the results more robust and make it available in its entirety to the scientific community?
Actually, that is easy. First, we have to separate exploration and confirmation more clearly from each other. In exploring, we are searching for new phenomena. You cannot plan everything in advance, e.g. sample sizes. You can change the direction the experiments are taking on the basis of the findings coming in. You have to give serendipity a chance. You don’t need test statistics; you only have to describe very well the data you have culled in terms of their distribution (e.g. confidence intervals) — just as you have to describe everything required to make your results sufficiently comprehensible and reproducible. The result of such discovery phases are the hypotheses. Because of the originality of hypotheses acquired in this manner and because of the low case numbers, you will necessarily produce many false positive results. The effect sizes will also be overestimated (see post “How original are your hypotheses really?”). In a subsequent phase, the results and hypotheses must then be confirmed in a separate study in so far as you consider them interesting and important. This would entail sorting out the false positives and establishing realistic effect sizes. Then a preliminary formulation of the hypothesis has to be made, a proper sample size estimated so that the Type I and II error rates are acceptable etc. Before the experiments are started, you will draw up a detailed analysis plan and will no longer deviate from this or from the test statistics proposed in it. Should any deviations from the study and analysis plan come necessary in the course of the study, these will be accounted for and reported. Ideally, you should register this study before it starts (e.g. with a time stamp (and if you wish embargo) at Open Science Framework until publication) so that you can demonstrate at publication that you haven’t told a “story”. With this you will have a clean separation of explorative and confirmatory studies, which could even published together, as suggested in Nature by Mogil and Macleod for all experimental studies in high-grade journals.
Such a simple separation in design, analysis and publication of explorative and confirmatory studies, together with preregistration, could significantly increase the transparence, validity and reproducibility in experimental biomedical research. It only has one disadvantage: We have to forego all spectacular (but then unreproducible) studies.
A German version of this post has been published as part of my monthly column in the Laborjournal: http://www.laborjournal-archiv.de/epaper/LJ_17_10/28/