We Wrote the Hypotheses Down First, and One Failed

A note on publishing a research design before the data, and on what the data did to it.

VM

Vadim Matyushkin Behavioral Scientist & Sociologist · ICDS

PublishedOctober 15, 2026

In April 2026 the ICDS series published Article 5, and the article contained no results. The search data for the FMCG corpus was still being collected, and the first page said so: the article was a pre-registration, a public commitment to the analytical design before anyone knew how the numbers would come out. It defined the variable (Balance of Expression), the instruments, three hypotheses with the mechanism behind each, and the statistical test for the third.

Why Commit Before the Data

A pre-registration has one function. It fixes which comparisons count before the outcome data can influence the choice. In a field where a correlation can be found in most datasets if enough categories and time windows are tried, that matters more than it might seem. The corpus was already assembled: 66 media peaks, 66 brands, 10 product categories. Only the search response was missing.

The decision to publish before the results had a simple reason for me. A hypothesis is worth discussing on its own. Theorems have waited centuries for their proofs, and the question they pose stays valuable in the meantime. The results matter, and until they exist there is good reason to put the formulation of the question up for discussion.

The Three Hypotheses

The three hypotheses were specific. The first said that in Spontaneous categories, such as snacks and drinks, a peak with very strong positive dominance would produce the highest immediate search response. The second said that in Cautious categories a moderate or strong positive dominance with a visible minority objection would produce a stronger delayed response than a one-sided positive field. The third was the one that carried the practical weight. It said there is no universally optimal balance, and that the best structure depends on how the audience processes information, which in turn depends on the category. Most reputation practice works from the default that more positive is always better. The third hypothesis predicted that this default holds for some categories and fails for others.

What the Data Did

Article 6 reported the results. The second and third hypotheses were supported in their main parts. The first was not. In Spontaneous categories (26 brands) no structure of the Balance of Expression predicted any search metric, and the correlation with search at the peak month was -0.16, not significant. The hypothesis that looked most natural, that impulse buyers respond best to a one-sided positive field, had nothing behind it in this corpus. Article 6 states this in a single sentence and moves on.

What the Data Also Corrected

The data changed a second thing, and it was not a hypothesis. The typology that assigns product groups to purchase types had been set from the expected typical purchase occasion. Article 6 checked each category against its observed search profile and revised three assignments. Confectionery moved from Spontaneous to Cautious, because search for the premium and gift-oriented brands kept growing by 28 index points after the peak month. Food products moved from Routine to Spontaneous, and alcoholic beverages from Spontaneous to Routine. The hypotheses themselves stayed exactly as written in Article 5.

On Being a Researcher

A first hypothesis is almost always a product of intuition built on experience and reading, and it often turns out that life is more interesting than one thought. At the start it seems that things will go the way intuition suggests. In the end the result is neither a refutation nor a confirmation. It is a correction and a complication. When I saw that the first hypothesis had not held, my first thought was: well, as usual. The next step is to change the point of view and look closely at what the data show and what they hide. I kept the hypotheses as a direction and left the details open to correction. I think that is what makes a person a researcher: Columbus knew the land lay ahead and did not know on which day the first birds would appear on the mast. The revision of the confectionery category came from earlier experience. I had run many surveys and analyzed secondary data, and that helped me find my bearings and move away from the template.

Two Limits

Two limits belong to this account. The first is that H2 and H3 were formulated for transactional search, and the corpus at that stage held only the aggregate brand search index, without the split into informational and transactional queries. The support is therefore at the level of aggregate search, and the test as pre-registered waits for the split. The second is that the category assignment was revised on the same corpus on which the hypotheses were then tested. The revised assignment is now a prediction in its own right and has to be checked on peaks outside the 66. The Cautious group has 18 observations, which is enough for a first analysis and short of what interaction tests need.

What the sequence shows

What the sequence shows is the working order of the series. Article 5 stays published as written, next to Article 6, so a reader can set what was expected against what was found. One of three hypotheses failed in public. The part that survived is narrower than the original proposal, and the result that matters for practice came from the hypothesis that said the right structure depends on the category.