It has been observed that a rising blast count during the waning moon and splenomegaly extending at least three fingers below the costal margin are associated with inferior outcomes, particularly among patients treated on a Friday or before a national meeting. Among 17 consecutive patients, chemotherapy with venetoclax for 12 days was associated with improved survival. Treatment duration was selected at physician discretion for reasons not entirely remembered. The survival benefit persisted on multivariable analysis after exclusion of two inconvenient subjects. Prospective validation is warranted but unlikely to be feasible, underscoring the enduring importance of the present series.
The traditional farmer’s almanac is a peculiar form of knowledge. It gathers observations about seasons, weather, planting, and recurring natural events, then uses those patterns to suggest future decisions. Its contents may be careful, accumulated over many years, and genuinely useful. Yet an almanac does not report the results of experiments. It records what has happened and forecasts what may happen again; it does not identify why events occurred or establish what would happen if one condition were deliberately changed.
Much of the hematology literature consists of retrospective cohorts, registries, prognostic models, molecular subgroup descriptions, and uncontrolled treatment series. These studies can be useful, especially in rare diseases. The risk is that they often adopt the language and appearance of hypothesis-testing science despite lacking a design capable of separating causal effects from selection, confounding, measurement, and chance.
The problem is not that hematology has too many almanacs. The problem is that we print them in the format of experiments and read them as though they were scientific proof.
What is almanac evidence?
Early modern English almanacs mixed calendars and weather predictions with medical guidance. In a review of 1,392 surviving English almanacs from 1640 through 1700, almost three-quarters contained preventive advice or advertisements for medical products and services. Modern biomedicine has even reclaimed the name: NCI-ALMANAC is a systematic screen of anticancer drug combinations explicitly intended to generate hypotheses for further testing.
That is the right way to use an almanac: as a catalog of observations and a starting point for tests, not their substitute. A transplant registry can describe relapse and non-relapse mortality. An AML cohort can associate a mutation with shorter survival. An institutional series can document responses to an off-label therapy. These are real observations.
The category error begins when an association becomes a mechanism, an outcome after treatment becomes an outcome caused by treatment, or a subgroup discovered in the data becomes a biological entity.
Some observational studies can support causal inference when the intervention, comparator, population, time zero, and assumptions are defined in advance. Most small retrospective series are not designed this way. A Kaplan-Meier curve, a multivariable Cox model, or a propensity score can improve the description and adjustment of recorded data. None can account for patients who were never referred, tested, treated, or entered into the dataset. Statistics can make an almanac more precise. They cannot turn it into an experiment.
Why we need almanacs
Hematology is unusually dependent on descriptive evidence. Many diseases are rare, molecular subgroups are small, therapies change faster than randomized trials can mature, and important questions may never attract funding for research studies. Waiting for definitive evidence may mean having no guidance at all.
Almanac evidence can identify toxicities, estimate event rates, describe natural history, reveal exceptional responders, and expose practice variation. Almanacs are a powerful hypothesis-generating tool. A good almanac may be more useful than a poor trial.
The distinction is not a ranking of worthy and unworthy work. It is a statement about the kind of knowledge produced. A descriptive study should say, plainly, “This is what occurred in our selected population.” It does not need to inflate that observation.
Show us every patient
If a study is functioning as an almanac, it should accept the obligations of a good one: record faithfully, disclose completely, and allow readers to inspect the entries.
For a retrospective cohort of 100 or fewer patients, publication of deidentified patient-level data should be the default. A baseline table, a multivariable model, and a Kaplan-Meier curve are not enough. Authors should provide one row for every patient, including the clinically relevant baseline characteristics, treatment and timing, response, duration of response, relapse or progression, subsequent therapy, transplantation, last follow-up, survival time, vital status, cause of death when known, and the timing and reason for censoring. Authors should include as much molecular information as possible, including specific mutations and variant frequency. The report should also account for every screened, excluded, and omitted patient.
A Kaplan-Meier curve remains useful, but it compresses fundamentally different clinical courses into the same downward steps and censoring marks. A patient censored alive after six months because follow-up was lost is not equivalent to one who remains in remission at their five-year follow-up appointment. Both may appear as small ticks on the curve. Readers should be allowed to see the difference clearly.
“All the data” does not mean releasing the electronic medical record or compromising patient privacy. It means providing every deidentified patient-level observation needed to reconstruct the reported results and understand every inclusion, exclusion, event, and censoring decision. Dates can be converted to intervals, and uncommon identifying combinations can be suppressed. If some information truly cannot be shared, authors should state exactly what was withheld and why.
The ledger can be a downloadable spreadsheet, accompanied by a compact patient-level table or swimmer plot for readability. For 100 patients, this is not technically difficult and would usually require only a few supplementary pages. Complete patient-level reporting also preserves value for methods not yet developed. Future AI systems may detect meaningful patterns across many small cohorts, but no model can recover clinical trajectories irreversibly compressed into medians, hazard ratios, and Kaplan–Meier curves.
Journals should stop accepting small retrospective series without the underlying patient-level ledger. If a dataset is small enough for investigators to examine one patient at a time, it is small enough for readers to do the same. Statistical compression should not conceal important details, and many studies simply do not need statistics.
Label the almanac
The abstract should say whether a study is descriptive, hypothesis-generating, predictive, or designed for causal inference. That label should govern the title and conclusion. It would not diminish exploratory research. It would make the boundary between observation and interpretation visible.
Calling a study an almanac is not an insult. Careful records have value. But an almanac should act like one: label its observations honestly and show us the entries in their entirety.
Neil Dunavin, MD is a hematologist-oncologist specializing in leukemia and stem-cell transplantation. He practices in the Chicago area. The author used ChatGPT to assist with drafting, editing, and condensing this essay.
