"We need a statistician" is a sentence said in a lot of departments and it almost never describes the problem. Sometimes what is missing is a person on the payroll; far more often what is missing is someone reading drafts before they go out, or doctoral students knowing when their design will not support the conclusion they want. Those are different needs and they are solved by different things.
I write this from the side of someone who sells one of the four options, so assume the bias and judge it by what it says about the other three. I have tried to make the drawbacks of my own the longest on the list, because they are the ones I know best.
Before choosing, measure three things
None of the four options is better in the abstract. The decision depends on three numbers you probably have not written down and that can be gathered in an afternoon by email:
- Volume. How many manuscripts with quantitative analysis leave the department each year, and how many theses are defended.
- Rhythm. Is it a steady trickle or peaks? A department that concentrates eight submissions in March and another eight in September has a capacity problem at specific moments, not a staffing problem.
- Type of work. Is the gap in designing analyses, running them, reviewing what has already been done, or teaching people to do it? Those are four different jobs and very few people do all four well.
Add a fourth question, which decides more than it looks: can the data leave the institution? If you work with clinical records or with minors, the answer conditions everything else and probably rules out half the list.
Option 1: a methodologist on the payroll
The solution everyone wants and the one that is hardest to get. When it happens, it works very well.
What it buys. Someone who learns your field, your instruments and your databases, and who is worth twice as much in year two as on day one. Continuous availability without a purchase file per question. Data governance: nothing leaves the building. And internal teaching, a badly underrated side effect: a staff methodologist ends up training the doctoral students by osmosis.
What it really costs. Not the salary but the salary plus employer contributions, plus space, plus the administrative time to create the post. And above all the lead time: between approving a post and the person starting, anywhere from nine months to two years depending on the institution and the contract type. If your problem is the manuscript going out in November, this option does not solve it.
The two drawbacks nobody mentions. The first is single-person risk: if all the department's methodological capacity lives in one head and that head leaves, retires or goes on sick leave, you are back to zero with less slack than before. The second is a career problem, and it is more serious: a technical methodologist who only provides service does not accumulate publications of their own, and without their own publications there is no promotion. Either you give them authorship when it is due and room for their own line, or in three years they leave. I have watched that full cycle twice.
When yes. Large departments, continuous flow, data that cannot travel, and the real ability to create the post. If you meet all four, this is the best option and the other three are patches.
Option 2: training
Workshops, seminars, doctoral school courses. The cheapest option per head and the one with the worst reputation, partly deserved.
What it buys. Scale: twenty people for the price of one consultation. It changes habits when it is done well. And it is the only option that attacks the cause rather than the symptom, because the person who learns to justify a sample size will do it again in the next project.
What usually goes wrong. Training without an immediate task does not transfer. A mixed models workshop in October, for someone who will need mixed models in April, is a wasted workshop: in April they will have to look it all up again. The practical rule is to deliver training when people have the data in front of them, not when it fits the department's calendar. A general "advanced statistics" course for a heterogeneous group is the most reliable way to spend a training budget without changing anything.
And a structural limit. Training cannot fix a design that has already been executed. If the data are collected and the allocation was not random, no workshop turns that into a trial. For the manuscripts currently in flight, training arrives late by definition.
When yes. For doctoral students, always, and the earlier in their trajectory the better. And for concrete needs that can be practiced the same day: preparing the dataset, reporting in APA 7, running a systematic review. The return there is high and cheap.
Option 3: external consulting
This is what I do, so here I am going to be particularly hard.
What it buys. It absorbs peaks without creating a post. Fixed, predictable cost, which is what a budget needs. And an outside view that is sometimes worth more than the analysis: someone who has not lived inside your project sees the holes the team stopped seeing three years ago.
What goes wrong. On day one they do not know your instrument, your population or the theoretical argument you are in the middle of, and you pay for that learning curve in meetings. It amortises if the relationship lasts; if you hire a different supplier every year, you pay it in full every year.
The timing mistake, which is the big one. Most consulting is bought late: data collected, manuscript written. At that point the room to maneuver is nearly zero. Nobody can turn a cross-sectional design into a longitudinal one, or recover the variable that was not measured. Buying at the design stage costs less money and is worth far more, and yet almost nobody does it, because at the design stage the problem does not hurt yet.
The dependency risk. If the supplier delivers results and no syntax, the department learns nothing and is in exactly the same position next year. Always ask for the code. A service that does not leave you reproducible syntax is selling you an output rather than a capability, and two years later you notice.
And the authorship point. Settle it at the start. If the person analyzing helps design, interprets and writes, they meet authorship criteria at most journals; if they execute a plan you already defined, acknowledgment is the right place. Both are legitimate. What creates conflict is not having discussed it until the manuscript is finished.
When yes. Peaks, design decisions with expensive consequences, pre-submission review and responses to reviewers. And as a bridge while a post that will take two years is being created.
Option 4: tools
Calculators, automated checkers, table generators, AI reviewers. The cheapest layer and the one most misunderstood in both directions.
What it buys. It is available at eleven on a Sunday night, which is when people write. It covers the mechanical layer well, and the mechanical layer is much bigger than it looks: half of the psychology papers using significance testing published between 1985 and 2013 contained at least one p value inconsistent with its test statistic and degrees of freedom, and one in eight an inconsistency serious enough to affect the conclusion (Nuijten et al., 2016). An automated check catches that in seconds.
What it does not do. A tool does not decide the design or interpret a result in its theoretical context. And automating a wrong analysis does not improve it: it makes it faster, prettier and harder to argue with, because now it comes with an APA 7 table. The real risk of this layer is not that it fails, it is that it creates confidence where none is warranted.
When yes. Always, as a floor, and never as a ceiling. On this site you have, free, reference verification against Crossref, PubMed, OpenAlex and Semantic Scholar, the APA 7 results formatter and the AI paper reviewer, which returns the objections of two simulated reviewers before the real one writes them. They are a good first filter for a whole department and they cost nothing.
What none of the four fully fixes
There is an underlying problem worth having in mind before signing anything, because it qualifies the promise of every option on this list, mine included.
Silberzahn et al. (2018) handed the same dataset and the same question — are football referees more likely to give red cards to darker-skinned players? — to 29 independent teams, 61 analysts in total. Analytic approaches varied enormously and the estimated effect sizes ranged from 0.89 to 2.93 in odds ratio units, with a median of 1.31. Twenty teams (69%) found a significant positive effect and nine (31%) found no significant relationship. Same data, same question, competent analysts.
The conclusion I draw is not "it does not matter who analyses". It is that what you should be buying is not "the right answer", which often does not exist as a single number, but declared and defensible decisions: a written record of why that model was chosen, which alternatives were discarded, and what happens to the result under other reasonable choices. That is what pre-registration, reproducible syntax and a second person arguing about the decisions before they become results are for. If pre-registration sounds like paperwork, the full procedure is in pre-registration on OSF, step by step, and it takes fifteen minutes.
The realistic mix, by size
Nobody picks one option. These are the combinations I see working:
- Small group (up to ~10 people, few submissions a year). Tools as a floor, an internal rule of cross-reading between colleagues — free and surprisingly effective — and external consulting only for peaks and for expensive design decisions. A post is not worth it, and specific training lands better from outside.
- Research line or mid-sized group. Tools, one annual workshop timed to the doctoral students' real calendar, and an external arrangement at a fixed price covering pre-submission review and reviewer responses for everybody. This is where an annual bundle for the group or the research line tends to come out cheaper than paying for individual consultations, and cheaper still in administrative effort: one purchase file a year instead of one per manuscript.
- Large department with continuous flow and sensitive data. A staff post, no question, plus tools. External consulting stays for what the post cannot cover: peaks, second opinions, and the specific methodological areas where that person is not the specialist, which always exist.
How to know whether the support is working
If you are going to spend money on this, measure before and after. Four indicators, all obtainable without surveying anyone:
- Mean months from first submission to acceptance.
- Mean number of journals per published article.
- Proportion of desk rejections out of all rejections. If it falls, journal choice is improving; if it does not move, the problem was never methodological.
- How often the same objection repeats across manuscripts written by different people. This is the indicator that best detects whether knowledge is staying in the group or only in the one-off consultation.
One year of data before and one year after is enough to have an adult conversation at a department meeting. Why those four indicators and not others is spelled out in what a rejection really costs at group scale. And if the conclusion ends up being to buy something, the administrative route — thresholds, budget lines, what your finance office will ask for — is in how to buy statistical support for a research group.
References
- Nuijten, M. B., Hartgerink, C. H. J., van Assen, M. A. L. M., Epskamp, S., & Wicherts, J. M. (2016). The prevalence of statistical reporting errors in psychology (1985-2013). Behavior Research Methods, 48(4), 1205-1226. https://doi.org/10.3758/s13428-015-0664-2
- Silberzahn, R., Uhlmann, E. L., Martin, D. P., Anselmi, P., Aust, F., Awtrey, E., ... Nosek, B. A. (2018). Many analysts, one data set: Making transparent how variations in analytic choices affect results. Advances in Methods and Practices in Psychological Science, 1(3), 337-356. https://doi.org/10.1177/2515245917747646