Statistical consultation

Our basic rule is that if the statistician contributes in any substantial way to a publication, with advice, planning, or analysis, co-authorship should be discussed. The statistician will in that case contribute by writing a description of methods used and will thereby be responsible for this part of the contents.

The interpretation is a matter of judgement, but a first meeting with us never obligates you to share the authorship.

Large language models (LLMs) can be helpful in statistical analysis if you have a clear research question and understand the details of your study design and data well. For example, they can help you produce informative figures and charts, correct syntax or generate ideas about how to continue an analysis.

If, however, you are not sure what to do with your data, it is very risky to let LLMs decide on models and analysis pathways unguided. Some commonly met problems are:

• LLMs do not inspect and validate your dataset. Any design structure and assumptions must be clearly communicated. For example:
  
  
  •  LLMs have no or little sense of different variable types. They might fit an ANOVA on 0/1 data instead of a logistic regression as they only recognise that the response is numerical, but not that it is binary. Similarly, they will not recognise categorical explanatory variables coded as numbers unless this is specified.
  • LLMs also do not understand which experimental design is present unless this is explicitly communicated. Otherwise, split-plot designs are commonly
    overlooked and analysed as standard factorial designs. For block designs, it is sometimes suggested to analyse blocks (farms, areas, …) separately instead of fitting block models. Similarly, clustered, hierarchical, and repeated-measures designs are usually not recognised.
    
    
  • Missing value codes, limitations due to sample size, and similar issues are not handled correctly unless clearly communicated.
    
    
•  LLMs often wrap the suggested models in loops or functions with the goal to produce nice tables with estimates and p-values. This can make the necessary model validation difficult, as the individual model fits are not easily accessible.

Also remember that you, as researcher, always remain fully responsible for the validity of your analysis and must ensure that the use of AI tools complies with institutional policies, confidentiality agreements, and data protection regulations. Sensitive or unpublished data should not be shared with external systems unless permitted.

If you come to Statistics@SLU for consultation, please inform us if you have used LLMs and for what purpose. We are happy to discuss output from LLMs, but we rely on you to have tried to understand and evaluate the output before a meeting. If you are unsure if you can trust the LLM results or do not understand them, it is often easier to have discussions without them.

In many consultations, the focus is on clarifying the research question, structuring the dataset appropriately, determining suitable sample size and power, and exploring appropriate analytical alternatives while assessing assumptions and limitations. Researchers are encouraged to prepare by conducting basic exploratory analyses, such as descriptive tables and simple plots.

The Centre for statistics does not offer statistical consultation for undergraduate students. However, statistical support related to degree projects is available for students at the S- and LTV-faculty without the centre's involvement. Undergraduate students who require statistical support for their thesis work can contact Adam Flöhr (for the LTV-faculty) and Hilda Edlund (for the S-faculty).