Combinations of factors not observed, non-full rank design matrix. How to explain to investigator?

The best approach depends on the goals of their (your) analysis! I provide general and detailed considerations and ideas below:

Concatenating Sex and Age into a new variable (V3) and running the model with only ~V3 is actually a good way to address the issue of a non-full-rank design matrix, as you pointed out. However, as you also correctly pointed out, this approach has limitations and potential drawbacks, which I explain more below:

By combining Sex and Age into a single variable, you lose the ability to separately examine the effects of Sex and Age on the outcome. The model will only capture the overall effect of the combined variable (V3: Sex and age), and any interpretation will be based on this combined effect. This may not be desirable if you (they) are specifically interested in understanding the individual contributions of Sex and Age.

If it is important for your (their) analysis to retain interpretability for Sex and Age separately, you might want to explore other ways of handling the missing combination. One practical option is to impute the missing combination based on the observed data or use other statistical methods for handling missing data.

You can also suggest them to consider analyzing subsets of the data based on Sex or Age separately. For example, they could analyze the data for males separately from females or for old individuals separately from young individuals. This way, they can still make valid claims within each subgroup, even if the overall model is affected by the missing combination. I know the sample size is too low, but this is a valid approach.

If the missing combination is primarily affecting the interaction term (sex * age), they might consider running the model with just the main effects (sex + age) and examining the interaction effect separately using additional analyses or visualizations. Moreover, if the missing combination is causing issues, they may also simplify their model. For example, they might drop the interaction term and focus on main effects only (sex + age).

Ultimately, the best approach depends on the goals of the analysis and the importance of retaining interpretability for Sex and Age separately. If the missing combination is a rare event in your (their) analysis and doesn't have a significant impact on the research question, simplifying the model or conducting subset analyses might be reasonable. However, if the missing combination is critical to your analysis, exploring imputation methods or alternative study designs may be necessary.

Hope this helps :-)