In my opinion, this claim "Why not use Python for statistical analysis" is rooted in:

• the fact that R (and originally S), was created as a dedicated tool just for statisticians, as a specialized purpose language, Domain-Specific Language (Python is a General-Purpose Language). It was a simplified language, while giving access to a specific data structure, like a frame (data.frame). For about 40 years, as multitudes of statisticians "grew up" on it, it became the lingua franca of statistics. New specialized methods were implemented in it , knowledge was exchanged by illustrating them with programs in R (CrossValidated, for example), articles and books using it were published. It's a bit like Latin in the past, English now and Turbo-Pascal in numerical methods in the 70-90s :)

• the availability of specialized statistical methods in R. While in the field of modern methods of so-called ML (Machine Learning; machine learning) Python seems to become the undisputed leader, in the field of classical (small-data) statistics, i.e. basic tests and models Python "makes it" (statsmodels is quite OK), already in the case of specialized statistics, such as clinical biostatistics, Python does not offer much.

• In this field, among OpenSource tools, only R is a reasonable alternative (because not competition) to SAS (SAS has been number one in clinical research since 1980 and is likely to be so for 2 decades[1] ). Probably for this reason, among others, Python is virtually non-existent in clinical research. One would also have to ask actuaries and econometricians - how about Python with coverage of the methods they use.
  

To confirm the second point, I have prepared a list of missing statistical tools in Python (as of 2020):

1. Lack of modern non-parametric methods for factorial and longitudinal (repeated measures) systems, including methods such as ATS (ANOVA-Type Statistic), WTS (Wald-Type Statistic), ART (Aligned-Rank Transform).

2. No frequentist procedure (there is only Bayesian, and in clinical research this is not the standard) of generalized and nonlinear mixed models, with a user-defined residual correlation structure. And this means little usefulness when analyzing data in repeated observations. In R, we have glmmTMB, glmmPQL, nlme (only general linear model) for this. The lme4 package, unfortunately, does not allow for this and is, on average, suitable for more serious analysis in this area, unless you fit "unstructured" covariance (convergence problems from lack of degrees of freedom) or "compound symmetry" mastics (but this is not a very realistic assumption). The successor to lme4 is precisely glmmTMB.

3. No vector generalized linear and additive models (VGLM, VGAM), and this is one of the basic classes in this field.

4. Lack of quantile linear mixed models (quantile linear mixed models). Lack of procedures to determine reference ranges.

5. Lack of confidence intervals for proportion differences (important vs. non-inferiority studies) Farrington and Manning's, Gart and Nam's, Miettinen and Nurminen's.

6. No advanced methods for determining sample size and power for clinical trials. Only basic, "textbook" methods available, practically of little use except in the simplest cases. What is really needed is missing, including 3-arm "Gold Standard" type studies, Williams MED, adaptive design, longitudinal studies (mixed models, GEE, WGEE).

7. Lack of advanced study design methods (adaptive, multi-arm (MAMS). Conditional power (conditional power) is lacking. This is an absolute foundation in modern research and at the same time one of the most complex issues in the field.

8. Lack of more sophisticated methods in survival analysis (multiple events - e.g. Andersen-Gill, competing risks (let's call them shorter, co-risks) - e.g. Fine-Gray, Cause-Specific Hazards). The Kaplan-Meier method does not have "confidence bands" (not to be confused with confidence intervals). I did not find the Beta Product Confidence Procedure (BPCP).

9. The Turnbull estimator for "interval-censored" observations (interval-censored) and Cox for such data (IntCox) are missing. The Pepe-Mori test for correlated co-risks is missing. No Wilcoxon-Prentice test for dependent data.

10. Lack of "joint-frailty" models (I don't undertake the translation) that allow for analysis of multiple events (e.g., recurrence) and co-risks (e.g., death) together in survival analysis.

11. I did not find multi-state (multi-risk) models, commonly used in survival analysis in clinical trials.

12. No "tipping point" type analysis - "tipping point" (sensitivity analysis in the context of missing data)

13. Lack of comparably (to R) advanced meta-analysis.

14. Lack of more advanced confidence intervals for the median, such as Nyblon, as well as an exact method (there is only a basic one, based on a binomial distribution).

15. Lack of more complex contrasts (relevant to clinical trials): Sequen, AVE, Changepoint, Williams and Umbrella-protected Williams, Marcus, Mc7Dermott.

16. I haven't found a comparably advanced framework for permutation linear models (in R it's e.g. lmPerm, permuco), although there are some first steps in scikit-bio.

17. I haven't found an implementation of LS-means (in R called EM-means; estimated marginal means), which are the absolute standard in clinical trials for modeling and reporting effects.

18. I'm not sure if marginal effects are available for the entire GLM class or only for logistic regression, but before performing analyses using them, I would suggest making sure in advance (or counting them yourself).

19. I have not found more sophisticated methods of multiple comparisons, especially those relevant to clinical trials: hierarchical hypothesis testing (relevant when we have hypotheses ordered according to a certain hierarchy, e.g. non-inferiority → superiority); Fixed sequence procedure; Fallback procedure; Gatekeeping (serial, parallel), e.g. Truncated Hochberg, tr. Holm, tr. Hommel so-called sequential online testing graphical approach.

20. Non-parametric methods for these contrasts: Tukey, Dunnett, Sequen, Williams, Changepoint, McDermott, Marcus, Marcuss-Umbrella protected.

Some of the items on this list could be implemented fairly quickly, but many of them are very complex, requiring a lot of time, expertise and access to other specialized packages, numerical validation. These are all tasks for multi-person teams of specialists, and for a long time (especially if it's an after-hours job). Besides, there are problems with this in R itself. Not to mention validation (required by regulations) against SAS (reference software in pharma) and possibly Stata.

To be honest, I would not be able to work effectively in my profession using Python alone. I would either have to implement the missing methods myself, or refer to R.

If a person does not need any of the above methods to be happy, he can successfully use Python's arsenal. However, it can be seen that, at least in terms of specialized (non-ML) methods, R offers much more.

Note, this does not mean that Python is generally "bad" in this area, it will successfully fulfill its role in many applications. It's just - there are areas where R provides a larger set of tools, and since, many of these tools are "interdisciplinary", hence the general statement, "R is better for statisticians", but no longer for "data scientists" (I don't know how to translate it sensibly :) , who use other methods (ML, AI) on a daily basis. The latter seem to strongly prefer Python.