I recently came across an article in Forbes on the salaries of Data Scientists. The piece summarized findings from the just-published 5th-annual Burtch Works Study: Salaries of Data Scientists. The study is generally well-done, providing insight into trends and breakdowns of compensation for Data Science professionals.
I like that Burtch Works contrasts Data Science (DS) with Predictive Analytics (PA), limiting its focus to DS. For Burtch Works, the two disciplines share quantitative and analytics skills. What distinguishes DS from PA are the emphases of the former on more complicated data and computation. I agree with BW that data and computation are critical for DS, even as some industry influencers equate Data Science with Modeling and Machine Learning.
My take on the purview of DS is actually a bit more expansive than Burtch Works’, with Data Science-defining categories Technology/Computing, Quantitative Methods/Models, Substance/Business, and Science/Research Methodology. Technology/Computing and Quantitative Methods/Models are similar to Burtch Works’ Computer Science/Coding/Unstructured Steaming Data and Quantitative Skills/Analyze Data respectively. Implicit in the Burtch Works’ scheme for gleaning insights is my Substance/Business. My Science/Research Methodology that includes knowledge of Research Methodology/Design, and both expertise in the conduct of research and the delivery of readily shareable and reproducible results, is perhaps a step beyond.
One complaint I have about most industry surveys is the lack of evidence that the sample represents the population of interest and, in the absence of random sampling, redresses sources of selection bias. It’d certainly be nice to understand how the BW sample of “399 of the approximately 4,000 Data Scientists with whom Burtch Works maintains contact” represents the 4000 on axes such as geography, age, and experience. Are there biases in the sample revolving on the selection criteria that: “Professionals were included in the sample only if (1) they satisfied Burtch Works’ criteria for Data Scientists, and (2) Burtch Works obtained complete information about that individual’s compensation, demographic, and job characteristics.” Alas, it seems I include a paragraph like this one in every survey review blog I write.
A report analysis that caught my eye was a comparison of Data Scientist salaries on the intersection of experience and education, with experience levels of 1-3 years, 4-8 years, and 9+ years, and educational attainment of Masters vs PhD. Not surprisingly, there’s an over 25% difference in median salary by category of advancing experience, along with a consistent differential between Masters and PhD of roughly 10%. Makes sense to me: obviously, experience is critical and PhD’s may well bring more of the Science/Research Methodology component of DS to the table from the get-go than do most Masters-trained practitioners.
The educational difference findings don’t imply, though, as more than a few of my LinkedIn connections have suggested, that PhD Data Scientists out-earn Masters peers over the course of their careers. In fact, you could make a persuasive argument that the opposite is likely true.
Figuring conservatively, it takes 2 years to earn a Masters and 5 to complete a PhD. Consider identical twins who start grad school at 22 years of age. M will complete her schooling and enter the job market at 24, while P will be studying until she’s 27. M thus gets a 3-year head start on P in the Data Science job market, where she’s not only drawing a salary while P isn’t, she’s also gaining important experience that will elevate her to the next level 3 years sooner than P. And M’s compensation with 3 year’s experience at age 27 will exceed P’s starting comp. If both then progress on “average”, P may never catch up in career compensation to M, even with an annual 10% PhD differential.
On the other hand, with her PhD, P may be more skilled and progress more rapidly than M, exceeding M’s career compensation in time. Unfortunately, we cannot draw any conclusion on this possibility from the survey. Much as some might conjecture, the question of career earnings for Master vs PhD-trained Data Scientists cannot be resolved with the Burtch Works’ data.
My cautions with this are 2-fold: first, the Data Scientist must be methodologically rigorous in formulating her theories relating to the areas of inquiry. And second, she must implement a design and gather data that can conclusively support resolving the hypotheses derived from those theories. She cannot speak louder than her data permits.
When all’s said and done, it may well be the rigorous Science/Research Methodology that most differentiates Data Scientists and drives compensation.