Hire me

Work/PRJ-11/Segmentation · US workforce

Workforce diversity: which US industries are the least diverse?

Construction, agriculture and mining. They are 88–92% white, and construction is 89% men. White workers are the majority in every sector (71–92%).

Data
US labour force statistics by sector and industry, 2020–2023
Coverage
13 sectors, 110 industries
Tools
Python, pandas, scikit-learn, seaborn
Code
github.com/BoredMongoose/strata-clustering ↗
Business question

Which US industries most need diversity hiring?

Recommendation

Start diversity hiring in construction, agriculture and mining, and focus on hiring women in construction, mining and transportation. Track the same numbers every year: mining's Hispanic/Latino share rose 5 points in two years, so they can move.

Estimated impact89%

Of construction workers are men, the most male-dominated sector (mining is 85%).

01Results

Construction is 89% men; education and health is 75% women.

Construction is 89% men; education and health is 75% women. Average share of women in each sector, 2020–2023. Only education and health is above 50/50.
Fig. 1Construction is 89% men; education and health is 75% women. Average share of women in each sector, 2020–2023. Only education and health is above 50/50.
White workers are the majority in every sector. From 71% in transportation and utilities to 92% in agriculture. Hispanic/Latino overlaps with the other groups, so the bars don't add up to 100%.
Fig. 2White workers are the majority in every sector. From 71% in transportation and utilities to 92% in agriculture. Hispanic/Latino overlaps with the other groups, so the bars don't add up to 100%.
Three groups of sectors. K-Means put construction, agriculture and mining in a group of their own: mostly white, mostly men. Most other sectors are about half women.
Fig. 3Three groups of sectors. K-Means put construction, agriculture and mining in a group of their own: mostly white, mostly men. Most other sectors are about half women.
110 industries, four groups. 70 mostly white industries, 21 with more Black workers, 14 with more Hispanic/Latino workers (restaurants, landscaping) and 5 with many more Asian workers (electronics, nail salons).
Fig. 4110 industries, four groups. 70 mostly white industries, 21 with more Black workers, 14 with more Hispanic/Latino workers (restaurants, landscaping) and 5 with many more Asian workers (electronics, nail salons).
Bigger sectors are only a little more diverse. Each dot is a sector. The correlation between diversity and size is 0.42, a weak link with 13 sectors.
Fig. 5Bigger sectors are only a little more diverse. Each dot is a sector. The correlation between diversity and size is 0.42, a weak link with 13 sectors.

02How I did it

From raw data to a decision.

  1. Cleaned the data: removed blank and total rows, fixed percentages typed as whole numbers (87.5 instead of 0.875), and dropped one impossible value.
  2. Answered six questions with pandas groupby: gender balance, ethnic mix, change from 2021 to 2023, and size against diversity.
  3. Grouped sectors and industries with K-Means after standardising the columns, picked the number of groups with the elbow method, and drew them with PCA.
  4. Scored diversity with Shannon entropy: higher means workers are spread more evenly across the groups.

03What could make this wrong

The limits, said out loud.

  • 01Only four years of data.
  • 02The percentages are survey estimates, so small industries jump around from year to year.
  • 03K-Means always finds groups, even weak ones: the groups describe the data, they don't prove anything.

04What I learned

  • 01Clean first: some percentages were typed as 87.5 instead of 0.875, and one Asian share was above 100%.
  • 02Standardise the columns before K-Means, otherwise the biggest numbers decide the groups.
  • 03Check written findings against the output: my first write-up had the industry groups mixed up.

05What I'd do next

  • 01Add pay data, to see whether the least diverse industries also pay differently.
  • 02Use more years, to tell real changes from survey noise.
  • 03Look at job level (entry level vs management), not just the industry.
See the code on GitHub ↗All projects