Probability of Automation of Occupations 2036

We used a similar CBTC methodology to Frey and Osborne (2017), Autor and Dorn (2013), and Caines et al. (2017). The US Department of Labor’s Dictionary of Occupations Titles (DOT) classification system and the 2010 Standard Occupation System (SOC), which allowed for the respective data sets to be cross-referenced. The 2010 SOC system classified 840 detailed occupations with similar job duties, skills, education and training into 461 broad occupations, 9 minor occupations and 23 major groups (U.S. Bureau of Labor Statistics, 2010). We also used a secondary data set, namely US Occupational Employment Statistics (OES) data as it contained detailed occupational descriptions and employment statistics per occupation for the US economy, broken down per industry according to the NAICS classification system. The NAICS is an industry classification system that grouped organisations into industries based on similarities in production processes. It used a six-digit coding system to classify all economic activities in the US across 20 industry sectors and 1 057 detailed industries (U.S. Office of Management and Budget, 2017). Employment data was classified according to the 2010 SOC system that contained 820 occupations. With the emergence of new roles and reclassification of the SOC system over the same duration, certain classifications needed to be cross-referenced (“crosswalked”) to provide comparable figures across different years.

In their studies, Frey and Osborne (2017), Caines et al. (2017) and Autor and Dorn (2013) aggregated the occupations slightly differently. For example, Frey and Osborne (2017) aggregated specific “postsecondary teaching” occupations into a single category and omitted occupations containing “all other”, whilst Caines et al. (2017) omitted selected “farm and agricultural” occupations. In total, Frey and Osborne (2017) calculated the probability of job automation for 702 occupations. Autor and Dorn (2013) calculated job routineness for 330 aggregated occupations. Caines et al. (2017) calculated job complexity for 315 aggregated occupations. Our approach combined and cross-referenced all of the above-mentioned data sets to produce a total of 291 occupations for the analysis.
In projecting the change in workforce structure from 2016 to 2036, the 2016 OES employment numbers were adjusted for the “probability of job automation” at an occupational level, and aggregated to illustrate the relative change in workforce structure for the entire US economy. Occupations which did not have a corresponding measure of “probability of job automation” were omitted from the analysis. After omissions, this analysis represented 96.3% of the US workforce in 2016.

本研究采用了与Frey和Osborne（2017年）、Autor和Dorn（2013年）以及Caines等（2017年）相似的中国铁路列车控制系统（CBTC）方法论。美国劳工部职业名称词典（DOT）分类体系和2010年标准职业分类（SOC）系统得以应用，以便于相应数据集的交叉参照。2010年SOC系统将840个具有相似工作职责、技能、教育及培训要求的详细职业划分为461个宽泛职业类别、9个次要职业类别和23个主要职业群体（美国劳工统计局，2010年）。此外，我们还使用了次级数据集，即美国职业就业统计（OES）数据，因为它包含了美国经济中每个职业的详细职业描述和就业统计数据，并根据北美行业分类体系（NAICS）按照行业进行细分。NAICS是一种行业分类体系，它根据生产过程的相似性将组织划分为不同的行业，采用六位数的编码系统对美国所有经济活动进行分类，涵盖20个行业部门及1,057个详细行业（美国管理预算办公室，2017年）。就业数据按照2010年SOC系统进行分类，该系统包含820个职业。随着新角色的出现以及SOC系统在同一时间段内的重新分类，某些分类需要交叉参照（即“映射”），以在不同年份提供可比数据。在他们的研究中，Frey和Osborne（2017年）、Caines等（2017年）和Autor和Dorn（2013年）对职业的聚合方式略有不同。例如，Frey和Osborne（2017年）将特定的“高等教育教学”职业合并为一个类别，并省略了包含“所有其他”的职业，而Caines等（2017年）省略了某些“农业和农业相关”的职业。总计，Frey和Osborne（2017年）计算了702个职业的自动化概率。Autor和Dorn（2013年）计算了330个聚合职业的工作常规性。Caines等（2017年）计算了315个聚合职业的工作复杂性。我们的方法结合并交叉参照了上述所有数据集，从而产生了291个职业用于分析。在预测2016年至2036年劳动力结构的变化时，根据职业层面的“自动化概率”调整了2016年OES的就业数据，以展示整个美国经济的劳动力结构相对变化。没有相应“自动化概率”度量标准的职业被排除在分析之外。在排除这些职业后，该分析代表了2016年美国劳动力的96.3%。