big data – Page 2 – Economic Thinking In Action

Do employers transmit revenue changes to employee wages? What does BLS big data say about it?

This paper (ASSA 2016 link below) looks to study revenue and sales volatility at the firm level and how that relates to employee level of wages. The main take away is that employee wages tend to be positively related to revunue shocks. That is, employers tend to keep employee wages steady and increasing over time regardless of the specific shocks that the firm may be experiencing at any given time.

ASSA 2016 paper

Using Microlevel BLS data to study aggregate wage dispersion

Can Microlevel BLS data be used to study how and why employees are paid differently at US employers ? This paper and the work ultimately looks to provide a method to use the Microlevel, i.e. individual level survey observations, to match dispersion measures like, the standard deviation, in big data BLS employment data. The first step for the researchers is to try and match the aggregate numbers to the micro numbers.

Perhaps one of the most useful STATA commands out there…

Working with wage and hour data and employment data, like we do on a daily basis involves the analysis of very large data sets. Big data and employment data are often one in the same. STATA has a very useful command that allows you to load in large Excel 2007/2010 spreadsheet files. It is:

set excelxlsxlargefile on

This simple command allows the user to bypass the pre-set limit on spreadsheet size.

Just remember, STATA and your computer will be unresponsive during the load. So be patient and let it all load up.

Employment and Wage & Hour Statistics Focus: American Time Use Survey

The American Time Use Survey (ATUS) is commonly used to estimate the amount of time spent on household services for valuing losses in personal injury and wrongful death cases. ATUS is sponsored by the U.S. Bureau of Labor Statistics and administered by the Census Bureau. It measures the amount of time people spend doing various activities, such as working for pay, caring for children, volunteering, commuting, and socializing.

Data collection for the ATUS began in January 2003, and estimates are released annually. The most recent survey was conducted in 2014, and approximately 11,600 individuals were interviewed. The participants were asked to account for every hour between 4 a.m. the day before the interview and 4 a.m. the day of the interview. Using this data, it is possible to construct the average number of hours spent on household services based on demographic characteristics. The categories most often used to measure household services are time spent:

Performing inside housework
Cooking food and cleaning up after a meal
Caring for pets, performing household maintenance, and maintaining vehicles
Managing the household
Shopping
Obtaining services
Traveling for household activities
Caring for and helping household children
Traveling to care for and help household members

The valuation of lost household services is performed by calculating how much it would take, in dollars, for the family to replace the services that were provided by the injured or deceased person. In other words, how much it would cost for the family to go out into the marketplace and pay for each of the individual services, such as having a person mow the lawn or take care of the financial records?

Once the market value of the service, expressed as an hourly rate, is determined, the final replacement cost of each service is calculated by multiplying the hourly market rate by the average number of hours spent performing the services, as reported in the ATUS. For example, in the 2003 ATUS it was reported that the average 24-year-old female spends 2.99 hours per week on food cooking and cleanup. The replacement cost for these services is $9.60 an hour. Therefore, the replacement cost for a week of food cooking and cleanup is $28.70 (2.99 x $9.60) and the annual replacement cost is $1,492.61 (2.99 x $9.60 x 52). The calculation is then performed for all household services for each remaining year of the person’s life expectancy. Finally, the summed loss of household services value must be discounted back to the present day.

Big data question: How big of a random sample is big enough in a wage in hour case?

That’s a question that comes up a lot in wage and hour land employment lawsuits. Typically the question is how many employees do I need to look at to have a statistically significant sample?

In some instances it’s not feasible to collect data or get all the records for
all the employees of a particular company. Sometimes the data is kept
in such a way that it takes a lot of effort to get that information. In
other instances it is a matter of the limitations of imposed by the court.

In any event, that’s a question that comes up a number times in wage and hour lawsuits particularly ones involving class or collective actions. So what’s the answer?

Generally, the size of the sample needs to be sufficiently large so that it is representative of
the entire employee population. That number could be relatively small say 40 employees or relatively large say to 200 employees depending on the number of employees at the company and the characteristics of the employee universe that is being analyzed.

For example if there are no meaningful distinctions between the employees in the universe, that is
it is generally accepted that all the employees are pretty much all
similarly situated, then a sheer simple random sample could be
appropriate.

That is, you could simply draw names from a hat, essentially. A simple random sample typically requires the smallest number of employees.

If there are distinctions between employees that need to be accounted for, then
either a larger sample or some type of stratified sampling could be appropriate.
Even if there are distinctions between employees, if the sample is sufficiently large then distinctions between the employees in the data could take care of themselves.

For instance, assume that you have a population of 10,000 employees and they are
divided into four different groups that need to be looked at differently.

One way to do a sample in this setting is to sample over each of the different groups of employees separately. The main purpose of the individual samples is to make sure that you have the appropriate number of employees in each of the different groups. That is, to make sure that the number of employees in the different samples are sufficiently representative of the distribution of the different groups of employees in the overall population.

Another way to do this is to simply just take a large enough sample so that the distinctions take care of themselves. If the sample is sufficiently large then the distribution of the different groups of employees in the sample should on be representative of the employee population as a whole.

So in this example, if there is a sufficiently large sample it could be okay to use a simple random sample and you would get to the same point as a more advanced stratified type of approach.

The key however is to make sure that the sample is sufficiently large that of course depends on the overall population and the number of groups of employees being studied.

Big BLS employment data, disability, and worklife expectancy

Big Data. Bureau of Labor Statistics. Survey data. Employment Big Data. Those are all things that calculating worklife expectancy for U.S. workers requires. Worklife expectancy is similar to life expectancy and indicates how long a person can be expected to be active in the workforce over their working life. The worklife expectancy figure takes into account the anticipated to time out of the market due to unemployment, voluntary leaves, attrition, etc.

Overall the goal of our recent work is to update the Millimet et al (2002) worklife expectancy paper and account for more recent CPS data. In addition we also wanted to supplement and expand on a few additional topics. The additional topics included looking at different definitions of educational attainment, adding in reported disability, and looking at occupational effects on worklife expectancy.

Finding: We also looked at the worklife expectancy for individuals with and without a reported disability. Disability was not covered in the Millimet et al. (2002) paper. As has been well reported, the disability measure in the BLS data is very general in nature. Accordingly the applicability of the BLS disability measure to litigation is somewhat limited. However it is interesting to note that there is a substantial reduction in worklife expectancy exhibited by individuals who reported have a disability. On average the difference is about 10 years of work life. This is consistent with other studies on disability that a relied on the BLS data. Other factors such as occupation and geographical region do not appear to have much impact on WLE estimates.

Younger workers today have slightly less attachment to the workforce than younger workers in the past

The goal of our recent work is to update the Millimet et al (2002) worklife expectancy paper and account for more recent CPS data. Their paper uses data from the 1992 to 2000 time period. Our goal is to update that paper using data from 2000 to 2013 and see if estimating the Millimet et al (2002) econometric worklife models with more recent data changes the results in the 2002 paper in any substantive way.

Finding: Overall, the worklife expectancy estimated using more recent data from 2000-2013 is shorter then in the earlier time period (1992-2000) data set. This is true for younger worker (18-early 40’s); younger workers from the more recent cohorts have a shorter expected work life then younger workers in the earlier cohorts. Conversely, while older workers in their 40s and 50s have a slightly longer worklife expectancy in the later time period data set. We are in the process of determining the statistical significance of these differences.

Table 4. Comparsion of Worklife Expectancy for 1992-2000 and 2001-2013 Time Periods
	1992-2000		2001-2013
Age	Less than High School	High School	Less than High School	High School
18	31.469	38.410	30.569	37.314
19	30.926	37.846	30.128	36.833
20	30.306	37.180	29.603	36.237
21	29.670	36.493	29.021	35.590
22	29.027	35.787	28.419	34.917
23	28.365	35.054	27.809	34.231
24	27.685	34.293	27.205	33.539
25	27.007	33.518	26.588	32.830
26	26.319	32.728	25.964	32.108
27	25.643	31.939	25.357	31.387
28	24.958	31.123	24.736	30.646
29	24.271	30.304	24.110	29.892
30	23.590	29.481	23.491	29.136
31	22.892	28.640	22.866	28.371
32	22.191	27.796	22.237	27.599
33	21.487	26.944	21.606	26.819
34	20.783	26.097	20.970	26.034
35	20.095	25.254	20.327	25.239
36	19.400	24.408	19.685	24.446
37	18.707	23.560	19.039	23.648
38	18.018	22.714	18.392	22.850
39	17.324	21.864	17.737	22.044
40	16.627	21.014	17.085	21.242
41	15.944	20.169	16.421	20.432
42	15.264	19.328	15.764	19.627
43	14.595	18.494	15.110	18.825
44	13.931	17.664	14.456	18.024
45	13.272	16.840	13.798	17.220
46	12.616	16.018	13.154	16.429
47	11.972	15.204	12.520	15.641
48	11.328	14.398	11.886	14.859
49	10.682	13.593	11.259	14.081
50	10.053	12.803	10.642	13.311
51	9.432	12.020	10.030	12.550
52	8.802	11.239	9.429	11.798
53	8.199	10.477	8.843	11.057
54	7.593	9.723	8.270	10.333
55	6.996	8.980	7.709	9.618
56	6.422	8.263	7.152	8.912
57	5.872	7.564	6.618	8.230
58	5.339	6.883	6.095	7.560
59	4.812	6.216	5.587	6.908
60	4.307	5.578	5.097	6.280
61	3.840	4.979	4.624	5.677
62	3.400	4.415	4.181	5.112
63	3.024	3.918	3.782	4.593
64	2.708	3.485	3.428	4.128
65	2.422	3.093	3.109	3.700
66	2.180	2.756	2.819	3.312
67	1.970	2.461	2.556	2.960
68	1.787	2.200	2.323	2.646
69	1.624	1.967	2.102	2.359
70	1.471	1.756	1.905	2.101
71	1.348	1.584	1.728	1.869
72	1.238	1.430	1.577	1.670
73	1.134	1.289	1.427	1.484
74	1.042	1.167	1.296	1.322
75	0.965	1.065	1.184	1.181
76	0.904	0.983	1.077	1.054
77	0.834	0.899	0.980	0.942
78	0.784	0.836	0.894	0.843
79	0.735	0.778	0.807	0.750
80	0.694	0.735	0.675	0.636

Notes:

The logistic equation includes independent variable for age, age squared, race, race by age interaction, race by age interaction squared, marital status, martial status by age, occupation dummies, year and year dummies.

The model is first estimated separately for each gender and education level combination for active persons. The model is then estimated again for inactive persons. The educational attainment variables used to estimate our model differ from that of Millimet et al. (2002) In our model, only individuals whose highest level of attainment is high school are included in the high school category. Millimet et al (2002) includes individuals with some college in the high school category.

Replication of the Millimet et al. (2002) work was sufficient and yielded similar results

Overall the goal of our recent work is to update the Millimet et al (2002) worklife expectancy paper and account for more recent CPS data. Their paper uses data from the 1992 to 2000 time period. Our goal is to update that paper using data from 2000 to 2013. The main goal of the paper is to see if estimating the Millimet et al (2002) econometric worklife models with more recent data changes the results in the 2002 paper in any substantive way

As for the results, overall there are several findings. First we were able to create a match CPS data set of 201,797 individuals where as the Millimet et al. (2002) found 200,916 matched individuals.

Overall we match their results very closely as well. For example Millimet et al. (2002) found that a male who was 26 years old with a less than a high school education had a 27.27 years WLE remaining while we found that person had 26.319 years remaining based on our replication of their work. They found that the same age person with a high school had 32.89 years remaining while we found 32.728 years remaining. The replication was particularly good for both less than high school and high school levels of educational attainment.

The WLE numbers are close but not quite as close for college and some college. This is primarily due to the fact that we use different definitions of some college and college then Millimet et al. (2002) did in their 2002 paper

Table 3. Comparsion of Millimet et al. (2002) and Steward and Gaylor (2015) Active to Active Worklife Expectancy Probabilities
	Millimet et al (2002)		Steward and Gaylor (2015) Replication
Age	Less than High School	High School	Less than High School	High School
18	32.331	38.944	31.469	38.410
19	31.801	38.239	30.926	37.846
20	31.247	37.522	30.306	37.180
21	30.684	36.794	29.670	36.493
22	30.080	36.058	29.027	35.787
23	29.450	35.294	28.365	35.054
24	28.766	34.513	27.685	34.293
25	28.035	33.711	27.007	33.518
26	27.270	32.890	26.319	32.728
27	26.495	32.052	25.643	31.939
28	25.710	31.201	24.958	31.123
29	24.923	30.341	24.271	30.304
30	24.131	29.477	23.590	29.481
31	23.345	28.606	22.892	28.640
32	22.556	27.735	22.191	27.796
33	21.775	26.862	21.487	26.944
34	21.006	25.989	20.783	26.097
35	20.233	25.112	20.095	25.254
36	19.452	24.240	19.400	24.408
37	18.681	23.370	18.707	23.560
38	17.921	22.504	18.018	22.714
39	17.178	21.641	17.324	21.864
40	16.459	20.782	16.627	21.014
41	15.734	19.928	15.944	20.169
42	15.031	19.081	15.264	19.328
43	14.333	18.242	14.595	18.494
44	13.669	17.410	13.931	17.664
45	13.020	16.588	13.272	16.840
46	12.381	15.775	12.616	16.018
47	11.758	14.974	11.972	15.204
48	11.144	14.185	11.328	14.398
49	10.538	13.409	10.682	13.593
50	9.952	12.646	10.053	12.803
51	9.379	11.898	9.432	12.020
52	8.836	11.167	8.802	11.239
53	8.299	10.459	8.199	10.477
54	7.775	9.772	7.593	9.723
55	7.265	9.107	6.996	8.980
56	6.767	8.456	6.422	8.263
57	6.261	7.829	5.872	7.564
58	5.800	7.236	5.339	6.883
59	5.397	6.678	4.812	6.216
60	5.016	6.153	4.307	5.578
61	4.678	5.672	3.840	4.979
62	4.350	5.225	3.400	4.415
63	4.060	4.815	3.024	3.918
64	3.797	4.420	2.708	3.485
65	3.574	4.061	2.422	3.093
66	3.395	3.741	2.180	2.756
67	3.224	3.445	1.970	2.461
68	3.047	3.162	1.787	2.200
69	2.873	2.886	1.624	1.967
70	2.691	2.621	1.471	1.756
71	2.528	2.401	1.348	1.584
72	2.362	2.196	1.238	1.430
73	2.170	1.999	1.134	1.289
74	2.002	1.829	1.042	1.167
75	1.898	1.672	0.965	1.065
76	1.743	1.533	0.904	0.983
77	1.592	1.449	0.834	0.899
78	1.514	1.339	0.784	0.836
79	1.461	1.274	0.735	0.778
80	1.374	1.172	0.694	0.735
81	1.273	1.046	0.661	0.687
82	1.222	0.993	0.631	0.656
83	1.121	0.912	0.604	0.623
84	0.874	0.755	0.569	0.585
85	0.433	0.355	0.522	0.532

Notes:

The econometric model described by Millimet et al (2002) and logistic regression equations by gender and education are used to calculate the worklife expectancy estimates. The model is estimated using matched CPS cohorts from 1992–2000 time period as described in the Millimet et al. (2002) paper. The logistic equation includes independent variable for age, age squared, race, race by age interaction, race by age interaction squared, marital status, martial status by age, occupation dummies, year and year dummies. The model is first estimated separately for each gender and education level combination for active persons. The model is then estimated again for inactive persons.

Steward and Gaylor (2015) Matched CPS Sample Sizes for 1993-2013 time period

Overall the goal of our recent work is to update the Millimet et al (2002) worklife expectancy paper and account for more recent CPS data.

The data for all years is shown below. Ultimately there were over 590,000 data points used in the analysis.

Table 2. Matched CPS Sample Sizes 1993-2013
	Female				Male
Year	Less than High School	High School	Some College	College	Less than High School	High School	Some College	College	Total

1993	3,766	7,326	4,898	3,452	3,376	5,619	4,280	3,935	36,652
1994	3,539	7,019	5,357	3,619	3,097	5,477	4,411	4,013	36,532
1995	3,082	6,161	5,086	3,545	2,664	4,815	4,086	3,938	33,377
1997	3,079	6,172	4,771	3,488	2,723	4,857	3,926	3,723	32,739
1998	2,839	6,113	4,873	3,672	2,694	4,952	3,995	3,834	32,972
1999	2,709	6,027	4,987	3,770	2,513	4,830	4,134	3,923	32,893
2000	2,692	5,930	5,009	3,915	2,463	4,899	4,052	4,204	33,164
2001	2,545	5,806	4,971	3,901	2,458	4,919	4,232	4,016	32,848
2003	1,096	3,218	2,579	2,411	1,019	2,701	2,122	2,470	17,616
2004	2,579	6,372	5,803	5,009	2,394	5,307	4,745	4,819	37,028
2005	2,039	5,378	5,146	4,673	1,867	4,632	4,270	4,285	32,290
2006	2,297	5,500	5,608	4,657	2,131	4,953	4,263	4,389	33,798
2007	2,147	5,730	5,466	5,060	2,076	5,133	4,344	4,592	34,548
2008	2,159	5,659	5,787	5,281	2,040	5,212	4,593	4,826	35,557
2009	2,027	5,637	5,780	5,556	2,023	5,062	4,776	4,976	35,837
2011	1,845	4,844	5,106	5,136	1,786	4,603	4,176	4,432	31,928
2012	1,733	4,849	4,930	4,956	1,779	4,693	4,151	4,616	31,707
2013	1,658	4,542	5,061	5,109	1,668	4,579	4,271	4,650	31,538

Total	43,831	102,283	91,218	77,210	40,771	87,243	74,827	75,641	593,024

Notes:

The CPS data was matched using the algorithm similar to Millimet et al (2002) and Peracchi and Welch (1995). Households in rotation 1-4 were matched using the household identifier number to the same household in rotations 5-8 of the following year. Individuals had to have the same sex, race and be a year older in rotation 5-8 to be determined a match.

Comparsion of CPS matched data sets – Millmet et al (2002) to Steward and Gaylor (2015)

Our approach is two fold. First we matched the BLS data cohorts based on the Millimet et al. (2002) and Peracchi and Welch (1995) papers. In a nutshell the CPS matching routine involves matching incoming and outgoing cohorts across a given year. Once the data is matched, we then look at the work status of the individuals to determine if they were active or in active across the year that they were interviewed by the BLS. . We were able to create a match CPS data set of 201,797 individuals where as the Millimet et al. (2002) found 200,916 matched individuals.

Table 1. Comparsion of CPS cohort matched data sets
Year	Millimet et al. (2002)	Steward and Gaylor (2015)
1992/93	37,709	36,652
1994/95	34,418	33,377
1996/97	31,691	32,739
1997/98	32,276	32,972
1998/99	32,083	32,893
1999/2000	32,739	33,164
Total	200,916	201,797

Notes: