This Tableau Public presentation visually communicates the results of my 2025 Masters in Data Science thesis, Regression Analysis of Remote Work Participation. The presentation is designed for a non-technical audience and goes through the research process, including:

• Background and problem
• Research question and hypothesis
• Data source
• Population selection process
• Comparison of target population to full dataset
• Linear regression results and findings
• Limitations
• Suggestions for further research
• Implications for business

Executive Summary 

Since 2019, businesses’ approach to telework has shifted dramatically. Conversations about whether to allow telework (either in a fully remote or hybrid posture) often center on team dynamics, job requirements, competitor policy, or company culture. Less considered is how telework options change the makeup of the labor pool itself. When deciding to allow telework, business leaders may assume a team is made of the same people whether they are at home or in the office. But is that true? The hypothesis of this study is that the availability of telework (including hybrid work) increases key populations’ participation in the workforce. In other words, for these populations, telework policies don’t just impact where they work, they impact how much or even whether they are participating in the workforce at all.  

The American Time Use Survey (ATUS) dataset is collected and maintained by the US Bureau of Labor Statistics (BLS) as part of the US Census (2024). Data from the 2003-2023 atuscps file is used comprising 298,841 observations and 92 variables. Observations are from 2019-2023 surveys. The variables include participant identifiers, demographics, work data, and personal information (Appendix A). Data preparation included the following steps: 

• Importing desired columns from atuscps file and dropping pre-2019 rows 
• Converting continuous variables for use in population selection (such as household size and age) into continuous variables based on distribution visualization 
• Maintaining a ‘people’ dataset with identifier columns that can be used to subset participants (e.g. all 2019 participants or all identified population participants) without having to load the full dataset 
• Renaming telework columns for clarity and creating combined “TELEALL” column 
• Creating calculated employment status field for population selection based on both current employment (full/part-time/unemployed) and desired employment to identify underemployed individuals 
• Creating calculated CORR field showing whether a participant’s employment and telework statuses are the same (counting underemployment with unemployment) 
• One hot encoding and sorting variable dimensions based on CORR using both MCA and groupby to select target populations (Appendix B) 
• Identifying variable dimensions that are >1 standard deviation above the mean (top16) or in the top quartile (top25) based on CORR value counts 
• Creating calculated field RANK to quickly separate top16 and top25 populations 

The research question is answered using an Ordinary Least Squares multilinear regression model (Appendix C). Using a 30/70 train/test split on the full dataset, PEHRUSLT (usual hours worked per week at all jobs; 0-150 continuous) is analyzed against rank, year, and telework variables. The model’s RMSE is 10.4 hours and the mean error of the training model’s predictions against the test set is 5.5 hours. According to this model, RANK is statistically significant in predicting hours worked across the dataset, meaning that the target populations do participate in the workforce at different rates than the whole population. 

Duplicate models run on the top16 and top25 (inclusive) subsets reveal that for the targeted populations, telework is a statistically significant predictor of hours worked, with TELECOVID (Apr 2020-Oct 2022) telework rates the strongest positive indicator. The R squared of the models varied from 0.002-0.008, meaning telework access is not the primary reason workforce participation varies across the dataset or subsets, but it has a larger effect on the target populations (TELECOVID_1 +4.6 hours for top16 populations vs.  +1.6 for full dataset).  

A limitation of this study is the type of telework data available. ATUS is representative of the entire United States population, but telework data as a standalone question has only been included since late 2022. Additional data from questions asked about teleworking during and before the COVID-19 pandemic are used in this study, but do not have as many valid responses. More detailed information about telework arrangements (hybrid vs. fully remote, temporary vs. long-term) would provide a more nuanced understanding of how telework impacts workforce participation. 

Suggestions for further study include: 

• The impact of telework on workforce participation by industry or job type 
• Projected candidate pool for a position if posted as telework vs. in-person
• Projected team makeup if telework is allowed vs. in-person only 

Across the board, positive telework status is associated with increased workforce participation at a statistically significant level. This effect is stronger in the key populations, with the largest effect among those who are disabled, White-Black or White-Asian, have a household >7 people, and married couples where at least one person is currently in the Armed Forces.  

These findings have several key implications for businesses. Expanding telework opportunities allows businesses to tap into a labor market that is otherwise limited or unable to access those positions. Fully remote positions allow companies to hire people who don’t live near their physical work location. Telework is more inclusive of disabled populations, which can increase retention, since anyone can become disabled at any time. A company that is prepared and open to flexible working arrangements is much better positioned to retain talent than one with limitations on how and where people work. Finally, many identified populations are underrepresented in corporations. Companies that are trying to be more representative of the populations they serve or incorporate perspectives that are different from the industry mainstream can find a different, more robust candidate pool with a telework position than they would with an in-person one. 

Appendix A: Variables

variable	type	used for data cleaning/formatting	used for population selection	used for regression	included for supplemental illustration only
TUCASEID	id	y
TULINENO	id	y
HRYEAR4	id	y	y	y
PEAFNOW	cat	y
PEMLR	cat	y	y	y
PEHRWANT	cat	y	y
PESCHENR	cat	y
PEDWWNTO	cat	y	y
TELE2019	cat	y	y	y
TELECOVID	cat	y	y	y
TELENOW	cat	y	y	y
EMPA	calculated	y
EMPB	calculated	y
EMPC	calculated	y
EMPD	calculated	y
EMPSTAT	calculated	y	y
TELEALL	calculated	y	y
CORR	calculated	y	y
GESTFIPS	cat		y
GTMETSTA	cat		y
HEHOUSUT	cat		y
HETENURE	cat		y
HRHTYPE	cat		y
HRNUMHOU_c	cont > cat		y
HUSPNISH	cat		y
PEABSRSN	cat		y
PEAFEVER	cat		y
PECERT1	cat		y
PECERT3	cat		y
PECYC	cat		y
PEDIPGED	cat		y
PEDISDRS	cat		y
PEDISEAR	cat		y
PEDISEYE	cat		y
PEDISOUT	cat		y
PEDISPHY	cat		y
PEDISREM	cat		y
PEERNCOV	cat		y
PEERNHRY	cont		y
PEGRPROF	cat		y
PEHSPNON	cat		y
PEMARITL	cat		y
PEMJOT	cat		y
PESEX	cat		y
PRCITSHP	cat		y
PRDISFLG	cat		y
PRERNHLY_c	cont > cat		y
PRMARSTA	cat		y
PRNMCHLD_c	cont > cat		y
PRTAGE_c	cont > cat		y
PTDTRACE	cat		y
PUAFEVER	cat		y
PUBUS1	cat		y
PEDW4WK	cat				y
PEDWAVL	cat				y
PEDWAVR	cat				y
PEDWLKO	cat				y
PEDWWK	cat				y
PEHRACT1	cat				y
PEHRACT2	cat				y
PEHRACTT	cat				y
PEHRAVL	cat				y
PEHRFTPT	cat				y
PEHRRSN1	cat				y
PEHRRSN2	cat				y
PEHRRSN3	cat				y
PEHRUSL1	cont				y
PEHRUSL2	cont				y
PEHRUSLT	cont			y
PEJHWKO	cat				y
PENLFACT	cat				y
PENLFJH	cat				y
PENLFRET	cat				y
PERET1	cat				y
PRABSREA	cat				y
PREMP	cat				y
PREMPHRS	cat				y
PREMPNOT	cat				y
PRFTLF	cat				y
PRHRUSL	cat				y
PRNLFSCH	cat				y
PRPTHRS	cat				y
PRPTREA	cat				y
PRSJMJ	cat				y
PRWKSCH	cat				y
PRWKSTAT	cat				y
PRWNTJOB	cat				y
PTCOVID1	cat	y	y
PTCOVID2	cat				y
PTCOVR1	cat	y	y
PTCOVR2	cat				y
PTCOVR3	cat	y	y
PTCOVR4	cat				y
PUABSOT	cat				y
PUDIS	cat				y
PUDIS1	cat				y
PUDIS2	cat				y
PURETOT	cat				y
PUWK	cat				y
rank	calculated	y		y

Appendix B: Target Populations

Top16 – Variables >1 standard deviation above mean in CORR=1 rate

Variable_dimension	% CORR=1	Description
PEDISDRS_1	86.03%	disability: difficulty dressing or bathing
PEDISOUT_1	81.29%	disability: difficulty running errands or going out alone
PEDISPHY_1	76.69%	disability: serious difficulty walking or going up stairs
PTDTRACE_6	72.85%	white-Black
PTDTRACE_8	70.22%	white-Asian
PEDISREM_1	68.69%	disability: serious cognitive or memory difficulties
HRHTYPE_2	67.19%	husband/wife primary family in Armed Forces
PRDISFLG_1	64.84%	any disability
HRNUMHOU_c_6	64.73%	household size >9
PEDISEYE_1	64.24%	disability: serious difficulty with vision
HRNUMHOU_c_5	63.38%	household size 7-8

Top25 – in addition to above; variables in the top quartile of CORR=1 rate

Variable_dimension	% CORR=1	Description
HRNUMHOU_c_4	61.49%	household size 5-6
PTDTRACE_7	61.31%	white-Native
HRHTYPE_4	59.41%	female civilian head of household
PRNMCHLD_c_1	58.83%	no children
HEHOUSUT_5	56.48%	mobile home or trailer with no permanent rooms
GESTFIPS_28	55.77%	MS
HEHOUSUT_6	54.73%	mobile home or trailer w/room(s) added
HETENURE_2	54.51%	rent home
PTDTRACE_2	54.38%	Black
PTDTRACE_3	54.13%	Native
GESTFIPS_56	54.10%	WY
GESTFIPS_2	54.00%	AK
HETENURE_3	53.99%	housing not owned and no cash rent paid
PRCITSHP_1	53.93%	US-born citizen
GESTFIPS_21	53.91%	KY
GESTFIPS_23	53.85%	ME
GESTFIPS_22	53.72%	LA
GESTFIPS_19	53.49%	IA
HRHTYPE_3	53.43%	civilian male head of household
GESTFIPS_5	53.29%	AR
PESEX_1	53.26%	male
GTMETSTA_2	53.20%	non-metro area
GESTFIPS_47	53.06%	TN
PEHSPNON_1	52.90%	Hispanic/Latino
GESTFIPS_40	52.83%	OK
GESTFIPS_20	52.78%	KS
HRNUMHOU_c_3	52.72%	household size 3-4

Appendix C: OLS Linear Regression Summaries

Training model on 30% of full dataset:

Top 16 model:

Top 25 model: