Using Paradata to Evaluate Youth Participation in a Digital Diary Study

Abstract

We analyze recruitment, access and longitudinal response paradata from a year-long intensive longitudinal study (mDiary) that used a mobile-optimized web app to administer 25 bi-weekly diaries to youth recruited from a birth cohort study. Analyses investigate which aspects of teen recruitment experiences are associated with enrollment and longitudinal response patterns; whether compliance behavior of teens who require multiple invitations to enroll differs from that of teens who enroll on the first invitation; and what personal and social circumstances are associated with different longitudinal compliance patterns. Latent Class Analysis (LCA) is used to derive longitudinal compliance classes. mDiary’s person-survey response rate of 70 percent is noteworthy considering reports that response rates for smartphone studies trail those administered via telephone or personal computers. Conditional on agreeing to participate, teens with texting capability were over six times as likely to enroll as their peers lacking access, and they also completed 6–7 more diaries. Youth who required multiple prods to register not only were less likely to enroll than their peers who registered at the first invitation, but also tended to attrite early. Compared with teens who completed all 25 surveys, those who attrited early had less access to texting capability, home Internet service, and had low-education mothers. Consistent with studies of adults, nonparticipants were disproportionately black males from socioeconomically disadvantaged backgrounds.

Increasing difficulty recruiting respondents for representative surveys coupled with declining survey response rates have fostered inquiries to understand the reasons for these trends and their implications for data quality and statistical inference (Couper, 2017; Czajka & Beyler, 2016; Groves et al., 2004). Widespread use of call answering and filtering systems that allow prospective respondents to screen telemarketing solicitations undermine both subject recruitment and data collection for representative surveys (Couper, 2000; Czajka & Beyler, 2016). Although the expansion of digital technology and broadened access to the Internet has amplified lowered the cost of data collection (Link et al., 2014), growing reliance on cell phones poses new challenges for survey researchers because increasing difficulty recruiting subjects increases risks of incurring both coverage and sampling error (Couper, 2017; de Leeuw, 2018; Jäckle et al., 2019).

Even as cross-sectional survey response rates continue a downward spiral, wave-to-wave longitudinal response rates appear to have remained steady, partly buoyed by use of paid incentives (Singer & Ye, 2013; Singer et al., 2000), identification with the study, and topic salience (Groves et al., 2004; Schoeni et al., 2013). For adults there is evidence that pre-paid financial incentives boost survey response rates across modes of administration (Singer et al., 1999; 2000), and there is some evidence of carry-over persistence across repeated surveys, particularly when topic salience is high (Schoeni et al., 2013; Laurie & Lynn, 2009). The rapid expansion of broadband Internet access has greatly expanded opportunities for administration of web surveys, but also introduced new challenges. In particular, differential access to Internet services, cellular data plans, and digital proficiency pose barriers for some socioeconomic and demographic groups ( Jäckle et al., 2019; Couper, 2017; Link et al., 2014; Buskirk & Andres, 2012).

Although teens are prolific users of digital technology (Anderson & Jiang, 2018; Rideout, 2015), evidence about their participation in panel studies is both limited and inconsistent, particularly for designs that involve frequent measurements over several months (Halpern et al., 1994; Boys et al., 2003; Powers & Loxton, 2010; Goldberg et al., 2019). Rideout (2015) reports that two-thirds of adolescents owned a smartphone, with ownership rates ranging from 65 percent for teens from families with low annual incomes (<$35,000) to 94 percent among youth from high income families ($100,000+). Because proficiency and access to digital technology differ among youth of varied socioeconomic backgrounds (Anderson & Jiang, 2018; Rideout, 2015; Lenhart, 2015; Hargittai, 2010), there is much to learn about variations in longitudinal survey compliance behavior in web-administered surveys (Lee et al., 2019; Jäckle et al., 2019; Couper, 2017). This is especially so for youth whose digital media habits are conducive to high survey engagement on the one hand, yet less frequently studied than the survey compliance behavior of adults on the other hand. Notable exceptions are college students (Lee et al., 2019) and youth who have reached the age of majority (Barber et al., 2016).

Furthermore, except for studies using ecological momentary assessments that take repeated daily measurements over several days (Wen, 2017; Hensel et al., 2012), youth are seldom the subjects of high frequency data collection for studies with time horizons spanning several months (e.g., Jaccard et al., 2004; Bergdall et al., 2012). A notable exception is the Relationship Dynamics and Social Life (RDSL) study, which administered a multi-year weekly diary study using both electronic and telephonic methods to track young women in order to investigate social experiences resulting in unintended pregnancy (Barber et al., 2016; Wagner et al., 2019). RDSL was restricted to women ages 18 and 19, which obviated the need for parental permission to participate in the study. To our knowledge, no study has administered a year-long diary (intensive longitudinal survey) to minors (e.g., youth ages 17 and under). Neither have researchers exploited paradata to understand variations in adolescents’ response behavior in studies that collect high frequency data.

Various types of paradata—process data compiled through subject recruitment and respondents’ navigation of survey instruments and web portals—have advanced understanding of response behavior in panel studies of adults (Kocar & Biddle, 2019; McClain et al., 2019; Bristle et al., 2019; Callegaro, 2013). In addition to recruitment indicators, such as modes and number of contacts to obtain assent that are customarily generated through face-to-face and computer-assisted telephone interviews (CATI), digitally-administered surveys produce access paradata that can advance understanding of compliance in web-administered surveys. Examples include types of devices used to access web portals, email providers, log-in attempts, and difficulty navigating web portals, among others (McClain et al., 2019; Kocar & Biddle, 2019; Callegaro, 2013). One important limitation of self-administered surveys compared with CATI and in-person interviews is the inability to obtain paradata based on interviewers’ observations, such as level of cooperation, that potentially bear on data quality.

Building on the classification schemes offered by Callegaro (2013) and expanded by McClain et al. (2019), we analyze data from the mDiary Study of Adolescent Relationships to investigate which aspects of subject recruitment facilitate enrollment in the study, and conditional on registering, which aspects of recruitment, digital access and respondent attributes are associated with persistence in a year-long, mobile-optimized web-administered diary study. Three broad questions guide the investigation. First, conditional on assenting to participate, which aspects of recruitment and digital access are associated with enrollment in the study? Second, do variations in recruitment experiences and digital access carry over to longitudinal response behavior? In particular, do teens who enroll only after receiving multiple invitations exhibit different longitudinal compliance behavior compared with teens who register after a single invitation? Finally, what respondent attributes and family circumstances are associated with enrollment and longitudinal compliance?

Background: Paradata and Response Behavior

Digital technologies are appealing for administering intensive longitudinal studies to adolescents both because they facilitate convenience in taking surveys and are well aligned with youth media habits (Anderson, 2015). For researchers, web-administered surveys not only permit timely data retrieval (Link et al., 2014; Raento et al., 2009), but also capture information about the data collection process (Callegaro, 2013; McClain et al., 2019). Several recent studies have furthered understanding about recruitment experiences and modes of Internet access among adults (Jäckle et al., 2019; de Leeuw, 2018), but inferences are not directly transferable to youth because their recruitment faces higher participation hurdles. This is because adult (usually parent) consent is necessary to request teen assent, and because of the stringent Institutional Review Board recruitment guidelines for studies involving youth. ⁱ

Barber and collaborators (Barber et al., 2016; Barber et al., 2011) avoided the stringent human subjects requirement for youth by targeting young women ages 18–19 years of age to field a 36-month web-based diary of sexual behavior to learn about unintended pregnancy. Their study used mixed modes for both subject recruitment and data collection (de Leeuw, 2018), but was not optimized for administration via mobile devices. Following a 60-minute face-to-face enrollment survey that collected extensive background information, respondents were invited to complete weekly journals (diaries) over 36 months either using CATI or on the web (desktop or laptop required). Over three-quarters of enrollees (78 percent) participated for at least 18 months, and 63 percent completed the final journal, albeit with intermittent nonresponse. Over the 36-month study period, 45 percent of all person-surveys were completed: the longitudinal response rate was achieved using cash bonuses, multiple reminders, and “refusal-conversion packets” to re-engage recalcitrant respondents. In addition to confirming established results that minority and lower economic status respondents completed fewer journals and had higher attrition rates, they report that contact information provided at the enrollment interview was associated with compliance behavior. Specifically, they find that respondents who provided both a phone number and email address had higher persistence rates, which they speculate indicates differences in commitment to participate at the outset. That pregnancy-related experiences were positively associated persistence and on-time participation suggests that topic salience boosted panel persistence.

Most intensive longitudinal studies involving youth under age 18 use signal-triggered methods that capture multiple daily measurements over short durations (Hensel et al., 2012; Runyan et al., 2013), but there is inconsistent reporting of compliance outcomes. Based on a meta-analysis of 42 studies that administered EMA protocols over 2 to 42 days to adolescents ages 18 and under, Wen et al. (2017) report average weighted compliance rates between 73 and 78 percent; however, they find no differences in response rates between studies that used only a mobile-EMA platform and those that used a mobile platform in addition to wearable devices. Rather, they find that sampling intensity, an indicator of respondent burden, is associated with compliance behavior. Although they stop short of an explicit call for examining paradata to better understand youth compliance in intensive data collection protocols, Wen and associates (2017) highlight the importance of considering whether nonresponse was due to technical issues, such as whether delivered prompts were actually received, or respondent recalcitrance. They also underscore the importance of considering social and demographic correlates of longitudinal compliance because of differential access to digital technology.

Irrespective of target population, survey administration mode or panel duration, there is evidence that noncompliance and attrition risks accumulate over waves in both CATI and web-administered surveys (Bristle et al., 2019; Wagner et al., 2019; Lugtig, 2014; Coyne, et al., 2017; Boys et al., 2003). The mechanisms generating nonresponse and attrition in panel studies change over the span of the study for many reasons, including extraordinary personal events (e.g., job losses, medical emergencies, and relocation disruptions), response fatigue, and topic salience, among others (Kocar & Biddle, 2019; Barber et al., 2016; Lugtig, 2014; Groves et al., 2004). Compared with EMA studies, however, information about adolescents’ compliance behavior in web-administered intensive longitudinal studies that span several months is scarce (Goldberg et al., 2019).

Correlates of nonresponse in web-administered surveys differ somewhat from face-to-face and CATI surveys because of variations in Internet proficiency and digital access according to age and socioeconomic status (de Leeuw, 2018; Jäckle et al., 2019; Hargittai, 2010); because automated notifications may go to spam folders (Couper, 2000; 2017; Callegaro, 2013); because of connectivity failures (Buskirk & Andres, 2012; Callegaro, 2010); or because respondents change service providers, email addresses, or cell phone numbers, among other reasons (Wen et al., 2017; McClain et al., 2019; Callegaro, 2013). Web-administered surveys that require access to a desktop or laptop also may dampen longitudinal compliance, particularly among groups that prefer mobile devices for social transactions (Turner et al., 1998; Barber et al., 2011; Link et al., 2014). For youth who interact with technology from very young ages, dubbed digital natives by Prensky (2001), smartphones have become a primary medium for entertainment and communication, including formation and maintenance of romantic relationships (Anderson, 2015; Anderson & Jiang, 2018; Goldberg & Tienda, 2017). This suggests that access to mobile devices will increase teens’ willingness to participate in a diary study about romantic partnerships, and potentially boost their longitudinal compliance.

To address whether, how much, and in what ways adolescents participate in a mobile-optimized diary study about romantic relationships, we draw upon insights about topic salience (Barber et al., 2016; Schoeni et al., 2013; Groves et al., 2004), established findings about the social and economic correlates of survey participation (Couper, 2017; Lugtig, 2014; Wen et al., 2017; Groves et al., 2001); and recent insights about the power of paradata to understand response behavior in adults (Bristle et al., 2019; McClain et al., 2019; Kocar & Biddle, 2019; Callegaro, 2013). Specifically, we hypothesize that there is a tradeoff between the amount of effort expended recruiting subjects and both the likelihood of enrolling in the diary study and longitudinal persistence. Further, we expect that access to mobile devices, home Internet services, and digital proficiency will boost longitudinal compliance. Finally, guided by prior research, we expect that response behavior will differ by teens’ demographic and social characteristics, with higher longitudinal compliance among girls, youth from more advantaged families, youth who live in homes equipped with Internet services, and youth who are tenacious in completing assigned tasks (Goldberg et al., 2019; Ryan & Lewis, 2017; Barber et al., 2016; Lugtig et al., 2014; Watson & Wooden, 2009).

Sample and Web Survey Administration

The empirical analyses draw on three sources of data: (1) response data collected by the Fragile Families and Child Wellbeing Study (FFCWS), a prospective birth-cohort study that followed almost 5,000 children from birth through approximately age 15; (2) response data from the mDiary Study of Adolescent Relationships (mDiary), which administered an intensive longitudinal survey over 52 weeks to a sample of the FFCWS youth aged 16 to 17 years old; and (3) paradata generated during the mDiary study recruitment, registration, and data collection process. The FFCWS birth cohort study followed a cohort of children born at the turn of the millennium in 20 medium-to-large U.S. cities (Reichman et al., 2001). By design, births to unmarried mothers were oversampled at baseline.

Primary caregivers were surveyed over six waves, most recently when their youth were approximately 15 years of age; the target youth were interviewed at ages 9 and 15. The FFCWS surveys provide rich information about target youths’ socioeconomic background, living arrangements, school behavior, and numerous age-specific measures of socio-emotional development. Our analyses use information from the baseline survey (FFCWS baseline) and Year 15 surveys conducted with primary caregivers (FFCWS Y15-parent) and target youth (FFCWS Y15-teen), as well as the mDiary enrollment survey and the 25 bi-weekly diaries.ⁱⁱ In addition to rich paradata about the process of subject recruitment and engagement, the mDiary study provides two additional advantages for understanding adolescents’ response behavior. These include a known sampling frame (Couper, 2017; 2000; Lee et al., 2019) and rich respondent and family background data collected over a 15-year period. The former is important for statistical inference and assessing coverage error (Couper, 2000) and the latter obviates the need to collect extensive background information about respondents, which shortens the length of the enrollment survey and mitigates response fatigue. Because the mDiary sample was drawn from an ongoing birth cohort study, there was limited age variation among respondents, whose median age at enrollment in the study was 16.7 years (mean = 16.7, s.d. = 0.358).

Recruitment for the mDiary study, which occurred over a 17-month period (November 2015 – April 2017) on a rolling basis, tracked the field operations of the FFCWS parent study by approximately one year. The only requirement to participate in the study was access to a private email address. Recruitment involved a mixed-mode strategy (de Leeuw, 2018; Couper, 2017) that began by sending welcome packages to respondents with valid contact information, followed by phone calls to obtain parental consent and teen assent. The welcome packages included information about the study, an unconditional $5 incentive for the primary caregiver, and a stamped envelope to return the signed consent and assent forms. Current email addresses and cell phone numbers were requested on the consent and assent forms, but many returned forms were incomplete, hence required phone contact.

Because only 23 percent of teens granted assent by returning the signed paper forms, numerous follow-up calls were required to obtain assent. Nevertheless, teen assent did not assure participation, as nearly one-quarter of teens who agreed to participate in the mDiary study failed to register on the website. Of 869 respondents with valid contact information, ultimately 689 (80%) assented to participate in the study, which is high by comparison to recruitment for web administered studies of adults and university students (e.g., Jäckle et al., 2019; Lee et al., 2019; Wen et al., 2017; Callegaro & DiSogra, 2009). Enrollment instructions, which invited teens to select a user name and password and to complete an enrollment survey, were provided via email; however, respondents that provided a cell phone number during the recruitment process also received a text message alerting them to view the instructions via email. Conditional on assent, over three-fourths of teens (531) registered for the study by completing an enrollment survey. All diaries opened on Sundays at 4 PM local time and remained open for one week. Diaries not completed by the end of the week-long response window were considered skipped.

mDiary did not use mixed-modes for survey administration (Lee et al., 2019; de Leeuw, 2018; Couper, 2017); rather, all surveys were administered via a custom, mobile-optimized website (mdiary.org) linked to the Qualtrics web survey platform via API calls. Respondents were permitted to complete the diaries on a device of choice—desktops, laptops, tablets or smartphones—provided mobile devices were not shared. mDiary was aligned with the design heuristics for effective smartphone surveys recommended by Antoun, et al. (2018, Table 3) and Callegaro (2010). The enrollment survey included several questions that replicated items in the parent study in order to gauge change during the intervening year, but these benchmark items were not asked in the diaries. An important example is about dating behavior: both the Year-15 teen survey and the mDiary enrollment survey asked whether respondents had ever dated.

Table 3.

Odds of Delayed Enrollment or Nonenrollment (vs.On-time Enrollment)

	Baseline						Full Model
	Delayed			Nonenrollment			Delayed			Nonenrollment
	Odds Ratio	S.E.		Odds Ratio	S.E.		Odds Ratio	S.E.		Odds Ratio	S.E
Recruitment
# Contact attempts	1.028	0.025		1.040	0.015	**	1.021	0.026		1.034	0.016	*
Assent/enrollment invite lag	1.032	0.019		1.024	0.013	*	1.035	0.020		1.026	0.013
Summer invite	0.521	0.283		0.819	0.232		0.567	0.316		0.867	0.261
Access
Uses school email	1.677	0.957		1.134	0.465		2.012	1.190		1.436	0.618
Receives text messages	0.486	0.317		0.144	0.045	***	0.438	0.293		0.133	0.046	***
Contacts recruitment team	4.092	1.668	***	0.792	0.345		3.851	1.698	**	0.698	0.320
Respondent attributes
Female (b)							0.525	0.172	*	0.433	0.090	**
Ever dated age 15(c)							0.498	0.151	*	0.685	0.148
Home Environment
Lives with both bio parents (d)							1.143	0.145		1.140	0.090
Has home internet service (d)							2.404	1.890		1.003	0.359
Home often chaotic (c)							1.470	0.785		0.635	0.271
Often spends time alone (c)							1.259	0.511		1.067	0.312
Mother’s Characteristics (b)
Race (nonHispanic white)							(ref)			(ref)
Black							2.192	1.000		1.739	0.477	*
Hispanic							1.336	0.682		0.687	0.216
Other							1.712	1.425		1.162	0.623
Education (< HS)							(ref)			(ref)
HS or equivalent							1.222	0.549		0.711	0.186
Some college							1.086	0.506		0.501	0.141	*
College graduate							1.134	0.747		0.363	0.154	*
Mother married at teen’s birth							1.290	0.557	1.180	0.325
R-sq	0.064						0.126
Constant	0.095			1.111			0.053	0.067	2.674	1.807

Sources: (a) mDiary paradata; (b) FF baseline; (c) FF Y15-teen; (d) FF Y15-Parent

Notes: follow-up contact attempts post welcome package

^***p<0.001

^**p<0.01

^*p <0.05.

To incentivize compliance, respondents received Amazon e-gift cards, disbursed via email or text, per indicated preference (Singer et al., 2000; Singer & Ye, 2013). Teens received a $5 gift card for completing the enrollment survey, $2 for each completed diary; and a bonus gift card of $10 for competing the last diary. The Amazon gift cards were delivered upon completing three ($6) or four ($8) consecutive diaries. Teens who registered for the study were randomly assigned to the 3 and 4 incentive groups, but there were no statistical differences in completion rates between the two groups.

Data and Analytic Strategy

1. mDiary Paradata

The process of recruiting and enrolling respondents for the mDiary study generated rich paradata (McClain et al., 2019; Durrant & Kreuter, 2013; Callegaro, 2013). Building on classifications proposed by Callegaro (2013) and expanded by McClain et al. (2019), we examine recruitment, access, participation and longitudinal response paradata. Table 1 defines the core measures and provides summary statistics for teens who did and did not register for the study after having agreed to participate.

Table 1.

Paradata Classification and Operational Definitions Means (s.d.) or Percentages (%)

			Assented	Enrolled	Not Enrolled
Paradata Types & Indicators	Operational Definition	Measures	N=689	N=531	N=158
Recruitment
Contact attempts	# contacts post welcome mailing to assenta	x̄/(s.d.)	6.1	5.8	7.2	*
			(6.2)	(6.0)	(6.8)
Assent-invite time lag	days between assent & 1st enrollment invite	x̄/(s.d.)	12.4	12.1	13.4
			(9.3)	(9.0)	(10.4)
Invitation season	first enrollment invite in summer	% summer	14.4	14.9	12.7
Access
Communication mode	receives text & email vs email only	% text + email	92.2	96.1	79.1	***
Email provider	school vs commercial	% school	5.8	5.8	5.7
Authentication difficulty	needed help enrolling	% contact team	7.0	7.7	4.4
Enrollment device type	smartphone vs other	% smart phone	NA	73.1	NA
Participation
Enrollment	enrollment survey completed	if enrolled	77.1	100	0
Enrollment timing	enroll after multiple invites vs. first invite	% delay	NA	8.9	NA
Longitudinal Response
Study compliance	# diaries completed	x̄	NA	17.6	NA
		(s.d.)		(9.3)
Study compliance	# of last completed diary	x̄	NA	19.4	NA
		(s.d.)		(9.1)
Study compliance	# of skipped diary spells	x̄	NA	1.0	NA
		(s.d.)		(1.2)
Study compliance	longest spell of consecutive skipped diaries	x̄	NA	6.7	NA
		(s.d.)		(9.0)

Source: mDiary paradata Note:

Notes:

^***p<0.001

^**p<0.01

^*p <0.05.

^aZero contact attempts indicates that family provided consent and assent via mailed welcome invitation

Recruitment paradata includes number of phone contacts to obtain assent following a mailed welcome package; the season of recruitment (summer versus other seasons); and both the date of assent and the date teens were invited to register on the mDiary web portal. In general, nonenrollees were harder to reach than their enrolled peers. Less than five percent of subjects granted both consent and assent via the welcome package materials sent via the U.S. postal service (or via email) and required no additional phone contact either to obtain assent or to furnish missing information. An additional 18 percent provided parental consent by return mail, but also required follow-up calls to collect missing information, including teen assent.

For the vast majority of recruited teens both consent and assent were obtained telephonically, but seldom simultaneously, which meant multiple calls to the Primary Caregiver’s home. Partly because of complex family and teen schedules (sports, school, work, social activities), multiple calls—averaging 6.1 calls per subject—were required to obtain teen assent. That enrolled teens averaged 5.8 calls to grant assent compared with 7.2 calls for their nonenrolled peers (p<.05) may indicate lack of interest (Bristle et al. 2019) or limited access to mobile technology (Anderson, 2015; Callegaro, 2010; Hargittai, 2010).

By design, assented teens were invited to register for the study within two weeks of assenting to participate, but owing to holiday schedules, extracurricular summer programs, and the need to collect updated email and/or phone contact information, some teens experienced slightly longer lags between the dates of assent and first invitation to register. Using the date assent was obtained, we calculate the time elapsed (in days) until the first invitation to register (assent-invite time lag) to assess whether longer spells lowered the odds of enrolling. The assent-invite time lag did not differ significantly by enrollment status. Because teens’ time commitments vary in accordance with academic calendars, we also recorded whether the first invitation to register for the study occurred during summer, when both time commitments and peer interaction patterns change.

Access paradata used to gauge respondents’ digital access and proficiency include teens’ notification preferences (text vs. email); whether teens use a school email address (versus commercial providers); whether teens’ contacted the survey team for assistance to access the website or to authenticate their login credentials; and conditional on enrolling, the type of device (smartphone versus other device) used to register. Access to an email account was required to participate because enrollment instructions were delivered to electronic addresses. The vast majority used commercial providers, but a small share relied on schools to access the Internet, which may restrict their ability to complete diaries during the summer. Although young people seldom check their email, most do so when required. Rather, adolescents prefer to communicate via text (Anderson, 2015; Coyne et al., 2017). Virtually all teens who reported access to text messaging registered for the study, but only 79 percent of assented teens who did not enroll had text messaging capability (p<.01). Ability to receive texts facilitated sending reminders and gift cards in accordance with teens’ preferred communication mode.

As a web-administered panel study, mDiary produced two additional classes of online paradata, namely participation and longitudinal paradata (Callegaro, 2013). mDiary paradata capture both whether teens registered on the web portal and whether enrollment occurred following a single or multiple invitations to do so. Although the vast majority of mDiary participants enrolled within a week of receiving their first invitation (on-time enrollees), about 9 percent required multiple invitations (delayed enrollees) (see Figure 1). Teens who failed to register within the enrollment window were added to the next cohort and re-invited to enroll up to five times before they were designated nonenrollees. Longitudinal paradata include both the total number and the last diary completed as well as patterns of skipped surveys and spells of nonresponse.

Figure 1.

Subject Recruitment and Compliance Behavior: mDiary Study

Following Callegaro (2013) we use total number of diaries completed (range = 25) and the last diary completed (range = 25) to portray longitudinal compliance. In addition to capturing diary-specific nonresponse, we also record the number of skipped diary spells and the length of the longest spell of consecutive skipped diaries. The lower panel of Table 1, which provides summary statistics for these longitudinal paradata measures, reveals that conditional on registering for the study, teens completed 18 diaries, on average, but the standard deviations reveal considerable variation resulting both from diary-specific nonresponse and attrition. Respondents averaged one spell of skipped diaries, but the length of the nonresponse spells averaged 6 diaries, with considerable variation owing to differences in the timing of attrition and patterns of skipped diaries.

2. Sample Characteristics

Measures about respondent and family circumstances, which are drawn from the FFCWS baseline survey, are informed by a vast literature about survey response behavior and attrition (Groves et al., 2001; Watson & Wooden, 2009; Barber et al., 2016; Kocar & Biddle, 2019). These include mother’s self-reported racial identification, educational attainment, and marital status at the birth of the teen. Parents reported on the availability of home Internet service at the Year-15 interview; teens reported on several aspects of their lives, including living arrangements (with both biological parents); the amount of time spent alone (often, sometimes, rarely or never); the level of chaos in their home; and their self-reported perseverance on tasks.

Table 2 summarizes characteristics for assented teens according to enrollment status. Several noteworthy differences between enrollees and non-enrollees include higher enrollment rates for girls as well as for teens with better-educated parents; those who reside with both biological parents; and those who reside in Internet-connected homes. These differentials are consistent with prior literature about survey response behavior (Schoeni et al., 2013; Watson & Wooden, 2009; Groves et al., 2001). Ethno-racial differences in enrollment rates reveal a mixed pattern. Among assented whites and Hispanics, the enrolled shares exceed the nonenrolled, but among blacks the obverse holds. Blacks comprise 40 percent of the sample, but over half of nonenrollees. The gender, racial and education differentials in participation are generally consistent with those observed for adult samples, with the noteworthy exception of Hispanics (Barber et al., 2016; Watson & Wooden, 2009; Groves et al., 2001).

Table 2.

Characteristics of Assented Teens by Enrollment Status (Percentages )

	Total (N=689)	Enrolled (N=531)	Not Enrolled (N=158)
Respondent/Parent Characteristics (a)
Female	51.4	55.2	38.6	***
Mother’s race
White, non-Hispanic	29.3	31.6	21.5	*
Black	39.9	35.4	55.1	***
Hispanic	26.6	28.6	19.6	*
Other	4.2	4.3	3.8
Mother’s education
Less than HS	26.1	23.5	34.8	**
HS or equivalent	28.9	27.9	32.3
Some college	29.8	31.5	24.1
College graduate	15.2	17.1	8.9	*
Mother married at teen’s birth	32.1	34.1	25.3	*
Topic Salience
Ever dated - Age 15 (b)	71.0	68.7	78.5	*
Ever dated - Age 16 (c)	NA	73.3	NA
Teen Persistence (b)
Sticks with plan to get something done	41.2	39.7	46.2
Finishes whatever begins	43.5	41.4	50.6	*
Home Environment (b)
Lives with both bio parents	35.7	39.0	24.7	***
Home often chaotic	6.8	7.2	5.7
Often spends time alone	14.5	14.7	13.9
Has home internet service (d)	92.4	93.8	88.0	*

Sources: (a) FF Baseline survey; (b) FF Y15-teen survey; (c) mDiary enrollment survey; (d) FF Y15-parent survey

Notes:

^***p<0.001

^**p<0.01

^*p <0.05.

Convenient access to the Internet facilitates compliance in web-administered studies (Couper, 2017; Link et al., 2014; Barber et al., 2011). Nationally over 85 percent of U.S. households subscribe to Internet services, and approximately 60 percent both subscribed to broadband services and owned multiple devices to access the Web (Ryan & Lewis, 2017). In alignment with these patterns, over 90 percent of assented teens had Internet service at home, with enrolled teens displaying a modest advantage over their non-enrolled counterparts (94 percent vs. 88 percent, respectively; p<0.05).

Prior research indicates that both respondent personality traits and interest in the survey subject can boost participation rates (Barber et al., 2016; Schoeni et al., 2013; Groves et al., 2004). mDiary was designed to track the emergence and evolution of adolescent romantic relationships; therefore, we measure topic salience as a binary indicator designating whether respondents had ever dated by age 15 or 16, as reported in the FFCWS Year 15-teen survey and in the mDiary enrollment survey, respectively. The FFCWS Year-15 measure is used to predict enrollment in the diary study, and the mDiary measure is used to predict longitudinal compliance. Over two-thirds of assented teens reported having dated at the Year-15 interview, with a 10-point difference between enrolled and non-enrolled teens that, surprisingly, was higher among the nonenrolled. The dating rate of mDiary enrollees inched up almost 5 percentage points between ages 15 and 16. Finally, we use two indicators from the FFCWS Year-15 teen survey to measure variations in persistence: (1) sticks with plans to get things done; (2) finishes whatever begins. These items were based on a 4-point scale indicating level of agreement: strongly agree, somewhat agree, somewhat disagree, and strongly disagree.

3. Analytical Strategy

The empirical analysis proceeds in two stages. To address whether there is a tradeoff between subject recruitment and participation, we estimate a multinomial logistic regression that reveals whether recruitment and access paradata predict the propensity of assented teens to register for the study. We sequentially model teen and family background characteristics associated with participation to evaluate the robustness of the baseline results (Watson & Wooden, 2009; Barber et al., 2016; Groves et al., 2001). Following Lugtig (2014), we use Latent Class Analysis (LCA) to derive a classification scheme that captures variation in longitudinal compliance patterns, and subsequently evaluate associations among latent response classes with recruitment and access paradata. Logistic regressions are conducted with Stata version 16.0MP using mlogit commands and the LCA analysis uses the gsem commands to generate class probabilities; a 0.5 cut-point is used to assign participants toclasses.

Results

1. Enrollment

We model the enrollment decision using a tripartite status outcome that distinguishes among assented teens who enrolled upon receiving the first invitation to register for the study (reference group); those who required multiple invitations to register (delayed); and those who, after multiple invitations, did not enroll (nonenrollment). The odds ratios for the baseline model, which includes the recruitment and access paradata, indicate that both the number of contacts to obtain assent and the days elapsed between the date of assent and the first invitation to enroll raised the odds of nonenrollment relative to enrollment.ⁱⁱⁱ Email provider (commercial vs. school was unrelated to enrollment odds; however, lacking texting capability significantly undermined participation. By contrast, delayed enrollment was largely driven by digital proficiency rather than recruitment experiences. Teens who contacted the research team for assistance registering on the web portal were four times as likely to enroll with delay than their peers who did not require assistance.

The associations between enrollment status and both the recruitment and access paradata indicators were only slightly attenuated after modeling teens’ demographic and social characteristics and family circumstances. Except for the days elapsed between assent and the first invitation to enroll, the odds ratios remain statistically significant, but the magnitude is slightly attenuated. None of the home environment circumstances influenced whether or not assented teens registered for the study. That access to the Web did not boost enrollment odds most likely reflects the widespread penetration of home Internet services and the proliferation of cellular devices (Ryan & Lewis, 2017; Couper, 2017; Lee et al., 2019). The adjusted R-squared values indicate that the recruitment and access paradata explain about 6 percent of the variation in enrollment status, which is comparable to the additional explanatory power resulting from adolescents’ demographic and family socioeconomic circumstances.

Girls were about half as likely as boys to enroll with delay and .43 times as likely to forgo enrollment altogether. There is mixed support regarding the importance of topic salience, proxied by prior dating experience, for participation. Teens who ever dated by age 15 were less likely to delay enrollment relative to their peers who never dated, but prior dating experience was unrelated to the odds of nonenrollment relative to enrollment. Compared with white teens, black youth were, respectively, 1.8 times as likely as Whites to forgo enrollment altogether, but were equally likely to enroll with delay. Socioeconomic status, proxied by mother’s educational status, was associated with whether teens registered for mDiary, but not the timing of enrollment. Youth whose mothers attended or completed college had lower odds of non-enrollment compared with teens whose mothers lacked high school diplomas, but mothers’ education was inconsequential for the timing of enrollment.

Overall, the results suggest that there is a tradeoff between the amount of effort spent assenting recalcitrant teens and their likelihood of registering for the study. Further, conditional on assenting to participate, access to texting capability significantly raises the odds of enrollment. But whether recruitment experiences and technology access carry over to longitudinal compliance is an empirical question that we address next.

2. Longitudinal Compliance

Collectively, enrollees responded to 70 percent of the 13,275 possible diaries (531 * 25), which is considerably higher than most web surveys (Jäckle et al., 2019; Lee et al., 2019), but comparable to many signal-triggered surveys that span shorter durations (Wen et al., 2017) and similar to the longitudinal compliance reported in the RSDL diary study (Barber et al., 2011). Nevertheless, mDiary teens exhibited considerable heterogeneity in their longitudinal response patterns. A handful of respondents (4.9 percent) completed the enrollment survey but not a single diary; a large plurality (44 percent) completed all 25 diaries; and the remainder (n=271; 51 percent) completed between 1 and 24 diaries, with varied response behavior over the study period.

Following Lugtig (2014), we modeled the longitudinal response data using latent class analysis, which sorts respondents according to the resemblance of their longitudinal compliance patterns. We first coded each diary using a binary scheme, 1 for completed (or partially completed) diaries and 0 otherwise (only 26 or 0.2 percent of all diaries were partially completed. Possibly because of the large number of respondents that completed all 25 diaries, the model failed to converge, which precluded predicting class membership for each respondent.^iv Instead, we generated probabilities for class membership using the four longitudinal compliance measures reported in Table 1 on the 271 respondents who skipped one or more diaries. The BIC statistics for the three- and four-class solutions, 5380.97 and 5224.76, respectively, indicate the latter solution better fits the data and, unlike the three-class solution, also generates clearly interpretable classes. In both solutions, the first and last class are similar—early attritors and highly engaged, but four-class solution provides more nuance to the pattern of nonresponse.^v

Table 4 summarizes the latent classes along with the groups that completed all 25 diaries (Loyalists) and who registered for the study but did not complete a single diary (No Shows). The four latent classes differ both in the number of diaries completed and their response patterns. The Highly Engaged are most similar to the Loyalists in their compliance behavior: these teens skipped very few diaries (approximately 2) and average less than two short nonresponse spells. Only 7 percent of the Highly Engaged teens enrolled with delay and the vast majority (90 percent) completed the last survey.

Table 4.

Longitudinal Compliance Response Classes Means (s.d.) or Percentages (%)

	25 diaries	LCA classes				0 diaries
	Loyalists	Highly Engaged	Churners	Occasionals	Early Attritors	No Shows
# Completed diaries	25	22.6	15.6	8.7	3.4	0
		(1.59)	(2.09)	(3.23)	(1.97)
Last completed diary	25	24.8	22.2	16.6	3.9	0
		(0.78)	(3.34)	(5.96)	(2.14)
# Skipped diary spells	0	1.7	3.2	2.1	1.3	1
		(0.82)	(1.37)	(1.03)	(0.59)
Longest spell skipped diaries	0	1.7	6.0	13.5	21.1	25
		(1.17)	(2.19)	(3.18)	(2.14)
% Enrolled with delay	3.0	7.2	20.0	12.5	16.9	19.2
% Completed last diary	100	89.7	51.4	26.8	0	0
N	234	97	35	56	83	26
Sample %	44.1	18.3	6.7	10.6	15.6	4.9

Sources: mDiary enrollment and diary surveys

Note: There were 25 diary surveys and one enrollment survey

At the opposite extreme are the No Shows and Early Attritors. Comprising 5 percent of all enrollees, No Shows registered for the study by completing the enrollment survey, but disengaged before answering a single diary. That nearly 20 percent of No Shows enrolled with delay—that is after multiple invitations to do so—suggests low interest in participating from the outset and calls into question the value of pursuing recalcitrant subjects that are disinclined to persist. Early Attritors averaged four diaries before disengaging, with few skipped diaries, as revealed by the long spells of skipped diaries. Roughly one in six Early Attritors enrolled with delay, signaling qualified interest in the study.

Response behavior of Occasionals and Churners is intermediate to that of the Highly Engaged and Early Attritors. Both groups experienced multiple spells of nonresponse, except that nonresponse spells were over twice as long for Occasionals as compared with Churners—13.5 vs. 6 diaries, respectively. Churners averaged about 16 completed diaries, compared with about 9 for Occasionals. Nearly one in five Churners enrolled with delay, signaling qualified commitment to participating, low digital proficiency, or both; however, only 12 percent of Occasionals required multiple invitations to enroll in the study. That half of Churners and over one-quarter of Occasionals completed the last diary reveals opportunistic response behavior because prior to the release of the final survey, participants were reminded about the $10 bonus incentive for completing the last diary.

To describe how recruitment and access carry over to longitudinal compliance patterns, following Lugtig (2014), we estimated a multinomial logit analysis that predicts respondent classification into the six latent classes. Both because they completed all diaries and are the modal response class (comprising 44 percent of enrollees), Loyalists serve as the baseline for all contrasts. Only statistically significant contrasts (p ≤ 0.05) are discussed.

First, none of the recruitment paradata indicators associated with enrollment timing and nonparticipation were associated with longitudinal compliance behavior. Partly this reflects variations in access, notably whether teens rely on schools for their email service and whether they have access to text-enabled devices, including smartphones. Variation in digital proficiency, indexed by whether teens required assistance to register for the study, was associated with whether or not youth registered for the study (Table 3), but not their longitudinal compliance. Substantively this indicates that the proficiency-linked constraint on enrollment affects response behavior by selecting out less digitally proficient teens.

Second, mode of web access is systematically associated with longitudinal response behavior, but differently across the response classes. For Occasionals, reliance on school email to receive notifications about the next available diary undermined longitudinal compliance. Compared to Loyalists, teens who relied on a school rather than commercial email provider were 6 times as likely to be classified as Occasionals. Inability to receive text messages is a key driver of early attrition; specifically, teens with email and texting capability were only .25 times as likely as Loyalists to be classified as Early Attritors. Furthermore, use of a smartphone to enroll also boosted longitudinal compliance. Over three-fourths of Loyalists registered for the study using a smartphone, compared with 62 percent of the Highly Engaged. This difference in access significantly lowered their odds of completing all diaries.

Consistent with studies of adults, females exhibit higher response rates than their male counterparts, but gender differences were particularly salient for No Shows. In the mDiary sample, girls represented 60 percent of Loyalists, but only 48 and 38 percent, respectively, of Early Attritors and No Shows. Partly because of minority group enrollment favoring whites (Table 3), racial differences in longitudinal compliance are not statistically significant with one exception. Race and socioeconomic background are associated with response behavior of Churners and to a more limited extent, that of Early Attritors. Relative to Loyalists, black teens are almost 3 times as likely as whites to be classified as Churners, and there is a clear educational gradient indicating that youth with lesser-educated mothers are more likely to exhibit churning behavior compared with Loyalists. Having a college-educated mother also significantly lowered the odds of early attrition.

Discussion

Rising nonresponse rates remain a vexing problem for survey researchers who simultaneously try to keep costs per survey low while minimizing coverage and nonresponse biases (Couper, 2017). Problems of coverage and nonresponse are not unique to web surveys, but digital administration of surveys poses new challenges, including for groups whose media habits are aligned with all things Internet (Rideout, 2015; Runyan et al., 2013; Prensky, 2001). Nonresponse begins at the recruitment stage, carries over to participation and longitudinal persistence. For studies that target youth, nonenrollment incurs formidable costs because of the extra effort required to obtain both parental consent and teen assent, which usually requires extra follow-up to recruit participants. Although mDiary respondents were drawn from an ongoing birth cohort study with a known sampling frame, 20 percent of eligible teens declined to participate and of those who assented to participate, 23 percent did not enroll for the study. Consistent with studies of adults, nonparticipants were disproportionately black males from socioeconomically disadvantaged backgrounds and unstable families.

Our findings indicate that there is a trade-off between the effort expended recruiting teens for the study and their odds of participating. Likely reflecting their interest in the study, teens who required fewer calls to grant assent had higher registration rates than their peers who needed multiple contacts. Furthermore, delayed enrollment presaged poor longitudinal compliance. Teens who enrolled upon receiving their first invitation to register completed more diaries and experienced lower attrition than their recalcitrant peers who enrolled only after receiving multiple invitations to do so. These results call into question the merit of pursuing youth who may be signaling a reluctance to participate, particularly for panel studies that call for repeated measurements.

Differential recruitment and response by socioeconomic status is not unique to web surveys, but rather is a longstanding challenge for survey research. The relatively small differences between the sample of assented and enrolled teens (Table 2) suggests modest representation bias toward nonminority youth with better-educated mothers. Overall response rates were likely boosted by the bias because 22 percent of Loyalists had mothers with college degrees as compared with 42 percent of Skippers. Also noteworthy is the availability of home Internet service, which significantly differentiated enrollees from nonenrollees, as well as Skippers and Loyalists. Internet access is crucial for web-administered surveys, particularly those targeting youth.

Response fatigue associated with participation in the Fragile Families birth cohort study is one plausible reason why 23 percent of teens declined to register after having assented to do so, and why 5 percent of enrollees failed to complete a single diary. Eligibility for the mDiary study was restricted to teens who completed the Year-15 FFCWS survey, which was conducted by phone and averaged 65 minutes. Recruitment for the mDiary study began approximately one year following the Year-15 teen survey. Although the median time for the diaries was under three minutes (Goldberg et al., 2019), the enrollment survey required a median completion time of 7.3 minutes, which may have dampened teens’ interest in the diary study. Furthermore, the enrollment survey included several questions about changes in respondents’ personal and family circumstances since the FFCWS Year-15 interview, but no questions about romantic partnerships, which may have dampened interest in the study. Unlike Barber et al. (2016), we find no evidence that topic salience boosts longitudinal persistence. One plausible reason is the coarse measurement of topic salience, which was indexed by having ever dated by age 15 for the enrollment analyses, and having ever dated by age 16 for the longitudinal compliance analyses.

Low cumulative nonresponse rates are a problem for most probability panel studies (Couper, 2017: 131) but evidence about both data quality and longitudinal response rates in web-administered surveys is mixed (Lee et al., 2019; Czajka & Beyler, 2016; Groves & Peytcheva, 2008). mDiary’s person-survey response rate of 70 percent is noteworthy considering reports that response rates for smartphone studies trail those administered via telephone or PCs (Lee et al., 2019; Callegaro & DiSogra, 2009). The high rates of smartphone ownership, texting capability and digital proficiency among recruited teens likely boosted longitudinal compliance relative to studies involving adults (Jäckle et al., 2019); however, differences both in capacity to send and receive texts and Internet access lowered participation and longitudinal compliance. Barber et al. (2011) suggest that respondents who provided both a phone number and email address exhibited higher longitudinal persistence, which they attribute to differential interest in the study. It is plausible, although unlikely given teens’ strong preference for texting over email for communication (Rideout, 2015), that mDiary teens who did not provide cell numbers also were disinclined to participate.

These findings have implications about the promise of paradata for future surveys involving youth and adults. For youth, convenient access to the web, preferably with a mobile device, is a sine qua non for both participation and persistence. Recruitment paradata also revealed variations in digital proficiency among youth, which is crucial for participation in web-administered surveys. A supplementary analysis that compared Loyalists with respondents who missed one or more diaries (Skippers) confirmed that reliance on schools for email access, lowered the odds of completing all 25 diaries; conversely, teens who enrolled with a smartphone were over 1.6 times as likely to complete all diaries compared with those who enrolled using other devices. Administration of web surveys should consider the implications of access to digital devices both for participation and longitudinal compliance, particularly for studies targeting youth.

Limitations

Our study has several limitations that warrant caution in drawing broad generalizations. These include the narrow age range of the sample as well as coverage and nonresponse bias. Problems of coverage and nonresponse are not unique to web surveys. mDiary, unlike many web surveys, had the advantage of drawing from a known sample frame; however, coverage error arose during both the recruitment and nonenrollment process, rendering a sample of youth that was less disadvantaged and comprised of fewer minority youth than the source population. Although statistical inference is possible only with probability-based sample designs, inferences from the mDiary study are not representative of the urban places sampled for the birth cohort study owing both to coverage and nonresponse bias (Reichman et al., 2001). Finally the low pseudo R-square values for the participation and compliance analyses indicate that additional covariates are needed to better understand adolescents’ longitudinal compliance.

Conclusions

Early skepticism about the promise of digital technology for administering surveys (Couper, 2000) has abated as access to digital technology has become widespread (see Couper, 2000 and 2017). Convenient access to Internet services and variation in digital proficiency remain formidable barriers for some population segments, such as seniors and low-income groups (Jäckle et al., 2019; de Leeuw, 2018). For digital natives, whose life experiences coincide with the emergence and expansion of smartphones (Goldberg & Tienda, 2017; Runyan et al., 2013; Prensky, 2001), mobile-optimized surveys administered via the web, including those involving repeated measurements, hold considerable promise. That the mDiary teens completed 70 percent of the possible diaries, and 44 percent completed all 25 diaries is impressive by comparison to recent studies based on university students (Lee et al., 2019) and adults (Jäckle et al., 2019). Whether larger incentives could have boosted participation rates further is an empirical question for future studies.

Evidence that teens prefer texting over email for digital communication (Anderson, 2015; Rideout, 2015) has important implications for youth recruitment and participation in web-administered surveys. Although access to an email account was the only requirement to participate in the study, this was a barrier for youth residing in homes lacking Internet service. Delivering enrollment instructions via text message or a social media platform like Facebook or Twitter rather than email could potentially increase enrollment not only because this strategy aligns with the target population’s media habits, but also because many teens lack private email accounts. Our results unequivocally show that access to devices with texting capability increases longitudinal compliance.

Nevertheless, there is much to learn about the correlates of nonresponse and longitudinal persistence. Our multivariate analyses explain only 12 percent of the variance in enrollment, with half attributed to recruitment and access paradata, and half to individual and family circumstances. Couper (2017) and others (e.g., Jäckle et al., 2019; Lee et al., 2019) call for continued focus on the reasons for nonresponse and a better understanding of the implications of nonresponse for data quality. Our focus on longitudinal compliance patterns depicted by the latent response classes is an initial foray to understand youth participation in a year-long high frequency survey. Given the mDiary study focus on the emergence and evolution of romantic relationships, the lack of associations between longitudinal compliance patterns and topic salience requires analyses that consider whether respondents were in romantic relationships throughout the study, and how response propensities were associated with the timing of partnership formation (Goldberg et al., 2019). This is the subject of our next inquiry, which evaluates wave-specific response behavior using methods suitable for time-varying outcomes (Kocar & Biddle, 2019). Time-varying methods can potentially clarify why topic salience was unrelated to overall longitudinal compliance and whether extenuating circumstances or major life events, such as a residential move, illness or death of a family member, are associated with nonresponse and early attrition.

Table 5.

Multinomial Estimates of Response Class Membership (vs. Loyalists)

	Highly Engaged		Churners		Occasionals		Early Attritors		No Shows
	Odds Ratio	S.E.	Odds Ratio	S.E.	Odds Ratio	S.E.	Odds Ratio	S.E.	Odds Ratio	S.E.
Recruitment (a)
# Contact attempts	0.996	0.023	0.992	0.033	1.044	0.025	1.028	0.022	0.966	0.045
Assent/enrollment invite lag	0.986	0.015	0.955	0.033	0.987	0.019	0.994	0.013	0.966	0.035
Summer invite	0.826	0.308	1.018	0.532	0.530	0.281	1.517	0.532	1.068	0.644
Access (a)
Uses school email	2.703	1.487	1.052	1.174	6.072	3.570 **	2.570	1.595	2.859	2.454
Receives text messages	0.542	0.385	0.295	0.276	0.305	0.231	0.251	0.167 *	--	--
Contacts recruitment team	0.869	0.476	2.351	1.395	1.437	0.831	1.399	0.686	2.437	1.568
Enrolled with smartphone	0.470	0.134 **	0.461	0.202	0.620	0.235	0.952	0.324	0.736	0.375
Respondent attributes
Female (b)	0.604	0.155 *	1.174	0.478	0.773	0.249	0.583	0.161 *	0.337	0.150*
Ever dated age 16 (a)	0.904	0.260	1.635	0.796	1.088	0.413	1.160	0.386	0.806	0.407
Teen Persistence
Sticks with plan to get something done	0.894	0.262	0.844	0.366	1.087	0.390	1.096	0.340	0.922	0.448
Finishes whatever begins	0.737	0.217	1.358	0.583	1.298	0.467	0.670	0.211	0.945	0.455
Home Environment
Lives with both bio parents (d)	0.976	0.108	1.012	0.163	1.224	0.150	1.154	0.124	1.178	0.212
Has home internet service (d)	0.463	0.276	1.827	2.041	0.355	0.218	0.332	0.178 *	0.471	0.415
Home often chaotic (c)	0.717	0.391	1.439	0.931	0.692	0.439	0.956	0.485	0.950	0.776
Mother's Characteristics (b)
Race (nonHispanic white)	(ref)		(ref)		(ref)		(ref)		(ref)
Black	1.750	0.593	2.848	1.562 *	1.798	0.786	1.181	0.428	2.284	1.223
Hispanic	1.585	0.563	2.135	1.212	1.393	0.631	1.399	0.506	0.524	0.394
Education (< HS)	ref
HS or equivalent	0.805	0.311	0.363	0.173 *	0.618	0.253	0.703	0.253	0.636	0.395
Some college	1.014	0.371	0.204	0.111 **	0.450	0.199	0.500	0.189	0.709	0.442
College graduate	0.844	0.423	0.185	0.149 *	0.512	0.335	0.349	0.193 *	0.563	0.493
Mother married at teen's birth	0.725	0.244	1.240	0.639	0.576	0.274	0.994	0.361	0.952	0.557
Constant	5.056	5.646	0.537	0.981	1.918	2.442	4.178	4.487	0.000	0.003

Sources: (a) mDiary paradata; (b) FF baseline; (c) FF Y15-teen; (d) FF Y15-Parent

Notes: follow-up contact attempts post welcome package; "other race" pooled with whites to avoid empty cells

^***p<0.001

^**p<0.01/p<0.05

^*p <0.05

Author Biographies

Marta Tienda is Maurice P. During ‘22 Professor of Demographic Studies and Professor of Sociology and Public Affairs at Princeton University. tienda@princeton.edu

Dawn Koffman is statistical programmer at the Office of Population Research, Princeton University. dkoffman@princeton.edu