Preferences for Accessing Electronic Health Records for Research Purposes: Views of Parents Who Have a Child With a Known or Suspected Genetic Condition

Abstract

Objectives:

The purpose of this study was to examine parental preferences for researchers accessing their child’s electronic health record across 3 groups: those with a child with (1) a known genetic condition (fragile X syndrome FXS), (2) a suspected genetic condition (autism spectrum disorder [ASD]), and (3) no known genetic condition (typically developing).

Methods:

After extensive formative work, a discrete choice experiment was designed consisting of 5 attributes, each with 2 or 3 levels, including (1) type of researcher, (2) the use of personally identifiable information, (3) the use of sensitive information, (4) personal importance of research, and (5) return of results. Stratified mixed logit and latent class conditional logit models were examined.

Results:

Parents of children with FXS or ASD had relatively higher preferences for research conducted by nonprofits than parents of typically developing children. Parents of children with ASD also preferred research using non-identifiable and nonsensitive information. Parents of children with FXS or ASD also had preferences for research that was personally important and returned either summary or individual results. Although a few child and family characteristics were related to preferences, they did not overall define the subgroups of parents.

Conclusions:

Although electronic health record preference research has been conducted with the general public, this is the first study to examine the opinions of parents who have a child with a known or suspected genetic condition. These parents were open to studies using their child’s electronic health record because they may have more to gain from this type of research.

Introduction

Electronic health records (EHRs) are a valuable source in obtaining real-world evidence for researchers, clinicians, regulators, and policy makers.^1,2 The volume of clinical data contained within EHRs as well as the large numbers of patients within EHR systems enable researchers to mine vast amounts of information on a scale that was previously unattainable. In addition to patient demographics, diagnosis and procedure codes, laboratory tests, patient history, and prescribed medications, EHRs contain clinical notes that can be analyzed with natural language processing.³ A broad range of studies have used EHRs, including retrospective observational, epidemiological, descriptive, and comparative effectiveness studies, among others.^4–6 There are noted challenges, however, with the quality of EHR data,^7–9 which has led some to call for and develop national standards or frameworks for their use in research.^10,11 Despite these issues, EHRs show promise as a useful health technology tool for clinical research.¹²

EHRs can be especially useful in genomic and rare disease research. If the disease or condition of interest has a low prevalence rate, EHRs are a practical, cost-effective method for gathering data on patients who are geographically dispersed. One way in which EHRs have been used in genomic or rare disease research has been to explore phenotypic variability. Using International Classification of Diseases, Tenth Revision codes, researchers can identify individuals with a specific genetic condition, then examine other data in the EHR to better understand healthcare utilization patterns or comorbid medical conditions.^13,14 A second application has been combining data from EHRs with genetic information gathered from biobank samples. For example, the Electronic Medical Records and Genomics Network has successfully used algorithms to identify patterns of patient phenotypes using EHRs and then conducted exploratory genome-wide association studies to diagnose diseases.^15,16 Finally, efforts are under way to combine information from EHRs with biospecimens as well as patient-reported data to further understand the burden of both common and rare diseases and their impact on quality of life.^17–20 In the United States, this is exemplified in the All of Us Research Program.²¹

Nevertheless, there are ethical considerations for using EHRs when conducting research in these special populations.²² Even if a study is exempt from federal guidelines that protect the rights of human participants in research, and thus does not require informed consent, patients and their caregivers from rare disease communities want to know how their or their child’s data are being used.²³ Some of this is owing to the risk of re-identification when analyzing data from small or unique samples.²⁴ To guide future EHR research on those with genetic or rare diseases and to aid in the development of responsive policies in this area, it is important to understand patient preferences and, in particular, whether they perceive that the benefits outweigh the risks.

Some preference studies on access to EHRs for research purposes have been conducted, but mainly among those in the general population and some groups with special healthcare needs.^25–27 Although a few have been quantitative, most studies have used focus groups or interviews. Overall, most individuals in these studies support EHR research, but many voiced concerns about the privacy and security of their information.^28–30 This was of particular concern when sensitive data, such as mental or sexual health information, were being accessed, or if pharmaceutical or insurance companies were seeking access, because patients worried about possible discrimination and stigmatization.^31,32 Participants in genetic biobank research have expressed similar privacy concerns.^33,34 Notably, the risk of re-identification was specifically mentioned.³⁵

No work, however, has been conducted to date to examine preferences of using EHR data to conduct research among those with, or who have a child with, a known or suspected genetic condition. This group has much to gain from EHR research, because it provides a method for further understanding genotype-phenotype associations and, ultimately, leading to personalized medicine.^36,37 Nevertheless, individuals, especially children, with known or suspected genetic conditions may need special protections given their genetic or disability status to ensure there are no negative consequences of using EHRs for research.^23,38

In this study, we sought to gather preference data from parents of a child with a likely or established genetic condition regarding research use of their child’s EHR using a discrete choice experiment. Three groups of parents were included: (1) parents of children with fragile X syndrome (FXS), (2) parents of children with autism spectrum disorder (ASD), and (3) a comparison group of parents with children who are typically developing (TD). In addition to intellectual disability, children with FXS often present with a range of co-occurring conditions such as attention problems, anxiety, aggression or self-injurious behavior, and seizures.³⁹ Children with ASD may also have intellectual disability, but core diagnostic criteria focus on communication challenges, social impairment, and restricted behaviors or interests.⁴⁰ Although ASD is not linked to a single gene, it has been associated with several genetic variants.⁴¹ Although parents of children with FXS or ASD do not represent the entire spectrum of parents who have a child with a known or suspected genetic condition, they are excellent prototypes for understanding patient preferences given their prevalence rate (1:58 for ASD, ~1:4000 for FXS) and heritability patterns. Understanding the preferences of parents who have a child with an established or likely genetic condition will lay the groundwork for ethical research practices and policies that take into account their perspectives on both risks and benefits of this type of research, regardless of whether consent is needed. We addressed 3 research questions:

1. Which factors drive parental decision making about research use of their child’s EHR?

2. Do parents’ preferences differ between those who have a child with a known or suspected genetic condition, and are either different than preferences of those who have a TD child?

3. What contributing factors, such as parental health literacy or severity of the child’s condition, are related to parents’ preferences regarding EHR research?

Methods

Participant Recruitment

We conducted an online discrete choice survey with 3 groups of parents. Parents of children aged 14 to 17 with FXS or ASD were eligible, as were legal guardians of an adult with FXS or ASD aged 18 to 40. Parents of TD children were eligible if they had a child aged 14 to 17 who did not have an intellectual or developmental disability or genetic condition, and also did not have a moderate or severe chronic health condition (eg, asthma, diabetes, epilepsy) or a psychiatric diagnosis (eg, anxiety, attention-deficit disorder). We targeted parents of adolescent children to not confound any potential ethical issues related to conducting research on younger children. Parents of children with FXS and ASD were a convenience sample recruited through research registries (eg, Our Fragile X World, Interactive Autism Network), parent advocacy organizations (eg, National Fragile X Foundation, Autism Society), and university partners (TEACCH Autism Program at University of North Carolina at Chapel Hill). Parents of TD children were recruited from an online survey panel (Qualtrics) with an effort to obtain a nationally representative sample. All participants resided in the United States.

Instrument

Discrete choice experiment

The design, analysis, and administration of the DCE followed current recommended guidelines in healthcare research.⁴² We conducted 2 steps of formative work to identify potential attributes and levels. First, we conducted a scoping review of research use of electronic health records and potential implications for those with genetic conditions.⁴³ Next, we held a series of focus groups with parents of individuals with ASD or FXS.⁴⁴ Based on these activities and discussions with expert advisers, we included 5 attributes, or characteristics, in the DCE. Each attribute took on 2 or 3 different levels, also informed by the focus groups. Combined, our final list was (1) who is conducting the research (for-profit, nonprofit, or government health researchers), (2) whether the research includes identifiable information (identifiable or not identifiable), (3) whether the research includes sensitive information (sensitive or not sensitive), (4) importance of the research topic (important to you or not important to you), and (5) how research results will be shared (individual, summary, or no results shared). A short description of each attribute and level was provided, including examples. For instance, for-profit companies, included pharmaceutical companies whereas nonprofits included universities and foundations. Sensitive information included mental health, substance abuse, or sexual health information, whereas not sensitive examples were data on height, weight, blood pressure, and allergies. A full description of the attributes and levels is available upon request. We used a paired comparison format in which respondents were asked to compare a series of hypothetical research studies, study A and study B, described by the attributes and levels. An example DCE comparison question is shown in Table 1.

Table 1.

Sample DCE item.

	Study A	Study B
Who is conducting research	For-profit researcher	Nonprofit researcher
Whether the research includes identifiable information	Identifiable	Not identifiable
Whether the research includes sensitive information	Sensitive	Not sensitive
Importance of research topic	Not important to you	Important to you
How research results shared	No results	No results
I would prefer to give permission for:	Study A	Study B

DCE indicates discrete choice experiment.

Before the DCE items were presented to participants, we provided an introduction to EHRs that included a definition, examples of what is typically included in an EHR, and how EHRs can be used in research. Next, we described each attribute and level. Participants were given 2 sample DCE questions; the first allowed participants to practice comparing only the first 3 attributes and the second included all 5 attributes and served as a validity test. The sample DCE item shown in Table 1 describes the choice options presented to participants in the validity test, where we assumed most participants would prefer study B. Approximately 90% showed preference for the expected choice in the second item.

Experiences with the healthcare system

After the DCE questions, participants were asked additional questions about their experiences with the healthcare system. Three questions, taken from the revised Health Care System Distrust scale,⁴⁵ asked participants about their level of trust of their healthcare provider, the healthcare system, and medical researchers (5-point Likert-type scale, 1 = strongly disagree [low trust] to 5 = strongly agree [high trust]). We created a trust in the healthcare system index (range 3–15) by summing these items (see Table 2).

Table 2.

Demographics of survey participants across 3 parent groups.

	Parent subgroups						Total (n = 1503)
	FXS (n = 397)		ASD (n = 611)		TD (n = 495)
	n	%	n	%	n	%	n	%
Sex
Male	28	7	62	10	205	41	295	20
Female	369	93	549	90	290	59	1208	80
Race/ethnicity
Non-Hispanic white alone	348	88	515	84	360	73	1223	81
Non-Hispanic black alone	11	3	39	6	55	11	105	7
Non-Hispanic other race	10	3	24	4	31	6	65	4
Hispanic	18	5	33	5	49	10	100	7
Missing	10	3	0	0.0	0	0.0	10	1
Education
High school or less	150	38	217	36	245	49	612	41
4-year college degree	128	32	198	32	152	31	478	32
Graduate degree	119	30	196	32	98	20	413	27
Income
Less than $50 000	73	18	153	25	185	37	411	27
$50 001 to $100 000	141	36	216	35	174	35	531	35
More than $100 000	169	43	215	35	133	27	517	34
Missing	14	4	27	4	3	1	44	3
Child sex
Male	332	84	486	80	262	53	1080	72
Female	65	16	125	20	233	47	423	28
Child age, M (SD)	23.22	(7.08)	17.10	(3.95)	15.57	(1.08)	18.21	(5.42)
Trust in healthcare system	11.28	(1.78)	11.09	(1.89)	11.45	(2.08)	11.25	(1.93)
Satisfaction with HCP	4.03	(0.74)	4.00	(0.77)	4.02	(0.80)	4.01	(0.77)
Health literacy	13.43	(1.83)	13.66	(1.75)	12.48	(2.38)	13.21	(2.06)
Healthcare decision making	4.14	(0.80)	4.07	(0.74)	4.07	(0.77)	4.09	0.77
Child’s co-occurring health conditions	4.39	(1.85)	3.71	(2.12)	0.45	(1.03)	2.82	2.43
W-ADL	20.67	(6.93)	22.27	(7.33)	29.41	(4.54)	24.20	7.42

HCP indicates healthcare provider; M, mean; W-ADL, Waisman Activities of Daily Living scale.

The next 2 questions, which were developed by the study team, asked participants to rate how well their child’s healthcare provider explains things so they can understand them and whether the provider considered and respected the healthcare choices the parent thinks are best for their child (4-point Likert-type scale, 1 = never to 4 = always). The final question asked participants if, overall, they were satisfied with the care their child receives from their healthcare provider (5-point Likert-type scale, 1 = strongly disagree to 5 = strongly agree). We rescaled the 2 4-point items to match the response range of the third question and created a satisfaction with the child’s healthcare provider scale by taking their average (see Table 2).

Health literacy

Three questions assessed a participant’s level of health literacy:⁴⁶ (1) how often they needed help reading health materials (5-point Likert-type scale, 1 = always to 5 = never), (2) how confident they are in completing medical forms (5-point Likert-type scale, 1 = extremely to 5 = not at all), and (3) how often they have problems learning about their child’s health because of difficulty understanding written information (5-point Likert-type scale, 1 = always to 5 = never). We summed these items to create a summary health literacy score (see Table 2).

Healthcare decision making

A single item developed by the study team assessed the participant’s views about how confident they were that their healthcare decisions aligned with their child’s preferences (5-point Likert-type scale, 1 = extremely to 5 = not at all) (see Table 2).

Child’s co-occurring health conditions

We also asked parents whether their child had ever been diagnosed or treated by a medical professional for 9 commonly co-occurring health conditions associated with ASD and FXS: attention problems, hyperactivity, aggressiveness toward others, self-injurious behavior, seizures, anxiety, depression, general developmental delay or mental retardation, and specific learning disability. The response options for each condition were 0 = no or 1 = yes. We created a summary score for these items representing the total number of co-occurring health conditions for each child (see Table 2).

Waisman Activities of Daily Living Scale (W-ADL)

We assessed the child’s level of independence using the 17-item W-ADL.⁴⁷ All items had a common question stem (ie, Rate your child’s level of independence in …), which was followed by a description of an activity varying in difficulty (eg, making his/her own bed; banking and managing daily finances). Each question had a 3-level response scale: 0 = does not do at all, 1 = does with help, and 2 = independent/does on own. We calculated a composite scale by taking the sum of all 17 items (mean = 24.20, SD = 7.42, range 0–51). Higher values on the W-ADL indicate greater independence (see Table 2).

Experimental Design

The attributes and levels describe 72 possible hypothetical research studies (3²·2³), or 2556 unique study pairs ([72·71] / 2). Nevertheless, only a small subset of well-chosen pairs needed to be shown to participants to achieve robust statistical identification.⁴⁸ We used NGene 1.2.0 software (Choice Metrics, Sydney, Australia) to select a D-efficient experimental design subset for the final study. Specifically, the design was optimized for main effects (no attribute interactions) multinomial logistic model with effects coded attributes and priors for the preference parameters based on study pretests. A constraint prohibiting a nonsensical combination of “not identifiable” with “individual” or “summary” results was also imposed. Candidate designs were compared for statistical efficiency, orthogonality (low correlation across attributes and levels), and level balance (near-evenly distributed combinations of attribute and levels). The final design consisted of 120 profiles or pairs, distributed in sets of 10 across 12 blocks, or versions, of the survey. In fielding, each respondent was asked to compare and indicate their preference over a set of 10 different study pairs. We randomly assigned each respondent to 1 of the 12 different versions (blocks). To eliminate any ordering effects, we randomized the order in which the 10 pairs were shown within a block and also randomized the left/right order of A and B.

Procedures

Prior to collecting data, we conducted cognitive interviews with participants (n = 10; 4 FXS, 3 ASD, 3 TD) to ensure that the survey instrument was understandable and that the practice items were helpful. Minor edits were made to the instrument based on participant feedback. Participants in the main study were invited to complete the survey online. The study team invited parents of children with FXS or ASD to participate and sent up to 2 reminders. These parents were given a $20 Amazon gift card as a thank-you for their participation. Our survey vendor managed recruitment, reminders, and incentives for the sample of parents with a TD child. Data collection took approximately 12 weeks. Institutional review board approval was received before any human participant research was conducted.

Statistical Analyses

We developed several models to compare preferences across 3 participant subgroups and to examine associations between individual characteristics and preferences. In all models, a binary-dependent variable indicated which alternative each participant selected in the 10 choice sets. We used effects coding for the attribute levels, which were entered as explanatory variables in the models.

Stratified mixed logit model

Discrete choice models are based on random utility theory, which assumes that the utility or net benefit an individual derives from choosing one alternative in a choice set consists of a systematic component and a random component.^49–51 The systematic component is made up of individual-specific parameters that account for differences in preferences and the observable attributes that differentiate choice alternatives. The random component is an error term summarizing all unexplainable factors that affect choice. We estimated stratified mixed logit models that allow for observed and unobserved heterogeneity in preferences.⁵¹ Equation 1 shows the utility function for a mixed logit model predicting the utility for individual n associated with alternative j in choice set s:

$$U_{njs}=\mathrm{\beta }{x}_{njs}^{\prime }+\left({\eta }_n{x}_{njs}^{\prime }+{\mathrm{\epsilon }}_{njs}\right),$$ (1)

where χ′_njs is a vector of observed explanatory variables (eg, attribute levels), b is a vector of unknown parameters to be estimated representing the relative contribution of attribute levels to the utility respondents assign to an alternative (eg, preference weights), η_n is a variance term, and ε_njs is a random error term. Individuals are assumed to choose alternatives that maximize utility, and when aggregated over many choices, we are able to use choice data to estimate relative preferences for various attribute levels. Mixed logit models allow preferences to differ among participants and produce estimates of this distribution, or heterogeneity, for each attribute level.⁵⁰ The model output includes a set of coefficients representing the mean preference weight for each attribute level and another set for the SD of preference weights of these attribute levels across participants. We used the mixlogit command in Stata 15.0 to estimate the stratified mixed logit model.^52,53

A large number of model parameters makes testing for subgroup differences in preference weights using interaction terms intractable, so we instead conducted a stratified analysis by estimating separate models for each of the 3 subgroups (TD, FXS, ASD). To determine whether differences across subgroups were owing to differences in the preference weights and the scale of utility functions, we applied the Swait-Louviere test procedure.⁵⁴ We used a likelihood ratio test⁵⁵ to compare the combined fit of these stratified models against an aggregate model that did not take the subgroups into account, to determine whether the stratified model fit the data better than the aggregate model despite a loss of parsimony. The formula for the likelihood ratio test statistic is

$$LR{\mathrm{\chi }}^2=-2\ast \left(L{L}_{\text{Aggregate}}-\left[L{L}_{\text{TD}}+L{L}_{\text{ASD}}+L{L}_{\text{FXS}}\right]\right),$$ (2)

where LL_Aggregate is the log-likelihood of the aggregate model, and (LL_TD + LL_ASD + LL_FXS) is the sum of the log-likelihoods from the separate models for the 3 subgroups. The degrees of freedom for this test are equal to the number of attribute levels in the models (ie, 7 in our case).⁵⁶ We used Wald tests to test equivalence of individual preference weights across subgroups.⁵⁷

Latent-class conditional logit model

An alternative to the continuous heterogeneity specified by mixed logit models is latent class analysis, which models preference heterogeneity as varying by discrete clusters (classes) of respondents with similar response patterns. We used latent class analysis to identify classes of participants within each respondent subgroup (TD, FXS, ASD) and tested associations of class membership with demographic factors. The utility function for these models allow for heterogeneity at the participant level (n):

$$U_{nj}={\mathrm{\beta }}_n+{\mathrm{\lambda }}_n^{\prime }{x}_j+{\mathrm{\epsilon }}_{nj},$$ (3)

where β_n is a vector of preference weights, λ′_n is a vector of individual characteristics that remain constant across alternatives and choice sets, χ′_j is a vector of attribute levels, and ε_nj is a random error term. The latent-class conditional logit model clusters together participants by common preference weights that are fixed within a class but heterogeneous across the classes. We estimated models with up to 10 classes and selected the optimum number of classes in each subgroup by comparing the Bayesian information criterion (BIC) and the consistent Akaike information criteria (CAIC) of these models. Both the BIC and CAIC apply penalty terms to more complex models with a large number of parameters, and lower values indicate a better fitting model.^58,59 Class membership was estimated with categorical demographic variables entered as binary dummy variables, including race/ethnicity (non-Hispanic white only, non-Hispanic black only, non-Hispanic other race, or Hispanic), annual household income ($50 000 or less, $50 001 to $100 000, or more than $100 000), and child’s sex (male or female). Interval-level variables included a trust in the healthcare system index (range 3 [low trust] to 15 [high trust]), health literacy (range 3 [low literacy] to 15 [high literacy]), parental confidence that healthcare decisions align with the child’s wishes (range 1 [not at all confident] to 5 [extremely confident]), and the number of health conditions the child has been diagnosed with other than FXS or autism (range 0–9).

We reasoned that parents’ preferences related to sharing their child’s EHR for research purposes may differ as a function of these variables. For example, perceived alignment between parents and their children regarding healthcare decision making was included because of the age of the children in the study. In some studies, parents can provide consent when their child is under 18, yet the study could last past the age when the child could consent himself or herself. Conceivably, research studies that use identifiable or sensitive information may be less appealing to parents who are not confident that their decision making aligns with their child’s wishes. We wanted to determine if preferences were related to perceived accordance of beliefs. We estimated the latent-class conditional logit model using the lclogit command in Stata 15.0.⁶⁰

Results

Description of Participants

A total of 1531 parents completed the survey. The final analysis sample (N = 1503; see Table 2) excludes participants who skipped 1 or more choice tasks (n = 21) or who completed the DCE in an amount of time that was ±3 SDs from the average time that participants took to complete it (n = 7). Most participants were female (80%) and non-Hispanic white (81%). Overall, most participants had a 4-year degree or more education (59%) and an income of $50 001 or more (69%). Most parents had a male child (72%) whose average age was 18.21 years (SD = 5.42).

Stratified Mixed Logit Model

Results from the first step of the Swait-Louviere test procedure revealed that preference weights for attribute levels differed significantly across parent subgroups, χ²_S-L (2) = 240.72, P < .001. Preference estimates from the final stratified mixed logit model are shown in Table 3 and also depicted in Figure 1. Preference weights within attribute levels with the same letter (eg, ‘a’) are not significantly different across subgroups at α = 0.05. The SDs of the preference weights revealed significant preference heterogeneity in each parent subgroup for most attribute levels; the only exception are preferences for summary results among parents of typically developing children (B_SD = −0.18, SE = 0.15, 95% CI [–0.48 to 0.13], z = −1.14, P = .252).

Table 3.

Mixed logistic regression analysis predicting parental choice for research use of their child’s EHR, stratified by group.

Factor	FXS		ASD		TD
	Coef	SE	Coef	SE	Coef	SE
Mean preference weights
Type of researcher
Nonprofit	0.79*	0.07	0.72*	0.06	0.24*	0.05
Government	−0.13	0.07	0.17†	0.06	−0.03	0.05
For-profit‡	−0.66*	0.08	−0.88*	0.08	−0.21*	0.05
Personally identifiable information
Not identifiable	0.32*	0.07	0.65*	0.07	0.40*	0.06
Identifiable‡	−0.32*	0.07	−0.65*	0.07	−0.40*	0.06
Sensitive information about child
Not sensitive	0.20*	0.04	0.43*	0.04	0.34*	0.04
Sensitive‡	−0.20*	0.04	−0.43*	0.04	−0.34*	0.04
Personal importance of research topic
Important	1.23*	0.08	1.18*	0.07	0.49*	0.04
Not important‡	−1.23*	0.08	−1.18*	0.07	−0.49*	0.04
Shared results
Individual results	0.40*	0.07	0.56*	0.07	0.16†	0.06
Summary results	0.33*	0.07	0.25*	0.06	0.08	0.05
No results‡	−0.74*	0.09	−0.81*	0.08	−0.25*	0.06
Standard deviation of preference weights
Type of researcher
Nonprofit	0.60*	0.08	0.81*	0.08	0.55v	0.06
Government	0.91*	0.09	0.78*	0.07	0.59*	0.06
For-profit§	–	–	–	–	–	–
Personally identifiable information
Not identifiable	0.70*	0.09	0.92*	0.08	0.74*	0.07
Identifiable§	–	–	–	–	–	–
Sensitive information about child
Not sensitive	0.38*	0.06	0.61*	0.05	0.62*	0.05
Sensitive§	–	–	–	–	–	–
Personal importance of research topic
Important	0.93*	0.08	1.11*	0.07	0.59*	0.04
Not important§	–	–	–	–	–	–
Shared results
Individual results	0.57*	0.09	0.88*	0.09	0.64*	0.07
Summary results	0.50*	0.12	0.55*	0.10	−0.18	0.15
No results§	–	–	–	–	–	–
Model
Log likelihood	−1997.25		−2996.68		−2899.90
AIC	3802.50		6021.36		5827.80
BIC	3900.21		6125.11		5928.60
χ2 7∥	588.04*		1098.04*		630.50*
N	397		611		495
Observations	7940		12 220		9900

Note. Likelihood ratio test of equivalence for stratified models: χ² (14) = 273.88, P < .001.

Coefficients in the means section of the table are the average preference weights for each attribute level in the designated subgroups. Significant coefficients in the SDs section indicate that there is preference heterogeneity for that attribute level and subgroup.

AIC indicates Akaike information criterion; ASD, parents of children with autism spectrum disorder; BIC, Bayesian information criterion; FXS, parents of children with fragile X syndrome; SE, standard error; TD, parents of typically developing children.

^*P < .01.

^†P < .001.

^‡Omitted reference level of attribute, computed from estimated coefficients.

^§Omitted reference level of attribute.

^∥Likelihood ratio test for the joint significance of the standard deviations, testing the null hypothesis that all standard deviations in the model are equal to 0.

Figure 1.

Average preference weights by attribute level and study population. Error bars denote 95% confidence intervals. Preference weights within attribute levels sharing a letter in common (eg, ‘a’) are not significantly different across subgroups at α = 0.05. Comparisons across subgroups assume no covariance between estimates.

Average preference weights for levels within attributes and subgroups are significantly different from one another (see Fig. 1), except that there are no pairwise differences between preferences for individual results versus summary results among parents of TD children (z = 0.89, P = .371) or among parents of children with FXS (z = 0.60, P = .546). Whereas parents of TD children or children with FXS appear to have made little distinction in their preferences between individual and summary results, parents of children with ASD were the only group that had a clear preference hierarchy on this attribute: no results was the level with the lowest preference weight, studies promising individual results had the highest preference weight, and preferences for summary results fell between the other 2 levels.

Preferences for several attribute levels also differed across subgroups. Compared to parents of TD children, parents of children with FXS (Wald χ² = 42.64, P = .000) and parents of children with ASD (Wald χ² = 37.26, P = .000) had higher preferences for studies conducted by nonprofit researchers; in line with this, parents of children with FXS had relatively lower preference for studies by for-profit researchers (Wald χ² = 20.23, P = .000) or ASD (Wald χ² = 51.56, P = .000). Parents of children with ASD also had higher preference for research conducted by government researchers, compared to parents of TD children (Wald χ² = 6.96, P = .008) or those with FXS (Wald χ² = 10.23, P = .001).

Parents of children with ASD had higher preferences for research using information that is not personally identifiable, compared to both parents of TD children (Wald χ² = 7.74, P = .005) or those with FXS (Wald χ² = 11.34, P = .001); however, this is only a difference of magnitude. On average, parents in all 3 groups preferred studies that rely on non-identifiable information to studies that use identifiable information.

Parents of children with FXS had lower preference for studies involving sensitive information about their child than did parents of TD children (Wald χ² = 6.50, P = .011); however, they still preferred studies that use non-sensitive information over sensitive. Parents of children with ASD showed higher preferences for studies involving nonsensitive information compared to parents of children with FXS (Wald χ² = 16.01 P = .000), but there was no difference compared to parents of TD children.

Compared to parents of TD children, parents of those with FXS (Wald χ² = 67.96, P = .000) or ASD (Wald χ² = 70.96 P = .000) had higher preference for studies with a research topic that was personally important to them or their family. Parents of children with FXS (Wald χ² = 22.09, P = .000) or ASD (Wald χ² = 31.89, P = .000) also had lower preference for studies that would not return results than did parents of TD children. Parents of children with FXS or ASD also had higher preferences for studies that would return summary results (FXS: Wald χ² = 8.56, P = .003; ASD: Wald χ² = 4.89, P = .027) or individual results (FXS: Wald χ² = 7.16, P = 07; ASD: Wald χ² = 19.54, P = .000) compared to parents of TD children.

Latent-Class Conditional Logit Model

Based on the BIC and CAIC criteria, models with 3 and 4 classes fit the data indicating significant preference heterogeneity, which is consistent with preference variation seen in the mixed logit results above (see Appendix Table in Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2020.06.016). The 3-class models were appropriate for parents with TD children (BIC = 6043.65, CAIC = 6086.65) and parents of children with FXS (BIC = 3830.11, CAIC = 3873.11). The 4-class model was best for parents of children with ASD (BIC = 6001.65, CAIC = 6062.65). The results of the latent-class conditional logit model for each parent subgroup are presented in Tables 4 to 6.

Table 4.

Estimated parameters for the latent-class conditional logit model among parents with a child with FXS (n = 373).

Factor	Class 1		Class 2		Class 3
	Coef	SE	Coef	SE	Coef	SE
DCE attributes
Type of researcher
Nonprofit	0.39‡	0.07	1.01‡	0.13	0.37‡	0.09
Government	0.28‡	0.07	−0.77‡	0.12	0.00	0.10
For-profit§	−0.67‡	0.09	−0.24*	0.11	−0.37‡	0.11
Personally identifiable information
Not identifiable	−0.03	0.07	0.57‡	0.11	0.37‡	0.10
Identifiable§	0.03	0.07	−0.57‡	0.11	−0.37‡	0.10
Sensitive information about child
Not sensitive	−0.03	0.04	0.38‡	0.07	0.22†	0.07
Sensitive§	0.03	0.04	−0.38‡	0.07	−0.22†	0.07
Personal importance of research topic
Important	0.28‡	0.05	0.11	0.06	1.45‡	0.10
Not important§	−0.28‡	0.05	−0.11	0.06	−1.45‡	0.10
Shared results
Individual results	0.39‡	0.08	−0.05	0.11	0.40‡	0.10
Summary results	0.24†	0.08	0.13	0.11	0.29†	0.11
No results§	−0.64‡	0.09	−0.08	0.12	−0.69v	0.12
Participant characteristics
Race/ethnicity
Non-Hispanic white alone∥	–	–	–	–	–	–
Non-Hispanic black alone	22.49	727.24	21.70	727.24	–	–
Non-Hispanic other alone	0.84	0.86	−0.01	1.42	–	–
Hispanic	−0.13	0.69	0.33	0.76	–	–
Income
Less than $50 000∥	–	–	–	–	–	–
$50 001 to $100 000	0.33	0.42	−0.27	0.46	–	–
More than $100 000	0.14	0.42	−0.21	0.44	–
Child sex∥
Male	–	–	–	–	–	–
Female	0.18	0.40	0.19	0.44	–	–
Co-occurring conditions	0.07	0.08	−0.15	0.09	–	–
Trust in healthcare system	0.05	0.09	−0.28†	0.09	–	–
Health literacy	−0.15	0.08	−0.10	0.09	–	–
Healthcare decision making	0.21	0.19	0.09	0.20	–	–
Intercept	−0.45	1.65	3.92*	1.71	–	–
FXS parents per class	32.2%		19.3%		48.5%

Note. Coefficients in the DCE attributes section of the table are the preference weights for each attribute level in the class. Significant coefficients in the participant characteristics section indicate variables that differentiate membership in the groups, where class 3 is the reference category.

DCE indicates discrete choice experiment; FXS, parents of children with fragile X syndrome; SE, standard error.

^*P < .05.

^†P < .01.

^‡P < .001.

^§Omitted reference level of attribute, computed from estimated coefficients.

^∥Omitted reference level of factor.

Table 6.

Estimated parameters for the latent-class conditional logit model among parents with a TD child (n = 492).

Factor	Class 1		Class 2		Class 3
	Coef	SE	Coef	SE	Coef	SE
DCE attributes
Type of researcher
Nonprofit	0.19	0.12	0.07	0.08	0.21‡	0.04
Government	0.19	0.11	−0.07	0.08	−0.02	0.05
For-profit§	−0.38†	0.13	0.00	0.08	−0.19‡	0.05
Personally identifiable information
Not identifiable	0.10	0.13	1.14‡	0.14	0.08	0.05
Identifiable§	−0.10	0.13	−1.14‡	0.14	−0.08	0.05
Sensitive information about child
Not sensitive	0.02	0.07	1.02‡	0.11	0.02	0.03
Sensitive§	−0.02	0.07	−1.02‡	0.11	−0.02	0.03
Personal importance of research topic
Important	1.14‡	0.14	0.22‡	0.05	0.08*	0.04
Not important§	−1.14‡	0.14	−0.22‡	0.05	−0.08*	0.04
Shared results
Individual results	0.43†	0.12	−0.02	0.09	0.12*	0.05
Summary results	0.24	0.12	0.02	0.09	0.10*	0.05
No results§	−0.67‡	0.15	0.00	0.10	−0.22‡	0.06
Participant characteristics
Race/ethnicity
Non-Hispanic white alone∥	–	–	–	–	–	–
Non-Hispanic black alone	0.07	0.44	−0.65	0.50	–	–
Non-Hispanic other alone	0.00	0.76	0.65	0.60	–	–
Hispanic	−0.35	0.52	0.21	0.43	–	–
Income
Less than $50 000∥	–	–	–	–	–	–
$50 001 to $100 000	−0.11	0.36	−0.09	0.32	–	–
More than $100 000	0.39	0.36	0.22	0.34	–
Child sex∥
Male	–	–	–	–	–	–
Female	0.12	0.31	0.22	0.28	–	–
Co-occurring conditions	−0.19	0.16	−0.23	0.18	–	–
Trust in healthcare system	0.09	0.07	0.03	0.07	–	–
Health literacy	0.16*	0.07	0.26‡	0.07	–	–
Healthcare decision making	0.08	0.20	−0.06	0.18	–	–
Intercept	−4.09†	1.41	−3.93†	1.22	–	–
TD parents per class	24.1%		28.8%		47.1%

Note. Coefficients in the DCE attributes section of the table are the preference weights for each attribute level in the class. Significant coefficients in the participant characteristics section indicate variables that differentiate membership in the groups, where class 3 is the reference category.

DCE indicates discrete choice experiment; SE, standard error; TD, parents of typically developing children.

^*P < .05.

^†P < .01.

^‡P < .001.

^§Omitted reference level of attribute, computed from estimated coefficients.

^∥Omitted reference level of factor.

Almost a third of parents of children with FXS were clustered into class 1. Parents in this class preferred studies conducted by nonprofit or governmental researchers and were dissuaded from participating in studies conducted by for-profit researchers (see Table 4). Members of this group were also drawn to studies that would share individual or summary results—with an aversion toward those that would not return any results—and preferred studies dealing with a personally important research topic. There were no differences in this group’s preferences for identifiability or sensitivity of information accessed. No defining family or child characteristics set this group apart from the comparison group (class 3). In the next subgroup, about a fifth of FXS parents (class 2) showed an especially high preference for participating in studies conducted by nonprofit researchers with corresponding low preference for government and for-profit studies. Participants who were clustered into this group also preferred studies that did not include personally identifiable information or sensitive information about the child (relative to those that did). Compared with participants in class 3, those who had lower trust in the healthcare system were more likely to be clustered into class 2 (B = −0.28, SE = 0.09, 95% CI [−0.46 to −0.10], z = −3.12, P = .002). The preferences of the remaining FXS parents (class 3), comprising nearly half the sample, were influenced to some degree by all 5 of the DCE attributes.

Among parents of children with ASD (see Table 5), there were 4 distinct subgroups. Approximately 17% were grouped together into class 1. These parents displayed an especially strong preference for participating in studies conducted by nonprofit and, to a lesser extent, governmental researchers. Members of this group also preferred studies that did not include personally identifiable information or sensitive information about their child, were on a personally important topic, and shared summary results. No defining family or child characteristics set this group apart from the comparison group (class 4). Members of class 2 make up a little more than a third of parents who have a child with ASD and are distinguished from the other groups by their particularly strong preference for research on a personally important topic. In other respects, the preferences of class 2 align most to those of class 1, with the exception of the shared results attribute; members of class 2 had a higher preference for individual results, were indifferent toward summary results, and had lower preference for studies that do not share any results. Compared to parents in class 4, parents of children with a greater number of co-occurring conditions (B = 0.12, SE = 0.06, 95% CI [0.00–0.24], z = 1.97, P = .048) or higher trust in the healthcare system (B = 0.19, SE = 0.07, 95% CI [0.05–0.32], z = 2.76, P = .006) were more likely to be grouped into class 2. A quarter of parents who have a child with ASD were clustered into class 3 and showed especially high preferences for participating in studies that shared individual or summary results and a corresponding aversion toward studies that would not return any results. Parents in class 3 were indifferent to the identifiability or sensitivity of information used in a study, and they showed a preference for studies conducted by nonprofit researchers over for-profit studies. Parents of children with ASD who had higher trust in the healthcare system were also more likely to be grouped into class 3 (B = 0.21, SE = 0.08, 95% CI [0.06–0.37], z = 2.66, P = .008) than class 4. The remaining parents of children with ASD who were grouped together in class 4 had especially high preferences for studies that did not include personally identifiable or sensitive information about their child (versus studies that relied on personally identifiable or sensitive information), preferred studies that dealt with a personally important research topic, and—similar to class 3—preferred studies by nonprofit over for-profit researchers. Members of class 4 were indifferent to whether results would be shared.

Table 5.

Estimated parameters for the latent-class conditional logit model among parents of children with ASD (n = 584).

Factor	Class 1		Class 2		Class 3		Class 4
	Coef	SE	Coef	SE	Coef	SE	Coef	SE
DCE attributes
Type of researcher
Nonprofit	1.28‡	0.14	0.38†	0.12	0.21†	0.07	0.25†	0.08
Government	0.27*	0.12	0.30*	0.12	0.02	0.09	0.02	0.09
For-profit§	−1.55‡	0.16	−0.68‡	0.16	−0.23*	0.09	−0.26†	0.10
Personally identifiable information
Not identifiable	0.25*	0.11	0.37†	0.13	0.08	0.07	1.23‡	0.15
Identifiable§	−0.25*	0.11	−0.37†	0.13	−0.08	0.07	−1.23‡	0.15
Sensitive information about child
Not sensitive	0.17‡	0.06	0.15*	0.07	0.03	0.04	0.82‡	0.09
Sensitive§	−0.17‡	0.06	−0.15*	0.07	−0.03	0.04	−0.82‡	0.09
Personal importance of research topic
Important	0.28‡	0.07	1.64‡	0.13	0.31‡	0.05	0.33‡	0.06
Not important§	−0.28‡	0.07	−1.64‡	0.13	−0.31‡	0.05	−0.33‡	0.06
Shared results
Individual results	−0.04	0.13	0.36‡	0.13	0.67‡	0.09	0.18	0.09
Summary results	0.26*	0.11	0.14	0.12	0.32‡	0.08	−0.01	0.08
No results§	−0.22	0.13	−0.50‡	0.15	−0.99‡	0.12	−0.17	0.10
Participant characteristics
Race/ethnicity
Non-Hispanic white alone∥	–	–	–	–	–	–	–	–
Non-Hispanic black alone	0.29	0.60	−0.13	0.56	0.18	0.60	–	–
Non-Hispanic other alone	0.30	0.80	0.05	0.71	−0.30	1.08	–	–
Hispanic	0.17	0.76	0.26	0.62	0.37	0.69	–	–
Income
Less than $50 000∥	–	–	–	–	–	–	–	–
$50 001 to $100 000	−0.08	0.46	−0.69	0.38	−0.88*	0.41	–	–
More than $100 000	−0.67	0.48	−0.82*	0.38	−1.22†	0.42	–
Child sex∥
Male	–	–	–	–	–	–	–	–
Female	0.33	0.39	0.30	0.32	−0.04	0.40	–	–
Co-occurring conditions	0.07	0.08	0.12*	0.06	0.10	0.07	–	–
Trust in healthcare system	0.01	0.08	0.19†	0.07	0.21†	0.08	–	–
Health literacy	−0.08	0.09	0.02	0.08	−0.10	0.09	–	–
Healthcare decision making	−0.09	0.22	−0.21	0.18	−0.12	0.20	–	–
Intercept	1.02	1.63	−1.05	1.44	0.10	1.64	–	–
FXS parents per class	16.8%		34.5%		25.8%		22.9%

Note. Coefficients in the DCE attributes section of the table are the preference weights for each attribute level in the class. Significant coefficients in the participant characteristics section indicate variables that differentiate membership in the groups, where class 4 is the reference category.

ASD indicates parents of children with autism spectrum disorder; DCE, discrete choice experiment; FXS, parents of children with fragile X syndrome; SE, standard error.

^*P < .05.

^†P < .01.

^‡P < .001.

^§Omitted reference level of attribute, computed from estimated coefficients.

^∥Omitted reference level of factor.

Among parents of TD children (see Table 6), roughly a quarter of participants (class 1) had an aversion toward studies conducted by for-profit researchers, showed an especially high preference for studies on research topics that were personally important to them, and preferred studies that would return individual research results—with a corresponding indifference toward participating in studies that would only return summary results and a disinclination to participate in studies where no results would be returned. Type of researcher, personal importance of the research topic, and the shared results attributes were also important to class 3, which made up nearly half of TD participants. Parents who were grouped into this class had higher preferences for studies conducted by nonprofit researchers and lower preferences for studies by for-profit researchers, preferred personally important research topics, and had low preference for studies that would not return any research results, favoring those that would return either summary or individual results. The remaining participants who were grouped into class 2 (almost 30%) were most concerned about privacy, data sensitivity, and the personal importance of research topics, showing higher preferences for studies that did not include personally identifiable or sensitive information about the child and dealt with a personally important research topic (compared to studies that used personally identifiable data, sensitive information, and did not address a personally important topic, respectively). Among the participant characteristic variables included in the model, participants with higher health literacy scores were more likely to be in class 1 (B = 0.16, SE = 0.07, 95% CI [0.02–0.30], z = 2.26, P = .024) or class 2 (B = 0.26, SE = 0.07, 95% CI [0.13–0.40], z = 3.90, P < .001) than in class 3. Race/ethnicity, income, child’s sex, co-occurring conditions, trust in the healthcare system, and perceptions of healthcare decision making were not associated with group membership.

Discussion

This study provides much-needed quantitative information about the views of parents who have a child with a known or suspected genetic condition about research use of their child’s EHR. Several factors have pushed the need to understand the opinions of this unique group of stakeholders, such as the precision medicine research initiative, which uses EHR data along with other health information to individualize treatment.²¹ Preference studies are essential to understanding the perspectives of these unique stakeholders, which can help to guide future research and inform policy decisions.

In keeping with our prior research findings,⁴⁴ parents had strong and clear preferences about trust. Parents of children with FXS or ASD favored nonprofit researchers to conduct studies using their child’s EHR. This was also the case for parents of TD children, but to a lesser degree. Nevertheless, 2 subgroups of parents had exceptionally strong preferences for nonprofit researchers: about 19% of parents with a child with FXS (class 2) and 17% of parents of a child with ASD (class 1). In contrast, just over one-half of parents of a TD child (class 1 and 2) were indifferent to who was conducting the research. The subgroup of parents of children with FXS had less trust in the healthcare system, which may have influenced their desire to have more trust in the researcher accessing their child’s EHR data. These results are consistent with earlier studies that found higher levels of trust for nonprofit research institutions or universities and lower levels of trust for pharmaceutical or insurance companies.^61–63 Thus, future EHR research sponsored or conducted by for-profit companies, such as pharmaceutical companies, may consider partnering with nonprofit organizations or academic research institutes.

A second key finding was that most parents of children with FXS or ASD and a subgroup of parents with a TD child preferred research use of EHRs that used non-identifiable and nonsensitive data. Nevertheless, this was especially true for a subgroup of parents of a child with ASD (class 1). This is an interesting twist on genetic exceptionalism (ie, the belief that genetic information is unique when compared with other types of medical or health information and as such deserves special protection).⁶⁴ It would not have been surprising if parents of children with FXS, who have a known genetic condition, espoused genetic exceptionalism and thus were more in favor of research using non-identifiable and nonsensitive data than parents of children with ASD, a suspected genetic condition. Nevertheless, the opposite was true. In fact, preferences of the subgroup of parents of a child with ASD were twice as high for non-identifiable and nonsensitive information than the preferences of parents of a child with FXS. Perhaps this finding is related to parents’ fear of possible discrimination against individuals with ASD who, despite having the same diagnosis, have quite variable functioning. The subgroup of parents with a TD child had similarly strong preference (class 2). Previous studies conducted with the general public have found mixed evidence for the idea of genetic exceptionalism as well.^65,66

It is unclear what factors are driving the difference between the groups with strong preferences for the privacy of their child’s data. Only a handful of the child and family characteristics (parent’s income, amount of trust in the healthcare system, level of health literacy, and severity of the child’s condition as reflected in the number of co-occurring health conditions) were related to the groups who preferred the use of non-identifiable and nonsensitive data, but there was no consistent pattern across the subgroups of parents. All of the subgroups of parents with a child with FXS or ASD, though, who prioritized the use of non-identifiable and nonsensitive information also had preferences for nonprofits to conduct the research. This may reflect their fear of privacy breaches and possible negative consequences for their child, such as stigmatization or discrimination. Other researchers have noted possible misuses of EHR data in their work as well.^35,67,68 For the subgroup of parents with a TD child, their concerns about privacy align with other research that has highlighted the need for the security of EHR data.^61,67,69

Not surprisingly, most parents of children with FXS or ASD had strong preferences for research that was personally relevant and returned results. Our previous research⁴⁴ indicated that desire to participate in personally relevant research that returns results is related to altruism, specifically a desire to help their child or others in their community. Altruism has been found to be a motivating factor for participating in research in earlier studies as well.^28,62,70 Those who have been diagnosed with a particular condition often want to pay it forward by participating in research that hopefully will lead to better treatment for those with the same condition.^71,72 Return of results in genetic research has similarly been supported by both those in the general public as well as individuals with genetic conditions.^71,72 Return of results is also seen as a personal benefit of research participation.^72–74 In studies of parents of individuals with ASD, some stated the return of results helps to alleviate guilt⁷⁵ or prepare for the future.⁷⁶ Finally, return of results in genetic studies often is seen as a benefit for other family members to help increase awareness and make more informed treatment or reproductive choices.^72,75

Implications

Taken together, these findings have implications for policy and practice. As echoed by others, there is a need to balance the privacy of individuals and the benefit that comes from mining EHRs.^37,77 The views of the general public as well as those expressed here by parents of individuals with known or suspected genetic conditions demonstrate a clear preference for maintaining trust, transparency, and relevance of research being conducted with data from EHRs. These beliefs are aligned with the ethical principles outlined in the Belmont Report, including respect for persons, beneficence, and justice. In practice, these principles are embodied in the informed consent process. Nevertheless, research studies that use deidentified data from EHRs are not classified as human participant research and therefore do not require informed consent. Despite this, a growing body of work^78,79 shows that there is a need for researchers to apply a social license framework, one that goes beyond the regulations for conducting research and embodies a broader set of ethical standards including reciprocity, nonexploitation, and service to the public good.⁸⁰ Models of informed consent that exemplify these ideals are opt-out, tiered, or dynamic consent models.^81,82 For European countries, these ethical principles have been incorporated into policy, specifically the General Data Protection Regulation, which governs the use of EHR data.⁸³ In the United States, the research use of EHRs falls under the policy umbrella of the Final Rule and the Health Insurance Portability and Accountability Act. These rules, though, are not as comprehensive as General Data Protection Regulation.⁸⁴ The results of this study highlight the need to use social license principles to guide the refinement of policies and practices for individuals with genetic conditions in the United States.

Limitations

The interpretation of results from this study should be tempered with the following limitations. First, although we had a large sample of parents across each of the 3 groups, they were not necessarily representative of the broader population. For instance, all participants were recruited through research registries or convenience sampling. In addition, most participants were non-Hispanic white and, in the FXS and ASD groups, more educated and had higher incomes. Given that these parents had participated in other research studies, their preferences for research use of their child’s EHR may differ from parents who have not participated. Additionally, the views of parents of children with FXS or ASD likely do not represent the views of parents whose children have other known or suspected genetic conditions. Finally, most participants had similar levels of trust in the healthcare system, health literacy, and satisfaction with their healthcare providers thus making it difficult to tease out differences between the subgroups of parents based on these characteristics. It is possible other unmeasured factors, such as the degree to which EHR research will make an impact on treatment options for their child, could be related to parents’ preferences.

Future Research

This study helps to shape the knowledge base of research use of EHR data from a unique population: parents of children with known or suspected genetic conditions. The preferences of these individuals have not been investigated previously, yet they are critical to understanding the motivations of parents to participate in the growing number of large-scale genetic research studies. Knowing the factors that affect parental decision making can help to shape not only the type of research being conducted but also how to better inform parents so they can make appropriate decisions for their child and family. Additional research should examine which types of studies parents are more inclined to participate in and whether they are willing to combine their child’s EHR data with other epidemiological (eg, lifestyle) or biological (blood or tissue samples) data to share with researchers.