Development of the Discrete Choice Experiment (DCE) Survey
Attribute development and level selection are cornerstones of DCE design, as misspecification can lead to biased or useless results. We based the development of the DCE on an initial qualitative component of this study including focus group discussions (FGDs) and in-depth interviews (IDIs) with women and IDIs with family planning providers. The methods and results of this qualitative phase are presented elsewhere (11).
We selected attributes and levels for the DCE separately for each country, first for India and later for Nigeria. For India, three team members independently reviewed analytical memos summarizing findings for attributes and levels discussed in the qualitative interviews. Next, the team compiled a reduced and prioritized list of attributes and levels based on their relevance for the qualitative interview participants, technical plausibility for a MAP, and methodological considerations limiting the number of attributes and levels that can be included in a DCE. The reduced list was then reviewed with the lead product developer at the Georgia Institute of Technology to determine whether the selected attributes and levels were realistic, credible, and pertinent to inform design decisions for the MAP. A similar process was used in Nigeria, although the memos reviewed as an initial step only covered half of the transcripts since findings were similar to those from India. Method side effects were not originally included as an attribute in the qualitative phase; however, frequent spontaneous discussions combined with technical considerations related to the implications of different possible hormonal formulations (combined estrogen and progestin or progestin-only) for bleeding patterns prompted us to add an attribute related to effects of MAP use on menstruation. The levels (wrist, kneecap, and top of foot) for the location of application attribute were suggested by the product developer because they offer a “harder” surface compared with other parts of the body, which may be important for more complete separation of the microneedles from the patch backing. The full list of attributes and levels included in the DCE is shown in Fig. 1.
Additional insights from cognitive interviews
Prior to implementation of the DCE survey in each country, two local consultants each paired with a study investigator conducted iterative rounds of cognitive interviews (CIs) with a convenience sample of women. The CIs tested clarity of wording in local languages, assessed comprehensibility and appropriateness of the attributes and levels and illustrative pictures, and checked participant understanding of the DCE task. Ten CIs were conducted across three rounds in India, and eight CIs across two rounds in Nigeria. Changes were made between rounds based on feedback received; changes were minimal by the final round in each country, indicating that saturation had been reached.
Eligibility and sampling for the DCE survey
Data collection occurred in November 2017 in India and between March-April 2018 in Nigeria. The DCE survey implemented in India contained all six attributes. Due to concerns that one of the attributes (effect on menstruation) may be dominant, two versions of the surveys were fielded in Nigeria: one with all six attributes and one with five, removing effect on menstruation.
In India, women were eligible for the study if they were married and aged 15–49. In Nigeria, both married women aged 15–49 and unmarried women aged 18–49 were eligible if they reported being sexually active in the last 30 days. Because women who have had experience using modern contraception may have different perspectives on desired method characteristics than women who have never used contraception, we stratified our sample by ever use of contraception. In India, due to the high prevalence of female sterilization, sterilized women were categorized as ever or never user based on contraceptive use prior to being sterilized. Women currently not using modern contraception in both settings were eligible if they reported not being opposed to contraceptive use.
Women were selected through a multi-stage random sampling process. Ten census enumeration blocks divided between urban and peri-urban were selected from three municipal corporations in and around Delhi in India (East, North and South Delhi). In Nigeria, 10 enumeration areas were randomly selected to include five from an urban locality (Agbowo in the Local Government Area (LGA) of Ibadan North) and five from a peri-urban locality (Ajibode in the Akinyele LGA). All households and eligible women within the selected areas were listed. Women were randomly selected within each population subgroup; no more than one eligible woman per household was selected. Up to three attempts were made to contact each sampled woman.
Sample Size
Despite some recent efforts (see for example, de Bekker-Grob et al., 2015), no consensus yet exists on the best way to estimate the sample size required for a DCE for obtaining meaningful, statistically robust parameter estimates given the multiple parameters and comparisons of interest (21). For our purposes, we considered the popular rule of thumb of Johnson and Orme
N > 500c/(t x a),
where c is the largest number of levels for any of the attributes, t is the number of choices sets to be given a respondent, and a is the number of options within choice sets (22, 23). Orme (24) cautioned, however, that the rule of thumb was intended as a minimum and recommended that researchers try to at least double this minimum sample size.
For this study, the largest number of levels for any attribute, c, was 3; the initial number of choice sets to be given a respondent, t, was 10; and the number of options within each choice set, a, was 2 resulting in a sample greater than 500*3/(10*2), or a minimum of 75 for each population group in each country. In India, we increased this minimum requirement to 125 women in each stratum (contraceptive use, urban/peri-urban setting), for a total of 500 women. In Nigeria, where we implemented two versions of the survey, given resource constraints and the addition of a third stratum (marital status), we aimed for 130 women for sample 1 (six attributes) and 100 for sample 2 (five attributes) for any population group defined by two stratifying variables.
Data collection
All DCE surveys were conducted with women in their homes by trained data collectors in the local language. Written consent was obtained from all participants. Sigma Institutional Review Board in India, the Oyo State Research Ethical Review Committee in Nigeria, and FHI 360’s Protection of Human Subjects Committee in the United States approved this study.
We used Sawtooth Software Lighthouse Studio v.9.5.2 to program and administer the DCE via handheld tablets. Respondents were presented 10 sets of two MAP designs. An additional fixed pair was presented to the respondent at the start of the survey to ensure comprehension of the choice process. Choice pairs were randomly generated to have balanced overlap between attribute levels. The efficiency of the DCE design was tested using the “test design” feature in Lighthouse Studio. We checked the estimated main effects and their standard errors in the test run to assess whether the design provides reasonable precision for the model estimates. We also checked the D-efficiency of the design, a commonly used metric to assess statistical efficiency.
Statistical analysis
Data were uploaded from the study tablets into Sawtooth Software’s Lighthouse Studio for analysis. The fixed choice set was not included in the analysis. We used Choice-based Conjoint with Hierarchical Bayes estimation to calculate individual utilities for each attribute level within each participant. Utilities are averages based on the frequency of choosing a contraceptive MAP with the given attribute level. We also examined first-order interactions between attributes. Interactions were included in the final model if they were statistically significant at the level of p ≤ 0.05, made intuitive sense, and improved the model fit. We also assessed associations between sociodemographic covariates including age, education level, and prior use of a modern method and attribute preferences and included in the final model if statistically significant defined as p ≤ 0.05. All models included age, education, and ever use of modern method use as covariates.
Using the final model, we computed the proportion of potential users in the three samples who would choose hypothetical MAP products using the market simulation function in Lighthouse. To demonstrate the potential gain in user desirability with changes to the MAP design, each hypothetical MAP product was compared with a reference MAP with the following attribute levels: a medium sized patch providing three months of pregnancy protection that would be applied to the wrist with pain similar to a light pin prick resulting in a three-day rash and an irregular period.