Abstract
Audio computer-assisted self-interview (ACASI) has been shown to reduce interviewer and social desirability bias related to sensitive questions, which can be especially important for studies of sexual behavior and HIV risk. Baseline demographic and HIV risk data were collected using ACASI for 849 adults aged 18–34 (423 males and 426 females) of unknown or HIV-negative status as part of an HIV-incidence cohort study in Kisumu, Kenya. ACASI questionnaires and possible responses were recorded being read either by a male or female voice in the three most commonly used languages (English, Kiswahili, and Dholuo). Participants were randomly assigned to hear either male or female voice. Dependent variables, including 12 sexual behavior questions, were evaluated using logistic regression. No significant differences in responses to the 12 sensitive questions were found for gender of the recorded voice on ACASI or for the interaction between gender of recorded ACASI voice and respondent gender.
Background
It is estimated that 1.42 million adults and children are currently living with HIV in Kenya, with the distribution of HIV infections varying greatly across the country. Prevalence remains highest in Nyanza Province (14.9% in persons aged 15–64), more than double the national prevalence estimate of 7.1% (National AIDS and STI Control Programme MoHK 2009). Because of the AIDS epidemic in Kenya and many other countries, there is an increased need to find the best ways to collect accurate sexual risk behavior data from individuals. However, sexual behavior questions are sensitive and potentially stigmatizing.
The literature shows that biased reporting can occur with face-to-face (FTF) survey interviews (Pollner 1998). This is especially the case when sensitive questions are asked about behaviors related to risk for HIV infection or activities that are prohibited or considered illegal (Catania et al. 1990). Previous studies examining the effect of the gender of the interviewer on survey responses regarding various topics have produced mixed results. Some have demonstrated differences in responses based on the gender of the interviewer (Pollner 1998; Schopper et al. 2009), while others have not (Johnson and DeLamater 1976; Johnson and Moore 1993; Kane and Macaulay 1993).
There has been recent interest in using the audio computer-assisted self-interview (ACASI) in developing country settings. The ACASI computer application allows respondents to read questions on the computer screen while simultaneously hearing the survey questions, prerecorded in a human voice, over a computer headset in privacy. Overall, studies have found higher reporting of sensitive behavior with ACASI as compared to FTF interviewer-administered surveys (Mensch et al. 2003; Tourangeau and Smith 1996; Turner et al. 1998; van der Elst et al. 2009). A systematic review of questionnaire delivery modes in developing countries found that computerized interviews lowered nonresponse to questions and increased rates of reporting sexual behaviors (Langhaug et al. 2010).
Reduced interviewer bias and social desirability bias will enable increased collection of accurate measures of premarital and extramarital sexual activity to evaluate HIV control programs (UNAIDS 2004). One component of ACASI that could potentially result in bias and that has received little attention is the gender of the recorded voice on ACASI heard by respondents. In one of the few studies assessing the effect of gender of voice in an ACASI study, the authors suggest that respondents may ascribe human traits to the computers and that characteristics such as gender of the computer voice may influence respondents to react to the computer as they would react to a FTF interviewer (Fahrney et al. 2005). If so, then the choice of gender for an ACASI voice may be as important a factor in ACASI design as choosing staff is for more traditional FTF interviews.
There is growing interest in and use of ACASI and other computer-based technologies in African research settings (Jaspan et al. 2007; Langhaug et al. 2010), however, to date, there are few published reports on the effect of the gender of the recorded voice on self-reports of sensitive behavior. Assessing whether gender of the recorded voice on ACASI influences responses to sensitive questions in HIV prevention studies is crucial to ensuring accurate reporting using this data collection method, especially in research settings in African countries where the sensitive questions may be asked in the context of stigma in the AIDS epidemic. The purpose of the study reported here was to determine whether gender of the recorded ACASI voice was a factor in responses to sensitive sexual behavior questions asked of adult males and females in an observational prospective HIV sero-incidence cohort study conducted in Kisumu, Kenya.
Methods
Participants
This substudy was carried out as part of the Kisumu HIV Incidence Cohort Study (KiCoS), an observational prospective cohort study to estimate incidence of HIV sero-conversion and to identify determinants of successful recruitment and retention in preparation for HIV vaccine or prevention trials. A total of 1,277 individuals were prescreened for KiCoS participation using the following criteria: residence within 5 km of the town of Kisumu, between 18 and 34 years of age, sexually active (must have had sex at least once during the past 3 months), and of unknown or HIV-negative status. The 849 persons (423 males and 426 females) meeting these four criteria agreed to complete ACASI assessment and undergo HIV and medical screening between March 2007 and March 2008.
During recruitment, potential participants were given explanations about the study, told that the information they provided would be kept confidential, and informed that their participation was voluntary so they could refuse to answer any questions. Written informed consent was obtained from participants. Centers for Disease Control Institutional Review Board and the Kenya Medical Research Institute Ethical Review Committee approval were obtained for the study. All participants received monetary reimbursement for their transportation costs.
ACASI Study Instrument
We used the Questionnaire Development System to develop an ACASI questionnaire for the KiCoS study. Respondents were randomly assigned to male or female gender of the recorded voice on ACASI before completing the ACASI (baseline) assessment at the screening visit for the incidence cohort study. The staff chose the persons who read the questions for the ACASI based on their ability to clearly pronounce words in the particular language. All respondents completed a hands-on orientation session led by a male and female staff member as well as a self-directed computerized tutorial on portable laptops on how to use the ACASI. Specifically, the orientation session included a power point presentation to explain the type of question and responses and how to maneuver on the computer during the interview. Respondents could choose the language they were most comfortable with (English: official language, Kiswahili: national language or Dholuo: native language among the local community). To ensure that respondents were ready to use the ACASI after the training, the last question in the tutorial offered the opportunity to repeat the training if they felt they had not mastered the process. Respondents were also told to ask for assistance at any time during the interview. Throughout the cohort study, respondents used the same gender voice.
Measures
We examined 12 dependent variables that were selected because of their importance in HIV transmission and their sensitive nature; yes or no were the choices for each response with the exception of age at sexual debut and lifetime number of partners: (1) regular current sex partner; (2) alcohol involved with first sexual intercourse; (3) drugs involved with first sexual intercourse; (4) age at sexual debut; (5) used condoms at first intercourse; (6) lifetime number of partners; (7) ever had anal intercourse; (8) ever had oral sex; (9) sexual intercourse at special occasion (e.g., funerals, planting time, and others) in last 3 months; (10) currently a person gives you money/gifts for sex; (11) ever had forced sex; and (12) ever treated for a sexually transmitted infection (STI). Specific questions are listed in Table 1. Independent variables included sex of the respondent, gender of the recorded voice on ACASI, and the interaction between sex of the respondent and gender of the recorded voice on ACASI. The continuous category “age at sexual debut” was categorized to reflect different age categories (preadolescence, adolescence, and late adolescence/adulthood) and the “lifetime number of sexual partners” was categorized to keep the zero to one category separate, and that at least 5% of the sample for each gender separately fell into the remaining categories.
Specific Questions from the Kisumu HIV Incidence Cohort Study Used for Analysis
Note. aResponses to this question were combined with the following question, “What best describes the first time you had sex?” so the variable reflected ever experiencing forced sex. One of the responses was “You did not want to have sexual intercourse, but you were physically forced.”
Analysis
To test for the influence of gender of the recorded voice on ACASI on respondent reporting of sensitive behaviors, we fit regression models in which a response to the sensitive behavior questions was the dependent variable and the sex of the respondent, gender of the recorded voice on ACASI, and the interaction between sex of the respondent and gender of the recorded voice on ACASI were the independent variables. For dichotomous variables, logistic regression was used, and for ordinal categorical variables, ordinal logistic regression was used and a test of the proportional odds assumption was conducted. If the interaction was not significant, a second regression model was fit, dropping the interaction term and testing the main effects. We adjusted for multiple testing by dividing the .05 α level by 12 and used the resultant .0042 α level as the level at which results were judged significant.
Results
A total of 849 respondents (423 males and 426 females) completed the ACASI questionnaire at screening for the incidence cohort study. Approximately one-half of respondents were assigned a matched gender of the recorded voice on ACASI (male respondent/male ACASI voice; n = 216; 25.4%; female respondent/female ACASI voice; n = 215; 25.3%). The other half were assigned an opposite gender of the recorded voice on ACASI (male respondent/female ACASI voice, n = 207; 24.4%; female respondent/male ACASI voice, n = 211; 24.9%). There were no major differences in demographic characteristics between the two groups of males or between the two groups of females (Table 2).
Demographic Characteristics of Persons Screened for the Kisumu HIV Incidence Cohort Study Participation by ACASI Voice
Note. ACASI = audio computer-assisted self-interview.
The n for each variable is not always the same due to refusal to answer.
Interaction of Gender of the Recorded Voice on ACASI and Sex of Respondent
Table 3 presents the results of the regression models for the 12 sensitive HIV risk sexual behavior questions. There were no significant interactions between gender of the recorded voice on ACASI and respondent gender, indicating that the effect of gender of the recorded voice on ACASI did not depend on respondent gender (i.e., the impact of ACASI gender of voice was the same for male and female respondents). For each of the 12 models, at least 90% of participants were included. For the remaining respondents, one or more of the outcome variables were missing; these were excluded from the regression models. Specifically, for the outcome variables, currently a person gives you money/gifts for sex and regular current sex partner, 80–81 responses were missing. For the outcome variables, sexual intercourse at special occasions and lifetime number of partners, 39–52 responses were missing. The remaining eight outcome variables were missing for less than 10 of the respondents. These numbers were too small to conduct analysis on differences.
Final Multivariable Analysis of the Influence of Respondent Gender, Gender of the Recorded Voice on ACASI and the Interaction of the Two on Sensitive Sexual Behavior Questions
Note. ACASI = audio computer-assisted self-interview. Kisumu HIV incidence Cohort Study Enrollment Period, 2007–2008 (N = 849).
ORA = odds ratio for gender of the recorded voice on ACASI (compares male voice to female voice).
ORG = odds ratio for gender of respondent (compares males to females).
PAG = p value for the interaction of gender of the recorded voice on ACASI and gender of respondent.
PA = p value for gender of the recorded voice on ACASI.
PG = p value for gender of respondent.
The n for each variable is not always the same due to refusal to answer, which was too rare to analyze for any gender-related trend.
*We adjusted for multiple testing by dividing the .05 α level by 12 and used the resultant .0042 alpha level as the level at which results were judged significant.
Influence of the Recorded ACASI Voice on All Twelve Responses
There were no significant differences in responses to the 12 sensitive sexual behavior questions from participants who heard a male ACASI voice versus those who heard a female ACASI voice (Table 3).
Response Differences by Participant Sex
There were significant differences between male and female respondents (independent of gender of the recorded voice on ACASI) in responses to seven of the twelve variables (Table 3). Compared to women, men reported an earlier age of sexual debut and a higher number of lifetime sexual partners. Men had greater odds of reporting sexual intercourse at a special occasion in the previous 3 months, a relationship in which a person currently gives them money/gifts for sex, and ever being treated for an STI. In addition, men had lower odds of reporting use of a condom at first intercourse and ever having forced sex. For age of sexual debut, the test of the proportional odds assumption was significant, indicating that the odds ratio differed significantly across the range of the dependent variable. However, we computed the odds ratios separately for each of the two possible two-level variables that could be created from the three-level variable and found that the odds ratios differed only in their magnitude, not their direction. Thus, we report only one odds ratio from the original ordinal logistic model.
Discussion
This is the first study we are aware of that assesses the impact of gender of the recorded voice on ACASI on responses to sensitive questions asked of research participants in a developing country setting. Our results showed that gender of the recorded voice on ACASI did not significantly affect responses when sensitive HIV sexual behavior questions were asked of Kenyan males or females. One unexpected respondent gender-related finding was that receiving gifts for sex was more frequently reported by males than females (regardless of gender of the recorded voice on ACASI); there is little literature about this but gifts or favors to boys from girls has been reported as part of the dating ritual (Nyanzi et al. 2001). Otherwise, our findings of respondent gender-related differences in reported sexual risk/behavior are consistent with other (non-ACASI) studies (Todd et al. 2009; Zaba et al. 2004) that show sexual risk/behaviors differ by gender of respondent. Our results showing no effect of gender of the recorded voice on ACASI provide evidence of ACASI’s value.
Our results were consistent with studies that found few differences in responses to sensitive topics based on interviewer gender in FTF and telephone interviews (Johnson and DeLamater 1976; Johnson and Moore 1993; Kane and Macaulay 1993). Furthermore, in one of the two previously reported studies of gender of the recorded voice on ACASI, there were no differences found in responses among men who have sex with men in the United States to sexual behavior questions asked by male versus female ACASI voices (Fahrney et al. 2005). The one exception was the report of unprotected receptive anal sex with an HIV unknown status partner, which was more frequently reported to a female ACASI voice than to a male (Fahrney et al. 2005). Preliminary results from the other study, which assessed gender of the recorded voice on ACASI among adults living in Baltimore, found that gender of voice did not have a large effect on responses (Rogers et al. 1996). Our study is generally consistent with preceding ones in that there were no significant differences in responses to sensitive questions based on the gender of the recorded voice on ACASI.
Studies on potential interviewer bias in developing countries deserve special attention because cultural and traditional values make open reporting of sensitive behaviors difficult (Mensch et al. 2003). For instance, adolescent sexual activity reported in FTF interviews has been judged inconsistent and implausible (Hewett et al. 2004), likely because of the influence of social norms and the inclination to reduce social distance or to be like the interviewer (Hewett et al. 2004). In an FTF survey of family planning in Nigeria, male interviewers were associated with inhibited responses to sexual questions because households in the state of Kano were Islamic and the male head of the family must be present when a male is conducting an interview (Becker et al. 1995). In another FTF household survey in Mexico, urban men gave different responses to female interviewers than did men living in other areas, suggesting the possibility of regional differences or differences between cities and rural areas in terms of response bias based on gender of interviewers (Flores-Macias and Lawson 2008). Finally, in a study of male and female sex workers in Mombasa, Kenya, sensitive behaviors such as intravenous drug use, rape, and group sex were more likely to be reported by persons responding to ACASI versus FTF interview (van der Elst et al. 2009).
Our results add to the literature regarding possible bias based on interviewer method/gender in developing countries. It is reassuring that gender of the recorded voice on ACASI did not influence responses to sensitive questions asked of male or female adults living in Kisumu, Kenya. One possible interpretation of this result is that the voice was perceived as disconnected from the person and considered only as a voice reading the questions and responses.
Our study had some limitations. First, our research participants were volunteers and may not be representative of the Kisumu population. Second, there may be response bias in that the respondents who refused to answer some questions were excluded from the analysis. However, generally, response rates were high, with most models including at least 90% of respondents. Finally, our study used a single male voice and a single female voice. The strength of our study, however, is that it included both males and females who underwent a well-balanced randomization that assigned them to gender of the recorded voice on ACASI; thus, we conclude that the absence of any effect of gender of the ACASI recorded voice on responses received is likely to be valid and meaningful. The fact that responses given by male respondents differed from those of female respondents in ways that are consistent with other published behavioral data provides further assurance about the validity of the study findings.
Conclusion
In our study in Kisumu, Kenya, we used ACASI to collect HIV risk behavior data and to investigate whether the gender of the recorded voice on ACASI affected responses received from male and female respondents. No differences based on gender of the recorded voice on ACASI were found. A recent review (Langhaug et al. 2010) reported that computer-assisted interviews like ACASI can improve data quality in developing countries. Our results lend further confirmation to the potential usefulness of ACASI in developing countries.
Footnotes
Acknowledgments
The authors thank all study participants, KICoS staff as well as Drs. Kayla Laserson and Peter McElroy for their assistance with study design and protocol development. They also thank the director of the Kenya Medical Research Institute/Center for Global Health Research (KEMRI/CGHR), Dr. John Vulule, for his support. This article was published with the approval of the KEMRI/CGHR director.
Authors' Note
The findings and conclusions in this report are those of the authors and do not necessarily represent the views of the Centers for Disease Control and Prevention. All authors contributed to the interpretation of the results and in editing and revising the manuscript. Sherri Pals and Sanjyot Shinde researched and conducted all statistical analyses and wrote the analysis section. Eleanor McLellan-Lemal provided input into variable selection for the models and reviewed the article. Deborah Gust drafted the article and provided input on variable selection. Tim Thomas, Wairimu Chege, Robert Chen, Lisa Mills, and Anne Gumbe provided critical review of the manuscript. Isaiah Oloo initiated the study, provided input into variable selection for the models, and contributed to the writing of the article.