Are condom advertisements targeting widowed Kenyan women?

I recently used the Kenya Demographic Health Survey (a nationally representative data set) and ran some numbers to see how condom marketing influences use amongst women. The findings were pretty bizarre. I was just interested to see if condom marketing actually translates in to condom use. As I was finishing up my study the Kenyan government put out this interesting condom commercial that was censored almost immediately.

Here’s what I found: 

Women, who are widowed, divorced, or not living with their partner and who were exposed to condom messaging in magazines had significantly greater odds of using condoms as a current form of contraception. Exposure to condom advertising on the radio, T.V, and billboards had no significant effects. This study uses a sub-sample of 4 155 Kenyan women who have been sexually active in the past four weeks. Don’t take this at face value though, have a read through my data, definition, and methods (and because this is super academicy, since I wrote it for university, skip to the tables and jump to “explanation of the numbers” : p ).

Where’d I get my data? How are my variables defined?

The data used comes from the 2008-2009 Kenya Demographic and Health Survey (KDHS). This is nationally representative dataset that provides updated information on the health situation in Kenya. The 2008-2009 KDHS is the fifth national survey conducted in Kenya. A sample of 4 155 women, who have been sexually active in the past four weeks is examined.

Outcome variable (Y. Variable)

The outcome variable is a binary variable, of those individuals who have been sexually active in the past four weeks and use a condom (coded 1) and those individuals who have been sexually active in the past four weeks and use other or no contraception (coded 0). This variable is labeled “condomnow”

(1) This variable was created by combining two variables: (a) What is your current contraceptive method? And the responses are: not using, the pill, IUD, injections, condom, female sterilization, male sterilization, periodic abstinence, withdrawal, other, norplant, lactional amenorrhea, and female condom. (b) Recent sexual activity, by timing of last sexual intercourse. And the responses are: never had intercourse, active in last four weeks, have not been active in the last four weeks due to postpartum, and have not been active in the last four weeks.

(2) The variable was created on selecting only those women who have been sexually active in the past four weeks. This was done because the condom-messaging questions ask if the participant has seen the messaging in the past several months. Therefore, this sample of women is selected to account for this exposure.  I am inferring that condoms were used after the participant was exposed to condom messaging.

Predictor Variables (X. Variables)

In addition to the combining and recoding noted below, please note that a new variable was created for each predictor variable that only included the sample of women that have been sexually active in the past four weeks.

(1) Main predictors

Condom advertising is done through four mediums: radio, T.V, magazines, newspapers, and billboards. There are four variables that ask participants if they have been exposed to condom messages through any of these platforms. The four variables are of main interest and will help answer the research question.

(a) Condomrad: heard about condoms on the radio in the past several months. 2 missing values were dropped. Variable was recoded, 0=no and 1=yes.

(b) Condomt: heard about condoms on T.V in the past several months. 2 missing values were dropped. Variable was recoded, 0=no and 1=yes.

(c)Condomag: heard about condoms in the newspaper or magazine in the past several months. 2 missing were dropped values were dropped. Variable was recoded, 0=no and 1=yes.

(d) Condobill: heard about condoms on billboard(s) in the past several months. 2 missing values were dropped. Variable was recoded, 0=no and 1=yes.

(2) Control variables

Five control variables of scientific importance were selected: age, region, education, religion, and marital status.

(a) Age, ranges from 15 – 49. Age was selected since condom use may vary by age. Variable name: v012.

(b) Region, was selected given the differences in HIV prevalence across Kenya. Since condoms provide some protection against HIV, perhaps condoms are used more aggressively in high risk regions. Furthermore, perhaps condom advertising is done more aggressively in high risk regions.

Region was combined in to three categories based on current H.I.V prevalence. The original variable has eight categories. Low risk region, this includes Eastern and North Eastern Kenya where H.I.V. prevalence is low around 0.9-3.5%. Medium risk region,  This includes Rift Valley, Central, Coast, Nairobi, and Western. H.I.V. prevalence ranges from 4.2-7%. Highrisk region, this includes Nyanza which has H.I.V. prevalence of about 14%. (KDHS Fact Sheet). Categories were merged to address the “perfect predictor” error in the logistical regression. After the error was displayed I noticed that certain regions had very few observations. As a result categories were combined and the error was managed. Variable name: region11.

(c) Education is included to control for the impact greater education has on knowledge of contraception. Variable name: educ.

(d) Religion is included to control for the impact religious teachings have on contraceptive use. Typically, Judeo-Christian and Islamic faith advocate against the use of condoms and other safe sex measures.

Religion was recoded in to four categories. “No religion” and “other religion” were combined. The original variable has five categories. The following categories are in the final variable: Catholic, Protestant/Other Christian, Muslim, and No religion/Other. Categories were merged to address the “perfect predictor” error in the logistical regression. After the error was displayed I noticed that certain categories had too few respondents (i.e. no religion and other religion). As a result categories were combined and the error was managed. Variable name: relig.

(c) Marital status, is used as a proxy for: (i) women trying to get pregnant[1] [i.e. not using condoms or other contraception], (ii) women using it as contraception, and (iii) women [or their partner] using it as H.I.V. protection who are having extramarital affairs. To my knowledge there is no question that asks if a woman is trying to conceive. Marital status is used as a proxy.

Marital status was recoded in to three categories. (i) Married and living together were combined to create a new category. (ii) Divorced, widowed, and not living together were combine to create another new category. The original variable had six categories. The following categories are in the final variable: never married, married/living together, divorced, widowed, and not living together. Categories were merged to address the “perfect predictor” error in the logistical regression. After the error was displayed I noticed that certain categories had too few respondents (i.e. not living together, and widowed). As a result categories were combined and the error was managed. Variable name: maritalstatus.

(3) Other variables of interest

(a) Media awareness related variables: The following variables were included to control for the frequency of media encountered. For instance, a person who watches T.V. almost every day has a higher probability of being exposed to condom advertising on T.V.

(i) Frequency of reading newspaper or magazine. This variable has four categories: not at all, less than once a week, at least once a week, and almost every day.  Nine missing values were dropped. Variable name: reading.

(ii) Frequency of listening to radio.  This variable has four categories: not at all, less than once a week, at least once a week, and almost every day.  Nine missing values were dropped. Variable name: listening.

(iii) Frequency of watching television.  This variable has four categories: not at all, less than once a week, at least once a week, and almost every day.  Nine missing values were dropped. Variable name: watching.

(b) Risk related variables: The following variables were included to control for knowledge about H.I.V/AIDS and perceived risk of H.I.V/AIDS.

(i) To avoid H.I.V/AIDS: abstain from sex. Responses: yes, no. Two missing values were dropped. Variable name: abstainhivkn.

(ii) To avoid H.I.V/AIDS: use condoms. Responses: yes, no. Two missing values were dropped. Variable name: condomshivkn.

(iii) To avoid H.I.V/AIDS: limit sex to one partner.  Responses: yes, no. Two missing values were dropped. Variable name: sexparthivkn.

(iv) To avoid H.I.V/AIDS: limit number of sex partners.  Responses: yes, no. Two missing values were dropped. This variable was not used in the final model.

(v) To avoid H.I.V/AIDS: avoid sex with prostitutes. Responses: yes, no. Two missing values were dropped. Variable name: prostithivkn.

(vi)To avoid H.I.V/AIDS: avoid sex with persons who have many partners. Responses: yes, no. Two missing values were dropped. This variable was not used in the final model.

(vii) To avoid H.I.V/AIDS: avoid sex with drug users. Responses: yes, no. Two missing values were dropped.  This variable was not used in the final model.

(viii) Do you think your chances of getting aids are:  small, moderate, great, no risk, and has aids. Variable name: hivrisksr.

Methods: What models did I use? What steps did I take to determine which variables I kept and discarded? 

Model One: Logistic Regression The following steps were taken before the main model was ran:

(1) The outcome variable, condomnow, was run individually with all predictor variables.

(2) All predictor variables that had at least one significant category (p <0.1) were kept. The following three variables were not included in the main model since they were insignificant (p >0.1): (i) To avoid H.I.V/AIDS, limit number of sex partners. (ii) To avoid H.I.V/AIDS, avoid sex with persons who have many partners. (iii) To avoid H.I.V/AIDS,  avoid sex with drug users.

(3) A final simultaneous logit model was run.

Model Two: Interaction, condom advertising (magazines)*marital status I was interested in running an interaction model with the main predictor variables and marital status. I hypothesized that marital status may moderate the outcome variable through condom advertising since, despite exposure, those married may not be motivated to use condoms.

(1) I first ran the interaction with all condom advertising variables [condomrad, condomt, condomag, and condobill].

(2) I re-ran the model with only the significance interaction term to see if the significance increases when the other interaction terms were eliminated.

My Findings!