Data source, sampling technique and sample size
This cross-sectional study used data from the Papua New Guinea Demography and Health Survey (PNGDHS) conducted from October 2016 to December 2018. This is the first demographic and survey conducted in PNG. The PNGDHS aimed to generate comprehensive data on demographic, maternal and reproductive issues such as fertility, family planning awareness and practices, breastfeeding practices, health behaviors, immunizations, domestic and intimate partner violence, among others. Through the Demographic and Health Survey (DHS) programme, technical support for the execution of the survey was provided by Inner City Fund (ICF), with the financial support of Papua New Guinea Government, Australian Government Department of Foreign Affairs and Trade, the United Nations Population Fund (UNFPA) and UNICEF . The sample for the 2016–18 PNGDHS was nationally representative and covered the entire population that lived in private dwelling units in the country. The survey used the list of census units (CUs) from the 2011 Papua New Guinea National Population and Housing Census as the sampling frame and adopted a probability-based sampling approach. Specifically, a two-stage stratified cluster sampling procedure was followed. The methodology and selection procedure details have been reported in the PNGDHS final report.
In summary, each province in the country was stratified into urban and rural areas, yielding 43 sampling strata, except the National Capital District, which has no rural areas. The division paid particular attention to urban–rural variations. Samples of census units were selected independently in each stratum in two stages. In the first stage, sorting the sampling frame within each sampling stratum to achieve implicit stratification and proportional allocation using a probability proportional-to-size selection was done. In the second stage of sampling, a fixed number of 24 households per cluster were selected with an equal probability systematic selection from the newly created household listing, resulting in a total sample size of approximately 19,200 households. To prevent bias, no replacements and no changes of the pre-selected households were allowed in the implementing stages. In cases where a census unit had fewer than 24 households, all households were included in the sample. A total of 17,505 households were selected for the sample, of which 16,754 were occupied. Of the occupied households, 16,021 were successfully interviewed, yielding a response rate of 96%. In the interviewed households, 18,175 women age 15–49 were identified for individual interviews; interviews were completed with 15,198 women, yielding a response rate of 84%. In this present study, the sample comprised 9,943 women who were in union (either married or cohabiting) during the survey. Thus, our analysis used data only on women who were in union during the survey.
Data availability and ethical consideration
The data have been archived in the public repository of DHS. The access to the data requires registration which is granted specifically for legitimate research purposes. Consent forms were administered at household and individual levels, in accordance with the Human Subject Protection. The dataset can be accessed at https://dhsprogram.com/data/dataset/Papua-New-Guinea_Standard-DHS_2017.cfm?flag = 0.
Main outcome and predictor variables
Current cigarette smoking was the outcome variable in this study. This was measured as having smoked cigarette in the last 24 h before the survey. Women in union were asked the question: Smoked cigarette in the last 24 h? Women in union current smoking status were classified as “No” (0): no current smoking in the last 24 h or “Yes” (1): smoking in the last 24 h. The key explanatory variable in this study was IPSV. This variable was derived from the optional domestic violence module, where questions are based on a modified version of the conflict tactics scale [14, 15]. Questions asked are in relation to physical, sexual or emotional violence experiences. In this study, the focus was on the experience of sexual violence. Three standard items including whether the partner ever physically forced the respondent into unwanted sex; whether the partner ever forced respondent into other unwanted sexual acts and; whether the respondent has been physically forced to perform sexual acts she did not want to were used to generate the experience of intimate partner sexual violence. For each of these items, the responses were ‘never’ ‘often’ ‘sometimes’ and ‘yes, but not in the last 12 months. However, for our analysis purpose, we created a dichotomous variable to represent whether a respondent had experienced sexual violence in the past 12 months. This was done by recoding the following responses: ‘never’ and ‘yes, but not in the last 12 months’ as “No” (0) and ‘yes’, ‘often’ and ‘sometimes’ as “Yes” (1).
Theoretically and empirically relevant demographic and socioeconomic variables were included as confounders. In all, we included twenty socioeconomic and demographic variables to adjust for in the modelling. These variables included age, region, religion, place of residence, highest educational level, literacy, marital status, residing with a partner, number of partner’s wives, partner’s age, partner’s education, health insurance cover, internet access, mobile phone ownership, watch television, listen to radio, read newspaper/magazine, occupation and wealth index. The selection of these variables was informed by their statistically significant associations with sexual violence and cigarette smoking in previous studies [1, 2, 4, 16, 17]. (See Table 1 for the details on the coding of the covariates).
Before the analysis, all missing data were removed. Both descriptive (frequencies, percentages, mean and standard deviation) and inferential (chi-square and modified Poisson regression) analytical frameworks embedded in STATA software version 13.0 (StataCorp LP, College Station, Texas, USA) were used. The statistical analysis followed some essential steps. We performed descriptive statistics such as frequencies and percentages to describe the sample. The Pearson’s Chi-square test was done to examine the differences in smoking cigarette by socio-demographic characteristics and IPSV. A modified Poisson regression, adjusting for demographic, social and economic variables, was also performed to model the association between IPSV and cigarette smoking, to estimate the relative risk (RR) of cigarette smoking [18, 19].
The study used the modified Poisson regression that incorporates the robust error variance procedure to optimize the accuracy of the estimates , as direct estimates of relative risk produce from modified Poisson regression modelling may be a preferred method for estimating population-level risk . We fitted four regression models. Model 1 included dependent and independent variables only; thus, was the base model. While adjusting for the theoretically relevant confounding variables, Models 2, 3 and 4, respectively introduced demographic and socioeconomic factors to investigate whether these variables play any role and might tamper the effects of IPSV on cigarette smoking. Before the regression analysis, diagnostics checks for multicollinearity were conducted using the variance inflation factor (VIF). In this analysis, none of the VIF scores exceeded the value of 2.38, suggesting no multicollinearity. The results of the regression analyses were presented as crude relative risk (CRR) and adjusted relative risk (ARR) at 95% confidence intervals (CIs). All the estimates provided in this study are derived by applying appropriate sampling weights supplied by PNGDHS, 2016–18. A statistical significance threshold of p ≤ 0.05 was selected.