Effect of education on attitude towards domestic violence in Nigeria: an exploration using Propensity score methodology

26 Background: Experimental studies remain the gold standard in making causal inference. 27 However, using experimental studies to estimate the effect of education on attitude towards 28 domestic violence (ATDV) is not feasible due to ethical issues. Propensity Score Methodology 29 (PSM) can be used to overcome this challenge. PSM is a statistical technique used in 30 observational studies to estimate the effect of interventions by accounting for covariates that 31 predicts the treatment. Therefore, PSM was used to investigate the effect of education on ATDV 32 among men and women in Nigeria. 33 Methods : A total of 14,495 and 33,419 records were extracted for men and women respectively 34 from the 2016-2017 Multiple Indicator Cluster Survey (MICS) in Nigeria. The outcome variable 35 was ATDV. The treatment variable was education while the covariates were age, residence, 36 geopolitical zones, marital status, ethnicity, parity, wealth index, alcohol use and media exposure

(use of television or radio).For the PSM analyses, selection bias was checked among the levels of education using the multinomial logit regression.Propensity scores (PS) and PS weights were generated for the treatment variable and average treatment effects (ATE) of ATDV were estimated using logistic regression that combined regression adjustment and inverse-probability weighting.Descriptive statistics, odds ratios and 95%CI were presented.

Results:
The mean age of men and women were 30.8±10.2 years and 29±9.4 years respectively.About 16% men 14% women had tertiary education.The proportion of men and women who justified domestic violence (DV) was 22% and 34.5% respectively.Results from the multinomial logit model showed the existence of selection bias between the covariates and level of education (p<0.05).The selection bias was effectively corrected (SD diff ≈ 0, Variance ratio ≈ 1) after estimation of PS. Results from the PSM showed that the odds of ATDV decreased as level of education increased.Men (AOR = 0.84, 95% CI: 0.78, 0.92) and women (AOR=0.94,95%CI: 0.80, 2.22) who attained tertiary education were less likely to justify DV in comparison to their uneducated counterparts.

Conclusion:
Education played a crucial role in ATDV among men and women in Nigeria.
Tertiary education was protective for ATDV among men and women.The use of PSM effectively controlled for selection bias in estimating the effect of education on ATDV.PSM will enable researchers make causal inference from non-experimental/ cross-sectional studies in situations where randomized control trials are not feasible.
Keywords: Propensity score, attitude towards domestic violence, treatment effect, selection bias.

Background
Experimental studies remain the gold standard when measurement of causal relationship is of interest.Scholars solely rely on Randomized Control Trials (RCT) to make causal inference in various fields of research.However, randomization, manipulation and intervention are impossible in some research especially in evaluating the effect of programs (Oliver et al., 2002).
For instance, it will be unethical for a researcher to deny some set of people access to education program because of research.Similarly, it will be unacceptable for a researcher to expose some women to violence and watch if their access to maternal health care will be poorer than those who were not exposed to violence (Kean et al., 1991;Sayar et al., 2019).However, analysis of observational data is an option, but the generalizability and the reliability of such findings are questionable especially in studies where causal factor is of interest.The major problem of nonexperimental study is "selection bias" which is known as the systematic difference between the treatment (exposed) and control (non-exposed) group based on any number of covariates (Rosenbaum & Rubin, 1984).This systematic difference (selection bias) was corroborated by Shadish in a study where study participants who self-selected themselves into training group performed better than those who were randomly assigned to the same training group (Shadish et al., 2006).Findings from Shadish study confirmed the claim of Rosenbaum and Rubin that participants who were not randomly assigned to treatment will tend to give better report on the treatment or the exposure.
All efforts to adjust and correct for selection bias such as structural equation modeling (SEM), adjusted regression and Randomized Control Trials (RCT) showed no improvement (Cepeda et al., 2003).Only the use of PSM can effectively control for the selection bias (Arikan et al., 2018).Many studies have used PSM to address the problem of selection bias in quasiexperimental and cross-sectional study designs (Feng et al., 2012;Rubin, 1997;Shadish et al., 2006;Yang et al., 2016;Yaya et al., 2019).PSM is a statistical method that has proven useful for evaluating treatment effect when using non-experimental or observational data (Guo & Fraser, 2015).PSM is used when researchers need to assess the effect of covariates on the outcome variable using survey data, census, administration data, and other observational data without any intervention by random assignment rules (Rubin, 1997).
The Nigerian Demographic and Health Survey (NDHS) report showed that 35 percent of women and 25 percent of men justified DV in Nigeria (DHS, 2013).ATDV has been identified as an indicator of the degree of social acceptance of DV and a known predictor of victimization and perpetration of DV.People's ATDV determines whether such violent acts will be reported or not (National Bureau of Statistics & UNICEF, 2017;Okenwa-Emegwa et al., 2016).A preponderance of educated men and women who justified DV was reported for reasons like; wife burns the food, argues with him, goes out without telling him, neglects the children, or refuses sexual intercourse with him (Okenwa-Emegwa et al., 2016).The magnitude, extent, and predictors of ATDV against women have been examined among men and women (Fawole et al., 2005;Okenwa-Emegwa et al., 2016).Factors such as Islamic religion, residency in the northern region, the South-South region, low levels of education and low household wealth have been reported to influence justification of DV.Of the reported associated factors of IPV and ATDV, studies have implicated education but majority of the evidence have been based on observational studies which has limitations when it comes to "causal inference".It is on this premise that the present study is aimed to employing PSM to investigate the effect of education on ATDV.
Further, the use of PSM to estimate the effect of drug use on violent behaviors while adjusting for selection bias among students in Southwest, Nigeria showed that drug use was associated with the likelihood of violent behavior.(Yusuf et al., 2014).Also, IPV has been linked as a risk factor for maternal health care utilization and poor pregnancy outcome using PSM (Yaya et al., 2019).
Studies have shown that higher level of education were protective against the risks of DV among men and women (Bates et al., 2004;Koenig et al., 2003;Okenwa-Emegwa et al., 2016;Wang, 2016).Since ATDV is an indicator of the degree of social acceptance of DV and a known predictor of victimization and perpetration of DV (National Bureau of Statistics & UNICEF, 2017), it was important to investigate whether education will also be a protective factor for ATDV among the general population so as to be able to make policies that will protect current and potential victims of domestic violence and enhance a protective ATDV among the perpetrators.
The objective of this study was to examine the effect of education on ATDV among men and women in Nigeria using PSM.We hypothesized that PSM will improve the estimation of the effect of educational level on ATDV among men and women.

Study design and setting
We used 2016-2017 Multiple Indicator Cluster Survey (MICS5), a cross sectional study carried out among adult (men and women) aged15 to 49 years in Nigeria.Nigeria is the most populous African country with estimated population of about 206 million inhabitants consisting 99.1 million females (Thomas & Crow, 2020;Worldometer, 2020).Nigeria has 36 states and a Federal Capital Territory (political divisions).Nigeria has more than 50 ethnic groups among which Yoruba, Hausa/Fulani, and the Igbo are the dominants while Islam and Christianity are the predominant religions practiced.

Study population and sampling procedures
The study population included men and Women who are between the ages of 15 and 49.The survey used the sampling frame to determine the enumeration areas (EAs), local government areas (LGAs), states, and zones in Nigeria as prepared in the 2006 Population Census of the Federal Republic of Nigeria.Further details of the sampling procedures were provided in the MICS5 report (National Bureau of Statistics & UNICEF, 2017).For this analysis, records of men and women who responded to the questions on ATDV were sorted giving a total of 14,495 and 33,419 records of men and women respectively.

Study variables
DV was defined as "any use of physical, sexual, psychological or economic violence of one family member, irrespective of person's age, gender or any other personal circumstance of the victim or the perpetrator of violence".The outcome variable was ATDV categorized as "DV justified" and "DV not justified".ATDV was measured by asking the respondents the following question.In your opinion, is a husband justified for hitting or beating his wife in the following situations: If she burns the food?(YES/NO).Any respondent who said yes to any of the five questions above was said to have justified DV and whosoever said no to all the five questions does not justify DV.The treatment variable was Educational level, while the covariates were age, religion, occupation type, residential type, geopolitical region, marital status, wealth index, ethnicity, number of children, age at first sex, alcohol use, tobacco use, and media use.

Data analysis
The demographic, socio-economic and lifestyle characteristics were described using frequency tables and percentages.Association between the treatment variable (educational level) and all the categorical variables were tested using the chi-square test.The PSM was thereafter used to estimate the effect of level of education on ATDV.

Techniques used in propensity score methods
The approach was in three stages.First, we checked for imbalance (selection bias) between the treatment variable and the covariates using multinomial regression.Each of the study covariate was used as the outcome variable in the model and the treatment variable (Educational level) as the explanatory variable in the model, Where πij is the probability of a response of the dependent that is greater or equal to a given category (i=2…4), πiJ is the probability of the response less than the given category (i=1), αj is a constant and βj is a vector of regression coefficients, for j=1,2,…,J−1.Xi is a vector of the .
Stage three was achieved by using the "tebalance summary" on "stata MP 14" to check if the standardized difference of the weighted scores are close to zero and the variance ratio for the weighted scores are close to one for all the covariates (SD diff ≈ 0, Variance ratio ≈ 1).If the result obtained satisfied the above criterion (i.e SD diff ≈ 0, Variance ratio ≈ 1), then selection bias has been corrected (i.e covariates are balanced) otherwise the selection bias has not been corrected.Lastly, we used the "teffect ipw" command on stata MP 14 to estimate the effect of the treatment (level of education).The "teffect ipw" command conducted a logistic regression that combined regression adjustment and inverse-probability weights between the study outcome variable ATDV and the propensity weight of the treatment variable.This provided the average treatment effect (ATE) which measures the effect of the PSW of educational level on ATDV.
Also, the potential outcome means (PO mean) which measures the effect of education on ATDV without the use of PS.Data were weighted to reflect educational level differentials in the population of men and women.Descriptive statistics, odds ratios and 95%CI were presented.All analyses were conducted at 5% level of significance using stata MP 14 (StataCorp, 2015).

Respondents profile
Information about the socio-economic, demographic characteristics of women were presented in

Selection bias
Result from the multinomial logit model fitted to check for selection bias was presented in additional table 1 and 2 for male and female respectively.The result revealed the presence of selection bias in men and women's data (p <0.05).

Weighted propensity scores
Table 3 shows the standardized difference and variance ratio of the weighted PS for men and women.The results showed that the standardized difference for the weighted are all close to zero, and the variance ratio are all close to one.This implied that selection bias has been addressed by PSM.Also, the similarity in the trends for each level of education presented in figure 1 and 2 showed that there is a good overlap in the estimated PS for educational level among men..5 -0.5 -0.4 0 0.9 0.03 1.0 -0.9 3-4 -0.6 -0.5 -0.5 0 1 0.01 1.0 -0.9 >4 0.31 1.9 0.27 1.8 0.41 2.2 -0.9 -0.9 0.09 1.0 Alcohol No 0.56 1 0.62 0.9 0.66 0.9 -1.1 0.04 1 0.14 1.0 Smoke No -3.8 -4.3 -0.1 2.7 0 0.9 -1.

Treatment effect for attitude towards domestic violence among men
Result from the logistic regression that combined regression adjustment and inverse-probability weighting was presented for men and women in table 4. In comparison with uneducated men, those who have attained tertiary level of education (AOR = 0.84, 95% CI: 0.78, 0.92) were less likely to justify DV.Similarly, the Yorubas (AOR = 1.12, p =0.141, 95% CI: 0.96, 1.31) were more likely to justify DV relative to Hausa men.The same pattern was observed for men from the rich wealth quintile (AOR = 1.07, 95% CI: 0.99, 1.17) compared to poor men.Also, men who were exposed to media (AOR = 1.02, p =0.076, 95% CI: 1.00, 1.03) were more likely to justify DV in relative to their unexposed counterparts.

Discussion
Effect of education on ATDV was assessed among men and women in Nigeria using a nationally representative data, where ATDV was the main outcome variable.Selection bias was detected in the data which led to the use of PSM since it's capable of minimizing selection bias in the data.
PSM has been proved to be effective and it has been used in previous studies (Yang et al., 2016;Yaya et al., 2019;Yusuf et al., 2014).This study showed that lower proportion of men justified DV compared to women, although this rate was higher than that of Ukraine and Ghana, but almost in line with the percentage reported in Moldova and Namibia (Sardinha & Catalan, 2018).
However, the disparity in the descriptive findings could be a result of the differences in the characteristics of the countries such as cultural beliefs and level of campaign against DV in the different countries.Arisi and Oromareghake reported that some cultures in Nigeria considered women as inferior beings, only useful in the kitchen, for pleasure and temptation (Arisi & Oromareghake, 2011).Also, it was known as common practice among men that women must kneel to beg their husbands when they are been beaten by their husband (Arisi & Oromareghake, 2011).Krause also corroborated the findings by further explaining that some cultures considered those acts of wife beating as a legitimate requital for a wife's defiance rather than seen it as 'violence' (Krause et al., 2016).Higher proportion of women justified DV in this study.This findings was similar to that of a report in Palestine and was buttressed that victims of DV were restrained from justifying DV to avoid marital separation as it could affect the children and their sustenance (Haj-Yahia, 2005).
Our results showed that only men who had primary education, secondary education and tertiary education were less likely to justify DV which was contrary to the previous finding where men who had primary and secondary education justified DV (Okenwa-Emegwa et al., 2016).This study and the previous study used a nationally representative data and the definitions of ATDV were similar, but the disparity could be as a result of the differences in the methods of analysis i.e the PSM that was used for this study has addressed the selection bias in the data thereby providing a better estimate (Cepeda et al., 2003).This paper has its limitations, PSM is only capable of adjusting for selection bias, and other types of bias such as measurement bias may not be addressed by PSM.However, this limitation does not erode the strength of this study as it added to knowledge about statistical methodology and alternative to improve findings from nonexperimental studies.

Conclusion
Education played a crucial role in ATDV among men and women in Nigeria.Tertiary education was protective for ATDV among men and women.The use of PSM effectively controlled for selection bias in estimating the effect of education on ATDV.PSM will enable researchers make causal inference from non-experimental/ cross-sectional studies in situations where randomized control trials are not feasible.

Supplementary Files
This is a list of supplementary les associated with this preprint.Click to download. AdditionaldocumentsTablesdec28.12.2020.docx [A] If she goes out without telling him?(YES/NO),[B]  if she neglects the children?(YES/NO),[C] If she argues with him?(YES/NO),[D]  If she refuses to have sex with him?(YES/NO),[E] covariates.At the second stage, we estimated generalized PS expressed as  ( , ) =  ( = | = ) which is the generalized PS of receiving treatment dose d for participants k with observed covariate X.The inverse of the PSW were obtained for participants.The inverse PSW is expressed as 1  ( , )

Figure 1 :Figure 2 :
Figure 1: Overlap plot for the propensity score of level of education (Men) equation modeling NDHS: Nigerian Demographic and Health Survey Oyindamola Bidemi Yusuf: Research idea, data analysis, manuscript writing and manuscript review.Rotimi F. Afolabi: Data extraction, manuscript writing and manuscript review.Olufunmilayo I. Fawole: Research idea, manuscript writing and manuscript review.

Figures Figure 1
Figures

Table 1
participated in this study, 22% justified DV.About 10.7% had no education while 17.3% had tertiary education.Close to half (48.2%) of the respondents were married.More than half