While measuring the number of people in modern slavery remains a challenge, substantial improvements have been made in this field in recent years. In 2017, the inaugural Global Estimates of Modern Slavery were produced by the International Labour Organization (ILO) and the Walk Free Foundation (WFF) in partnership with the International Organization for Migration (IOM). The regional estimates produced through this collaboration form the starting point for the estimation of national level estimates presented here.

Global Estimates of Modern Slavery

The Global Estimates were comprised of two sub-estimates: an estimate of forced labour and an estimate of forced marriage. The sub-estimate of forced labour was then further broken down into three categories: forced labour in the private economy, forced sexual exploitation, and state-imposed forced labour.

Figure 1Typology of modern slavery

As no single source provides data that are suitable for the measurement of all forms of modern slavery, a combined methodological approach was adopted for the Global Estimates of Modern Slavery, drawing on three sources of data to calculate the sub-estimates:

1. The central element of the methodology is the use of 54 specially designed, national probabilistic surveys involving interviews with more than 71,000 respondents across 48 countries. The estimates of forced labour in the private economy (excluding the sex industry) and forced marriage were derived from these surveys. Only cases of modern slavery that occurred between 2012 and 2016 were included in these estimates, and all situations of forced labour were counted in the country where the exploitation took place. In the five-year reference period for the estimates, while surveys were conducted in 48 countries, men, women, and children were reported to have been exploited in 79 countries.1
2. Administrative data from IOM’s databases of assisted victims of trafficking were used in combination with the 54 datasets to estimate forced sexual exploitation and forced labour of children, as well as the duration of forced labour exploitation. This involved calculating the ratio of adults to children, and also of “sexual exploitation” cases to “labour” cases in the IOM dataset, which contained information on 30,000 victims of trafficking around the world who had received assistance from the agency. These ratios were then applied to the estimates taken from the survey data on forced labour of adults to arrive at an estimate of the number of children in forced labour and another estimate of “sexual exploitation.”
3. As the surveys focused on the non-institutionalised population, meaning that people in prisons, labour camps or military facilities, and other institutional settings are not sampled, the surveys are not suitable for estimating state-imposed forced labour. Instead, the estimate of state-imposed forced labour was derived from validated secondary sources and a systematic review of comments from the ILO Committee of Experts on the Application of Conventions and Recommendations relating to state-imposed forced labour.

Each sub-estimate was initially calculated as a flow estimate; that is, the total number of persons who were victims of modern slavery during a specified period of time between 2012 and 2016. The flow estimate was then converted into a stock estimate; that is, the average number of persons in modern slavery at a given point in time during the 2012 to 2016 reference period. The stock estimate is calculated by multiplying the total flow by the average duration (the amount of time in which people were trapped in forced labour) of a spell of modern slavery. The average duration of modern slavery was determined from the database of the IOM, containing records of assisted victims of trafficking who were registered during or after 2012.

A detailed explanation of the methodology underpinning the Global Estimates of Modern Slavery is available online.2

From global and regional to national estimates

The national estimates presented in this Global Slavery Index were calculated7 using individual and country-level risk factors of modern slavery. The final set of risk factors were selected from an exhaustive list of variables to optimally predict confirmed cases of forced labour and forced marriage. The model was then used to generate average predicted probabilities of modern slavery by country. The regional totals in the Global Estimates of Modern Slavery were then apportioned based on each country’s average predicted probability of modern slavery. This process involved the following key steps:

1. Identifying risk factors of modern slavery. Using national surveys that included questions on experiences of forced labour and forced marriage to identify which variables were statistically associated with respondents in the survey who had been victimised, versus those who had not been victimised. This included using a series of statistical tests to identify relationships between instances of victimisation and other variables collected through the survey process (such as age, gender, marital status, education, urban/rural, employment, life evaluation, business ownership, and ability to live on current income). Country-level predictors of risk from the most recent Global Slavery Index Vulnerability Model were also included.
2. Predicting modern slavery. These risk factors were used to build a statistical model that best predicts occurrence of modern slavery at the individual level.
3. Estimating prevalence and aligning with Global Estimates of Modern Slavery regional estimates. Individual predictions were aggregated into risk scores at the country level. Whereas survey data on forced labour and forced marriage are not available for every country, a broader set of survey data covering variables such as age, gender, marital status and so on was available for 147 countries.8 Country risk scores were used to estimate country prevalence, based on the extent to which the country risk score deviated from the average regional risk scores. For example, if a country had the exact same risk score as the relevant region in the Global Estimates, then it was assumed that the prevalence in the country was the same as in the region.
4. Final calculation of estimated prevalence. Number of victims was then estimated by applying the estimated prevalence to population data for each country. To this “base” estimate, an estimate of state-imposed forced labour was added to determine the final estimated prevalence of all forms of modern slavery.

The process followed in each of these steps is detailed below:

1. Identifying risk factors of modern slavery

Data and variables

First, individual and country-level variables that have a significant relationship with forced labour or forced marriage at the individual level were identified. Data for this analysis were taken from Gallup World Poll (GWP) surveys conducted in 2014, 2015, and 2016,9 including a set of surveys with a module on modern slavery used to estimate the risk model, and a broader set of surveys used for prediction purposes, as well as country-level risk variables from the Global Slavery Index Vulnerability Model.

Estimation data and outcome variables

Estimation data were drawn from 54 surveys conducted in 48 countries which included a module on Modern Slavery, with a total sample of 71,158 individual interviews. This included:

• Fifty-three national surveys conducted through the GWP in 48 countries between 2014 and 2016, with a total sample of 57,158 individual interviews.
• A 2016 survey covering 15 Indian states: Andhra Pradesh, Bihar, Chhattisgarh, Himachal Pradesh, Jharkhand, Karnataka, Kerala, Madhya Pradesh, Maharashtra, Odisha, Punjab, Rajasthan, Telangana, Uttar Pradesh, and West Bengal, with a total sample of 14,000 individual interviews.

Cases of forced labour and forced marriage were identified with a series of screening and follow up questions as described earlier. On the basis of these questions, victims of forced labour were identified according to the following criteria:

• The work was involuntary (“Yes” to any of the screening questions), AND
• The work was under coercion or the menace of a penalty, AND
• The work occurred in the last five years.

Victims of forced marriage were identified according to the following criteria:10

• The marriage was involuntary (“yes” to the screening question), AND
• The marriage had occurred without their consent (forced marriage).

Prediction data and predictor variables

A broader set of data (“Prediction data”) was drawn from 433 GWP national surveys conducted in 155 countries between 2014 and 2016, with a total sample of 451,161 individual interviews. A total of 157 variables that could potentially be used to predict forced labour or forced marriage status were identified using the five dimensions of the Walk Free Foundation Vulnerability Model as an organising framework. These included:

• One hundred twenty-two individual-level variables from the GWP core questionnaire, which collects information on basic demographic variables such as the respondent’s age, gender, educational attainment, marital status, employment status, urban/rural location, and number of adults (15 and older) and children (under 15) in the household, as well as development-oriented topic areas including law and order, food and shelter, health, government and politics, business and economics, citizen engagement, education and families, environment and energy, social issues, religion and ethics, work, and well-being.
• Thirty-five country-level variables from the Walk Free Foundation Vulnerability Model.11

Not all GWP variables were fielded in each survey country12 during each of the three data periods (2014, 2015, 2016), which lead to varying levels of geographic coverage. A list of 19 independent variables with low levels of missing data was identified to maximize geographic coverage of a “base” driver model. In addition to the “base” model, four additional models were created with an increasing number of predictor variables, and corresponding decrease in geographic coverage.

2. Predicting modern slavery

Individual-level models

Several steps were undertaken, using the data noted above to identify a best-fitting prediction model. Forced labour and forced marriage were modelled separately as the two distinct forms of slavery are expected to be predicted by different subsets of variables. The probability of a given respondent13 reporting a case of modern slavery was estimated for each outcome (forced labour and forced marriage) separately, using a logit model14 of the form:

Equation 1

Where the logit of the probability of FL/FM for each individual i is a function of a constant term  (intercept), a vector of individual-level demographic control variables with values varying for each individual , and with unknown coefficients , a vector of individual-level predictor variables , with values varying for each individual , and with unknown coefficients , and an individual error term .

Of the 157 variables available, a subset of variables was selected based on statistical and theoretical criteria in order to enhance the predictive power of the model, while maintaining explanatory relevance. Variables were excluded on the basis of having no significant association with either forced labour or forced marriage:

• no multivariate significant association with either forced labour or forced marriage when entered within their respective geographic block, and
• a high degree of multicollinearity, as expressed by variance inflation factors of 3 and above.

Finally, when variables are collinear or multivariate insignificant, priority was given to variables with greater theoretical relevance. For example, “confidence in judicial system,” which relates to issues of regulatory quality that have a direct bearing on modern slavery, is preferred over “confidence in financial system,” which may only have an indirect relationship with modern slavery.

Variables were entered into six models (numbered 1-6 in Table 1) to allow for the inclusion of a successively more exhaustive set of predictors. These models are nested hierarchically, with each successive model including all variables in the prior models, running from the simplest model that includes only seven demographic factors, to an “extended plus” model including 33 predictors of forced labour and 29 predictors of forced marriage (see Table 1 for final list of variables).

The models were estimated using survey data from the 48 countries where the modern slavery module was included. In order to estimate risk of modern slavery in countries available in the GWP without a modern slavery module, the probability of a positive outcome for each individual in the prediction dataset is calculated and then aggregated into a weighted average predicted probability at the country level. Table 2 shows the sample sizes and number of countries included in each of the estimation and prediction models. The demographic factors-only models showed relatively poor performance, so they were not used for prediction purposes. The “base” models, including a relatively small number of variables, have the widest geographic coverage (152 countries). The “extended plus” models, with the largest set of predictors, have the narrowest geographic coverage (116 countries for forced labour and 110 countries for forced marriage).

The predictive performance of each model was evaluated using a broad set of post-estimation goodness-of-fit metrics,16 which were calculated on the same set of respondents (i.e. that had data available for all variables) to ensure comparability. Results indicated good predictive power (AUC values greater than .70) for all models. The base model is used as it is most useful for estimation, but the other models, with a greater complexity in terms of predictors but similar predictive performance, are useful to validate the robustness of the base model, which maximises geographic coverage (see Table 1).

Multi-level models

After identifying the “base” model as the best for prediction and maximising geographic coverage, multilevel models (MLM) were fitted to the data in order to enhance the predictions of the individual-level models and take into account the hierarchical nature of these data. MLMs allow for the extrapolation of model results beyond the sample of 48 countries.

All multilevel models were estimated using Bayesian17 applied regression modelling18. The individual-level base model was fitted before being expanded sequentially. First by allowing intercepts to vary across countries according to a random effect:

Equation 2

Equation (2) is the same as the individual-level regression Equation (1), with the addition of a subscript to classify individuals in countries, and an additional coefficient and its associated distribution, representing a random coefficient that is allowed to vary by country.

Next, by modelling country-level variation in order to improve our predictions for countries where there was no modern slavery survey with country-level predictor representing the vulnerability score , with values varying for each country, and with an unknown coefficient :

Equation 3

The Vulnerability Model is ideal for this purpose, as it incorporates a robust set of external datasets aggregated to explain or predict the prevalence of modern slavery. An examination of the association between country-level prevalence estimates and vulnerability scores confirmed a moderate correlation (Pearson r =.33). 

The individual-level models showed that owning a business significantly increases the risk of being a victim of forced labour. Members of the Walk Free Foundation’s Expert Working Group indicated that this result was surprising, as business ownership was expected to be a protective factor and hypothesized that the result could be driven by “necessity entrepreneurs”: individuals who are forced into starting a small-scale business because of a lack of other income-generating opportunities.19  A preliminary set of regional regression analyses confirmed that business ownership was only a significant predictor of forced labour in South Asia and Sub-Saharan Africa, where not owning a business has a protective effect. Finally a cross-level interaction effect for “business ownership” was introduced:

Equation 4

With   representing the interaction between business-ownership dummy , with a coefficient that varies for the two levels of regional dummy variable .

Model Performance

An examination of model performance20 shows that all MLMs perform similarly. In the case of the forced labour models, all MLMs perform similarly, and better than the fixed-effects model. In the case of the forced marriage models, differences were negligible.

Actual vs. fitted prevalence plots for the 48 countries with modern slavery survey data showed a very close fit, without any clear outliers. The introduction of country fixed (and then random) effects represented a major improvement above and beyond the individual-level models, which relied on regional fixed-effects. While the simplicity of the country fixed-effects model is attractive, this approach would not achieve our goal of estimating slavery in countries that are not included in the sample.

In order to exemplify the benefit of the more complex MLMs, country intercepts were removed from the predictions to simulate new data including countries not previously surveyed. The random intercepts model shows a poorer fit with the actual values than the other two models (a perfect fit is exemplified by the red dotted line). A random intercepts model with country level predictors and (in the case of forced labour) a cross-level interaction provides the most comprehensive framework to undertake these inferences and was the model on which estimates were based.

A fuller description of the process by which the final model was achieved is set out in a forthcoming publication.21

3. Estimating prevalence and aligning with regional estimates from The Global Estimates

As summarised above, several concomitant risk factors for modern slavery, including demographic factors such as age, gender, and employment status – but also a variety of socio-economic and psychographic risk factors such as feelings about household income, life evaluation scores, and negative experienced affect – were identified in the analysis. Based on these risk factors, as well as country-level vulnerability scores, a hierarchical Bayes modelling approach was used to accurately predict the forced labour and forced marriage status of individuals and the average prevalence of modern slavery at the country level.

Average weighted predicted probabilities were then calculated for forced labour and forced marriage using this best-fitting predictive model.22 The process to arrive at estimated prevalence of modern slavery was undertaken in several steps, as follows:

1. Disaggregate regional-level estimates of modern slavery from The Global Estimates into 12 homogeneous subregions (11 broad ILO sub-regions, plus split Southeast Asia & Pacific).
2. Calculate subregional level prevalence of modern slavery for each subregion (for example, South Asia = 0.77%).
3. Impute risk factors for 20 countries that are missing GWP data, as an average over several multiple imputation approaches (hot deck, amelia, glm, random forests).
4. Adjust country risk by country of exploitation. The basic premise is to take a region such as “Receiving- Southeast Asia,” with Singapore, Malaysia, and Thailand, and apply an adjustment factor equal to the ratio of victims identified in the national surveys in Singapore, Malaysia, and Thailand to total exploited victims in the region. If no national surveys were conducted in a given region, we estimate that the prevalence is equal to modelled risk multiplied by population. This is calculated using the following steps:
   1. Calculate number of victims identified by the country surveys who are exploited in a different country.
   2. Code countries as either “net sending” or “net receiving (see Table 3). This was done on the basis of available information from GSI 2016, UNODC Global Report on TIP 2016, US TIP 2017, GRETA, and ILO Global Estimates of Migrant Workers.23Sending versus receiving status was coded by two independent coders. If there was agreement, the common code was maintained. However, if conflicting codes were allocated, the decisions were jointly reviewed and often resolved. In the event that no agreement was reached, a third party would be consulted.
   3. Calculate aggregate number of victims by place of exploitation in sending and receiving areas.
   4. Adjust down the risk score of sending regions that have a lower number of victims being exploited in country.
   5. Adjust up the risk score of receiving regions that have a higher number of victims being exploited in country. For example, the risk in “receiving” Southeast Asia (Singapore, Thailand, Malaysia) is adjusted up by a factor of 1.58, while “sending” Southeast Asia is adjusted down by a factor of 0.94.
5. Taking adjusted country risks, estimate prevalence in a country based on the regional prevalence and the distance between the adjusted country risk and the weighted average regional risk score, following these steps:
   1. Normalise the adjusted and imputed country risk scores to a 1-100 range, with 1=min risk, 100= max risk.
   2. Multiply the normalised risk score by the country population.
   3. Calculate regional prevalence by dividing aggregates for total modern slavery (excluding state-imposed forced labour) over total population.
   4. Calculate average normalised regional score by dividing the sum of normalised risk scores by the country population.
   5. Calculate country prevalence by multiplying the regional average by the ratio of normalised country risk score over the average normalised regional score.
6. To simplify, since normalised modern slavery risk in Afghanistan (39.1) is 2.89 times higher than the average risk in the South Asia region (13.6), we estimate that prevalence in Afghanistan is 2.89 times greater than the regional average, or about 2.2 percent.
   
   Only one exception is made, for Mauritania, where the survey estimate (2.1 percent) is used rather than the modelled risk score due to the distinct nature of slavery in the country. The practice is entrenched in Mauritanian society with slave status being inherited and deeply rooted in social castes and the wider social system. Those owned by masters often have no freedom to own land and cannot claim dowries from their marriages nor inherit property or possessions from their families.24  Given the evidence available that supports the higher survey estimate, that estimate is take from Mauritania and other countries in Sub-Saharan Africa are adjusted down to ensure totals are aligned with the Global Estimates of Modern Slavery.
7. A final calculation is performed to incorporate existing estimates derived from multiple systems estimation (MSE) in the Europe and Central Asia region. The predictive models are built on information from countries where the World Poll, including the modern slavery module, was conducted face-to-face. Countries where the World Poll is only implemented using computer-assisted telephone interviewing (CATI) have zero chance of selection for a modern slavery survey, meaning that, at present, we are not able to test how the risk factors will behave in CATI countries. Despite this, there is also no evidence to suggest that the risk factors will not act in the same way, for example, being female is very likely to remain a risk factor for forced marriage and poverty a risk factor for forced labour.
   
   Further, MSE has emerged as a suitable methodology for estimation in countries where surveys are not used and the methodology has been endorsed by several governments and international organisations. Several considerations precluded the use of MSE-estimates alone, notably, (i) the methodology is at an early stage of use in the modern slavery space with several refinements underway, and (ii) some forms of modern slavery, for example, forced marriage are at present unlikely to be captured in administrative records meaning that MSE alone would lead to an underestimate.
   
   While survey-based estimates are subject to a high level of uncertainty due to sampling and non-sampling errors, the two available MSE estimates employed different approaches and therefore show large variability across countries with similar risk profiles. In light of the considerations set out above, the average of the model-derived prevalence estimates and MSE-based prevalence estimates for the United Kingdom 25and The Netherlands26were set as anchor points within the region. This was applied as follows:
   1. Countries within the Europe and Central Asia subregions were divided into ‘MSE’ and ‘non-MSE’ sub-regions. In practice this aligns with CATI vs. F2F countries in the Gallup World Poll. For example, both Spain and Greece are in “Southern Europe”, but Spain is CATI, and hence its estimate is based on extrapolation from MSE, while Greece is F2F, and hence its estimate is based on non-MSE extrapolation (more below on each type).
   2. Set anchor points and extrapolate to other MSE countries: (i) Average prevalence estimate for the UK was set as the anchor point for Northern Europe (MSE sub-region, i.e. excluding Baltic states) and (ii) the average prevalence estimate for the Netherlands was set as the anchor point for Western Europe (all countries) and Southern Europe (MSE sub-region).
   3. The corresponding anchor point was then extrapolated to other countries in the region based on ratio of risk in the anchor to risk in the extrapolation country. For example, if average prevalence in the UK is 0.20% and modelled risk is 0.775, we estimate that prevalence in Sweden is given by the ratio of its risk to the UK’s risk, multiplied by the UK prevalence, or (0.645/0.775)*0.20%= 0.17%.
   4. Adjust the prevalence of non-MSE countries in Europe and Central Asia to ensure the total aligns with the Global Estimate. The number of victims from the Global Estimates who are unaccounted for in Europe and Central Asia following the MSE adjustment were calculated, then prevalence in non-MSE countries was calculated proportional to the risk of each country relative to the non-MSE population adjusted regional average risk.

4. Final calculation of estimated prevalence, including state-imposed forced labour

The process outlined in steps 1 to 3 produces prevalence estimates for all forms of modern slavery except state-imposed forced labour. Given the nationally-specific manifestations of state-imposed forced labour where it does occur, it was excluded from the steps outlined above. The final step involves aggregating the estimate resulting from the process set out above, with the estimates of state-imposed forced labour. A final estimate of the prevalence of all forms of modern slavery is then calculated. The resulting estimates are presented in Table 4.

Data limitations

Limitations of the source data

As with all empirical research, there are some limitations of the data used to produce the Global Estimates of Modern Slavery, within which the findings of this Index should be interpreted.

First, the set of surveyed countries that was used to produce the 2017 Global Estimates of Modern Slavery was treated as a random sample of the world and the global figure was estimated directly from that (that is, without first calculating national estimates). However, the selection of the countries to be surveyed was not random as countries were selected for specific reasons, including:

• Countries where prevalence is expected to be higher, thereby increasing the chance of identifying cases through a household survey. This leads to the selection of more "developing" and/or "source" countries than "developed" countries as a random sample survey is unlikely to identify cases in the latter;
• Where the mode of delivery is through face to face surveys, as opposed to telephone interviews, and
• To ensure regional representation so that the surveys could facilitate extrapolation.

Second, while regional estimates of prevalence of modern slavery were presented in the Global Estimates of Modern Slavery, critical gaps in available data were noted. These are particularly problematic in the Arab States, where only two national surveys were undertaken, none of which was in the Gulf Cooperation Council (GCC) countries despite the incidence of forced labour reported by different sources in such sectors as domestic work and construction in the GCC. Further, measurement of forced marriage among residents of countries within the region is particularly problematic where there are no surveys. Taken together, these gaps point to a significant underestimate of the extent of modern slavery in this region.

Similarly, it is usually not possible to survey in countries that are experiencing profound and current conflict, such as Syria, Iraq, Yemen, Libya, South Sudan, and parts of Nigeria and Pakistan. Yet it is known that conflict is a significant risk factor – the breakdown of the rule of law, the loss of social supports, and the disruption that occurs with conflict all increase risk of both forced labour and forced marriage. The lack of data from countries experiencing conflict means that modern slavery estimates in regions in which conflict countries are situated will understate the problem.

Similar coverage gaps exist for the Central and East Asia subregions where a larger number of surveys (only one in East Asia) were not able to be conducted for reasons that included: (i) mode of delivery was only by telephone, (ii) limited survey languages, (iii) consent of national authorities to the module on modern slavery was not given or was highly unlikely. Further, for countries in these subregions, none were identified as the country where exploitation took place by respondents in the 48 countries where surveys were implemented. As a result of these gaps, the estimates for countries within these subregions are likely to be conservative.

Limitations of the risk modelling

This analysis is not without the limitations inherent to any cross-sectional research endeavour. Our selection of variables is driven by both theoretical and statistical considerations, but unfortunately the field of modern slavery lacks a unifying causal theory that can be used to inform variable selection. Finally, we have a limited sample size of confirmed individual cases, which limits the extent to which we can expand our predictive models and enhance the accuracy of our predictions. Further surveys will lead to an increase in our sample, thereby enabling the study of more complex effects and refinement of the modelling.

Comparability with previous estimates

Due to substantial differences in scope, methodologies, regional groupings, and expanded data sources, the 2018 Global Slavery Index is not directly comparable to the previous edition. These differences are due to the joint development of the Global Estimates of Modern Slavery and, accordingly, the changes in the estimated number of victims at the national level cannot be interpreted against the previous Global Slavery Index. It is important to note the key differences in how the Global Estimates, the 2018 Global Slavery Index national estimates, and the 2016 Global Slavery Iindex estimates were calculated, these include:

• What we count: In the 2016 Global Slavery Index we identified gaps in the measurement of children across all forms of modern slavery and adults in forced sexual exploitation. These gaps were addressed when developing the methodology for the Global Estimates, which drew on both survey and administrative data from IOM to calculate sub-estimates for forced sexual exploitation and the forced labour exploitation of children. In addition, a more systematic approach to the measurement of state-imposed forced labour was adopted for the Global Estimates and is used here.
  
  The 2018 Global Slavery Index represents a “stock” estimate; that is, people in slavery on any given day in 2016.
• Where we count, where exploitation happens: The 2016 Index had too few survey countries to consistently count victims where they were exploited, which is not the case in the 2017 Global Estimates, which are based on a much larger number of survey countries. This change had the impact of increasing the number of victims counted in developed countries, with the exception of the Arab States. As noted previously, measures in that region are hampered by insufficient data.
• How we measure: While nationally representative surveys remain central to the process, the collaboration on a global estimate necessitated a change from the “bottom-up” approach of first calculating national estimates and then aggregating to a global total.

In the 2017 Global Estimates, the countries surveyed were treated as a random sample of the entire world and global and regional totals were estimated directly from that without first calculating national estimates. In the 2018 Index, national prevalence is calculated on the basis of a predictive model that takes individual and country-level risk factors into account. The results are then weighted such that they aggregate to regional totals from the Global Estimates of Modern Slavery.