Trade liberalization, social policies and health: an empirical case study

FsQCA is carried out in three steps. First, outcome indicators, cases and causal conditions are identified. Included cases are then assigned membership scores for each of the outcomes and causal conditions. In a conventional QCA, cases are either members of the set created by the indicator (with a membership score of 1) or not (with a membership score of 0). In fsQCA by contrast, cases can have partial membership anywhere in the range of 0-1 [25]. It was decided to use fsQCA instead of conventional (i.e., ‘crisp-set’) QCA since cases in this study are better characterized through their degree of membership in the causal conditions under consideration.

In the second stage of fsQCA, examinations of necessity and sufficiency are undertaken. Here a truth table is constructed which outlines the empirical instances of configurations, as well as their relationship to the outcome indicators. With fsQCA there are 2^k possible configurations, where k represents the number of causal conditions.

Because it is rare for cases to conform precisely to either a necessary or sufficient relation, the concepts of consistency and coverage are used to measure how well this is achieved [28, 29]. Consistency measures the degree to which a necessary/sufficient relation is met. In terms of necessity, consistency measures the degree to which an outcome is a subset of a causal condition. If all instances of the outcome display the causal condition, consistency will be high. In terms of sufficiency, consistency measures the degree to which a causal condition (or combination of) is a subset of an outcome. If all instances of the condition display the outcome, consistency will be high. Scores are calculated by the fsQCA software and range from 0 (no consistency) to 1 (perfect consistency). The minimum basis on which a necessary (sufficient) relation can be claimed is 0.90 (0.75) [29, 30].

Coverage by contrast, measures empirical relevance [28, 29]. For necessity, coverage measures the frequency with which an outcome occurs relative to a causal condition. Here very low coverage scores indicate that a causal condition is present in almost all cases, regardless of whether they display the outcome. In this scenario, a necessary condition would be deemed trivial. For sufficiency, coverage indicates the degree to which a condition (or combination of conditions) explains all occurrences of an outcome. If coverage scores are very low this would indicate that the causal condition explains only a limited set of the cases with an outcome. Coverage scores are also calculated by the fsQCA software and range from 0 (no coverage) to 1 (full coverage). It is suggested that when testing for necessity, coverage scores should not be lower than 0.5 and that no cause should be deemed necessary, independent of a theory that recognizes it as a relevant cause [29]. Minimum coverage scores are not suggested for sufficient relations since configurational methods recognize that an outcome may be produced via different combinations of conditions.

The final fsQCA stage involves a process of ‘logical reduction’ where a simplified statement is made about which conditions are necessary/sufficient for an outcome (termed a solution path). In a conventional QCA, this is achieved through Boolean Algebra. For example, if two combinations of conditions are found to be sufficient, one with causal conditions A, B, and C and the other with causal conditions A and B (but not C), we could reduce this to one configuration: AB, since the outcome occurs whether condition C is present or absent. In fsQCA, an equivalent process is undertaken by the software using the Quine-McCluskey algorithm. This algorithm takes into account the more complex features of fsQCA, including consistency scores [21].

Overall consistency and coverage scores are used to describe the logically reduced solution paths. These measures are a calculation of how well an outcome is explained when all of the reduced solution paths are considered. Generally speaking, overall consistency is an average of the consistency scores of each of the individual solution paths found for an outcome. Overall coverage is a measure of how well the cases displaying an outcome are covered by the logically reduced solution paths.

This study examines two outcomes: adult female mortality rates (AFM) and infant mortality rates (IMR). The former was chosen since most T&C workers are female. The latter was chosen for its rapid response and sensitivity to macro-level policy changes [31‐33]. IMR was conceptualized to have been potentially impacted both directly, through T&C workers having children, and indirectly, if the phase-out influenced health important conditions at the national-level. Two national-level conditions highlighted by the EMCONET framework are material deprivation and economic inequality [14]. Both of these conditions may have been impacted through shifts in T&C employment (e.g., through additional provision or loss of wages) and both have been previously associated with IMR (e.g., [34, 35]).

There is a relatively robust body of literature which finds evidence for changes in national-level health outcomes following changing macro-economic conditions [36‐39]. Of particular relevance here is the evidence for the health impacts of job loss (e.g., [39]). While fewer studies have measured the health effect of employment growth [40], there are many pathways through which we can expect it to impact health at the national-level [9]. Furthermore, although much of the public health literature surrounding changing macroeconomic conditions is focused on the developed world, evidence suggests that such changes also have important implications for national-levels of health in poorer countries [41, 42], where much of the T&C sector is concentrated.

AFM and IMR were obtained from Rajaratnam and colleagues [43, 44]. Historically, the usefulness of adult mortality data has been hindered by a range of well-known weaknesses [45, 46]. Models have often extrapolated adult mortality from child mortality. Ambiguity in both the sources of data and the methods used has also hindered replication of results [46]. Documenting short-term fluctuations and linking them to changing socio-economic contexts requires far greater detail than past methods have provided [47].

The authors of the data this study utilizes, by contrast, estimate AFM (IMR) through a variety of sources including vital registration systems, sample registration systems, and nationally representative survey/census data [43, 44]. These methods demonstrate a higher predictive validity and are transparent and replicable [43, 44, 46]. Moreover, the authors specifically acknowledge that a main advantage of their data is that it can be linked to changes in socio-economic contexts. AFM is summarized by the probability that an individual who is 15 years old will die before age 60. IMR is summarized by the probability of death before age 1, conditional on surviving to 1 month.

Countries were included in this analysis if, between 2000 and 2004, employment in the T&C sector (as a proportion of total manufacturing employment) was greater than 10 %, given that more than 10 % of the working population was employed in manufacturing. Total manufacturing and T&C employment figures were obtained from the United Nations Industrial Development Organization (UNIDO) [20]. Data on the proportion of the working population employed in industry were obtained from the World Bank [48]. While 53 countries were initially identified as reliant on the sector (Table 1), only 32 countries were ultimately used for the analysis (Table 2). Inclusion of countries was limited by the quality of mortality data and the availability of data used to operationalize the causal conditions (Table 3). Countries were excluded if mortality data was characterized by relatively high and/or erratic levels of uncertainty. While excluded countries were comprised of both highly developed and less developed countries, it is unclear how their inclusion might have impacted the results of the analyses. This work thus reiterates calls for better quality cross-national health and social policy data. Despite this limitation, the number of cases included in this study well exceeds the minimum number of cases below which there is a high chance that a fsQCA will find an association due to random variation [49].

Five causal conditions were selected for inclusion in the fsQCA: countries’ level of development; (2) labour market protection; (3) welfare state protection; and (4) T&C employment loss or (5) growth after the phase-out. There are a variety of approaches that can be used to select causal conditions for a fsQCA [50, 51]. Here conditions were selected in direct response to the research question. A development indicator was included to contextualize how employment changes impacted health in countries of different levels of development and to group countries with similar health profiles together. Since the expectation is that the chosen causal conditions will combine in different ways to impact the health, they can also be seen as selected via the conjunctural approach [50, 51]. This approach is described in QCA literature as best aligned with the characteristics of a fsQCA analysis [50]. Specific hypotheses concerning these conditions were not made since the nature of this study tends towards theory development rather than theory testing.

Fuzzy-set membership scores are assigned through a process called calibration [21]. Calibration refers to the transformation of outcome indicators and causal conditions into membership sets. This procedure requires the use of theoretical and substantive knowledge to denote meaningful differences in the data in order to define cases’ degree of membership in the set created by an indicator. Calibration methods can be either direct or indirect. In the direct method, three thresholds are specified which correspond to the qualitative breakpoints of full membership (1), the cross-over point (.5), and full non-membership (0). At the crossover point there is maximum ambiguity of whether a case is more “in” or “out” of a set. Once these breakpoints are specified, fuzzy membership scores are assigned by the fsQCA software. Generally speaking, the software calculates scores by translating variable scores into the metric of log odds [21]. A strength of this method is that it is able to calculate precise fuzzy-set scores when there is similarly precise variation in the data.

The indirect method, by contrast, relies on a broad grouping of cases into a number of categories which represent different degrees of membership. This method is generally used when it is difficult to translate data using the three qualitative breakpoints or when the data is better aligned with a smaller number of membership categories (e.g., when there is less precise variation in the data).

In this study, the direct method was used to transform the health outcome indicators, countries’ level of development and employment growth and loss after the MFA phase-out. This is because data associated with these conditions could be anchored to the three qualitative breakpoints and because using the direct calibration resulted in more precise fuzzy-set scores. An indirect calibration method was used to transform the causal conditions of countries’ labour market and welfare state protection. This is because the data used to operationalize these conditions was not aligned with the direct calibration method and better transformed through the indirect method, as will be made clearer below.

External standards with which to calibrate the conditions included in this study do not yet exist. As a consequence, calibration thresholds were established based on the structure of the data and careful considerations of what meaningful thresholds would require in terms of best representing the condition. Sensitivity analyses were carried out which evaluated the impact of lower and higher thresholds and demonstrated little difference in fuzzy-set scores and final results. Further details of the calibration process for each of the conditions are noted below. The raw data and corresponding fuzzy-set scores for each of the outcomes and causal conditions can be found in an additional file (Additional file 1).

For each outcome indicator, AFM and IMR, a ‘Health Improving’ and ‘Health Worsening’ membership set was constructed. Relative changes in mortality rates were calculated based on the five-year period preceding (2000–2004) and following the phase-out (2005–2009). Data for these calculations are displayed in Tables 4 and 5. Although this is a relatively short time to examine changes in population health, it is consistent with studies which show an association between unemployment and adult mortality after a similar time lag [41, 52‐56]. In relation to FMR (IMR), the qualitative breakpoints for the health improving set were conceptualized respectively as a 3 % (4 %) increase in mortality rate reduction, a 0 % change in mortality rate reduction and a 3 % (4 %) decrease in mortality rate reductions. Scores in the sets of ‘health worsening’ were taken to be the negation of health improving scores and calculated by subtracting a country’s score in the health improving set from 1. In terms of AFM, 10 of the 27 analyzed countries experienced health improvement after the MFA phase-out. In terms of IMR, 17 of the 29 analyzed countries experienced health improvement.

The United Nation’s Human Development Index (HDI) was used to assign scores in the set of ‘Highly Developed Countries’ [57]. This data reflects conditions in countries in 2004. Data was directly calibrated in a way which aligned with the Index’s rating of countries into the categories of High, Medium and Low Human Development. The qualitative breakpoints were conceptualized as 0.9, 0.8 and 0.5, respectively. The cross-over point was chosen at 0.8 since below this point, countries are deemed as having medium human development. Countries receiving a HDI score of lower than 0.5 are deemed by the Index as having Low Human Development.

Countries’ labour market protection was indirectly calibrated based on the number of Fundamental ILO Conventions ratified by a country [58]. Here a six-value fuzzy-set [29] was used to assign scores in the set of ‘Protective Labour Market Policies, taking into account the number of Conventions ratified before the MFA phase-out, as well as additional ratifications made prior to 2009. Table 6 further demonstrates this calibration process. Since these Conventions represent minimum standards, relatively strict thresholds were set for countries to be characterized as having protective policies.

Welfare state protection was measured and calibrated using the ILO Income Security Index [59]. This Index uses a range of input, process and outcome indicators and categorizes countries into one of four clusters. ‘Pacesetting’ countries are characterized as scoring highly across all indicators. ‘Conventional’ countries score highly only on input and process indicators. ‘Pragmatists’ score high on outcome indicators and ‘Much-to-be-done’ countries score low across all indicators. These categorizations were used to assign scores in the set of ‘Protective Welfare State Policies’ since they delineate important qualitative features of countries. Another option would have been to use the individual index scores to directly calibrate fuzzy-set memberships; however, index scores do not directly align with the qualitative clusters. For instance, a Conventional country might score lower on the Index than a Much-To-Be-Done country. Using the direct calibration method thus would have clouded important qualitative differences between countries. For this reason, scores were indirectly assigned as follows: Pacesetters (1), Conventionals (.67), Pragmatists (.33), and Much-To-Be-Dones (0).

A direct calibration method was used to assign fuzzy-set scores in the employment growth and loss membership sets. Qualitative thresholds were chosen with a consideration of the variation of change across countries and with a consideration that changes would need to be somewhat significant to influence health at the population level. Employment growth and loss were treated separately, rather than as a single employment change condition, since the qualitative breakpoints of a single membership set were tasked with meeting two conditions found to be at odds with each other. Specifically, a single membership set would need both to differentiate between countries experiencing employment growth and loss (conditions which have different implications for health) and denote meaningful changes in employment (i.e., changes that would have feasibly made an impact on population health at the national level). For a single membership set to differentiate between countries experiencing employment growth and loss, the cross-over point (of 0.5) would need to be set at a 0 % change in employment. However, this would mean that countries experiencing a small change in employment, for example a 5 % increase, would be characterized as largely ‘in the set’ of employment change. This was seen as problematic since small changes are unlikely to result in discernable changes in national levels of health. Using two membership sets however, allowed for meaningful changes in employment to be more accurately accounted for. This is because the cross-over point for each of these sets could be set at a 5 % change in employment loss/growth. In this way, countries with a small change in employment are characterized as only somewhat in the membership set. Fuzzy-set scores were therefore calibrated across two membership sets based on percent changes in T&C employment between 2004 and 2008 (or the closest year for which data was available). For the Employment Growth (Loss) membership set, the qualitative breakpoints were conceptualized at a 15 % increase (decrease), a 5 % increase (decrease), and 0 % increase (decrease). Employment figures were obtained from UNIDO [20].

Tables 7 and 8 respectively summarize the fuzzy-set scores for the AFM and IMR membership sets, as well as for the five causal conditions. These tables demonstrate ample variation between countries both in terms of the outcomes and causal conditions.

Ideally, all cases included in an fsQCA would be studied in-depth; however, for this study this would require a prohibitively large number of studies. Therefore, twelve countries were selected for in-depth analysis so that each of the fsQCA solutions could be explored by at least one typical case (i.e., one that is characterized both by the configuration and outcome of the necessary/sufficient relation). When a fsQCA solution was characterized by multiple typical cases, a comparative approach was undertaken since our confidence in a causal mechanism is increased if it is found to be in place across multiple typical cases [26]. A comparative study design was also undertaken to take advantage of deviant cases. These cases are members of a configuration characterized by a logically reduced solution, but are not members of the associated outcome. As such, these cases provide evidence against a necessary/sufficient relation but represent a potentially useful opportunity to understand the fsQCA results. For example, the most likely reason for a deviant case in a sufficient relation is the omission of a causal condition of which the deviant case is not a member but the typical cases are [26].

In line with process tracing literature [27], evidence was collected to build a narrative about the overall structure of each country’s T&C sector (e.g., its workers, how employment changed after the phase-out, alternative employment opportunities) and its labour market and welfare state policies. Next, the aim was to inductively work backwards in search of a plausible causal mechanism that might help explain the fsQCA results.

Because selection bias is particularly acute in process tracing research [27], an attempt was made to minimize this bias by using a systematic process to search for evidence. A preliminary search strategy found that traditional databases, such as the Applied Social Science Index and Abstracts database, returned a dearth of relevant material; therefore, Google and Google Scholar were used to locate sources of evidence. Search keywords included the country name, ‘Multi-Fibre Arrangement, ‘health’, ‘employment’, ‘textile and clothing sector’, ‘apparel’, and ‘garments’. Once a narrative of a country’s T&C sector was constructed, material regarding labour market and social policies was searched for, particularly across international organizations including the International Labour Organization, World Bank and Asian Development Bank.