Introducing the Jordan Labor Market Panel Survey 2016

As the second wave of the JLMPS longitudinal study, JLMPS 2016 both followed the 2010 panel and added a refresher sample. For the panel component of the data, we attempted to recontact all households that were included in the 2010 wave. Among the households that were found, we also followed any split households. Split households occur when one or more individuals from 2010 leave their 2010 household to form a new household. For example, an individual who was the son of the household head in 2010 might marry and form a new household. The entire new household is included in our sample, including members who were not a part of the 2010 sample. The refresher sample over-sampled neighborhoods in Jordan that, as of the 2015 Census, had a high proportion of non-Jordanians. The sample weights account for this sampling strategy to generate nationally representative statistics, allowing for analyses not only by nationality but also for the Jordanian labor market overall (Assaad et al., 2019; Assaad and Salemi, 2019; Krafft et al., 2019; Malaeb and Wahba, 2019). The final JLMPS 2016 sample is made up of 7,229 households, including 3,058 that were part of the original 2010 sample, 1,221 split households, and 2,950 refresher households. The JLMPS 2016 sample captured a total of 33,450 individuals. We discuss the sampling strategy and the creation of the sampling and attrition weights in detail below.

The questionnaires for JLMPS 2016 build on those used in 2010, as indicated in Table 1. The questionnaires include a household questionnaire, an individual questionnaire, and a questionnaire administered at the household level that elicits information about household enterprises and current migrants and remittances. The household questionnaire includes the identifying information for the household, a household roster and information on housing conditions, access to public services, ownership of durable goods, and use of household help. The individual questionnaire includes (i) a personal biography, which elicits information about marriage history, entry and exit from school, start and end of jobs, and residential mobility, (ii) modules on father's, mother's, and siblings’ characteristics, (iii) self-reported health, health insurance, and health-seeking behavior, (iv) educational status and detailed educational history, (v) employment in the short (1 week) and long (3 months) reference periods, including keyword questions to detect employment as defined by the 19th International Conference of Labour Statisticians, (vi) unemployment and job search, (vii) subsistence and domestic work, (viii) detailed characteristics of the primary job, (ix) characteristics of the secondary job, (x) labor market history, (xi) fertility, (xii) women's decision-making, mobility, and acceptability of intimate partner violence, (xiii) costs and characteristics of marriage, (xiv) women's employment, maternity leave, and childcare, (xv) wage earnings in primary and secondary jobs, (xvi) return migration, (xvii) in-migration for non-Jordanians, (xviii) use of information technology, (xix) savings and borrowing behavior, and (xx) gender attitudes. The current migration and household enterprise questionnaire elicits information about (i) current migrants abroad, (ii) remittances, (iii) sources of non-labor income, (iv) household non-farm enterprises, including sections on hired labor, expenditures, assets, and revenues,

Due to problems in the implementation of appropriate screening questions and skip patterns, a non-representative sub-sample of the self-employed and employers was captured; the resulting data will not be publicly released.

Public use microdata from the 2016 wave of the JLMPS, as well as all previous waves of ERF LMPSs, are available through ERF's Open Access Microdata Initiative (OAMDI). Researchers can access the microdata free of charge from the ERF Data Portal

www.erfdataportal.com

Data collection for the 2016 wave proceeded in two phases. First, enumeration was undertaken to track and, if possible, locate the 5,102 households included in the 2010 wave, including any households formed by individuals splitting from 2010 households. Second, fielding was undertaken with located households from 2010 as well as a refresher sample of 3,000 households that over-sampled neighborhoods with a high proportion of non-Jordanian household heads, as ascertained by the 2015 Population Census. The enumeration phase lasted from June 5, 2016 until November 14, 2016 and the main data collection phase lasted from December 10, 2016 until April 27, 2017.

Additional data collection to capture individuals or households missed in initial fielding continued until September 27, 2017.

The 2010 sample was a nationally representative sample designed to represent urban and rural areas in the three regions of Jordan: North, Middle, and South. For sampling purposes, the sample was stratified into 30 strata based on a combination of the 12 governorates of Jordan and 5 different location classifications within them: (i) basic urban, (ii) rural, (iii) large central city urban in Amman, Zarqa, and Irbid governorates, (iv) suburban Amman and Zarqa, and (v) exurban Amman. The 2010 sample captured 5,102 households and 25,953 individuals.

A few individuals, during 2016 fielding, were determined to have been incorrectly included in the 2010 sample; for example, guests visiting were included but should not have been. These individuals were removed from the revised 2010 data.

A refresher sample was added to the 2016 wave of the JLMPS to capture the large influx of non-Jordanian populations to Jordan. The refresher sample was designed to over-sample neighborhoods with high concentrations of non-Jordanian household heads as reported by the 2015 Population Census in order to ensure national representativeness in the JLMPS 2016, as well as sufficient observations for analysis of key groups such as Syrian refugees.

For the panel data, tracking households from 2010 to 2016, a key issue is sample attrition. There are two points in time when attrition can occur: between the 2010 wave and 2016 enumeration and between the 2016 enumeration and 2016 fielding. There are also two types of attrition that can occur: Type I attrition occurs when we cannot locate a 2010 household at all, while Type II attrition occurs when we can locate a 2010 household, it has a split, and we cannot locate the split household.

Since attrition could occur at two points in time, there were some cases in which the original household and a split were both found during enumeration, but only the split was found in fielding. In these cases, we reclassified the split to be the original household and the original household, not found in fielding, to be the split so that attrition could be modeled and the households reclassified from splits to original included in the data.

We model sample attrition for two reasons. First, we wish to examine whether attrition is random or related to household characteristics. Second, if there are differences in attrition related to observable characteristics, we account for these differences by creating appropriate weights. While we cannot correct for attrition related to unobserved household characteristics, we try to reduce their influence by including as many observable characteristics as possible as predictors in the model. Table 6 presents odds ratio estimates from a logit model, on the household level, for Type I attrition. Households that naturally attrited are excluded from the model, resulting in a universe of 4,943 households from 2010 at risk of Type I attrition. Characteristics are, necessarily, those measured in 2010.

We use a logit model rather than a probit model because it produces odds ratios that are more readily interpretable than probit coefficients. The logit model is commonly used to model attrition (e.g., Himelein, 2014). We do not include nationality of household head in attrition models because it would assume no migration occurs. Furthermore, we calculate our weights by nationality to reflect the population distribution by nationality observed in the 2015 Population Census.

There are some significant predictors of Type I attrition. In terms of household composition, households with more working age and especially more elderly (65+) females were significantly less likely to attrite. Households composed of all males, compared to mixed sex households, were significantly more likely to attrite. There were no significant differences by the geographic strata that were used in 2010 for stratifying the sample, which is encouraging for sample representativeness. Households in the top wealth decile were significantly more likely to attrite than the poorest, but there were no significant odds ratios for other deciles. The higher attrition in the top decile was driven by urban areas.

Although the tenth decile rural interaction is insignificant, it shows lower odds of attrition, cancelling the higher odds of tenth decile main effect. The higher Type I attrition in the top decile of urban areas is presumably due to a higher rate of refusals among this category of respondents, which are generally known to be less cooperative in face-to-face household surveys (Hlasny and Verme, 2018).

In terms of governorates, there are significantly lower odds of Type I attrition for (the reference, urban) Karak, Tafileh, and Ma’an, but higher odds of attrition in urban Aqaba (all in the South region), compared with Amman. Karak and Ma’an's interactions with rural are near one and insignificant, so the lower odds of Type I attrition pertain to rural areas of these regions as well. There are significant interactions with rural, lower odds of attrition, for Zarqa, Irbid, and Aqaba, compared with their urban counterparts, and significantly higher odds in rural Tafileh, compared with urban Tafileh. Homeownership as opposed to renting predicts significantly lower attrition. There were no significant differences by head age group, or sex, although households headed by females aged 25–34 years were significantly more likely to attrite. There were no significant differences by marital status, but there was a significantly higher probability of attrition for households with divorced female heads.

Households with more educated heads were more likely to attrite, significantly so for secondary and higher education, as compared to less than basic. There were few significant labor market characteristics of the household head, which bodes well for the labor market representativeness of our panel. Households with unemployed household heads were significantly less likely to attrite, while those out of the manpower basis (disabled or elderly) were more likely to attrite than the reference household head, who was employed in the public sector. There were no significant labor market status interactions with the rural dummy. Overall, the model had a pseudo-R² of 14.7%. In this case, a low pseudo-R² is a good outcome, as it means that only a limited portion of the probability of attrition can be explained by this long list of observable characteristics and that much of the rest is presumably random.

For the Type II attrition model (Table 7), the sample is restricted to those 2,386 splits with heads aged 6+ years (who have individual characteristics from 2010).

Essentially, we are assuming that split households with heads less than age 6 years are unlikely to be “real” split households, but rather that we are simply missing information about these individuals (for instance, that they died). For those few splits with “heads” aged 0–5 years, we use the mean predicted probability of attrition while weighting.

Female-headed splits are significantly less likely to attrite, although the female and 25–34 years age group interaction indicates significantly higher odds of attrition for females in that age group. Compared to those splits with less than basic educated heads, other categories are less likely to attrite, all but higher education significantly so. This may be related to the ability of the remaining household members to accurately communicate the new location. Compared to splits with household heads who were public sector workers in 2010, splits with formal private wage, informal wage, and out of labor force heads are more likely to attrite. Overall, the pseudo-R² of the Type II attrition model is 10.0%, again indicating that most of the Type II attrition is not systematically related to observable characteristics.

Weights are initially constructed at the household level. Two essential inputs are the results of the Type I and Type II attrition models. For an original (2010) household, denote the household as h.

The start of this section draws on the concepts and notation used for the ELMPS 2012 (Assaad and Krafft, 2013).

Or, alternatively and equivalently: Pr(Ahs)=1−Pr(h found & s found)=1−Pr(h found)*Pr(s found |h found) \matrix{ {\Pr \left( {{A_{hs}}} \right)} \hfill & = \hfill & {1 - \Pr \left( {h\,{\rm{found}}\,\& \,s\,{\rm{found}}} \right)} \hfill \cr {} \hfill & = \hfill & {1 - \Pr \left( {h\,{\rm{found}}} \right)*\Pr \left( {s\,{\rm{found}}\,|\,h\,{\rm{found}}} \right)} \hfill \cr }

We then can compute a response adjustment factor, r_h for original households as: (2) rh=11−Pr(Ah) {r_h} = {1 \over {1 - \Pr \left( {{A_h}} \right)}}

For the split sample the response adjustment factor is: (3) rhs=1[1−Pr(Ahs)]*cs {r_{hs}} = {1 \over {\left[ {1 - \Pr \left( {{A_{hs}}} \right)} \right]*{c_s}}}

Here the adjustment includes c_s, the number of component households. Essentially, component households are the number of different originating households in the population (not the sample) that contribute to a split household. If the split contains only individuals from a 2010 household, there is only one component household. If the split contains individuals not from a 2010 household, there are two component households. Accounting for component households in this manner maintains population representativeness.

For panel analyses using only observations present in both 2010 and 2016, the specific panel weight does not include the division by component households.

The calculation of our panel sample weights also brings in the weights from 2010. Denote as e the expansion weight from 2010.

There was some individual variation in weights within households in 2010 for some households, which was averaged out in the revised data for JLMPS 2010 and to enable generation of weights for 2016.

The refresher sample weights both stand alone, should someone want to analyze the refresher sample only (for instance, as an additional validation check), and act as inputs into the combined sample weights. As with the panel sample, weights are calculated on a household level. We first account for nonresponse at the PSU level, p, with a weight, w_p, based on the number of observed households, m_p as: (4) wp=15mp {w_p} = {{15} \over {{m_p}}}

Essentially, the observed households in clusters with nonresponse are weighted up to represent the planned 15 households. Otherwise, this weight is one.

Recall that the refresher sample was stratified along governorate, urban/rural/camps and high/low non-Jordanian lines. Denote governorate as g, urban/rural/camps (location) as l, and high/low non-Jordanian strata as s. We undertake ex-post weighting along nationality, n, lines. We therefore identify the response-corrected number of households, h_g,l,s,n, of each nationality n in each governorate g in location l and high-low stratum s by summing across PSUs as in: (5) hg,l,s,n=∑p=1pwp*hp,g,l,s,n {h_{g,l,s,n}} = \sum\limits_{p = 1}^p {{w_p}*{h_{p,g,l,s,n}}}

We then draw on the census data for the nationality-specific number of households in the same governorate, location, and strata, c_g,l,s,n. After accounting for PSU-level nonresponse, this implies an expansion weight of: (6) wp,g,l,s,n=wpcg,l,s,nhg,l,s,n {w_{p,g,l,s,n}} = {w_p}{{{c_{g,l,s,n}}} \over {{h_{g,l,s,n}}}}

Although at this point the weight is conceptually correct, two empirical problems arise. First, there are some extreme outliers within the weight distribution, which are particularly problematic when analyzing subgroups. These are primarily driven by instances in which we observed only a few households of a particular non-Jordanian nationality in a combination of governorate, location, and strata. We therefore winsorized the high end of our weight, at the 95th percentile. We denote the winsorized weight as wp,g,l,s,n′ w_{p,g,l,s,n}^\prime .

The other empirical problem is that we do not always have, in our sample, individuals from all nationalities in each combination of governorate, location, and strata. For Jordanians, we are very close to the national population. However, we under-represent other nationalities to an increasing extent as their numbers diminish. Additionally, the winsorizing reduces numbers somewhat. Therefore, to ensure that our statistics represent nationalities appropriately within the country, we revise the weights, multiplying up to represent higher levels of aggregation. Specifically, we use from the 2015 Population Census the number of households of a nationality within a region, r (there are three regions), and area, specifically inside or outside of camps, which we denote as a. We use the census number of households relative to the weighted, winsorized, observed households as a multiplier on our weights, aggregating over governorates (from 1 to G) within a region, locations (from 1 to L within a governorate and area), and strata, to generate an adjusted weight: (7) wp,g,r,l,a,s,n=wp,g,l,s,n′*cr,a,n∑g=1G∑l=1Lg,a∑s=12wp,g,l,s,n′ {w_{p,g,r,l,a,s,n}} = w_{p,g,l,s,n}^\prime*{{{c_{r,a,n}}} \over {\sum\nolimits_{g = 1}^G {\sum\nolimits_{l = 1}^{{L_{g,a}}} {\sum\nolimits_{s = 1}^2 {w_{p,g,l,s,n}^\prime} } } }}

Here the denominator includes only those governorates within the region and only the locations within the area.

This approach generates mathematically identical numbers of households at the national level. On the individual level, it does very well for representing Jordanians (expanding to around 6,603,000, very close to the census) and Syrians (expanding to around 1,268,000, also very close to the census). Other groups have fewer individuals, likely due to the fact that many of these individuals are immigrants who may live in collective housing, rather than private households. We expand to around 546,000 other Arabs (underestimated by around 272,000), 185,000 Egyptians (underestimated by around 451,000), and around 58,000 other nationalities (underestimated by around 139,000). Ultimately, our sample expands to 8,661,000 individuals, whereas there were 9,293,000 individuals in private households according to the 2015 Census. Given the sizes of households we observe and those implied in the census data, it is likely that the sampling frame was more stringent in identifying private households, and thus under-represents those other nationalities living collectively.

An additional problem on the individual level arises in that, in addition to household-level consent, we collected individual-level consent and had some refusals (69 in the refresher sample and 172 in the panel sample) as well as some individual questionnaires that were only partially completed (12 in the refresher sample and 58 in the panel sample). We removed the partial individual data and treated both cases as individual nonresponse. The only characteristics we have for these individuals are the basic characteristics from the roster. Upon examination, we determined that there were sex- and age-specific patterns of individual nonresponse. To address both the under-representation of certain types of individuals and individual nonresponse, we created individual-level weights. The starting point for these individual-level weights was the household-level weight. We adjusted this household-level weight by the age, e, and sex-specific, x, interacted nonresponse rate, r_e,x as: (8) wp,g,r,l,a,s,n,x,e=wp,g,r,l,a,s,n1−re,x {w_{p,g,r,l,a,s,n,x,e}} = {{{w_{p,g,r,l,a,s,n}}} \over {1 - {r_{e,x}}}}

We then summed the household-weighted number of individuals with valid responses and multiplied it by the ratio of that number to the number in the population on the region, areas, and nationality levels from the 2015 Census, which we denote as j_a,r,n. Essentially, we calculated an individual expansion weight, i, of: (9) ip,g,r,l,a,s,n,x,e=wp,g,r,l,a,s,n,x,e*jr,a,n∑r,awp,g,r,l,a,s,n,x,e {i_{p,g,r,l,a,s,n,x,e}} = {w_{p,g,r,l,a,s,n,x,e}}*{{{j_{r,a,n}}} \over {\sum\nolimits_{r,a} {{w_{p,g,r,l,a,s,n,x,e}}} }}

These individual weights (mathematically) generate expanded numbers of individuals that are identical to the census on the region, area, and nationality levels.

The calculation of weights for the combined sample starts with the normalized panel and refresher sample weights (meaning weights that average to one in each sample). As with the refresher sample weights, we do ex-post weighting on a nationality, governorate, and location basis. Since the initial sample was not stratified by high/low non-Jordanian that dimension is not incorporated explicitly into the combined sample weights, but is built in implicitly through the refresher sample weights. Denote the normalized weights for a household in governorate g, location l, and nationality n as w˜g,l,n {\tilde w_{g,l,n}} . We calculated the initial combined sample weights as: (10) wg,l,n=w˜g,l,n*cg,l,n∑w˜g,l,n {w_{g,l,n}} = {\tilde w_{g,l,n}}*{{{c_{g,l,n}}} \over {\sum {{{\tilde w}_{g,l,n}}} }} where c_g,l,n are the number of households from the 2015 Census at the governorate, location, and nationality levels. As with the refresher weights, we winsorize at the 95th percentile to address outliers. We denote the winsorized weight as wg,l,n′ w_{g,l,n}^\prime . We then need to account for the absence of some nationalities in combinations of governorate and location. Again, we use the number of households of a nationality within a region, r, and area (inside or outside of camps), a from the 2015 Census. We use the census number of households relative to the weighted, winsorized, observed households as a multiplier on our weights, aggregating over governorates (from 1 to G) within a region and location (from 1 to L) within governorates, to generate an adjusted weight: (11) wg,r,l,a,n=cr,a,n∑g=1G∑l=1Lg,awg,l,n′ {w_{g,r,l,a,n}} = {{{c_{r,a,n}}} \over {\sum\nolimits_{g = 1}^G {\sum\nolimits_{l = 1}^{{L_{g,a}}} {w_{g,l,n}^\prime} } }}

As was the case with the refresher weights, we find that this under-represents individuals, particularly non-Jordanians, and also we again have an issue of non-response on the individual level, so for the combined sample we also generate individual weights that account for age and sex-specific individual non-response, r_e,x, as: (12) wg,r,l,a,n,x,e=wg,r,l,a,n1−re,x {w_{g,r,l,a,n,x,e}} = {{{w_{g,r,l,a,n}}} \over {1 - {r_{e,x}}}}

We then adjust this by the census number of individuals in a specific area and region of the same nationality to generate an individual weight, i: (13) ig,r,l,a,n,x,e=wg,r,l,a,n,x,e*jr,a,n∑r,awg,r,l,a,n,x,e {i_{g,r,l,a,n,x,e}} = {w_{g,r,l,a,n,x,e}}*{{{j_{r,a,n}}} \over {\sum\nolimits_{r,a} {{w_{g,r,l,a,n,x,e}}} }}

The individual weights should be used when trying to generate individual-level statistics, particularly those incorporating any data from the individual questionnaire.

The distribution of (5-year) age groups is quite comparable between the JLMPS surveys and other data sources. Figure 1 compares the JLMPS 2010 and EUS 2010. The JLMPS sampled slightly more children in the 0–9 age group than the EUS in 2010. The JLMPS also sampled slightly fewer individuals in the 20–24 age group. The few differences are small and quite plausibly the result of sampling variability. Figure 2 compares the 2015 Population Census and EUS 2016 with JLMPS 2016. The JLMPS 2016 results are quite close to the 2015 Census, closer even than the EUS 2016 (which has a larger sample). The EUS in particular appears to have fewer young children, while both JLMPS 2016 and 2015 Census show a more modest inflection of the population pyramid.

The changing population structure of Jordan and particularly the resumption of fertility decline after a decade of stall is explored further in Krafft and Sieverding (2018).

Figure 1

Figure 2

Correctly defining and identifying households is a challenging part of household-based survey fieldwork. JLMPS 2010 shows a comparable distribution of household sizes among Jordanians to the EUS 2010 survey (Figure 3). JLMPS 2010 finds slightly more households with just two people by a percentage point or so, as well as slightly more households with four people compared to the EUS, and slightly fewer households of larger sizes. There are more substantial differences comparing the 2015/2016 data (Figure 4). Here, the Census data are all nationalities living in private households, as household size split by nationality was not available. JLMPS 2016 and the 2015 Census find a similar share of one-person households, but the EUS 2016 captures more one-person households. The JLMPS generally finds more small households than the other two data sources, but the difference with the Census is likely to be driven by Syrian refugees having larger households (Krafft et al., 2019). All three data sources align for households of six and more.

Figure 3

Figure 4

In terms of the marital status of respondents aged 15 years and older, the data sources are quite consistent (Figure 5). The JLMPS 2010 finds one percentage point fewer never married individuals and one percentage point more married individuals than the EUS 2010. They identify similar shares of divorced or separated (1%) and widowed (4%) individuals. The results over 2015/2016 are slightly more varied, with the 2015 Census and JLMPS 2016 both identifying 37% of individuals as never married, but the EUS 2016 40% and correspondingly fewer married. While all three sources still find around 1% of individuals are divorced or separated, fewer are widowed in the 2015 Census (3%) than JLMPS 2016 (4%) or EUS 2016 (6%).

Figure 5

While most demographic indicators are quite close across data sources, there are more substantial differences in terms of educational attainment (Figure 6). The share of the population in the 25–64 age group, of an age to have finished all education, classified as having a post-graduate education is comparable (2–3%) across time and data sources. The share with a university degree is 16% in EUS 2010 and 15% in JLMPS 2010, quite comparable, and rises to 18–19% across 2015/2016 sources. Similar alignment occurs for 2-year post-secondary degrees, 12–13% in 2010 across surveys, and 11% across 2015/2016 data sources. Where the differences occur are at the secondary and lower levels of education. While the JLMPS 2010 identifies 15% of Jordanians as secondary educated, the EUS 2010 finds 18%. The Census in 2015 finds 25% of Jordanians to have a secondary education, compared with 13% in the EUS 2016 and 17% in the JLMPS 2016. Some of these differences may be driven by difficulties in actually classifying attainment, the degree or level completed. The 2015 Census asked attainment categorically (Department of Statistics (Jordan), 2015f), which may have been interpreted as the highest level attended, whereas the JLMPS specifically asks for the highest level successfully completed. Since the secondary examination (tawjihi) pass rate is only 50% (Ministry of Labor and Ministry of Planning and International Cooperation, 2012), this definitional difference could explain much of the disparity around secondary. Definitional differences, particularly with the change in the structure of basic from 6-year primary and 3-year preparatory to 10-year basic, may also be contributing to disparate definitions and differences in classifying individuals as basic educated or read and write (which is higher in the JLMPS 2016). Estimates of illiteracy are fairly comparable across JLMPS 2010 (7%) and EUS 2010 (6%) as well as the 2015 Census (6%) and JLMPS 2016 (7%), with the EUS rates in 2016 being lower (4%). Notably, there is substantial consistency across JLMPS 2010 to JLMPS 2016, after accounting for expected rises in higher education, suggesting that classification and definitional differences are the cause for disparities in educational attainment across sources.

Figure 6

Having established the consistency of the JLMPS data with other Jordanian sources in terms of demographic representation, in this section we turn to analyzing labor market statistics, trends over time, and their alignment. We include 95% confidence intervals from JLMPS estimates in order to assess whether differences in indicators exceed what we might expect due to sampling variability. We add to our comparisons Q1 of 2017 from the EUS (noted as 2017), since JLMPS 2016 was fielded primarily in 2017 and there appear to be substantial differences in the EUS results between 2016 and Q1 2017.

Figure 7 examines the labor force participation rate among Jordanians by sex. In 2010, the JLMPS detected slightly higher labor force participation than the EUS. The estimates from the EUS for the total rate and the female rate fall within the JLMPS 2010 confidence interval, but the male rate is higher—70% in the JLMPS 2010 and 67% in the EUS 2010 although the JLMPS 2010 estimate is not far from the 2009 EUS rate of 69%. Estimates from the 2015 Census appear to be substantially different (much higher) than EUS or JLMPS 2016, which may be due to different definitions or differences in the application of the criteria for participation. The JLMPS 2016 data is in line with the EUS, particularly with the results of Q1 of 2017, which align better with the timing of most of the JLMPS data collection. The total is exactly in line, and while the confidence intervals of the JLMPS 2016 estimates by sex fall slightly below (for men) and above (for women) the EUS estimates for 2016, they include the EUS estimates for Q1 of 2017. Based on EUS data, this was a time of rapid change in labor force participation for Jordanians aged 15 years and older. Male labor force participation increased from 58% in Q4 2016 to 63% in Q1 2017, while female rates increased from 13% to 18% in the same period (Department of Statistics (Jordan), 2021). After this large jump between Q4 of 2016 and Q1 of 2017, participation rates for Jordanians aged 15 years and older continue along their previous downward trend, falling to 57% for males and 15% for females by Q1 2018. We can, therefore, not rule out that the jump observed between Q4 2016 and Q1 2017 was due to changes in data collection and sampling methodology in the EUS. The overall trend, discussed further in Assaad et al. (2019) of a very low and declining rate of labor force participation in Jordan remains a major concern.

Figure 7

Turning to employment rates (Figure 8), we see similar patterns as with labor force participation. JLMPS 2010 has a higher estimate of the employment rate for men than EUS 2010, and the total as well as the male rate has confidence intervals that do not include the EUS 2010 estimate. The census again has higher employment rates than the EUS 2016 or JLMPS 2016. JLMPS 2016 confidence intervals for men and women include the EUS 2016 estimates, although the total is a little low, and the JLMPS 2016 estimates are closer to the EUS Q1 2017 estimates.

Figure 8

Since the unemployment rate is calculated as a share of the labor force, necessarily the JLMPS estimates have larger standard errors and confidence intervals for unemployment than for employment or labor force participation (Figure 9). Comparing 2010 data, the JLMPS 2010 estimates of unemployment are slightly lower and have confidence intervals that just marginally exclude the EUS estimates for men and the total, but include the EUS estimates for women. The EUS 2016 estimates of unemployment are not included in the JLMPS 2016 confidence intervals for women or the total; however, the JLMPS 2016 confidence intervals include the 2015 Census estimate and the 2017 Q1 estimate for all groups, a high rate of unemployment for women at around 35% in the JLMPS 2016. Overall, the JLMPS results appear to be very close to other estimates, especially those of Q1 2017 of the EUS, which was implemented at about the same time. This was a period of rapid change in unemployment rates for Jordanian women. While the unemployment rate for men remained stable at 14% for men between Q4 2016 and Q1 2017 according to the EUS, it jumped sharply for women from 25% to 33%, in line with the JLMPS estimate. The overall unemployment rate also increased by 2 percentage points. The increase in unemployment for women continued through Q2 2017, when it reached 34% before declining to 28% by Q4 2017 (Department of Statistics (Jordan), 2021). Again, we cannot rule out that the jump from Q4 2016 to Q1 2017 in the female unemployment rate in the EUS is due to some change in methodology.

Figure 9

As with any longitudinal survey, attrition from one wave to another is an issue. We consider in this paper two types of attrition. Type I is when a household interviewed in the previous wave is not found in its entirety and Type II is when individuals who are no longer in the household due to splits cannot be found. The probability of both types of attrition was relatively high in JLMPS 2016 at 38.1% (Type I) and 50.5% (Type II). Factors that contributed to higher Type I attrition presumably include the relatively high proportion of non-national households living in Jordan and the generally higher residential mobility of households in Jordan (moves and mobility are linked to attrition in other surveys as well (Thomas et al., 2012; Vaillant, 2013)). The relatively high Type II attrition resulted from fieldwork-related issues at the fielding stage, which prevented individuals newly identified to be splits from being tracked. In any case, modeling of the two types of attrition reveals that both were for the most part random, although some observables systematically affected attrition. For instance, being in the highest urban decile of wealth increased attrition of both types, a pattern that can be linked to the higher refusal rate among wealthy households (urban and higher income individuals are more likely to attrite in other surveys as well (Himelein, 2014; Thomas et al., 2012)). Other correlates of Type I attrition were also fairly predictable, such as the higher attrition of all male households, which tend to be temporary in nature, and the higher attrition among households that rent as opposed to own their dwelling. Households with more educated household heads also had higher attrition rates, presumably because of their higher residential mobility. Type II attrition was driven in part by location, with splits from smaller urban areas being easiest to locate. Females were easier to locate in general when they split, but this does not apply to young females (25–34 years) who presumably move from their natal households when they marry. Finally, private sector workers appear to be more difficult to locate if they split than public sector workers. We created both panel and cross-sectional weights to account for these differential rates of attrition and maintain the representativeness of our sample, but it should be kept in mind that attrition based on unobservables can continue to be a source of bias.

Although the JLMPS included important innovations in the labor market history, a number of challenges arose in fieldwork that are limitations of the data and point to important directions for future LMPSs. First, we are no longer undertaking separate enumeration rounds, as this again led to attrition and added costs. In ELMPS 2018, eschewing a separate enumeration round substantially reduced attrition (Krafft et al., 2019). The software we used in JLMPS 2016 was limited in its ability to handle multilevel information, requiring separate household and individual data and leading to a non-representative sample of household enterprises in JLMPS 2016. Going forward, we are using the ODK-X tools (Brunette et al., 2017) that allow for multilevel data and appropriate skips and constraints far more effectively. This will also allow us to ask who in the household engaged in various enterprises, improving our detection of women's work (Krafft et al., 2019).

For future LMPSs, we will also be ensuring that both men and women are asked about gender role attitudes, as well as adding a brief contraception and pregnancy question series to the fertility history. Given the importance of the “second shift” to women's challenges reconciling market work and domestic responsibilities (Assaad et al., 2017; ERF and UN Women, 2020; Selwaness and Krafft, 2021), we recommend adding a 24-h our time diary to future LMPSs. In light of both the ongoing struggles in the MENA region and especially Jordan to integrate youth into the labor market, challenges only exacerbated by the COVID-19 pandemic, we also recommend adding questions on engagement with active labor market policies, trainings, internships, and apprenticeships.

The comparability of the JLMPS data to other statistics on Jordanians makes clear that the results of the survey are generalizable, and the rich, detailed data on a variety of subjects represent the Jordanian population well. The comparisons of the data from 2010 and the 2016 retrospective data about 2010 show substantial improvements in the questionnaire design and resulting retrospective data quality. Additionally, the over-sampling of non-Jordanians in the refresher sample of 2016 provides a unique opportunity to study both refugees and other migrant groups in Jordan. Already, the data provide important insights into issues as diverse as the well-being of refugees (Krafft et al., 2019), the impact of the refugee influx on Jordan's labor and housing markets (Al-Hawarin et al., 2021; Fallah et al., 2019), social insurance reforms (Alhawarin and Selwaness, 2019), and marriage and fertility trends (Krafft and Sieverding, 2018; Sieverding et al., 2019). We look forward to seeing additional research with the JLMPS as it is now publicly available, and comparing Jordan with the other countries with LMPSs.