Introducing the Egypt Labor Market Panel Survey 2018

Each wave of the survey attempts to maintain consistency for the indicators measured in previous waves while adding additional modules and questions to examine new issues or to allow more in-depth examination of existing issues. Accordingly, the 2018 wave devoted more attention to the measurement of the instability of employment, focusing in particular on job turnover among casual workers. It also provided more detailed information on health, gender role attitudes, food security, hazardous work, community infrastructure, and the cost of housing. It incorporated specific questions on vulnerability, coping strategies, and access to social safety net programs.

The 2018 wave had two primary questionnaires: a household questionnaire and an individual questionnaire. The modules in these two questionnaires are listed in Table 1. Most of the parts in the modules are the same as those in the previous waves of the survey with a few exceptions. The “tracking splits” module in the household questionnaire allows interviewers to ascertain whether the composition of the household has changed since the 2012 wave and inquire about new members present in the household as well as those who may have split to form new households. The “shocks and coping module” is also new in the 2018 wave and enquires about both idiosyncratic and community-level shocks that the household may have been exposed to, household food security, and coping mechanisms that the household may have used to respond to shocks. The main changes in the individual questionnaire relative to the 2012 wave were a substantial expansion of the health module, a reconfiguration of the labor market history module to better capture past periods of non-employment,

See Assaad et al., (2018) for a discussion of ways to improve the collection of retrospective labor market data. See Section 4 for an examination of these improvements.

The collection of data was tablet-based. The ODK-X (previously named ODK2) tools were used, given their ability to easily handle hierarchical data structures, such as multiple births to individuals within households, and validate across these different structures (Brunette et al., 2017). Training of the trainers was held at CAPMAS in January 2018. Enumerators were trained in April 2018, and data collection began in end-April. The bulk of data collection was completed by July 2018, but some teams continued to collect data until November 2018, to complete the sample. For fieldwork, teams were governorate-based and composed of a supervisor and from three to five enumerators (numbers vary depending on the sample size in the governorate). All the enumerators were women.

Based on experience from past surveys, we eschewed a distinct enumeration round (a phase of locating and listing individuals interviewed in the previous wave before fielding). From our experience in the 2012 wave, we found that the data from the enumeration phase could not simply be used directly in fielding, as individuals may have split between enumeration and fielding, and locating households twice added to the cost, time, and attrition. In fielding the ELMPS 2012, 1,680 individuals who were enumerated in 2011 were simply not found in the main fieldwork phase in 2012, and we lacked information on whether they split together, died, moved abroad, or otherwise (Assaad and Krafft, 2013). A similar problem arose in Jordan in 2016, where, although we designed the questionnaire to track splits at fielding as well as enumeration, many such splits were not actually fielded, leading to the loss of 616 split households and 647 individuals from the sample (Krafft and Assaad, 2021a).

The repetition of enumeration in fielding did allow us to determine that 208 individuals had naturally attrited, i.e., died, migrated, or moved to group housing between enumeration and fielding (Krafft and Assaad, 2021a).

During fieldwork, field quality control took place. We undertook an innovative quality control process where, rather than asking the quality control teams to entirely redo certain questionnaires, we randomly deleted a certain number of modules from a sample of completed questionnaires and asked the quality control teams to repeat the data collection for those modules. More critical modules had a higher probability of being selected for deletion. For example, we took a 5% deletion rate in Section 4.0, which covers assets and housing characteristics, as these questions primarily have yes/no answers and presented fewer data quality challenges. More central modules, such as unemployment detection, were assigned a higher probability of deletion. Because certain variables such as economic activities and occupations were entered in text form and post-coded at the office, this provided a further opportunity for quality control by checking for inconsistencies and possible data entry errors.

A key goal of the ELMPS 2018 was to track ELMPS 2012 households and all their members, to be able to generate a panel of households and individuals (a panel that now spans 1998, 2006, 2012, and 2018). To maintain a representative sample (Beegle et al., 2011; Himelein, 2014; Vaillant, 2013), households and individuals were tracked, even if they migrated (so long as they remained within Egypt). Two types of attrition could occur during this process. First, entire households that were included in 2012 might be lost in 2018. We refer to the inability to locate entire households from the 2012 wave when fielding in 2018 as Type I attrition. If a 2012 household was found, some of its members may have left to form a split household (for example, young persons may marry and leave their natal household to form a new household). Although fieldworkers collected as much information as possible from the original 2012 household (as located in 2018) to locate split households, sometimes split households could not be located. We refer to the inability to locate the individual(s) in 2018 who were part of a 2012 household as Type II attrition; the main household was found but a split household was not. In this section, we describe and then model these two types of attrition. We then use these models to estimate the predicted probability of each type of attrition as a function of 2012 observable characteristics and use these probabilities to create weights to correct for attrition, as detailed in Section 3.

The ELMPS waves track individuals over two decades, including a large number of individuals who were in multiple waves. Table 6 presents the sample, with an observation being a unique individual, and shows the number and percentage of individuals in different combinations of waves, e.g., in 1998 and 2006 and 2012 and 2018 (i.e., in all waves) or in 2018 only (i.e., 2018 refresher). There were 10,145 individuals tracked from 1998 through 2019 (out of 23,997 in the 1998 wave). Since refresher samples were added in each wave, the number of individuals in multiple waves increases over time. For instance, there were 12,756 individuals in 2006 and 2012 and 2018 (but not in 1998), who can be analyzed as a 2006–2018 panel along with the 10,145 individuals in the 1998–2018 panel.

In addition to tracking the households that were present in the 2012 round into 2018, the ELMPS 2018 added a refresher sample. The refresher sample was focused on over-sampling the 1,000 poorest villages in Egypt, as a key theme of the 2018 wave was economic vulnerability. A poverty map created by CAPMAS in 2013 with the assistance of the World Bank and UNDP was used to identify the 1,000 poorest villages.

The 2013 poverty map was never formally published but was made available to the ELMPS team through CAPMAS.

Table 7 presents the distribution of the refresher sample clusters across these various strata. It is important to note that the 1,000 poorest villages were all in rural areas and only in some governorates. We sampled the poorest villages from all governorates that had them. Likewise, there are no rural areas in Cairo, Alexandria, Port Said, or Suez. Since the ELMPS sample frame excludes the frontier governorates, the refresher sample likewise excludes them. Overall, we planned for a refresher sample of 2,000 households distributed over 200 geographic clusters (primary sampling units or PSUs). Of those, we allocated 60 clusters to urban areas, 100 to poor rural areas, and 40 to non-poor rural areas. Within strata, sampling was carried out according to the principle of probability proportional to size.

Each cluster was designed to sample 10 households. A list of 12 households per cluster was generated in case some households could not be located and 10 were to be collected. There was some nonresponse. Overall, of a planned 2,000 refresher households, 1,953 were sampled. The response rate ranged from a low of 20% (only 2 of 10 households collected, which happened in 2 clusters), followed by 3 clusters with a rate of 60%, 5 at 70%, 14 at 80%, 26 at 90%, 117 at 100%, and 16 clusters at 110% along with 17 clusters at 120%, where the “additional” households were fielded beyond the intended 10. The response rate (97.7% overall) was slightly lower in urban (response rate of 94.2%) than in rural areas (response rate of 99.1%), but no other nonresponse patterns were discernable.

The main idea behind the weights for the panel component of the sample is to generate weights that weight “up” remaining households whose observable characteristics were similar to those households that attrited. From the Type I attrition model, we estimate Pr(A_h), the probability of Type I attrition (attrition of the entire household) for 2012 household h. There may also be splits from that household; refer to a split household as s. For such households we calculate: (1) Pr(Ahs)=1−Pr(h found and 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}}\,{\rm{and}}\,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 }

To adjust for attrition among the 2012 households that were found in 2018, we compute a response adjustment factor, r_h, for original households: (2) rh=11−Pr(Ah) {r_h} = {1 \over {1 - \Pr \left( {{A_h}} \right)}} and r_hs for split households: (3) rhs=1[1−Pr(Ahs)]*cs {r_{hs}} = {1 \over {\left[ {1 - \Pr \left( {{A_{hs}}} \right)} \right]*{c_s}}} where c_s is the number of component households. Component households are the number of originating households in the population (not the sample) that contribute individuals to the new, split household. For example, when two individuals leave their natal households to get married, they come from two households that existed in 2012; this would be a case of two component households. Essentially, this household has “double” the probability of selection and this must be accounted for in weighting by using a share correction (dividing by component households) (Himelein, 2014). If a split contains only members from a 2012 household or born since 2012, there is only one component household.

When calculating panel weights rather than wave-specific cross-sectional weights the share correction was omitted.

The calculation of the 2018 weights for 2012 households accounts for both the household weight in 2012 and attrition. Denote the 2012 weight (whether cross-section or panel) as w₂₀₁₂. We therefore calculate 2018 panel weights as w₂₀₁₈ = w₂₀₁₂* r_h(s). We normalize the weights (dividing by the mean weight) to have a mean of one. When we subsequently normalize the refresher weights, this allows 2012 and refresher households to contribute equally to the sample, on average.

In addition to the 2012 sample, the 2018 wave included a refresher sample over-sampling the 1,000 poorest villages. This subsection describes the creation of the refresher sample weights. These weights can stand alone, and thus serve as further validation of the full sample, as well as feed into the 2018 combined sample weights. Since our sampling strategy was to sample households in a two-stage stratified cluster sample, weights were initially calculated at the household level.

Each stratum in governorate g, urban/rural location l, and poor vs. non-poor area s is made up of a certain number of clusters P_g,l,s as indicated in Table 7. Given that each cluster p in the refresher sample is designed to have 10 households, the total number of households per stratum is given as: hg,l,s=∑p=1Pg,l,s10 {h_{g,l,s}} = \sum\limits_{p = 1}^{{P_{g,l,s}}} {10}

To incorporate deviations from the planned 10 households per cluster, we accounted for the cluster level nonresponse rate. Households start with weight, w_p, such that they represent the planned number of households per cluster, based on the observed number of households in the cluster, m_p: (4) wp=10mp {w_p} = {{10} \over {{m_p}}}

We have population counts from the April 2017 Egyptian Population Census for each of our strata, which we use to weight our sample. Denote the census count of the population in a given stratum as c_g,l,s. We, therefore, calculate household weights as: (5) wp,g,l,s=wpcg,l,shg,l,s {w_{p,g,l,s}} = {w_p}{{{c_{g,l,s}}} \over {{h_{g,l,s}}}}

Using the Population Census to ex-post weight households necessarily generates expansion weights that yield the same number of households as in the population. However, they do not necessarily yield the same number of individuals. We also allow for individual nonresponse (refusal) as part of the individual consent process (71 individuals in the refresher and 490 in the panel refused or could not be completed). Individuals would still be listed in the roster, collected at the household level, but not have the individual interview data. To account for potentially age-group (e) and sex-specific (x) nonresponse, we calculated an age and sex-specific nonresponse rate of r_x,e. We then adjusted the household weight by this nonresponse to get an individual weight, as: (6) wp,g,l,s,x,e=wp,g,l,s1−rx,e {w_{p,g,l,s,x,e}} = {{{w_{p,}}_{g,l,s}} \over {1 - {r_{x,e}}}}

To combine the refresher and panel samples into a single panel with a unified set of weights, first we divided the weights in each group (refresher and from 2012) by their means to have a mean of one. The process for the combined weights follows a very similar structure to the latter part of the refresher weight construction. If w˜g,l {\tilde w_{g,l}} is the normalized weight for a household (either refresher or panel) in a particular governorate-urban/rural location combination, we then calculate the household combined sample weight as: (7) wg,l=w˜g,lcg,l∑w˜g,l {w_{g,l}} = {\tilde w_{g,l}}{{{c_{g,l}}} \over {\sum {{{\tilde w}_{g,l}}} }} where c_g,l is the number of households in the 2017 Census for that same governorate-urban/rural location combination. Again, we allow for sex and age-group-specific nonresponse (r_x,e) and adjust to create individual weights as: (8) wg,l,x,e=wg,l1−rx,e {w_{g,l,x,e}} = {{{w_{g,l}}} \over {1 - {r_{x,e}}}}

In this section, we compare the demographics of ELMPS 2018 with other sources. We start with the size of the population, and then discuss its household size, age structure, marital status, and educational attainment. The 2017 Population Census enumerated 23.1 million households (within the governorates covered by ELMPS), which contained 93.2 million individuals (48.1 million men and 45.1 million women). Given the design of the weights, both the overall ELMPS and the refresher sample generate the same number of households. From the whole ELMPS sample, when we calculated the expansion-weighted population numbers, after adjusting for individual nonresponse, we found 88.6 million individuals (44.0 million men and 44.6 million women). When we calculated the refresher sample expansion-weighted population numbers, after adjusting for individual nonresponse, we found 91.4 million individuals (45.7 million men and 45.8 million women). Both the refresher sample and the full sample found fewer individuals, and especially fewer men, than the Census.

One of the reasons for different numbers of individuals across data sources may be different definitions of households implemented in the different data sources, an issue observed in ELMPS 2012 as well (Assaad and Krafft, 2013). As Figure 3 shows, ELMPS 2018 sampled more households of one person (11%) than the Census (5%), with the LFSs between the two (8%). Likewise, while the ELMPS found 16% of households had one individual, the Census found 8% of households did so, and the LFSs 13–14%. Correspondingly, the Census found that 23% of households had three people, while the ELMPS found 16% and LFSs 13–14%. Likewise, the Census found that more households (27%) were four persons than the ELMPS (20%) or LFSs (21–23%). The sampling of households with five members or above in ELMPS 2018 was almost identical to that of the Census.

Egypt's population is young and has a well-known “youth bulge,” whereby a decline in mortality in the early 1980s followed by a delayed decline in fertility led to a particularly large youth population (Assaad and Roudi-Fahimi, 2007; Krafft and Assaad, 2014; Robinson and El-Zanaty, 2006). The youth bulge were young adults by 2018 (Krafft et al., 2019). Figure 4 presents the distribution of the population by 5-year age groups for the ELMPS 2018, Census 2017, and LFS 2014 and 2017 rounds. While for ages starting around 40, the data had similar age distributions, at younger ages they diverged substantially. The ELMPS 2018 and Census 2017 had a similar share of 30–39-year-olds, but the LFS 2014 and 2017 had slightly fewer. The LFS 2017 was strikingly different than other sources at ages 15–24, showing a large youth bulge at those ages, which was not reflected in the LFS 2014. The LFS 2014 was somewhat more similar to the ELMPS 2018, with some more 20–24-year-olds and somewhat fewer children under 10. The 2017 Census showed less of a “trough” at ages 20–24 than the ELMPS, slightly fewer children 5–9, and more 0–4. These different age compositions are important to keep in mind when considering labor statistics that are strongly related to age, for instance, unemployment rates.

Figure 5 compares the age distribution of the sources in Figure 4 with the 2006 Population Census. To compare the LFS and ELMPS, we calculate what individuals’ ages would have been in 2006 and exclude those who were born after 2006. Because we only have reports, not microdata, for the Census 2017, we subtract two age groups (10 years, so off by 1 year). The spike in ages 15–24 observed above in the LFS is prominent as a spike in ages 5–9 in 2006, and while larger for the 2017 wave, it was also visible in the 2014 wave. The 2006 and 2017 censuses also diverged in terms of the share of the population at younger ages, with the 2017 Census having a higher share from ages 0 to 14 in 2006, but not thereafter. Both the Censuses and the ELMPS show a very similar youth population, starting at 15–19 years in 2006. The Census 2006 shows it as a youth bulge, with less youth thereafter, whereas the higher share of 0–14 years means the bulge is less pronounced in Census 2017. Overall, the ELMPS 2018 is very consistent with the 2006 Census, and the 2017 Census for ages 15+ (in 2006), but diverged at younger ages.

Figure 3

Figure 4

Figure 5

Figure 6 presents the distribution of the population aged ≥18 years by marital status. ELMPS 2018 and Census 2017 align in terms of the share of the population married (68–69%). ELMPS found more divorced (2.0%) individuals than the Census (1.2%) and more widowed (8.9% ELMPS vs. 6.4% Census). These variations are similar to those across LFSs. As a result, ELMPS 2018 found fewer never-married individuals (20.1%) than the Census 2017 (24.4%), LFS 2014 (24.2%), or LFS 2017 (32.4%). These differences may, however, be due to differences in the age composition of each data source, as, for instance, LFS had far more 15–24-year-olds. Differences may also be due to difficulties in distinguishing between different categories of unmarried individuals in large-scale data collection exercises that do not collect the information from the individual person, such as the census and the LFSs.

Figure 6

Figure 7 presents the distribution of the population aged ≥10 years by educational attainment and data source. Patterns are generally similar. For instance, ELMPS 2018 sampled individuals with no certificate (35%) slightly less than the Census (37%). ELMPS found a similar share of university graduates (13%) as the Census (12%) and more than the LFSs (10%), although this may be driven by age composition differences. The share in secondary was similar across ELMPS 2018, Census 2017, and LFS 2017 (28–29%), although LFS 2014 was lower (25%). Differences were largest in terms of primary and preparatory education, 19% in the Census, 23% in ELMPS 2018 and LFS 2014, and 27% in LFS 2017.

Figure 7

In this section, we turn to comparisons of labor market outcomes between the ELMPS and LFS. We first look at trends over time and comparisons across data sources. We then, given some differences in labor market aggregates as well as demographics, show labor market outcomes by age and sex, to disentangle composition differences from differences in outcomes after accounting for composition.

Krafft et al., (2019) explore patterns of labor supply in detail using the ELMPS.

5.2.1

Labor force participation (LFP)

Figure 8 presents the LFP rate of the population aged 15–64 years by sex and in total. The figure includes LFS 2001–2017 rates and the rates and CIs from ELMPS 1998, 2006, 2012, and 2018. While the total LFP rose from 1998 to 2006, it plateaued in the late 2000s and began falling thereafter. The drop in LFP has been particularly pronounced in the LFS in recent years, dropping from 49% in 2016 to 47% in 2017. The ELMPS 2018 found a similar LFP (47%) as the 2017 LFS. Since the LFS published statistics are for 15+ they are not directly comparable; however, comparing the 2018 and first two quarter of 2019 numbers with those of the 2017 LFS shows a continuing decline in LFP rates among ages 15+ (CAPMAS, 2019a, b). Examining the results by sex, the ELMPS female LFP rate is slightly below and the male rate slightly above the trend from the LFS. The ELMPS has historically found higher participation rates than the LFS, which may be in part due to the focus on an individual responding himself or herself, rather than using a proxy respondent (Assaad and Krafft, 2013). The CIs from the refresher sample for women and for the overall rate, but not for men, overlap with the full ELMPS point estimates, suggesting that there may be some small differences in the representativeness of the sample.

Figure 8

Figure 9 presents the LFP rate by age, sex, and data source. Generally, the pattern for men is similar across sources. For men, LFS 2017 had a slightly lower participation rate for ages 15–20 and a slightly higher participation rate for ages 25–45 than ELMPS 2018. In the LFS 2014, participation was higher for all ages. The pattern across time, from 2014 to 2017 and 2018 shows a clear decline in participation across ages. For men, differences in the overall rates were likely driven by differences in age composition, given how similar patterns are by age. For women, participation from LFS 2017 was higher than ELMPS 2018 for ages 15–35, while the ELMPS participation rate was higher than the LFS, but only slightly so, for ages 35–60. We return to these differences for women below, in examining their types of employment.

Figure 9

5.2.2

Employment rates

Figure 10 presents employment rates by data source. Employment rates follow a similar pattern to LFP rates in rising through approximately 2010 and then declining thereafter (the declining trend continued into 2018/19 (CAPMAS, 2019a, b)). The decline, particularly among men, is a notable new trend, although it must be interpreted with some caution given the composition patterns by age in the LFS. The results with the ELMPS 2018 compared with LFS again show similar women's employment, slightly higher total employment, and higher men's employment. The refresher CIs overlap the full sample point estimates in 2018 for the total and for women's employment rates, but not for that of men, as in the case of LFP. For men's employment and total employment ELMPS 2006 and 2012 were also higher than LFS.

Figure 10

In Figure 11, we see that employment rates for men, by age, were very similar in the ELMPS 2018 and LFS 2014 and 2017. The drop in prime-age male employment observed in ELMPS 2018 (Krafft et al., 2019) is thus confirmed by other sources. There were only slight differences with the LFS 2017 having lower youth employment rates for men, and the LFS 2014 having higher employment rates for older men. For women, while the employment rates at ages 35 and older were generally similar, with ELMPS 2018 finding slightly more employment at these ages than LFS 2017, there was again a difference at younger ages in employment rates, with ELMPS 2018 estimating lower employment rates for young women, a point we revisit in examining the nature of employment, below.

Figure 11

5.2.3

Types of employment

Examining the structure of employment by sex, industry, and data source illustrates important patterns that may explain differences in the overall labor market indicators. Figure 12 examines the structure of employment by industry, specifically comparing agriculture vs. non-agriculture and employment status (wage vs. non-wage). All data sources agree that the share of workers in agricultural non-wage work has declined over time. However, when the ELMPS data are overlapped with the time trends for the LFS, they show different employment statuses and industries, especially for women in recent years. In particular, recent LFSs have classified far more women as agricultural wage workers and fewer as agricultural non-wage workers. In general, fewer individuals were classified as agricultural non-wage workers, especially among women. It may be that the LFS employment detection questions are picking up more subsistence work as market work.

Difficulty in capturing the correct employment status of women engaged in agriculture and livestock production has plagued Egyptian data collection efforts for some time. See Anker and Anker (1995); Assaad (1997); Langsten and Salem (2008).

This phenomenon may explain why more young women were classified as employed in the LFS than ELMPS.

Figure 12

Figure 13 shows the structure of employment for those employed outside of agriculture, by sex, comparing data sources. Slightly fewer years of the LFS are included (2008 is excluded), due to different industry coding systems being used over time. Within non-agricultural activities, the LFS and ELMPS show similar distributions of employment by industry. The results were most similar overall and for men, but differences among women may be due to the relatively low employment rates of women and thus more sampling variability. Women were most likely to be employed in education, followed by the “other” sector, which includes several services. The vast majority of women were employed within these two sectors. Men were distributed across a wider variety of sectors, with a rising share in construction particularly.

Figure 13

5.2.4

Unemployment rates

Figure 14 presents the unemployment rates by sex and data source. It is important to keep in mind here that differences in age composition as well as employment and unemployment can contribute to differences in unemployment rates. Unemployment fell from 1998 to the 2000s, and further in the 2006–2008 period, before rising in 2009 and thereafter with the start of the global financial crisis and Egypt's 2011 uprising. Since peaking in 2015, unemployment rates have been declining in the LFS. ELMPS 2018 shows consistently lower unemployment rates than LFS 2017. These rates were, however, consistent with declines in the 2018 LFS and in the first two quarters of 2019, which found that overall unemployment dropped from 10.6% in Q1 of 2018 to 9.9% in Q2, 10.0% in Q3, 8.9% in Q4, 8.1% in Q1 of 2019, and 7.5% in Q2 of 2019. The ELMPS 2018 estimate of 8.2% unemployment overall (8.7% in the refresher sample) is in line with this trend. It is, however, important to keep in mind that discouraged unemployment – where individuals were ready and want to work but were not searching – rose between 2012 and 2018 (Krafft et al., 2019). Women have consistently had higher unemployment rates than men and have experienced smaller declines in unemployment over time; they also had more growth in discouraged unemployment (Krafft et al., 2019).

Figure 14

Figure 15 examines the unemployment rate by age. Historically, unemployment in Egypt has been an educated new entrant phenomenon (Krafft and Assaad, 2015; Krafft et al., 2019), and that generally continues to be the case. Here, ELMPS 2018 and LFS 2017 show similar patterns of women's unemployment, with only small differences at older ages, while LFS 2014 shows a higher rate of young women's unemployment from 15 to 24. The ELMPS and LFSs both show a rise in younger (aged 15–19) men's unemployment, suggesting some shift among men of unemployment toward the younger and less educated (Krafft et al., 2019). LFS 2014 clearly shows higher unemployment among the youngest for both men and women.

Figure 15

Although the 2018 wave of the ELMPS made substantial progress in both questionnaire design and reducing attrition, there remain several limitations to the data that must be kept in mind. While attrition has been reduced and appears to be mostly random, we can only measure and account for how characteristics were observed in 2012 relating to attrition. Dynamic processes that occurred between 2012 and 2018 as well as unobservable characteristics can still bias the resulting sample, which is one of the reasons we carefully compare the results with other data sources in this paper.

Improvements in the labor market history module's design improved the detection and timing of statuses such as unemployment. However, with this and all retrospective data, recall biases remain an issue. Even in the contemporaneous data, measurement error, whether stemming from social desirability bias when answering gender role attitudes questions or challenges with survey fatigue or enumerator data entry errors remain potential problems. Continuing to update the design of the questionnaire and ensure best practices for fielding remains an important issue for the LMPS series. For example, better capturing the care economy and care work, not just market work, is an important area for future improvements, potentially including the addition of a 24-h time use diary (ERF and UN Women, 2020).

The longitudinal, retrospective, and rich data of the ELMPS have already begun to illustrate important developments in Egypt's economy and society, such as fertility, after rising in 2012–2014, once more decreasing in 2018 (Krafft et al., 2019). The rise in non-participation in the labor force (Krafft et al., 2019), and especially the high and increasing share of youth not in education nor employment (Amer and Atallah, 2019), are concerning developments that can be better understood through rigorous research using the ELMPS. These are just a few of the ways that ELMPS can contribute to our understanding of Egypt's economy and society.