Measurement of labour market transitions

The large amount of research undertaken in recent decades on the dynamics of job loss and hiring underscores the importance of the measurement of labour market transitions, often in the literature referred to as flows. Early work on the topic focused on gross flows of workers between unemployment and employment. Hall (1972) and Feldstein (1973) highlighted the crucial role of turnover and transitions in understanding unemployment. In their work, the unemployed are not seen as a constant pool of individuals who want to work but are not doing so, even though they would be available for employment and are actively seeking work.2 Instead, unemployment is seen as the result of changing rates of entry and exit of workers from and into jobs. Most individuals are unemployed for a short period of time, and only a few experience longer unemployment spells.

A vast literature seeks to understand the underlying determinants of the cyclical fluctuations in unemployment. Darby et al. (1986) find that increases in unemployment during recessions are attributable to a rise in the rate at which workers separate from their jobs. In contrast, Hall (2005) argues that adverse shocks increase unemployment through a decline in job finding rates rather than an increase in separations. Consistent with this latter argument, Shimer (2012) finds that most of the variation in the unemployment rate is due to changes in the job finding probability, as opposed to changes in the employment exit probability.

The literature then evolved to allow for multi-state transitions including, for example, inactivity or direct job-to-job transitions. Adding non-employment, Elsby et al. (2015) find that the labour force participation margin is important for understanding fluctuations in the unemployment rate. They conclude that analyses that focus on trends in stocks only, such as the share of the population that is inactive and hence out of the labour force, ignores the cyclicality of the underlying flows, such as the flows between unemployment and inactivity. Fallick and Fleischman (2001, 2004) focus on workers’ movements across employers that occur without a non-employment spell. They similarly argue that overlooking these flows understates the dynamism in the labour market.

A related literature then analyzed the contribution of worker heterogeneity to job finding and separation probabilities. The “heterogeneity hypothesis” posed by Darby et al. (1985) is that during a downturn, the unemployed pool shifts towards a larger share of non-temporary layoffs. These workers have lower job finding probabilities. Thus, it is not that recessions are periods of particularly low outflows from unemployment, but rather that they are periods of high separation rates for a set of workers who are more likely to be unemployed for longer. The composition of the unemployment pool, and not changes in the overall job finding probability is what drives the changes in unemployment rates. Instead, Nakamura et al. (2020) argue that recessions lead to a composition effect that shifts the pool of unemployed workers towards higher skilled individuals, for whom it might be easier to move out of unemployment. This is because job separation rates increase even for this group of workers during downturns.

In the context of developing economies, the analysis of labour market transitions often focuses on transitions between formality and informality, and vice-versa. Given the vast prevalence of informal employment in many developing countries (ILO, 2018), which to a large extent determines workers’ access to social protection, formal worker representation and other aspects of job quality and decent work, this margin of transition is particularly relevant for these countries. Bosch and Maloney (2007) consider five mutually exclusive states across which workers transit in Mexico between 1987 and 2002: inactive, unemployed, informal self-employed, informal salaried and formal salaried. They find that while the cyclical properties in job finding and job separation probabilities are similar in Mexico and the US for formal salaried workers, the pattern is reverse for self-employment. Herrera et al. (2005) explore labour market transitions across different formality statuses in Peru. Tanzel and Ozdemir (2019) analyze the case of Egypt, while Gutierrez et al. (2019) focus on Bangladesh. These analyses underscore the usefulness of a flow approach, that considers transitions between formal and informal jobs, as opposed to a stock approach, which focuses on formality and informality rates at given points in time to study the evolution of the labour market.

Biases caused by attrition and misclassification error

Abowd and Zellner (1985) highlighted two issues that arise when estimating gross labour flows using rotating panel surveys. First, by construction, labour market transitions cannot be observed for individuals who are not matched on consecutive survey waves. Regardless of whether the inability to match observations across time arises as a result of temporary sample drop-outs, incomplete survey responses, or incorrect identifiers, these individuals’ transitions are excluded from measures of labour market transitions. This gives rise to attrition or rotation group bias if the probability of its occurrence is not random. Second, measurement or classification error can arise when workers misreport their labour market situation, or when interviewers introduce errors when recording the survey response. Misclassification error usually cancels out or has a minimal effect on stocks but can substantially affect flow estimation. For example, if one worker who was unemployed in one survey wave is incorrectly classified as employed in the next survey wave, even though she has actually not changed her status, and if the reverse is true for another worker, the overall error that occurs when measuring stocks might cancel out. These two classification errors, however, do contribute to an over-estimation of transitions between employment and unemployment and vice-versa.

The existing literature developed various approaches to address attrition bias. The simplest approach is to rely on a “missing-at-random” (MAR) assumption. If there is no correlation between attrition probability and labour market transitions, then it is possible to simply drop the unmatched observations to estimate unbiased flows. However, analyses of various countries and time periods indicate that the missing at random assumption is unlikely to hold in most settings.3

There are different approaches to deal with non-random attrition bias. Donovan et al. (2020) use post-stratification of the sampling weights so that the distribution of the longitudinally matched sample resembles the distribution in the cross section. Fujita and Ramey (2009) propose an approach that minimizes the differences in the stocks derived from official cross-sectional data and stocks imputed from worker flows. Bleakley et al. (1999) and Fallick and Fleischman (2001), among others, use an approach similar to that employed by the Bureau of Labor Statistics (BLS) to reweight observation in the United States’ Current Population Survey. The approach relies on estimating the attrition probability for different groups in the sample, and then using the inverse estimated match rate to reweight the observation in the sample.

Also classification error can be an important source of bias in labour market flows. Clark and Summers (1979) and Poterba and Summers (1986) discuss the potential concerns of spurious transitions arising from inconsistent reporting in survey data. They estimate the bias due to classification error and correct the bias using data from re-interviews.4 Assaad et al. (2018) find important inconsistencies in cross-time reporting across different panel waves, even for time-invariant information, suggesting that classification error can be an important source of bias in some instances.

Biases caused by time aggregation

Shimer (2012) highlights the importance of time-aggregation bias when analyzing labour market flows measured in discrete time intervals. Time aggregation bias occurs because data on a worker’s situation in the labour market are not observed continuously, but only at discrete time intervals such as quarters in the case of a quarterly survey. Transitions that occur within this time interval are not observed. In the context of transitions into and out of employment, “Ignoring time aggregation will bias a researcher towards finding a countercyclical employment exit probability, because when the job finding probability falls, a worker who loses her job is more likely to experience a measured spell of unemployment” (Shimer, 2012, p. 129). Shimer (2012) corrects for this bias by explicitly working in a continuous time model in which data are available at discrete intervals.

Data

This paper next estimates gross worker transitions and transition probabilities for Viet Nam in 2011-19.5 The paper uses micro-level data on individuals’ employment status and job type in consecutive quarters, available from Viet Nam’s quarterly labour force survey, to construct time series for the gross transitions of workers into and out of a given labour market status.

Viet Nam’s labour force survey is run as a rotating panel, in which each quarter includes data from newly surveyed households for about half of the observations, while the remaining half of the observations correspond to households that were first surveyed in the previous quarter. However, individuals can only be tracked between different quarters within a year, not across years. Transitions can hence be calculated between Q1 and Q2, between Q2 and Q3, and between Q3 and Q4 for every year; they cannot be calculated between Q4 of one year and Q1 of the next year. In line with ILO definitions, individuals who during the first of the two survey waves are under 15 years old are excluded. Furthermore, individuals which for any reason cannot be unambigously tracked over two different quarters are excluded6, as well as those who are followed in non-consecutive waves, or who appear to be followed for more than two waves in the survey.

The number of individuals that can be tracked across different quarters varies slightly over time. On average across all quarterly pairs of surveys, the number of individuals that can be tracked is about 65,000. The number varies between 56,000 and 85,000 in 2011-19.7

Gross transitions

The paper examines gross labour market transitions across four different labour market categories. These categories, and the corresponding, mutually exclusive labour market statuses that encompass them, are:

1. Labour force status (LFS): employed, unemployed, out of the labour force.

2. Formal job (FNJ): employed with a formal job, employed with an informal job, not employed.8

3. Industry (IND): 86 2-digit or 524 4-digit industry classification categories, plus a not employed group.

4. Occupation (OCCU): 11 2-digit or 568 4-digit industry classification categories, plus a not employed group.

Let $C\in$ {LFS, FNJ, IND, OCCU} be the set of these 4 categories, labelled as c. Each c is composed of s_c mutually exclusive statuses. For each possible pair of statuses (j,k) in category c, the paper calculates flows from state j in quarter t to state k in quarter t+1 as the weighted sum of all individuals who transition between these two states

The quarterly transition probabilities out of state j into state k in period t are obtained by dividing the weighted flows by the weighted sum of all individuals in state j in period t.

Weights

In all cases, the analysis in this paper considers two possible weight: sampling weights and attrition revised weights. The former uses the sampling weights provided by Viet Nam’s General Statistics Office as frequency weights, without any further adjustment. These weights are meant to reflect the occurrence of each observation in the general population of Viet Nam. Tables and figures presented in this paper show the sampling weights estimates.

For the attrition revised weights, the paper follows Fallick and Fleischman (2001) and uses attrition weighted inverse probability weights. The approach relies on estimating the attrition probability for different groups in the sample, and then using the inverse estimated match rate to reweight the observation in the sample. To do so, in a first step, probit models of attrition are estimated for each quarter in the data using equation (2) where $attr{it}_{i,t}$ is an indicator variable equal to 1 if individual i who is in the survey’s sample on quarter t cannot be matched to an observation in period t+1, and zero otherwise. ${\gamma }_g^t$ are a set of fixed effects for the 3-way interaction between i’s gender, labour force status9, and age group in 10-year age bins. These effects are allowed to vary by quarter.

Then, the predicted attrition probability for each group g is used to estimate a relative attrition probability and scale the sampling weights using the relative attrition probability.

where ${\omega }_i$ is individual i’s sampling weight and $attri{t}_t$ is equal to 1 minus the share of matched observations in quarter t. By rescaling the sampling weights in this manner, our estimation gives more weight to transitions by individuals in groups who are more likely to drop off the sample thus compensating for their higher attrition rate.

Given that the Viet Nam labour force survey only tracks households for two consecutive quarters, plus the lack of data on re-interviews, this paper does not attempt to assess the extent of classification errors or make any corrections to account for their existence. In this regard, it is just worth to point out that estimates of labour market transitions that involve more disaggregated labour market categories (such as 4-digit industry or occupation classifications) are likely to suffer more from this type of bias. However, there is no unusual increase in gross flows between categories across different levels of disaggregation, which to some extent assuages concerns of misclassification.

Time aggregation

This paper follows Gomes (2015) and uses the quarterly labour force survey data to extrapolate monthly transition rates across labour market states. Continuous transitions are calculated by taking the limit of the extrapolated transition matrix, as time elapsed between observations tends to zero.

Following Gomes (2015), let $n_q$ be the $s_c\times s_c$ square, quarterly transition matrix obtained from quarterly data for labour market status category C. Then, the monthly transition matrix, $\widehat{n_m}$, is estimated as follows:

where ${\mu }_q$ is the diagonal matrix of eigenvalues for matrix $n_q$ and $p_q$ is a matrix of its eigenvectors. We then obtain the continuous time transition rates from the off-diagonal elements of the matrix  $\widehat{\lambda }=\underset{\mathrm{\Delta }\to 0}{\lim }\frac{p_q{\mu }_q^{\mathrm{\Delta }}{p}_q^{-1}-I}{\mathrm{\Delta }}$ where $I$ is the identify matrix.

Ideally, in order to assess the accuracy of this approach, one would like to compare these extrapolated monthly transition rates to those estimated from data collected on a monthly basis. As Viet Nam’s labour force survey data have quarterly frequency, there is no possibility to make such a comparison with data from a monthly-level database.

Labour force status

The first and second rows of Figure 1 show, respectively, inflow and outflow probabilities from employment in Viet Nam from 2011 to 2019. Each panel shows quarterly transition probabilities into and out of employment, from and into unemployment and inactivity, respectively. Figure 1 also shows the extrapolated monthly probability calculated using equation (4) above, as well as the continuous transition rate.

Individuals flow into employment from either unemployment or inactivity. Not surprisingly, comparing panels (a) and (b) in the first row of Figure 1, indicates that the job finding probability is approximately four times larger for unemployed individuals than for those who are inactive. But for a short-lived decline in 2016, the quarterly job finding probability from unemployment was relatively stable throughout the period at a mean value of 40 per cent and a standard deviation of 5 percentage points without taking into account seasonal variation. In other words, workers that are unemployed in one quarter have a probability of 40 per cent of being in employment in the next quarter.

Job separations to inactivity are more likely than separations to unemployment, although it is important to note that outflows to unemployment may be underestimated if unemployment spells are shorter than three months and hence unobserved at a quarterly frequency. In other words, there could be some individuals that upon losing their job become unemployed, but find a new job in less than three months time, which makes them appear in the labour force survey as continuously being in employment. After 2015, quarterly job separations into unemployment declined from a peak of 0.6 per cent to under 0.2 per cent by the end of 2018.

Figure 2 depicts transitions probabilities between unemployment and inactivity. The quarterly transition probability from unemployment to inactivity declined from 27 to around 10 per cent between 2011 and 2015, before rising again more recently to above 20 per cent. The probability for an inactive worker to become unemployed has been on a declining trend.

Figure 1. Job finding and job separation probabilities

Notes: The data series that are shown use (unadjusted) sampling weights. See section 2 for more details. Breaks observed for some data series between 2018 and 2019 might be driven by a major revision of sampling weights.

Figure 2. Transition probabilities between unemployment and inactivity

Notes: See figure 1 notes.

Table 1 shows the average quarterly transition probabilities, the mean continuous transition rates and their standard deviations. Note that because they are calculated by taking the limit as time intervals shrink to zero from extrapolated monthly rates (see equation (4)), the estimated continuous transition rates are closer to the transitions that occur per month, rather than per quarter. Multiplying the continuous rate by 3 makes both rates more directly comparable.

Table 1. Transition probabilities (per cent) and volatility in labour market status, 2011-19 (average)

Notes: The data points that are shown were calculated on the basis of (unadjusted) sampling weights. See section 2 for more details. They correspond to the simple average of quarterly transition probabilities and continuous transition rates in 2011-19.

Formality

This section focuses on formality as defined by the nature of the job, which is for example determined by whether the worker has access to social security schemes. Figure 3 shows a relative stable overall employment rate between 2011 and 2018. However, the share of employed workers in formal jobs increased steadily from 20 to close to 30 per cent between 2013 and 2018.

The increase in formalization was not equally experienced by all groups in the workforce. Figure 4 separately calculates the share of formal employment by education levels for men (panel (a)) and women (panel (b)). Women with basic and intermediate education observed the largest increase in formalization (of 9 and 8 percentage points, respectively). The same education groups also had the largest increase for men (6 and 7 percentage points).

Figure 3. Employment and formality rates

Notes: The employment rate was calculated as the share of employment in the working-age population. The formality rate was calculated as the share of formal employment in total employment.

Figure 4. Formality rates by sex and level of educational attainment

Notes: The formality rates shown in this figure were calculated as the share of formal employment in total employment by sex and level of educational attainment.

Figure 5 shows the quarterly transition probabilities (using sampling weights) from and to formal employment. Panel (a) presents transitions between formal and informal jobs without an observed non-employment spell, while panel (b) shows transitions between formal jobs and non-employment. These time series indicate that the increase in the share of workers in formal employment shown in Figure 3 can be attributed to a decrease in the separation rate of formal workers to informal jobs (from 9 to 7 per cent between 2013 and 2018) and an increase in inflows to formal jobs from informal employment (from 3 to 4 per cent within the same period). Formal jobs became more stable, and more likely to find for informally employed individuals.

Panel (b) shows that the same is not true for unemployed and inactive workers. The probability of transitioning to a formal job from non-employment remained stable around 1 per cent, as did separations from formal jobs to non-employment. It is important to note that these findings may be affected by time aggregation bias. It may be the case that transitions between informal and formal jobs shown in panel (a) are mediated by unobserved periods of non-employment. However, the quarterly frequency in the survey data plus the lack of granular information on unemployment spell durations does not allow us to correct for this further. Regardless of whether the transitions are mediated through unemployment spells or not, it is interesting to see that individuals with an informal employment spell a quarter prior are more likely to transition to a formal job than those without it.

Figure 5. Formal job finding and job separation probabilities

Notes: The data series that are shown were calculated on the basis of (unadjusted) sampling weights. They show quarterly transition probabilities.

Industry

Workers’ industry or sector of employment is grouped into categories using the 2 and 4-digit level of the International Standard Industrial Classification of All Economic Activities (ISIC). At the 2-digit level, there are 86 industries of which 10 categories consistently encompass three-quarters of the employed population in Viet Nam between 2011 and 2019.

This section provides an overview of labour market transitions between economic sectors at the 2-digit level. The section considers outflows to a different economic sector that occur without an observed non-employment spell (cross-sector job-to-job transitions), and outflows to non-employment. Table 2 presents the average quarterly outflow probabilities from each sector s to a different sector -s,  ${outflow}_{s,-s}$, and the outflow probability from sector s to non-employment, ${outflow}_{s,nonemp}$, which are respectively calculated as follows:

${outflow}_{s,-s}=\frac{1}{T}\sum _t\left[1-\left(\frac{\sum _{i\in \{t\&t+1\}}{\omega }_i\times I\left[{ISIC2}_{i,t}=s \&{ISIC2}_{i,t+1}=-s\right]}{\sum _{i\in \{t\&t+1\}}{\omega }_i\times I\left[{ISIC2}_{i,t}=s\right]}+outflow rat{e}_{s,nonemp}^t\right)\right]$

${outflow}_{s,nonemp}=\frac{1}{T}\sum _{t}outflow rat{e}_{s,nonemp}^t=\frac{1}{T}\sum _t\frac{\sum _{i\in \{t\&t+1\}}{\omega }_i\times I\left[{ISIC2}_{i,t}=s \& LF{S}_{i,t+1}\ne emp\right]}{\sum _{i\in \{t\&t+1\}}{\omega }_i\times I\left[{ISIC2}_{i,t}=s\right]}$

where $i\in \{t\&t+1\}$ refers to the set of individuals that we can match in the labour force survey in the consecutive quarters t and t+1, and ${\omega }_i$ corresponds to the sampling weight for individual $i$. ${ISIC2}_{i,t}$ denotes the ISIC 2-digit level sector in which individual $i$ is employed in quarter $t$. ${LFS}_{i,t+1}$ denotes the labour force status of individual $i$ in quarter $t+1$, where $emp$ stands for employed.

On average, 13.3 per cent of workers transition to a different sector in consecutive quarters without an observed unemployment or inactivity spell. “Financial service activities, except insurance and pension funding”, “Education”, and “Human health activities” are the sectors with the lowest probabilities of transition to another economic sector. At the other end of the spectrum, 68 per cent of all workers in “Information service activities” on average transition to a different sector in the following quarter. Various factors are likely to account for differences in the probablities to transit to another sector, including skills requirements, sectoral differences in wages, informality rates or other job quality indicators. For the vast majority of 2-digit economic sector categories (82 out of 86) the most likely average outcome for a worker is to remain in his or her current economic sector.

Table 2. Quarterly job-to-job transitions to a different sector and quarterly transitions to non-employment (per cent), 2011-19 (average)

Notes: The data points that are shown use (unadjusted) sampling weights. See section 2 for more details. They correspond to the simple average of quarterly transition probabilities in 2011-19.

To provide an example of the type of analysis that can be done with transitions, this section looks at whether transitions out of an economic sector covary with wage growth in the sector. This analysis can provide some preliminary evidence on whether transitions across sectors accompany wage growth, with workers leaving sectors in decline for sectors experiencing wage growth. To do so, the median and average monthly wage for salaried employees is calculated in each 2-digit sector per quarter. To prevent outliers from having too much weight on the average wage, wages are winsorized at 1 per cent.

Figure 6 panel (a) indicates that there is a negative correlation between the 2011-18 growth of the average wage of a sector and the average cross-sector outflow probability. Consistent with workers seeking opportunities in sectors with better wage growth prospects, the sectors that experienced smaller wage growth in the period are those that on average have higher outflow probabilities to other sectors. The average probability of transitions to non-employment does not appear to be correlated with wage growth (panel (b)).

Figure 6. Sectoral outflow probabilities and sectoral wage growth

Notes: Mean outflow probability corresponds to the average quarterly outflow probability in 2011-18, calculated on the basis of (unadjusted) sampling weights. See section 2 for more details. Wage growth in 2011-18 was calculated as the change in the natural logarithm of the average sectoral wage between 2011 Q3 and 2018 Q3. Each dot corresponds to one sector at the ISIC 2-digit level.

Occupation

Workers’ occupation is grouped into categories using the 2 and 4-digit level of the International Standard Classification of Occupations (ISCO-08). At the 2-digit level, there are 10 occupational categories plus a “Not Elsewhere Classified” group. In the following, this section provides an overview of labour market transitions out of different occupations at the 2-digit level.

Table 3. Quarterly job-to-job transitions to a different occupation and quarterly transitions to non-employment (per cent), 2011-19 (average)

Notes: The data points that are shown use (unadjusted) sampling weights. See section 2 for more details. They correspond to the simple average of quarterly transition probabilities in 2011-19.

As shown in Table 3, on average 15.3 per cent of workers transition to a job in a different occupation without an observed unemployment spell. Cross-occupation transition probabilities range from 37 per cent for clerical support workers, to 10.4 per cent for service and sales workers. Table 4 presents the average quarterly transition matrix across all occupation categories. The shaded diagonal shows the average percentage of workers who remain in the same occupation in the following quarter. The most common outcome for an employee is to remain in the same occupation. Conditional on switching occupations (without a non-employment spell), there is some indication that workers seek occupations that are to some extent similar to their previous one. For example, the most like cross-occupation outflow for managers is to the professionals category. The most likely transition from agricultural, forestry, and fishing occupations, conditional on switching, is towards elementary occupations, and vice versa.

Table 4. Quarterly occupational transition matrix, 2011-2019 (average)

Notes: The data points that are shown use (unadjusted) sampling weights. See section 2 for more details. They correspond to the simple average of quarterly transition probabilities in 2011-19. Each cell shows the transitions from one occupation in one quarter to the same or another occupation in the next quarter. All rows and columns add up to 100 per cent; any difference is due to rounding.