Valuing Formal Jobs and Formality in Labor Market Trajectories Using Discrete Choice Experiments: Evidence from Colombia

This study uses a survey-based discrete choice experiment (DCE) to elicit worker preferences for formal employment attributes, formal labor market trajectories and policy uncertainty of social benefits in Colombia. The experimental variation comes from randomly generated hypothetical scenarios presented to respondents.

The survey contains two conjoint modules:

JOB CONJOINT: Respondents are asked to choose between pairs of hypothetical job offers. Each job is described by 11 attributes, with levels drawn randomly and independently for each profile:
(1) Salary: anchored to the respondent's own income, varying from 60% less to 60% more than their current gross monthly income (drawn from a discrete normal distribution centered on the respondent’s current income);
(2) Health insurance: none, free EPS coverage, or EPS plus private medical plan;
(3) Retirement plan: no plan, employer contribution of 8%, 10%, 12%, or 15% of monthly salary to individual account;
(4) Workplace injury insurance: none, half salary during partial disability, or full salary during permanent disability;
(5) Housing subsidy; none, 25 SMMLV, 50 SMMLV;
(6) Work weeks needed to access pension: 350/400, 650/700, 950/1000, 1250/1300;
(7) Life insurance: none, or beneficiaries receive full retirement fund upon death;
(8) Severance pay: none, 10, 20, or 30 days' of salary per year worked upon dismissal;
(9) Vacation: none, 7, 15, or 21 paid vacation days;
(10) End-of-year bonus (prima): none, 15 days, 30 days, or 45 days;
(11) Unemployment severance fund (cesantías): none, employer contribution of 4%, 8%, or 12% of monthly salary to individual account.

In the Job Conjoint only arm, each respondent completes 8 choice tasks in which they accept one of the two job offers (rounds 1-7 with unique profiles and round 8 as a reliability check that reverses the column order of round 2). Attribute order is randomized across respondents but fixed within respondent. After completing these 8 tasks, respondents answer a series of new outcome questions about the same 8 choice tasks: lower dismissal risk (2 tasks), more task independence (2 tasks), more schedule flexibility (2 tasks), and ease of finding job (1 task).

LIFE-HISTORY CONJOINT: Respondents compare pairs of imaginary persons' complete labor market biographies spanning ages 25-70. Each life history is described by 10 attributes:
(1) Number of jobs (5, 10, 15, 22, or 30);
(2) Job type descriptions (informal, formal, or mixed sector jobs);
(3) Formality share (0%, 25%, 50%, 75%, or 100% formal);
(4) Employment continuity (never unemployed, short unemployment spells, or significant unemployment);
(5) Income trajectory (start and end monthly income, varying by formality);
(6) Weeks contributed to social security, pension savings plan type, and retirement fund accumulation;
(7) Health event (workplace injury at 50, diabetes diagnosis at 45, or childbirth complication at 35);
(8) Health care coverage during health event (covered by EPS, SISBEN if fully informal or not covered/paid out-of-pocket);
(9) Total retirement savings at age 70/monthly retirement income;
(10) Presence of policy uncertainty on minimum social security contribution requirements; tax burden from subsidizing defined benefit pension plans; delays on receiving care from public system; and changes in minimum retirement age.

In the life-history conjoint, health care coverage is randomized independently of formality. Informal-covered profiles use the subsidized regime (SISBÉN) rather than contributory EPS. Row order is randomized across respondents but fixed within respondent. Each respondent completes 8 choice tasks (7 unique + 1 reliability check).

In the “Both” conjoint arm, respondents participate in four independent choice tasks of each: four tasks of the job conjoint (offer acceptance outcome only) followed by four tasks of the life-choice conjoint.

All respondents also answer pretreatment background questions on employment, wages, benefits, health, retirement expectations, and demographics. After the assigned arm there is a section on the perceived uncertainty behind social security benefits and policies.

2026-06-22

2026-07-26

(1) Binary choice (selected job or life history) in each conjoint task — the dependent variable for estimating average marginal component effects (AMCEs) and marginal willingness to pay.

(2) For the Job Conjoint: the marginal rate of substitution (MRS) between each nonwage benefit attribute and salary, interpreted as the respondent's implicit willingness to pay for that benefit.

(3) For the Life-History Conjoint: the marginal rate of substitution between formality-related attributes (formality share, EPS coverage, health-event coverage), policy uncertainty realizations, and two numeraires: (i) the present discounted value of lifetime income, computed under a linear trajectory assumption between the randomized start and end monthly income levels, and (ii) monthly retirement income, which is independently randomized. These MRS estimates capture the willingness to pay for formal labor market trajectories in both lifetime-income and retirement-peso units.

(4) Perceived job security: binary choice indicating which job in the pair has lower dismissal risk (asked twice in the Job Conjoint).

(5) Perceived task independence: binary choice indicating which job offers more freedom in choosing tasks and how to perform them (asked twice in the Job Conjoint).

(6) Perceived schedule flexibility: binary choice indicating which job offers more freedom in choosing work hours (asked twice in the Job Conjoint).
(7) Perceived ease of finding: binary choice indicating which job offer would be easier to find in real life (asked once in the Job Conjoint)

The primary outcome variable is a binary indicator equal to 1 if the respondent chose the profile and 0 if they chose the other profile in the same choice task. This variable is regressed on indicator variables of the attribute levels in the profile using linear probability models (OLS), following the standard conjoint analysis framework (Hainmueller, Hopkins, and Yamamoto, 2014). AMCEs are estimated as the coefficient on each attribute level relative to a reference category. The same procedure applies for the secondary outcomes: perceived job security, task independence, schedule flexibility, and ease-of-finding. The MRS between a nonwage attribute and salary is computed as the ratio of the AMCE of that attribute to the AMCE of a unit change in salary (in pesos), yielding a peso-denominated willingness to pay. In the Life-History conjoint, the MRS is estimated using i) the present discounted value of wages over the employment history (25-70 years old) and ii) the monthly retirement income.
We calculate willingness to pay (WtP) for benefit k using the standard formula:
WTPk = – (βk / βwage),
where βk is the AMCE of attribute level k and βwage is the estimated coefficient on the wage attribute in the choice model when wage is measured continuously as a percent of respondent’s current wage.
To measure the cost of institutional uncertainty, we randomly assign an “institutional-uncertainty shock” to life histories. A story is assigned uncertainty over at most one of four possible policy dimensions: (i) an increase in the minimum number of weeks required to qualify for a pension, (ii) an increase in the minimum retirement age, (iii) a forced diversion of a share of an individual-account contributor's wages to the public fund, or (iv) a delay (in months) before publicly covered health care is received, or no shock at all (the stable baseline). Because the shock is randomized orthogonally to the pension, health, income, and formality attributes and does not mechanically alter the displayed benefit outcomes, its effect is identified separately from the value of the benefits themselves. We estimate the AMCE of being exposed to an uncertainty shock relative to a stable story, both pooled across dimensions and separately for each of the four dimensions, using the same linear probability (OLS) and conditional logit specifications described above. The peso-denominated cost of uncertainty is then recovered as the MRS between the uncertainty indicator and salary (and, alternatively, monthly retirement income), computed as the ratio of the uncertainty AMCE to the AMCE of a unit change in the corresponding monetary attribute. Because each shock is also assigned a randomized magnitude (e.g., the increase in required weeks or retirement age, the percentage of wages diverted, or the length of the health-care wait), we additionally estimate whether the cost of uncertainty increases with the size of the rule change, regressing the choice indicator on the shock magnitude within each dimension.

(1) Reservation wage for self-employed workers: the monthly salary that would induce the respondent to close their business and accept salaried employment.

(2) Reservation pension: the monthly pension amount that would induce the respondent to stop working at age 60.

(3) Willingness to pay for an annuity: the maximum lump sum the respondent would pay at age 60 for a lifetime monthly income stream equal to their stated reservation pension.

(4) Subjective probability of main health risk: the respondent's perceived probability of their self-identified main health risk occurring.

(5) Expected medical costs: the respondent's estimate of the annual cost of treating their main health risk, a heart attack, a car/workplace accident, and major organ problems, for an uninsured person.

(6) EPS service quality expectations: ordinal measure of whether the respondent believes EPS would provide timely and effective treatment in a serious health emergency.

(7) Perceived policy uncertainty: respondents' subjective beliefs about the stability of the pension system, measured along three dimensions: (a) the extent to which administrative frictions, bureaucracy, or corruption reduce the effective value of a pension that is formally owed; (b) the perceived likelihood and timing of the private individual-account system (AFP) being eliminated so that all workers must contribute to the public pay-as-you-go system (Colpensiones); and (c) the perceived likelihood and timing of the government increasing the number of contribution weeks required to qualify for a pension.

Reservation wages and pensions are elicited as continuous variables (in pesos). The willingness to pay for the annuity is elicited using a multiple-price-list-style slider in millions of pesos. Health risk probabilities are elicited on a visual probability scale. Expected medical costs are entered as continuous peso amounts.
The policy uncertainty questions include three dimensions. First, benefit erosion: respondents indicate on a three-point scale (agree / neither agree nor disagree / disagree) whether a worker who meets all requirements receives their full pension without losses or delays from administrative, bureaucratic, or corruption-related frictions. Those who disagree then report, on a 0–100% slider, the share of the pension's value they believe would be lost. Second, cross-subsidization of the defined-contribution pillar: respondents rate, on a five-point likelihood scale (from "not at all likely" to "it will definitely happen"), how probable it is that the private defined-contribution system will be eliminated so that everyone instead has to contribute to the public pension system. Those who view this as at least somewhat likely then report the expected time horizon (next year; 1–5; 6–10; 11–20; or more than 20 years). Third, contribution-requirement tightening: respondents rate, on the same five-point likelihood scale, how probable it is that the government increases the number of contribution weeks required to qualify for a pension and then report the expected time horizon on the same brackets.
These secondary outcomes will be analyzed descriptively and used to explore heterogeneity in the primary conjoint estimates (e.g., whether WTP for health insurance varies by health risk perceptions or current EPS enrollment).

DESIGN OVERVIEW

This is a survey-based discrete choice experiment with three experimental arms, implemented as an online survey on Qualtrics and distributed through the Netquest panel in Colombia.

TREATMENT ARMS AND ASSIGNMENT

Respondents are randomly assigned to one of three arms after answering the pretreatment background questions:
Arm 1 — Job Conjoint only (30% of sample, ~1,200 respondents)
Arm 2 — Life-History Conjoint only (30% of sample, ~1,200 respondents)
Arm 3 — Both conjoints (40% of sample, ~1,600 respondents)

Assignment is performed by the Qualtrics randomizer at the block level.

WITHIN-RESPONDENT RANDOMIZATION

Within each conjoint, every choice task presents a fresh pair of randomly generated profiles. Attribute levels are drawn independently for each profile using JavaScript executed in the respondent's browser at the beginning of each experimental arm (about 60% of the way through the questionnaire). For the Job Conjoint, salary levels are anchored to the respondent's own reported wage or income and drawn from a distribution that overweights values near the current wage (using repeated entries). For the Life-History Conjoint, health-event coverage is randomized independently of formality (informal-covered use SISBÉN, not contributory EPS), and income trajectories are correlated to the formality category with overlap in the ranges. In the Life-History Conjoint, each profile is also independently assigned an institutional-uncertainty shock, capturing the risk that the rules governing formal benefits change over a worker's lifetime. Each profile draws at most one “unstable” (i.e. changing) dimension, or none (the stable baseline), from four possibilities: (i) an increase in the number of contribution weeks required to qualify for a pension; (ii) an increase in the minimum retirement age; (iii) a forced diversion of part of an individual-account (AFP) contributor's wages to the public fund; and (iv) a delay before publicly covered health care is received. Each shock is assigned a randomized magnitude (e.g., the size of the increase in required weeks or in the retirement age, the share of wages diverted, or the length of the health-care wait). The shock is decoupled from the benefit outcomes: it is presented to the respondent as narrative text folded into the relevant existing attribute cell, but it does not recompute the profile's pension or health outcomes. Randomizing the shock independently of the benefit levels keeps perceived uncertainty orthogonal to benefit generosity, allowing us to identify the willingness to pay to avoid institutional uncertainty separately from the value of the benefits themselves.

The order of attribute rows is randomized once per respondent (Fisher-Yates shuffle) and held fixed across all choice tasks for that respondent.

NUMBER OF CHOICE TASKS

Job Conjoint: 8 rounds per respondent (7 unique pairs + 1 reliability check) querying offer acceptance, followed by 7 rounds querying ease-of-finding (1), perceived job security (2), task independence (2), and scheduling flexibility (2).

Life-History Conjoint: 8 rounds per respondent (7 unique pairs + 1 reliability check).

Combined arm: 4 job rounds + 4 life-history rounds.

The reliability check (round 8) presents the same two profiles from round 2 but with swapped column positions.

SURVEY STRUCTURE

All respondents complete the following modules before the conjoint tasks: informed consent, age screening, employment screening (based on ENOE methodology), job characteristics, industry, wages/income, informality and benefits, secondary employment, demographics, household composition, retirement expectations, and health status/risk perceptions. Three attention checks are embedded throughout the survey. After their conjoint block they answer questions on the perceived policy uncertainty.


TARGET SAMPLE AND ELIGIBILITY

Target: 4,000 completed surveys.
Eligibility: Adults aged 25-65, currently working at least 30 hours per week in Colombia (either salaried or self-employed), recruited from the Netquest online panel.
In addition, respondents who fail the first attention check will be immediately screened out of the survey. As a robustness check, we will conduct analyses dropping speeders, who we define as respondents who complete the survey in less than 1/3rd the median duration to completion.
We use the following quotas to achieve representativeness (of full-time workers between 25-65) on these variables: sex (2,460 men / 1,640 women), age (2,050 for ages 25-40; 2,050 for ages 41-64), and socioeconomic strata (1: 640; 2: 1160; 3: 1373; 4: 440; 5: 284; 6: 108).

Not available

Randomization is performed by computer using two mechanisms:
(1) Assignment to experimental arm (Job Conjoint, Life-History Conjoint, or both) is implemented via the Qualtrics survey flow randomizer, which assigns respondents to blocks with pre-specified probabilities (30%, 30%, 40%).
(2) Within each conjoint, attribute profiles are generated by JavaScript code executed in the respondent's browser. The code uses a Fisher-Yates shuffle algorithm and Math.random() to independently draw attribute levels for each profile in each choice task. A check ensures that the two profiles in any pair are not identical across all attributes. Attribute row order is also randomized once per respondent using the same shuffle algorithm.

Individual respondent (for arm assignment) and individual choice task (for profile generation within arms)

No

N/A — randomization is at the individual level.

4,000 respondents. At the profile level: approximately 1,200 x 7 x 2 = 16,800 independent observations from Arm 1 (Job Conjoint only); 1,200 x 7 x 2 = 16,800 from Arm 2 (Life-History Conjoint only); 1,600 x 4 x 2 = 12,800 for each from Arm 3. Total profile observations: approximately 29,600 for Job Conjoint and 29,600 for Life-History Conjoint.

Arm 1 (Job Conjoint only): ~1,200 respondents (30%)
Arm 2 (Life-History Conjoint only): ~1,200 respondents (30%)
Arm 3 (Both conjoints): ~1,600 respondents (40%)
Total: 4,000 respondents.

For conjoint experiments, statistical power depends on the number of respondents (N), the number of choice tasks per respondent (T), the number of profiles per task (J=2), and a design effect based on intra-cluster (respondent) correlation (ICC), following the framework in Cohen 1988. For the Job Conjoint (Arms 1 and 3 combined): The number of profiles is 29,600. Assuming an ICC of .10, the minimum detectable effect (MDE) for each attribute level (at 80% power) is 1.96 percentage points. Assuming an ICC of .00, the MDE is 1.63 percentage points. For the Life-History Conjoint (Arms 2 and 3 combined): The power calculations are analogous. The number of profiles is 29,600. Assuming an ICC of .10, the MDE for each attribute level (at 80% power) is 1.96 percentage points. Assuming an ICC of .00, the MDE is 1.63 percentage points.

PAP_DCE_Colombia