The health capacity to work at older ages in urban China

Abstract

Tapping into the older workforce is a potential economic solution to population aging, but its feasibility depends on the health capacity to work among older people. Existing estimations in OECD countries involve establishing the relationship between work and health on a younger cohort, extrapolating the relationship to older individuals, and deriving the excess health capacity as the difference between predicted and actual employment rates. However, benchmarking on the younger cohort is sub-optimal because the observable retirement-health relationship changes with age. The dual nature of the Chinese social security system provides us with a relatively neat benchmark, allowing us to estimate the excess health capacity among urban workers benchmarking on rural residents in the same age range. Using the China Health and Retirement Longitudinal Study, this choice, combined with other fine-tuning, yields significantly lower but still substantial excess capacity among older urban workers than benchmarking against younger cohorts. Altogether, among urban Chinese aged 45–69, 31.2 million extra workers can potentially be added to the workforce.

1. Introduction

Population aging poses significant challenges for an economy by reducing the number of workers relative to retirees. In a county dominated by a pay-as-you-go social pension scheme, a worsened support ratio forces an economy to either raise taxes on workers or to risk snowballing deficits, threatening macroeconomic stability. China is one of the fastest aging countries in the world. For every older person (defined as 60 and older), the number of working-age persons (defined as aged 20–59) is 3.62 in 2018, and this measure of support ratio will quickly decline to 1.32 in 2050.1

One effective way to alleviate the pressure of population aging is to delay retirement. Many countries have already done so, but the process is often contentious (Coile, Milligan, & Wise, 2017a). One of the most commonly used counter arguments has been that older people are unhealthy and thus incapable of working. Indeed, it is inhumane to force an ailing person to work, no matter how much economic sense it makes. Therefore, to gauge the potential of postponing retirement, it is crucial to examine the work capacity among older persons based on health considerations alone.

A series of highly influential studies organized by the NBER estimated the health capacity to work of older people in OECD countries, including the U.S., Canada, Japan, the United Kingdom, Germany, and several other European countries (papers collected in Wise 2017). Using surveys in the family of HRS (Health and Retirement Study) around the world which contain rich and comparable information on health and employment, sizeable excess capacity is found in every studied country. For example, unutilized work capacity among men aged 65–69 is 40% in Japan, 31% in the U.S., 57% in the U.K., and 84% in Germany. For women in the same age group, the additional work capacity is 31% in Japan, 29% in the U.S., 57% in the U.K., and 68% in Germany.

It is worth emphasizing that these studies define the health capacity to work mainly on health status, aside from very limited demographics. Naturally, an individual’s decision to retire is also affected by considerations such as social security, private pension, and disability insurance programs, which have contributed to the declining labor force participation of older workers in OECD countries in the 20th century (Gruber & Wise, 1999, Wise, 2016). The stabilization or even rising labor force participation of older workers in recent decades is also attributable to policy changes (Wise, 2017). Caring responsibilities (for parents and grandchildren) also affect retirement decisions. By focusing on health, this definition of work capacity deliberately measures physical capacity alone, casting aside individual preferences for leisure or home production, as well as employer considerations under existing institutions. Therefore, it measures the size of the potential workforce that can be tapped into by altering institutions or incentives without appearing to be inhumane.

The procedure of the estimation goes as follows. First, pick a benchmark group whose decision to work is free from institutional distortions and estimate the relationship between employment and health. Secondly, use the estimated model to predict employment among older people who qualify for social security retirement. Finally, compare actual and predicted employment rates to derive the excess health capacity to work. The procedure sounds simple, but the devil lies in how to choose the benchmark group. In the NBER series, the benchmark group is a younger cohort just before they reach the retirement age, so essentially they assume the retirement-health relationship to be the same between younger and older cohorts.

Unfortunately, this assumption may be flawed because the relationship between observed work and health may change with age. For example, chronic disease at an older age may be more debilitating as it may come with other unobserved functional declines. Thus, using younger cohorts as the benchmark group may overstate the extra capacity to work.

While in OECD countries, finding a suitable benchmark group may be difficult, the dual-retirement system in China offers such an opportunity. In urban areas, workers enjoy generous retirement pensions at relatively young ages, but rural people do not have this privilege and usually work until advanced ages (Giles, Lei, Wang, & Zhao, 2015). Even though the New Rural Pension Program was introduced in 2009 and has provided universal social pension coverage, the amount of pension benefit is minuscule (Lei, Zhang, & Zhao, 2013) and insufficient to induce early retirement. Existing studies have found that although urban Chinese residents generally have better health status than rural residents (Gong et al., 2012; Strauss et al., 2010), the retirement rate is much higher among urban residents (Giles et al., 2015). The capability of rural workers to participate in the labor market under worse health conditions suggests that the excess work capacity of urban workers is likely substantial. It is thus our goal to estimate the health capacity to work among older urban workers. In the process, we evaluate the merits of choosing various benchmark groups and estimation methods.

The structure of this paper is as follows. Section 2 briefly describes the Chinese retirement system in urban and rural areas to illustrate the rationale for using rural residents as the benchmark group. Section 3 discusses the literature, lays out the estimation strategy, and describes our data. Benchmark regressions and prediction results are presented in Section 4, followed by education-stratification in Section 5. Finally, Section 6 summarizes our findings.

2. Background: the Chinese retirement system

The Chinese retirement system has a dual structure separable by urban and rural sectors, mirroring that of the economy. The formal sector, which covers the majority of urban workers, enjoys generous social pensions payable upon reaching the statutory retirement ages (Zhao & Zhao, 2018). The statutory retirement ages, set in the 1950s, are 50 for blue-collar women, 55 for white-collar women, and 60 for all men. After processing retirement, although a worker is free to seek new employment, it is often difficult due to the loss of job-specific human capital and discrimination against older workers. The mandatory nature of retirement, commonly seen in government institutions and state-owned enterprises, reflects the unwillingness of employers to keep older workers, probably because senior workers receive salaries that are higher than justifiable by their productivity due to a rigid salary scale. On top of early statutory retirement ages, the urban retirement system also allows for early departure from the workforce. Those who work in occupations that are dangerous, harmful to health, or require heavy physical labor can retire and receive pensions up to five years before the statutory retirement ages. Overstaffed firms also have an “internal retirement system,” which lets underemployed workers retire early on a minimum salary before they become eligible for social pensions.

Many countries have raised the retirement age in recent decades in response to population aging, but China has not. One important reason is that the rural sector has served as a bumper to rising pressure on social security. Rural residents, defined by agricultural household registration (hukou in Chinese), do not have a formal retirement scheme. Before the late 1980s, China restricted labor migration from rural to urban areas (Zhao, 1999). After the relaxation, large scale migration from rural to urban areas occurred, allowing young rural workers to replenish the social pension pools, alleviating the pressure of funding gaps caused by population aging. In 2009 the Chinese government rolled out a rural social pension scheme (New Rural Pension Scheme) to cover rural residents. This program is composed of two parts, a non-contributory social pension paid by general fiscal revenue and an individual pension account accruable through contributions, both payable at age 60. The rural social pension allowance is low – 50 yuan a month initially, rising to 90 yuan a month in 2019. The rural pension scheme is not employment-based, and the amount of pension income is well below the subsistence level. Therefore, it provides little disincentive to work past age 60.

Older rural Chinese have traditionally kept working for as long as their health permits (Dwayne, Loren, and Fan (2003). Using data from the first three waves of CHARLS (2011–15), it is clear that the labor force participation rates of rural residents are substantially higher than their urban counterparts (Fig. 1). Giles et al. (2015) show that while retirement hazard at statutory retirement ages displays sharp spikes for urban workers, the age pattern of actual retirement is very smooth among rural residents. This unique urban-rural divide in the Chinese retirement system justifies the use of rural residents as the benchmark group to study the excess work capacity of urban residents. It is reasonable to assume that conditional on demographics (age, sex, education), urban and rural people with the same observed health conditions have the same physical capacity to work.

Fig. 1.

Labor Force Participation Rate of the Older Population, 2011–2015.

Data: China Health and Retirement Longitudinal Study (2011, 2013, 2015 pooled).

3. Literature, data and estimation strategy

The literature has used two main methods to estimate excess health capacity to work. One compares declines in cohort mortality (as a proxy for health improvement) and changes in the employment of older people over a long period (Milligan & Wise, 2015). Excess work capacity thus appears when the health of a certain age improves, whereas the employment rate does not rise to the same extent. This method illustrates long-term trends in excess work capacity, but can not give a precise estimation due to the crude nature of mortality as a measure of health conditions.

We adopt the approach by Cutler, Meara and Richards-Shubik (CMR) (Cutler, Meara, & Richard-Shubik, 2013), which relates employment to comprehensive indicators of individual health conditions available in the Health and Retirement Study. As mentioned earlier, this method applies the retirement-health coefficients estimated from a benchmark group to make out-of-sample predictions of retirement rates of the older group, and take the difference between actual and predicted retirement rates as the excess work capacity. The approach is applied to the U.S. (Coile, Milligan, & Wise, 2017b), Canada (Milligan & Schirle, 2017), Japan (Usui, Shimizutani, & Oshio, 2017), the United Kingdom (Banks, Emmerson, & Tetlow, 2017), Germany (Jürgen, Thiel, & Rsch-Supan, 2017), and many other European countries (Bingley, Datta, & Pedersen, 2017; Blanchet, Caroli, Prost, & Roger, 2017; Johansson, Laun, & Palme, 2017; Jousten & Lefebvre, 2017).

The relationship between work and health is assumed to take the following form:

$$\text{work}it={\beta }_0+{\beta }_1\text{Health}it+{\beta }_2\text{X}it+{\epsilon }_{it}$$ (1)

where i denotes individual, and t denotes wave. Control variables (Xit) include age group dummies, marital status, education, provincial fixed effects, and wave fixed effects. The number of non-health control variables is deliberately small to maximize the explanatory power of health status.

Applying coefficients from estimating Eq. (1) on a benchmark group to predict the counterfactual work status of a study group requires certain assumptions. To illustrate, we decompose the residual in Eq. (1):

$${\epsilon }_{it}=s_{it}+u_{it}+{\delta }_{it}$$ (2)

where s_it is unobserved institutional factors such as the social security retirement system that entices a worker to exit the labor market. The counterfactual state is when s_it = 0. The literature, by using the pre-retirement age cohort (50–54) as the benchmark group to simulate for the policy influence-free world decision among older cohorts (55–74), indeed (largely) satisfies this condition.

The second challenge is represented by the term u_it, which is health conditions unobserved by the researcher but correlated with observed health (Healthit). As is well known, the biological aging of a body is multi-dimensional, so that multi-morbidity is more common among older persons. Therefore, those who fare worse in observed health conditions are also likely to be worse in unobserved conditions. Not controlling for u_it causes the health coefficient (β₁) to be smaller among younger than older cohorts, rendering the benchmark group relationship invalid.

The Chinese rural-urban distinction provides a cleaner benchmark group – older workers in the rural sector overlap with urban workers in age, but do not benefit from the generous social pension programs afforded to urban people. Therefore, we have an opportunity to provide a better estimation of the counterfactual work status for urban people at older ages.

Following the literature, we estimate excess work capacity for each five-year age group – 45–49, 50–54, 55–59, 60–64, and 65–69. Our first set of estimation uses rural people aged (45–69) as one benchmark to estimate the counterfactual work status of urban people of the same age range, but by five-year age groups. Because the work-health relationship is age-specific, which we confirm using the rural sample, as the second set of estimation, we also use rural residents in the same 5-year age groups to benchmark urban residents.

Using rural residents as the benchmark has a drawback. Due to the lack of social protection, personal wealth and family support, rural people may stretch their work-life well beyond reasonable health limits, causing the predicted excess work capacity among urban people to be too high. To be “fair” to urban people, we trim the rural sample by excluding some people who work too hard.²

To see how sensitive our benchmark method is in comparison to that used in the literature, as the third benchmark estimation, we use as benchmark groups younger urban men five years before the statutory retirement age (45–54).

We use data from the China Health and Retirement Longitudinal Study (CHARLS). Similar research on other countries has used surveys in the Health and Retirement Study (HRS) family, such as the English Longitudinal Study of Ageing (ELSA) and Survey in Health, Ageing and Retirement in Europe (SHARE). As the CHARLS survey is harmonized with the HRS family of surveys, our results are broadly comparable when similar estimation techniques are used. CHARLS is a nationally representative longitudinal survey of the middle-aged and elderly population (45+) in China (Zhao, Hu, Smith, Strauss, & Yang, 2014). The survey includes detailed information on demographics, health conditions, and employment to enable the study of health capacity to work.

The CHARLS baseline survey of 2011–12 contains about 17,700 individuals. Follow-up surveys took place in 2013 and 2015. In this paper, we pool the CHARLS 2011, 2013, and 2015 waves to increase our sample size.³ The total sample size of this study is 13,226 for rural men, 14,250 for rural women, 4955 for urban men, and 4614 for urban women. As one person may appear in multiple waves, we cluster standard errors at the individual level.

The dependent variable (Workit) is a dichotomous variable, which takes value 0 if a person did not work and 1 if worked. The CHARLS survey, due to agricultural seasonality, defines farm work as working for at least ten days in the past year and defines non-farm work as working for at least one hour in the last week. We drop from our analysis people who have never worked in their entire life.

The “health” variable in Eq. (1) takes two alternative forms. The first is a comprehensive list of health indicators, and the second is a single index. The list of health indicators includes self-reported health, self-reported doctor diagnoses of chronic conditions (diabetes, lung disease, hypertension, heart disease, strokes, psychiatric disorder, cancer, and arthritis), depression (CESD-10) score, episodic memory score, mental intactness score, number of difficulties in activities of daily living (ADLs),⁴ number of difficulties in instrumental activities of daily living (IADLs),⁵ having one physical limitation, having more than one physical limitations,⁶ having body pain, having poor eyesight, and having poor hearing. Smoking and BMI categories (underweight, overweight, and obese) are also included.⁷

To fully capture the health status in the estimation, the above method uses as much health information as the survey can provide. However, many health conditions coexist in one individual, causing multicollinearity problems and making the interpretation of the coefficients difficult. In our second specification, instead of employing a full list of health indicators, we use one variable—the first principal component index of the indicator set—to sufficiently represent all the twenty health variables. This index, proposed by Poterba, Venti, and Wise (known as the PVW index), is strongly correlated with mortality and future health events such as strokes in other countries (Poterba, Venti, & Wise, 2013). To construct the index, following Poterba et al. (2013), for a combined sample of men and women from each wave, we first extract principle components of the health indicators, then rank each individual by their first principle component score.⁸ The index is thus an ordinal scale from one to one hundred indicated by the ranking percentile of each respondent: the higher the index, the healthier. When we trim the rural sample to avoid overestimating the urban work capacity, we use this index by excluding from benchmark groups rural residents who were working and had the lowest 5 or 10% scores of our constructed health index.

Because the PVW index has the advantages of simplicity and also allows for trimming samples, we use models based on this indicator more intensively than the full list of health variables.

Eq. (1) is estimated as linear probability models, reporting robust standard errors. Following the literature, all of the above analyses are done separately by sex, so that rural women serve as the benchmark for urban women and rural men for urban men.

4. Main results of excess health capacity to work

4.1. Descriptive statistics

Tables 1 and 2 show summary statistics among rural and urban residents, separately. In each table, the left panel shows statistics of men, and the right panel shows statistics of women, all by five-year age groups.

Table 1.

Descriptive statistics, rural residents, mean values.

	Men						Wome
Variables list	45–49	50–54	55–59	60–64	65–69	45–69	45–49	50–54	55–59	60–64	65–69	45–69
Working	0.942	0.928	0.89	0.832	0.743	0.875	0.868	0.793	0.758	0.698	0.620	0.759
PVW Index	66.177	61.153	56.410	51.998	47.352	57.146	55.096	48.128	44.657	40.504	36.072	45.781
SRH Very Good	0.155	0.116	0.11	0.096	0.086	0.114	0.098	0.089	0.075	0.067	0.058	0.079
SRH Good	0.184	0.169	0.162	0.14	0.122	0.157	0.138	0.118	0.123	0.115	0.099	0.121
SRH Fair	0.513	0.539	0.513	0.516	0.515	0.519	0.533	0.521	0.491	0.494	0.493	0.508
SRH Poor	0.119	0.14	0.176	0.199	0.219	0.168	0.191	0.216	0.254	0.248	0.280	0.234
SRH Health Very Poor	0.029	0.036	0.039	0.049	0.059	0.041	0.041	0.057	0.057	0.075	0.069	0.059
CESD score (0–30)	6.471	6.809	7.394	7.707	8.254	7.276	8.399	9.093	9.575	10.089	10.564	9.432
Episodic memory score	4.318	4.105	3.823	3.676	3.436	3.895	4.237	3.978	3.683	3.384	3.136	3.736
Mental intactness score	9.119	8.86	8.606	8.551	8.345	8.715	8.242	7.861	7.264	7.002	6.789	7.529
1 Physical limitation	0.053	0.08	0.104	0.144	0.178	0.107	0.101	0.137	0.189	0.236	0.317	0.185
>1 Physical limitation	0.014	0.022	0.031	0.048	0.063	0.034	0.026	0.042	0.057	0.091	0.130	0.064
Any ADL limitations	0.058	0.086	0.126	0.172	0.223	0.127	0.103	0.150	0.190	0.242	0.293	0.187
Any IADL limitations	0.088	0.112	0.161	0.216	0.272	0.163	0.156	0.208	0.286	0.367	0.414	0.274
Body pain	0.229	0.268	0.279	0.291	0.336	0.277	0.351	0.396	0.413	0.431	0.466	0.406
Diabetes	0.03	0.045	0.053	0.066	0.065	0.051	0.041	0.062	0.075	0.091	0.086	0.069
Lung disease	0.092	0.113	0.145	0.175	0.206	0.142	0.067	0.093	0.105	0.136	0.159	0.108
Hypertension	0.159	0.21	0.241	0.284	0.328	0.239	0.155	0.230	0.264	0.355	0.394	0.269
Heart	0.045	0.064	0.074	0.103	0.133	0.081	0.092	0.116	0.135	0.169	0.178	0.134
Stroke	0.014	0.014	0.023	0.032	0.04	0.023	0.010	0.013	0.017	0.023	0.033	0.018
Psychiatric disorder	0.016	0.017	0.015	0.021	0.02	0.018	0.020	0.027	0.024	0.033	0.033	0.027
Cancer	0.007	0.006	0.008	0.008	0.007	0.007	0.017	0.019	0.014	0.010	0.013	0.015
Arthritis	0.236	0.287	0.32	0.355	0.384	0.312	0.302	0.381	0.421	0.458	0.473	0.400
Back pain	0.056	0.064	0.07	0.074	0.092	0.07	0.098	0.111	0.119	0.142	0.158	0.123
Eyesight poor	0.192	0.292	0.323	0.314	0.328	0.289	0.303	0.380	0.400	0.418	0.445	0.383
Hearing poor	0.062	0.08	0.103	0.153	0.2	0.115	0.090	0.107	0.125	0.144	0.206	0.128
Normal weight	0.622	0.646	0.69	0.684	0.699	0.668	0.536	0.534	0.559	0.575	0.586	0.556
Underweight	0.026	0.032	0.048	0.066	0.093	0.052	0.027	0.034	0.061	0.060	0.086	0.051
Overweight	0.31	0.276	0.235	0.219	0.189	0.247	0.358	0.351	0.312	0.298	0.276	0.322
Obese	0.042	0.045	0.027	0.031	0.018	0.033	0.080	0.081	0.067	0.067	0.051	0.070
Current smoker	0.596	0.616	0.633	0.574	0.558	0.598	0.038	0.041	0.052	0.073	0.073	0.053
Former smoker	0.137	0.16	0.17	0.207	0.223	0.176	0.009	0.013	0.017	0.019	0.028	0.016
Married	0.95	0.954	0.934	0.906	0.873	0.927	0.968	0.943	0.921	0.867	0.758	0.903
High School and above	0.117	0.204	0.142	0.051	0.035	0.114	0.045	0.087	0.043	0.012	0.004	0.041
Observations	3621	3397	3809	3608	2412	16,847	4328	3766	4053	3874	2557	18,578

Data: China Health and Retirement Longitudinal Study (2011, 2013, 2015).

Table 2.

Summary statistics, urban residents, mean values.

	Men						Wome
Variables list	45–49	50–54	55–59	60–64	65–69	45–69	45–49	50–54	55–59	60–64	65–69	45–69
Working	0.906	0.836	0.695	0.387	0.313	0.624	0.720	0.497	0.305	0.225	0.150	0.402
PVW Index	68.890	64.744	60.175	55.902	52.316	60.169	64.676	54.960	49.903	45.505	40.658	52.012
SRH Very Good	0.153	0.121	0.115	0.102	0.082	0.114	0.126	0.100	0.078	0.066	0.072	0.090
SRH Good	0.252	0.249	0.206	0.178	0.179	0.211	0.235	0.175	0.161	0.139	0.119	0.170
SRH Fair	0.470	0.494	0.524	0.564	0.556	0.524	0.507	0.569	0.560	0.581	0.563	0.555
SRH Poor	0.096	0.110	0.128	0.123	0.146	0.121	0.119	0.137	0.158	0.180	0.196	0.155
SRH Health Very Poor	0.029	0.025	0.027	0.032	0.037	0.030	0.013	0.019	0.043	0.034	0.050	0.031
CESD score (0–30)	5.348	5.565	5.730	5.679	5.799	5.634	6.325	6.654	7.268	7.515	7.678	7.040
Episodic memory score	4.831	4.743	4.524	4.518	4.410	4.598	5.162	5.006	4.780	4.549	4.280	4.796
Mental intactness score	9.928	9.858	9.712	9.680	9.546	9.738	9.645	9.447	9.150	9.122	9.053	9.301
1 Physical limitation	0.031	0.047	0.066	0.079	0.104	0.065	0.050	0.095	0.119	0.160	0.225	0.121
>1 Physical limitation	0.009	0.014	0.021	0.029	0.041	0.023	0.009	0.022	0.034	0.041	0.073	0.032
Any ADL limitations	0.041	0.064	0.079	0.111	0.146	0.088	0.038	0.085	0.128	0.151	0.208	0.115
Any IADL limitations	0.055	0.061	0.093	0.118	0.157	0.097	0.068	0.093	0.149	0.183	0.223	0.136
Body pain	0.162	0.184	0.176	0.176	0.166	0.173	0.253	0.268	0.307	0.314	0.339	0.293
Diabetes	0.071	0.110	0.101	0.140	0.140	0.113	0.039	0.084	0.109	0.168	0.183	0.110
Lung disease	0.089	0.084	0.125	0.156	0.192	0.129	0.067	0.079	0.114	0.137	0.105	0.100
Hypertension	0.265	0.275	0.299	0.385	0.404	0.326	0.166	0.252	0.322	0.388	0.516	0.313
Heart	0.076	0.105	0.148	0.188	0.211	0.147	0.097	0.151	0.241	0.279	0.327	0.210
Stroke	0.017	0.024	0.029	0.045	0.052	0.033	0.008	0.027	0.024	0.035	0.044	0.026
Psychiatric disorder	0.020	0.011	0.019	0.016	0.019	0.017	0.018	0.024	0.019	0.031	0.016	0.022
Cancer	0.003	0.006	0.015	0.014	0.016	0.011	0.007	0.020	0.019	0.017	0.021	0.017
Arthritis	0.166	0.193	0.231	0.238	0.238	0.214	0.212	0.266	0.342	0.391	0.423	0.318
Back pain	0.035	0.030	0.040	0.034	0.041	0.036	0.065	0.085	0.101	0.103	0.093	0.089
Eyesight poor	0.175	0.233	0.219	0.244	0.219	0.220	0.216	0.281	0.327	0.321	0.327	0.292
Hearing poor	0.031	0.034	0.076	0.080	0.128	0.071	0.048	0.039	0.076	0.086	0.123	0.071
Normal weight	0.481	0.528	0.556	0.540	0.539	0.532	0.547	0.535	0.518	0.510	0.475	0.519
Underweight	0.031	0.021	0.036	0.030	0.031	0.030	0.017	0.030	0.019	0.015	0.041	0.023
Overweight	0.401	0.382	0.352	0.376	0.376	0.376	0.349	0.364	0.351	0.392	0.385	0.367
Obese	0.086	0.068	0.056	0.054	0.054	0.062	0.088	0.071	0.112	0.083	0.099	0.091
Current smoker	0.545	0.562	0.534	0.498	0.427	0.515	0.019	0.036	0.044	0.041	0.041	0.036
Former smoker	0.133	0.145	0.208	0.249	0.292	0.206	0.011	0.010	0.013	0.010	0.020	0.012
Married	0.958	0.949	0.953	0.948	0.934	0.948	0.936	0.916	0.875	0.833	0.795	0.878
High School and above	0.546	0.599	0.447	0.285	0.319	0.435	0.449	0.520	0.340	0.181	0.196	0.349
Observations	918	953	1097	1124	863	4955	1025	961	1056	953	619	4614

Data: China Health and Retirement Longitudinal Study (2011, 2013, 2015).

We first note that the rate of employment goes down monotonically with age. It is, on average, 94.2% among rural men aged 45–49, down to 74.3% at 65–69. Rural women have lower rates of employment than rural men, 86.8% at 45–49, and 62% at 65–69. Urban employment is significantly lower than that of rural people, especially at older ages. Urban men aged 45–49 have an employment rate of 90.6%, down to 31.3% at 65–69. Urban women have the lowest employment rate of all four groups: 72% at 45–49, declining to 15% at 65–69.

Health indicators generally deteriorate with age as expected. Our comprehensive index of health, PVW, whose higher value indicates better health, is closely associated with age. Rural men aged 45–49 score 66, declining to 47 at 65–69. Women have lower scores at every age group, and urban men and women both have higher scores than their rural counterparts. Urban men aged 60–64 are as healthy as rural men aged 55–59. The gap in health conditions is even larger between urban and rural women.

The rest of these tables show how specific measures of health vary across these age groups. For example, 15.5% of rural men and 9.8% of rural women aged 45–49 report that their health status being “very good,” while this is only true for 8.6% of rural men and 5.8% of rural women aged 65–69. For mental health, measured by depression symptoms according to CESD-10 items, similar trends also exist across these four samples. In general, men have better depression scores than women, and urban residents have better depression scores than rural residents. These are as expected and documented in the literature (Lei, Sun, Strauss, Zhang, & Zhao, 2014). Smoking is predominantly a male phenomenon. For the whole sample, 51.5% (59.8%) of urban (rural) men currently smoke, while only 3.6% (5.3%) of urban (rural) women do. The rates of male smoking peak at 50–54 among urban men and 55–59 among rural men due to higher rates of smoking cessation at older ages. For BMI, both underweight and obese are relatively rare. Only 5.2% (3%) of rural (urban) men and 5.1% (2.3%) of rural (urban) women are underweight, and 3.3% (6.2%) of rural (urban) men and 7% (9.1%) of rural (urban) women are obese. As shown, rural people have higher rates of underweight and lower rates of obesity than urban people for both sexes. Significant number of people are overweight – 24.7% (37.6%) of rural (urban) men and 32.2% (36.7%) of rural (urban) women. Evidently, the problem of overweight and obese is more serious among women.

4.2. Benchmark regressions

4.2.1. Rural residents, aged 45–69

We first run regressions of Eq. (1) separately on rural men and women aged 45–69. Regression results using all health variables are in Table 3, and those using the single PVW index are in Table 4.

Table 3.

Benchmark regression on all health variables, rural residents. Linear probability model; dependent variable: working = 1.

	45–69 Men	45–69 Women
	A	B
SRH Good	−0.0023 (0.0085)	0.0173 (0.0134)
SRH Fair	−0.0057 (0.0075)	−0.0069 (0.0116)
SRH Poor	−0.0612*** (0.0110)	−0.0571*** (0.0137)
SRH Health Very Poor	−0.1459*** (0.0200)	−0.1164*** (0.0194)
CESD score (0–30)	0.0003 (0.0006)	0.0021*** (0.0006)
Episodic memory score	0.0030 (0.0017)	0.0038 (0.0020)
Mental intactness score	−0.0051*** (0.0013)	−0.0065*** (0.0015)
1 Physical limitation	−0.0804*** (0.0125)	−0.0624*** (0.0108)
>1 Physical limitation	−0.1671*** (0.0248)	−0.1400*** (0.0177)
Any ADL limitations	−0.0296** (0.0104)	−0.0231* (0.0099)
Any IADL limitations	−0.0723*** (0.0089)	−0.0401*** (0.0081)
Body pain	0.0085 (0.0067)	0.0148 (0.0076)
Diabetes	−00545*** (0.0159)	−0.0562*** (0.0143)
Lung disease	0.0013 (0.0081)	−0.0083 (00112)
Hypertension	−0.0173* (0.0070)	−0.0366*** (0.0084)
Heart	−0.0570*** (0.0123)	−00559*** (0.0114)
Stroke	−0.1562*** (0.0243)	−0.0716** (0.0277)
Psychiatric disorder	−0.0426 (0.0269)	−0.0490* (0.0221)
Cancer	−0.2125*** (0.0451)	−0.0866*** (0.0289)
Arthritis	0.0366*** (0.0059)	0.0184*(0.0073)
Back pain	0.0281* (0.0119)	0.0251* (0.0110)
Eyesight poor	0.0072 (0.0060)	−0.0001 (0.0069)
Hearing poor	0.0226* (0.0090)	0.0125 (0.0100)
Underweight	−0.0259 (0.0150)	0.0266 (0.0158)
Oveweight	−0.0225** (0.0072)	−0.0328*** (0.0079)
Obese	−00373* (0.0185)	−0.0979*** (0.0158)
Current smoker	0.0083 (0.0065)	−0.0012 (0.0163)
Former smoker	−0.0201* (0.0089)	0.0244 (0.0265)
Married	0.0976*** (0.0130)	0.0793*** (0.0125)
High School and above	−0.0094 (0.0076)	−0.0150 (0.0170)
50–54	−0.0041 (0.0058)	−00550*** (0.0085)
55–59	−0.0312*** (0.0065)	−0.0873*** (0.0091)
60–64	−0.0721*** (0.0076)	−0.1314*** (0.0099)
65–69	−0.1473*** (0.0103)	−0.1948*** (0.0122)
year = 2013	0.0052 (0.0064)	0.0096 (0.0077)
year = 2015	−0.0035 (0.0061)	−0.0182* (0.0077)
Constant	0.9172*** (0.0225)	0.9312*** (0.0247)
Observations	16,624	18,072
R-squared	0.169	0.132

Data: China Health and Retirement Longitudinal Study (2011, 2013, 2015).

Notes: Standard errors are clustered at the individual level. Significance levels indicated by

^*p < 0.05,

^**p < 0.01, and

^***p < 0.001. Provincial fixed effects included.

Table 4.

Benchmark regression on the PVW index, rural residents and younger male urban residents. Linear probability model; dependent variable: working = 1.

	45–69 Rural	Exclude PVW < 5%	Exclude PVW < 10%	45–54 Urban
	A	B	C	D
Men
PVW Index	0.0023*** (0.0001)	0.0026*** (0.0001)	0.0030*** (0.0001)	0.0025*** (0.0004)
Married	0.1038*** (0.0140)	0.1075*** (0.0142)	0.1068*** (0.0144)	0.1080* (0.0550)
High School and above	−0.0182* (0.0080)	−0.0186* (0.0081)	−0.0189* (0.0081)	−0.0049 (0.0179)
50–54	0.0031 (0.0062)	0.0040 (0.0062)	0.0050 (0.0063)	–
55–59	−0.0270*** (0.0069)	−0.0272*** (0.0069)	−0.0273*** (0.0070)	–
60–64	−0.0717*** (0.0079)	−0.0714*** (0.0080)	−00729*** (0.0081)	–
65–69	−0.1488*** (0.0105)	−0.1495*** (0.0107)	−0.1517*** (0.0108)	–
year = 2013	−0.0009 (0.0062)	−0.0019 (0.0063)	−0.0014 (0.0063)	0.0782*** (0.0210)
year = 2015	−0.0118 (0.0061)	−0.0129* (0.0061)	−0.0126* (0.0062)	0.0744*** (0.0206)
Constant	0.7014*** (0.0192)	0.6837*** (0.0195)	0.6542*** (0.0199)	0.5614*** (0.0725)
Observations	15,360	15,186	14,878	1598
R-squared	0.095	0.102	0.114	0.079
Women
PVW Index	0.0022*** (0.0001)	0.0027*** (0.0001)	0.0034*** (0.0001)	–
Married	0.0725*** (0.0132)	0.0726*** (00134)	0.0761*** (0.0136)	–
High School and above	−0.0282 (0.0180)	−0.0299 (0.0183)	−0.0309 (0.0184)	–
50–54	−0.0497*** (0.0088)	−0.0494*** (0.0089)	−0.0490*** (0.0091)	–
55–59	−0.0798*** (0.0092)	−0.0808*** (0.0094)	−0.0814*** (0.0096)	–
60–64	−0.1292*** (0.0100)	−0.1314*** (0.0101)	−0.1349*** (0.0103)	–
65–69	−0.1972*** (0.0124)	−0.2029*** (00126)	−0.2099*** (0.0128)	–
year = 2013	0.0044 (0.0076)	0.0028 (0.0077)	0.0030 (0.0079)	–
year = 2015	−0.0218** (0.0077)	−0.0246** (0.0078)	−0.0249** (0.0079)	–
Constant	0.7518*** (0.0193)	0.7213*** (0.0198)	0.6761*** (0.0203)	–
Observations	16,764	16,359	15,795	–
R-squared	0.090	0.101	0.120	–

Data: China Health and Retirement Longitudinal Study (2011, 2013, 2015).

Notes: Standard errors are clustered at the individual level. Significance levels indicated by

^*p < 0.05,

^**p < 0.01, and

^***p < 0.001. Provincial fixed effects included.

Table 3 shows that the association of employment with most of the variables are statistically significant with correct signs. Self-reported poor or very poor health, having worse memory, having one or more physical limitations, having at least one ADL or IADL difficulties, having diabetes or hypertension, having heart disease or stroke, cancer, being underweight for men and overweight or obese for women are all correlated with lower levels of employment. Several variables have coefficient estimates with wrong signs, such as CESD scores, mental intactness scores, arthritis for men, and back pain. The wrong signs may be due to multicollinearity among many health variables entered simultaneously.

In Table 4, where a single PVW health index replaces the long list of health variables, we find that the PVW coefficient is 0.0023 for rural men and 0.0022 for rural women, both statistically significant. When compared with similar studies from OECD countries (Wise, 2017), our coefficients are among the smallest. When estimated on men aged between 50 and 54, the corresponding coefficient is 0.0059 for the U.K., 0.0062 for the U.S., and 0.0045 for SHARE countries, including Belgium, Denmark, France, Germany, Italy, the Netherlands, Spain, and Sweden. For women, this coefficient is 0.0049 for the U.K., 0.0048 for the U.S., and 0.0038 for SHARE countries. The relatively small coefficient of our PVW health index means that the work decisions of Chinese rural residents are less responsive to health conditions, which is not surprising and consistent with the findings that rural Chinese have inadequate wealth to support retirement (Chen, Mao, Wang, Zhang, & Zhao, 2019).

Given that rural older people often continue to work despite ill health, we decide to be more conservative by excluding those who remain in the workforce but have the lowest PVW scores, below 5 or 10%iles.⁹ The results are in Columns B and C in Table 4. As expected, the PVW coefficients are larger – the more we exclude, the larger the coefficient. When we exclude 5% of the worst health but working individuals, the coefficient becomes 0.0026 for men and 0.0027 for women; when 10% we excluded, the coefficients go up to 0.0030 for men 0.0034 for women.

4.2.2. Age-specific rural groups

The above analyses have estimated one employment-health coefficient for a broad age group (45–69) in the benchmark regression while controlling for age-group fixed effects. Although this method has improved over one that uses younger cohorts to benchmark older-ones, arguably using the same age group is superior. Table 5 shows age-group specific regressions for rural residents using the PVW method, and also versions of the regressions that exclude workers with worst health index values, each row showing three sets of regressions using one age group. Looking horizontally, for each age group, as we exclude more unhealthy but working individuals, the health gradient goes up, as expected. For each specification, moving down the columns shows the effect of age. Interestingly, the health coefficients increase in age for both sexes. For example, the coefficient for rural men aged 45–49 is 0.0015, and it rises to 0.0036 for rural men aged 65–69. The rising health gradient with age implies that as people age, the decision to work is more responsive to observed health conditions, which justifies our decision to use rural people of the same age group to benchmark urban people.

Table 5.

Benchmark regression coefficients of the PVW index in age-group specific regressions, rural residents. Linear probability model; dependent variable: working = 1.

	Sample size	PVW Index	Robust SD	PVW Index (exclude 5%)	Robust SD	PVW Index (exclude 10%)	Robust SD
	A	B	C	D	E	F	G
Men
45–49	3245	0.0015***	(0.0002)	0.0016***	(0.0002)	0.0018***	(0.0003)
50–54	3060	0.0019***	(0.0002)	0.0020***	(0.0002)	0.0022***	(0.0003)
55–59	3482	0.0020***	(0.0002)	0.0023***	(0.0003)	0.0026***	(0.0003)
60–64	3313	0.0028***	(0.0003)	0.0030***	(0.0003)	0.0036***	(0.0003)
65–69	2260	0.0036***	(0.0004)	0.0040***	(0.0004)	0.0047***	(0.0004)
45–69	15,360	0.0027***	(0.0001)	0.0029***	(0.0001)	0.0034***	(0.0001)
Women
45–49	3914	0.0016***	(0.0002)	0.0018***	(0.0002)	0.0021***	(0.0002)
50–54	3422	0.0019***	(0.0003)	0.0022***	(0.0003)	0.0028***	(0.0003)
55–59	3657	0.0022***	(0.0003)	0.0027***	(0.0003)	0.0034***	(0.0003)
60–64	3489	0.0024***	(0.0003)	0.0030***	(0.0003)	0.0039***	(0.0003)
65–69	2282	0.0030***	(0.0004)	0.0040***	(0.0004)	0.0052***	(0.0004)
45–69	16,764	0.0027***	(0.0001)	0.0032***	(0.0001)	0.0039***	(0.0001)

Data: China Health and Retirement Longitudinal Study (2011, 2013, 2015).

Notes: Each row represents a regression. Model specification is the same as in Table 4. Standard errors are clustered at the individual level. Significance levels indicated by

^*p < 0.05,

^**p < 0.01, and

^***p < 0.001.

4.2.3. Younger urban workers

Finally, to show how our results differ when following the literature in benchmarking, we also use younger (45–54) urban residents to benchmark older urban people. We choose this age group because, although age 60 is the statutory retirement age for men, many urban men were able to take early retirement. Due to the early retirement age of Chinese urban women, no proper benchmark group exists for urban women; thus, we only study urban men with this method. Column D in the upper panel of Table 4 shows the regression results for urban men. The PVW coefficient is 0.0025, which is larger than if using the same aged rural residents as the benchmark (0.0015 for ages 45–49 and 0.0019 for ages 50–54, shown in Table 5).

4.3. Calculating excess capacities

Using PVW coefficients derived from Tables 4 and 5 and substitute with mean values from the urban sample for the right-hand side variables, we predict the employment rates of urban people. The actual and predicted employment rates are in Table 6, men on the top panel, and women on the lower panel. Column A reports the sample size for each age group. Actual employment rates are in Column B. For each age group (row), we report eight sets of predicted employment rates. Column C comes from the model containing all health variables as covariates, and the rest of the columns all use the PVW index. Columns D–F come from benchmarking all rural residents aged 45–69, and Columns G-I use age-group specific groups as benchmarks. Each of the above two age grouping choices has three columns: the entire sample, trimmed by the bottom 5 or 10% of the PVW health index while working. The last column J uses urban younger cohorts as the benchmark for men.

Table 6.

Actual and predicted employment rates using various benchmark groups.

	Observations	Actual employment rate	Predicted employment rate
				Benchmark: rural 45–69			Benchmark: age specific rural groups			Benchmark: urban men 45–54
			All health variables	PVW index	Exclude PVW < 5%	Exclude PVW < 10%o	PVW index	Exclude PVW < 5%	Exclude PVW < 10%o
	A	B	C	D	E	F	G	H	I	J
Men (Urban)
45–49	883	0.906	0.918	0.934	0.933	0.932	0.951	0.950	0.949	–
50–54	918	0.836	0.906	0.923	0.922	0.921	0.936	0.935	0.934	–
55–59	1070	0.695	0.869	0.886	0.883	0.878	0.873	0.870	0.865	0.852
60–64	1087	0.387	0.810	0.826	0.823	0.815	0.821	0.817	0.811	0.841
65–69	840	0.313	0.724	0.742	0.737	0.727	0.714	0.711	0.696	0.839
Women (Urban)
45–49	976	0.720	0.828	0.853	0.855	0.855	0.893	0.892	0.890
50–54	914	0.497	0.744	0.766	0.764	0.757	0.787	0.783	0.778
55–59	988	0.305	0.697	0.720	0.715	0.704	0.708	0.703	0.690
60–64	884	0.225	0.639	0.668	0.659	0.642	0.620	0.616	0.601
65–69	580	0.150	0.550	0.586	0.572	0.548	0.557	0.546	0.527

Data: China Health and Retirement Longitudinal Study (2011, 2013, 2015).

Notes: Column C makes predictions based on the estimation results in Table 3. Columns D–F are based on the estimation results in Column A-C of Table 4. Columns G-J are based on the estimation results in Columns B, D, F of Table 5. Column J is based on the estimation results in Column D of Table 4.

It turns out that when using a full list of health variables, the predicted rates of employment are slightly lower (compare Columns C vs. D). For men, the difference ranges from 1.5 to 1.8 ppts (percentage points); for women, it ranges from 2.2 to 3.6 ppts. A lower predicted employment rate means lower excess capacity; thus, a benchmark regression that is based on the full list of health variables leads to slightly less estimated health capacity to work. The difference is small enough to let us focus on models based on the PVW index, which is simple and enables us to trim the data.

Among all models that use the PVW index, the ones that benchmark on broad age band (Columns D–F₎ generally yield lower predicted employment for younger age groups than those using age-specific benchmark groups (Columns G-I). In comparison, the reverse is true for older age groups. In other words, age-specific benchmarking raises predicted employment for younger age groups and it for older ones. As expected, trimming the benchmark sample decreases predicted employments, and the reduction is more noticeable among older groups because the health gradient of employment is the steepest at older ages (as shown in Table 5).

When comparing actual and health-predicted rates of employment, apparently predicted employment rates decrease as people age, just like actual employment rates do, but the decline for predicted employment is much more gradual than the actual employment rate. For urban men, the health-predicted employment rate decreases from 91.8% to 72.4% for the age group from 45 to 49 to 65–69 (Column C), but the actual employment declines from 90.6% to 31.3% for the same age groups. For urban women, while the predicted employment rate decreases from 82.8% at ages 45–49 to 55.0% at ages 65–69, the actual employment declines from 72.0% to 15.0%. Similar is true when using other benchmark groups.

When we use urban men aged 45–54 as the benchmark group (Column J), predicted employment rates are quite high for older age groups. For instance, the predicted rate for ages 65–69 is 83.9%, in contrast to 74.2% when using the entire 45–69 rural men as benchmarking, or 71.4% when using the same age group. One consequence of adopting this benchmarking is that the predicted rates are very stable across age groups, 85.2% for ages 55–59, 84.1% for ages 60–64 and 83.9% for ages 65–69. Given health declines with age, the stability of predicted employment is concerning.

To compute excess health capacity to work, we subtract the actual employment rates from the health-predicted ones and show the results in Table 7. As these are the key results of this paper, we discuss them in turn by three ways of benchmarking. To simplify discussions, in the rest of the paper, we focus on results using the PVW index that allows for trimming of the least healthy rural workers in benchmarking.

Table 7.

Excess work capacity of urban residents using various benchmark groups.

	Observations	Excess work capacity
		Benchmark rural 45–69				Benchmark age specific rural groups			Benchmark urban men 45–54
		All health variables	PVW index	Exclude PVW < 5%	Exclude PVW < 10%	PVW index	Exclude PVW < 5%	Exclude PVW < 10%	PVW index
	A	B	C	D	E	F	G	H	I
Men (Urban)
45–49	883	0.012	0.028	0.027	0.026	0.045	0.044	0.043
50–54	918	0.070	0.087	0.086	0.085	0.100	0.099	0.098	-
55–59	1070	0.174	0.191	0.188	0.183	0.178	0.175	0.170	0.157
60–64	1087	0.423	0.439	0.436	0.428	0.434	0.430	0.424	0.454
65–69	840	0.411	0.429	0.424	0.414	0.401	0.398	0.383	0.526
Women (Urban)
45–49	976	0.108	0.133	0.135	0.135	0.173	0.172	0.170
50–54	914	0.247	0.269	0.267	0.260	0.290	0.286	0.281
55–59	988	0.392	0.415	0.410	0.399	0.403	0.398	0.385
60–64	884	0.414	0.443	0.434	0.417	0.395	0.391	0.376
65–69	580	0.400	0.436	0.422	0.398	0.407	0.396	0.377

Data: China Health and Retirement Longitudinal Study (2011, 2013, 2015).

Notes: Column B makes predictions based on the estimation results in Table 3. Columns C-E are based on the estimation results in Column A-C of Table 4. Columns E-G are based on the estimation results in Columns B, D, F of Table 5. Column I is based on the estimation results in Column D of Table 4.

4.3.1. Benchmarking on rural residents, aged 45–69

For urban men aged 45–49,the actual employment rate is 90.6%, while the predicted employment rate is 93.4%, yielding 2.8% additional work capacity (Column C). The estimated excess capacity for urban men aged 50–54, 55–59, 60–64 and 65–69 is 8.7%, 19.1%, 43.9% and 42.9%, respectively. Note there is a large increase around age 60, the statutory retirement age for urban men.

When we exclude rural workers with the worst health, the predicted work capacity for urban residents becomes lower, as expected, but the change is not large (Columns D-E). The additional work capacity for ages 45–49 and 50–54 is 2.7% (2.6%) and 8.6% (8.5%) by excluding the bottom 5 (10) percentile of the PVW health index of workers, in contrast to 2.8% and 8.7% before trimming. For urban men aged 55–59, 60–64 and 65–69, the excess capacity after excluding the bottom 5 (10) percentile of the PVW index is 18.8% (18.3%), 43.6% (42.8%) and 42.4% (41.4%), respectively. To gauge the magnitudes of these numbers, we take the ratio of the excess capacity to the actual employment rates. Using the lowest numbers coming from trimming workers with 10% lowest PVW index, the excess capacity represents an increase in the rates of employment by 10.2% (8.5/83.6) for the age group 50–54, 26.3% for those aged 55–59, 110.6% for the aged 60–64 and 132.3% for the aged 65–69.

For urban women aged 45–49, the actual employment rate is 72.0% and the predicted employment rate is 85.3% based on the PVW health index, implying an additional work capacity of 13.3% for this group. If excluding workers with the bottom 5 (10) percentile of the PVW health index, the additional work capacity is 13.5% (13.5%). Repeating the same exercise for urban women aged 50–54, 55–59, 60–64, and 65–69, the excess capacity is 26.7% (26%), 41.0% (39.9%), 43.4% (41.7%), and 42.2% (39.8%) after excluding the bottom 5 (10) percentile of the PVW index, respectively. Using the lowest numbers coming from trimming workers with 10% lowest PVW index, the excess capacity represents an increase in the rates of employment by 52.3% (26.0/49.7) for those aged 50–54, 130.8% for those aged 55–59, 185.3% for the aged 60–64 and 265.3% for the aged 65–69.

4.3.2. Benchmarking on rural residents of the same 5-year age groups

Using the entire rural people aged 45–69 as the benchmark underestimates the ability to work for younger people and over-estimates the ability to work among older workers. Therefore, when using age-specific rural groups as benchmarks (Columns F–H of Table 7), the predicted excess capacity to work is higher for younger age groups (men under 60 and women under 55), and lower for older age groups than benchmarking against the entire rural residents aged 45–69 (Columns C-E). Our discussions below pertain to results from benchmarking against specific age groups, which is our preferred benchmarking method.

Among urban men aged 45–49, the estimated excess capacity is 4.5% higher than the actual employment rate. Trimming the rural sample by workers with 5 (10) percent of the lowest health PVW scores, the excess capacity is reduced to 4.4% (4.3%). For urban men aged 50–54, 55–59, 60–64 and 65–69, the excess capacity after excluding the bottom 5 (10) percentile of the PVW index is 9.9% (9.8%), 17.5% (17.0%), 43.0% (42.4%) and 39.8% (38.3%), respectively. Using the lowest numbers coming from trimming workers with 10% lowest PVW index, the excess capacity represents an increase in the rates of employment by 11.7% (9.80/83.6) for the age group 50–54, 24.5% for those aged 55–59, 109.6% for the aged 60–64 and 122.4% for the aged 65–69.

For urban women aged 50–54, 55–59, 60–64 and 65–69, the excess capacity after excluding the bottom 5 (10) percentile of the PVW index is 28.6% (28.1%), 39.8% (38.5%), 39.1% (37.6%) and 39.6% (37.7%), respectively. Using the lowest numbers coming from trimming workers with 10% lowest PVW index, the excess capacity represents an increase in the rates of employment by 56.5% (28.1/49.7) for the age group 50–54, 126.2% for those aged 55–59, 167.1% for the aged 60–64 and 251.3% for the 65–69 age group.

Overall, although using rural people of five-year specific age groups as benchmarks results in different estimations, the results are not vastly different.

4.3.3. Benchmarking on urban younger people

When benchmarking against urban younger men (45–54) older men (Column I in Tables 6 and 7), the estimated excess work capacity is 15.7% for men aged 55–59, 1.3 ppts lower than our preferred method of using rural residents of the same age group with trimming at the bottom 10 percent. However, for the age group 60–64, the estimated excess work capacity is, 3.0 ppts higher than our preferred benchmarking method. For those 65–69, the estimated excess capacity to work is 52.6%, 14.3 ppts higher than our preferred benchmarking method with trimming. To summarize, using younger age cohorts as the benchmark under-estimates the excess capacity among younger people and over-estimates the excess work capacity for older people, and the margins of error are quite large.

4.4. Excess work capacities by education

To introduce more heterogeneity in the benchmark regression, we stratify our sample by two groups, a high-education group is high school and above and the low-education group is middle school and below. Using five-year age-specific rural residents as benchmarks is ideal, but the sample size for the high-education group among older cohorts becomes too small; therefore, we use rural people aged 45–69. We take comfort in the fact that the differences between these two methods are not large, as Section 3.3.2 shows. The left panel of Table 8 shows the results for men, and the right panel is for women. Comparing predicted rates of employment across education, it is evident that in almost all age groups, those with high education have lower predicted rates of employment. However, the differences dwarf those of actual employment. Thus, education difference in the work capacity is essentially a measure of the difference in actual work participation between education levels.

Table 8.

Prediction results stratified by education for urban residents.

		Urban, Middle School and below				Urban, High School and above
		Obs	Actual working rate	Predicted working rate	Excess work capacity	Obs	Actual working rate	Predicted working rate	Excess work capacity
Men (Urban)
45–49	401		0.888	0.939	0.051	482	0.921	0.930	0.009
50–54	364		0.835	0.921	0.086	552	0.835	0.925	0.090
55–59	595		0.682	0.895	0.213	474	0.713	0.874	0.161
60–64	775		0.406	0.832	0.426	312	0.34	0.813	0.473
65–69	566		0.346	0.749	0.403	273	0.245	0.729	0.484
Women (Urban)
45–49	537		0.683	0.931	0.248	438	0.765	0.912	0.147
50–54	432		0.498	0.905	0.407	481	0.497	0.882	0.385
55–59	646		0.331	0.848	0.517	340	0.253	0.836	0.583
60–64	721		0.232	0.789	0.557	162	0.198	0.783	0.585
65–69	462		0.158	0.704	0.546	116	0.121	0.675	0.554

Data: China Health and Retirement Longitudinal Study (2011, 2013, 2015).

Notes: Calculations are based on benchmark regressions using all rural people 45–69 and the PVW health index, stratified by gender and education.

Among urban men aged 45–49, 50–54 and 55–59, the high-education group had excess work capacity of 0.9%, 9.0% and 16.1%, respectively, less than the low-education group (5.1%, 8.6% and 21.3%). Among men older than age 60 (60–64 and 65–69), the order is reversed, with high-education groups having more excess capacity (47.3% and 48.4% vs. 42.6% and 40.3%). The same reversal is found among urban women but occurs five years earlier than men. Among women 45–49 and 50–54, excess work capacity among the high-education group is less than the low-education group (14.7% and 38.5% vs. 24.8% and 40.7%); among women older than 55 (55–59, 60–64, 65–69), excess work capacity among the high-education group is higher than the low-education group (58.3%, 58.5% and 55.4% vs. 51.7%, 55.7% and 54.6%).

Interestingly, the reversals occur around the statutory retirement ages, age 60 for men and age 55 for white-collar women. This is not surprising because less educated people have a higher tendency to work in jobs that qualify for early retirement – those that require hard manual labor, are harmful or dangerous. After the statutory retirement age, those who have better education tend to have more retirement benefits, are wealthier, and thus may afford to exit work earlier.

Our post-retirement-age results in education are similar to those in OECD countries, which showed more excess work capacity among the better educated (Wise, 2017). Due to the lack of a proper benchmark group for before-retirement-age cohorts, the OECD studies were not able to estimate before-retirement-age excess capacity or differences by education. However, due to widespread disability claims among younger cohorts in these countries, a reversed relationship is also likely.

5. Summary and discussions

As one of the fastest aging countries in the world, China faces tremendous financial pressure in supporting the large and growing number of retirees, just like many developed countries. Currently, China has the youngest statutory retirement age for its urban workers, implying a potentially large amount of unused capacity to work. Making a precise estimation of this capacity can help a country to plan for future changes in retirement policy. Such research has been carried out for many OECD countries organized by the NBER, but not for China. The idea is to designate a certain level of health as being physically capable of working and define those who have better health but not working as having excess health capacity. The standard procedure is first estimating a relationship between employment and health status on a benchmark group, then making out-of-sample prediction for those of interest, and taking the difference between actual and predicted employment rates to get the excess capacity to work. OECD country studies have all used younger cohorts as the benchmark to estimate the excess work capacity for older workers. The choice of younger workers is based on the assumption that their decision to work is free from the enticement of retirement incentives; thus, the relationship between health and employment reflects the true physical ability to work. The assumption is violated if the relationship between observed health conditions and work changes with age due to unobserved health conditions that worsen with age. Using the China Health and Retirement Longitudinal Study (CHARLS), a survey that is harmonized with the HRS family of data used in OECD country studies, we show that, indeed, the observed employment-health relationship strengthens in older age groups.

We improve the estimation by selecting a cleaner benchmark group, rural Chinese residents who, under the dual economy, enjoy minimal retirement benefits and generally work to their full capacity. This selection allows us to apply rural samples of the same age groups to benchmark urban people. We further exclude from the benchmark samples extreme cases of rural people who continue to work despite having the worst health conditions. Our estimated excess capacity is substantially smaller by 14.3 percentage points for urban men aged 65–69 than if using younger cohorts as the benchmark group, implying that the OECD studies probably over-estimated their extra work capacity.

According to our most conservative estimation, which trims 10% of the hardest working rural people as benchmarks, the amount of extra health capacity to work is still substantial in China. Among age groups 45–49, 50–54, 55–59, 60–64 and 65–69, the additional work capacity among men (women) are 4.3%, 9.8%, 17.0%, 42.4% and 38.3% (17.0%, 28.1%, 38.5%, 37.6% and 37.7%). The excess capacity is higher among younger women than men because of the very early statutory retirement age (50 for blue-collar women vs. 60 for all men). Sizable excess capacity before the statutory retirement age (17.0% for men aged 55–59 and 17.0% for women 45–49) implies a significant amount of early-retirement granting. Some of these excess capacity is large relative to the existing employment rate. For instance, using our preferred benchmarking method with trimming, among Chinese urban men (women) aged 60–64, 42.4% (37.6%) can physically work but are not, representing 109.6% (167.1%) of the current employment rates.

We find interesting patterns in education. While among pre-retirement-age cohorts, the excess capacity is larger among the less educated, in the post-retirement-age cohorts, there is more excess capacity among the better educated, indicating a waste of more educated human resources caused by the retirement age policy.

To evaluate the reliability of these numbers, we compare with direct questions asked in the CHARLS survey in 2018, in which respondents answer whether their health limits the ability to work. The question is: “Please evaluate if the following statement fits your situation: Due to disability or health reasons, I am unable to perform normal work.” There are three choices: A. I am completely unable to perform normal work; B. I can not work for a long time; C. I do not have a problem. This question is simple and direct, but answers may suffer from justification bias, in which a person who has stopped working may be more inclined to state having health limiting the ability to work. Nonetheless, it provides an independent check to our estimated capacity to work. Table 9 compares our predicted employment rate (using age-specific benchmarking trimming 10% of those working too hard) with those reporting not having any health issue limiting the ability to work (shown in Fig. 2). The predicted employment rates are significantly lower for men under 55, indicating possible justification bias in self-evaluates. For urban men 55 and older, the predicted work capacities match that of self-reports very well: the predicted (self-reported) employment rates for ages 55–59, 60–64, 65–69 are 86.5% (84.8%), 81.1% (76.8%), and 69.6% (68.1%). The predicted rates are somewhat higher (1.5–4.3 ppts) because the self-reported work capacity has a higher standard of having no limits at all, and some of those reporting partial limitations may still be able to work part-time. For urban women, the predicted employment rates are generally lower than those reporting no health conditions by a larger margin. This under-performance may be explainable by rural women shouldering many family responsibilities, causing their employment rates to be systematically lower than their true capacity.¹⁰ Overall, by and large, our predicted work capacities are conservative and generally reliable.

Table 9.

Predicted employment, self-reported work ability and aggregate excess work capacities for urban China.

	Predicted working rate (%)	Self-reported work ability (%)	Aggregate work capacities (in million)
	A	B	C
Men (Urban)
45–49	94.9	87.8	0.80
50–54	93.4	84.3	1.52
55–59	86.5	84.8	1.96
60–64	81.1	76.8	4.87
65–69	69.6	68.1	3.09
Total			12.23
Women (Urban)
45–49	89.0	85.7	3.05
50–54	77.8	83.1	4.25
55–59	69.0	77.8	4.28
60–64	60.1	67.3	4.33
65–69	52.7	63.4	3.04
Total			18.95

Note: Column (A) is from Table 6, Column (X); Column (B) is derived from Fig. 2; Column (C) extrapolated excess capacity in Column (X) of Table 7 to national statistics.

Fig. 2.

Health Limiting Work Ability..

Source: China Health and Retirement Longitudinal Study 2018.

To convert ratios to sizes, we multiply the urban population by rates of excess health capacity to work, all by sex for each 5-year age group, then aggregate across age groups (Table 9).¹¹ The total number of urban people who are not currently working but are capable of working judged solely by their health status is 12.2 million men and 19.0 million women. The great majority of the potential male workers are in the age range of 60–64, while for women they are 55–59 and 60–64.

Our results indicate that China has substantial untapped human resources in urban areas. However, having a large reservoir of under-utilized older workers does not automatically guarantee that they can be readily tapped. There are many other legitimate considerations in retirement decision-making (Giles et al., 2015). For example, care facilities for children under age three are nearly non-existent, and affordable care service for aged care is also scarce, causing many people to choose to retire to take care of their grandchildren or aging parents. Developing a well-functioning social care industry can free up older workers from their care duties and take part in the labor force.

Appendix A

Table A1.

First principal component index of health indicators.

Health measure	Wave 2011	Wave 2013	Wave 2015
Difficulty stooping/kneeling/crouching	0.35	0.35	0.34
Difficulty with an ADL	0.34	0.35	0.34
Difficulty climbing stairs	0.34	0.34	0.33
Difficulty getting up from chair	0.34	0.34	0.33
Difficulty walking several blocks	0.33	0.32	0.31
Difficulty lifting/carrying	0.31	0.31	0.30
Difficulty reaching/extending arms up	0.27	0.25	0.25
Difficulty picking up a coin	0.22	0.19	0.19
Self-reported health fair or poor	0.19	0.20	0.20
Ever diagnosed arthritis	0.18	0.21	0.22
Back problems	0.18	0.16	0.22
Ever diagnosed heart problems	0.14	0.15	0.17
Ever diagnosed high blood pressure	0.13	0.13	0.14
Ever diagnosed stroke	0.12	0.10	0.11
Ever diagnosed lung disease	0.12	0.13	0.13
Doctor visit	0.12	0.14	0.11
Hospital stay	0.11	0.13	0.14
Ever diagnosed psychological problem	0.08	0.06	0.08
Ever diagnosed diabetes	0.08	0.09	0.10
Ever diagnosed cancer	0.03	0.03	0.03
N	16,179	15,450	17,551