Self-Selection and Recruit Quality in Sweden’s All Volunteer Force: Do Civilian Opportunities Matter?

ABSTRACT

This paper studies how local labour market conditions influence the quality composition of those who volunteer for military service in Sweden. A fixed-effects regression model is estimated on a panel data set containing IQ scores for those who applied for military basic training across Swedish municipalities during the period 2010 to 2016. The main finding is that low civilian employment rates at the local level tend to increase the mean IQ score of those who volunteer for military service, whereas the opposite is true if employment rates in the civilian labour market move in a more favourable direction. As such, the results suggest that the negative impact of a strong civilian economy on recruitment volumes is reinforced by a deterioration in recruit quality.

KEYWORDS:

• Military labour market
• military recruitment
• self-selection
• enlistment test
• Roy model

JEL CLASSIFICATION:

• J01
• J20
• J45
• H56

Introduction

Sweden abandoned compulsory conscription in favour of voluntary recruitment to the enlisted ranks in 2010. Relying on volunteers, rather than conscripts, to staff the military force means that recruitment volumes become sensitive to disturbances from the outside world. In particular, the general attractiveness of the civilian alternatives available to those thinking about whether or not to join the military becomes important in determining the number of people willing to work within it. Indeed, empirical studies, on how the willingness to serve in the military differs between local labour markets, typically find that fewer individuals wish to join the military when conditions in the civilian labour market improve (see, for example, Bäckström 2019; Asch et al. 2010; Asch, Heaton, and Savych 2009; Simon and Warner 2007; Warner, Simon, and Payne 2003; Ellwood and Wise 1987; Brown 1985). Little is known, however, about how changes in the economic circumstances facing potential recruits influence quality aspects of military recruitment. Do conditions in the civilian economy influence the way in which people select themselves into the military? Is the negative impact of a strong civilian economy on recruitment volumes exacerbated by a deterioration in recruit quality?

This paper explores these questions in the context of voluntary recruitment to the Swedish Armed Forces. The purpose is to study how local labour market conditions influence the quality composition of those who volunteer for military service in Sweden. I aggregate data from the military enlistment test across Swedish municipalities to examine how IQ scores for those who volunteered for military basic training over the period 2010–2016 differ between local labour markets. The basic question is this: Does an increase in the civilian employment rate in a municipality alter the quality composition of the flow of voluntary recruits originating in that municipality? The empirical analysis relies on fixed-effects panel regression in order to estimate the effect from labour market conditions on the quality composition of volunteers, while controlling for unobserved heterogeneity across municipalities.

This is a topic that has largely remained unexplored in the economics of military manpower literature. Most previous research that relates military recruitment to economic circumstances in the civilian sector focuses on the number of individuals who are willing to serve in the military, rather than on the type of individuals who make up the recruit pool. As such, the current paper adds to the literature by explaining disparities in recruit quality, rather than recruitment volumes, across geographical regions in Sweden. Knowledge of how the composition of the recruit flow reacts to the economic environment is important for future military recruitment strategies, as well as for a deeper understanding of how regional economic conditions influence occupational choices made by young individuals.

The results from the empirical analysis show that poor labour market opportunities at the local level tend to increase the mean IQ score of those who volunteer for military service, whereas the opposite is true if conditions in the civilian labour market move in a more favourable direction. The application rate from individuals that score high on the IQ test is more responsive towards the employment rate in the municipality of origin, compared to the application rate from individuals that score low. As such, the results from this paper suggest that the negative impact of a strong civilian economy on recruitment volumes is reinforced by a deterioration in recruit quality.

The rest of this paper is structured in the following way. Section 1 provides a simple framework for thinking about self-selection and recruit quality in relation to the civilian labour market. Section 2 discusses the data and some aspects of the enlistment test procedure in Sweden. The empirical strategy is presented in Section 3, while Section 4 presents the main result. Section 5 summarizes the findings and concludes the paper with a discussion.

Theoretical Framework

The aim of this section is to provide a simple framework for thinking about self-selection and recruit quality in relation to the civilian labour market. Self-selection has been a core topic in labour economics ever since Roy’s 1951 paper on how workers select between hunting and fishing. A key aspect of the Roy model is the notion that rational actors will not sort themselves randomly across employment opportunities; a starting point in an economic analysis of the enlistment decision must be that those who desire to enlist have a higher expected utility in the military than in the civilian labour market, whereas the opposite is true for those who choose to remain in the civilian sector.

Following the outline of the Roy model presented in Borjas (1987), Chiquiar and Hanson (2005), and Borjas (2015), consider an economy that consists of just two sectors: the military and the civilian. Earnings in either sector depend on a single skill that can be used in both sectors. Assume that the level of skill that an individual has, denoted s, is a continuous random variable that follows some frequency distribution in the population of workers.

A key assumption in this application of the Roy model is that individuals that rank high in in the earnings distribution of one sector would also rank high in the earnings distribution of the other sector. If an individual chooses the civilian sector, indexed by c, his or her wage is given by the equation:

(1) $\mathrm{E}\left({\mathrm{w}}_{\mathrm{c}}\right)=\omega \left({\mu }_{\mathrm{c}}+{\delta }_{\mathrm{c}}\mathrm{s}\right)$(1)

where $\mathrm{E}\left({\mathrm{w}}_{\mathrm{c}}\right)$ is the expected wage in the civilian sector; ${\mu }_{\mathrm{c}}$ is the base wage in the civilian sector; s is an individual’s skill level; ${\delta }_{\mathrm{c}}$ is the return to skills in the civilian sector; and $\omega$ is the probability of finding a job. If an individual chooses the military sector, indexed by m, he or she is paid in accordance to the wage equation:

(2) ${\mathrm{w}}_{\mathrm{m}}={\mu }_{\mathrm{m}}+{\delta }_{\mathrm{m}}\mathrm{s}$(2)

where ${\mathrm{w}}_{\mathrm{m}}$ is the wage for a soldier in the military sector, µ_m is the base wage in the military, and ${\delta }_{\mathrm{m}}$ is the return to skills in the military sector.

Individuals make their enlistment decision by comparing potential earnings in the two sectors. An individual enlists whenever the earnings that he or she can expect to get in the military are larger than those in the civilian sector:

(3) ${\mathrm{w}}_{\mathrm{m}}-\mathrm{E}\left({\mathrm{w}}_{\mathrm{c}}\right)>0$(3)

Using Equations (1)(1) $\mathrm{E}\left({\mathrm{w}}_{\mathrm{c}}\right)=\omega \left({\mu }_{\mathrm{c}}+{\delta }_{\mathrm{c}}\mathrm{s}\right)$(1) and (2(2) ${\mathrm{w}}_{\mathrm{m}}={\mu }_{\mathrm{m}}+{\delta }_{\mathrm{m}}\mathrm{s}$(2) ) in Equation (3)(3) ${\mathrm{w}}_{\mathrm{m}}-\mathrm{E}\left({\mathrm{w}}_{\mathrm{c}}\right)>0$(3) , we see that military enlistment is preferred by an individual whenever:

(4) $\left({\delta }_{\mathrm{m}}-\omega {\delta }_{\mathrm{c}}\right)\mathrm{s}>\left(\omega {\mu }_{\mathrm{c}}-{\mu }_{\mathrm{m}}\right)$(4)

Equation (4)(4) $\left({\delta }_{\mathrm{m}}-\omega {\delta }_{\mathrm{c}}\right)\mathrm{s}>\left(\omega {\mu }_{\mathrm{c}}-{\mu }_{\mathrm{m}}\right)$(4) gives rise to four possible scenarios of how the individuals sort themselves between the military and the civilian sector, depending on the relative magnitudes of ${\delta }_{\mathrm{m}},\omega {\delta }_{\mathrm{c}}$, $\omega {\mu }_{\mathrm{c}}$ and ${\mu }_{\mathrm{m}}$. These four possible scenarios are summarized in Table 1. When $\omega {\mu }_{\mathrm{c}}<{\mu }_{\mathrm{m}}$ and $\omega {\delta }_{\mathrm{c}}>{\delta }_{\mathrm{m}}$, so that the returns to skills are higher in the civilian sector, Equation (4)(4) $\left({\delta }_{\mathrm{m}}-\omega {\delta }_{\mathrm{c}}\right)\mathrm{s}>\left(\omega {\mu }_{\mathrm{c}}-{\mu }_{\mathrm{m}}\right)$(4) implies that $s<\left(\omega {\mu }_c-{\mu }_m\right)/\left({\delta }_m-\omega {\delta }_c\right)$. In this case, those who join the military will be negatively selected in terms of skills and the recruits will come from the lower part of the skill distribution. In contrast, when $\omega {\mu }_{\mathrm{c}}>{\mu }_{\mathrm{m}}$ and $\omega {\delta }_{\mathrm{c}}<{\delta }_{\mathrm{m}}$, so that returns to skills are higher in the military sector, we have that $\mathrm{s}>\left(\omega {\mu }_{\mathrm{c}}-{\mu }_{\mathrm{m}}\right)/({\delta }_{\mathrm{m}}-\omega {\delta }_{\mathrm{c}})$. In this case, those who join the military will be positively selected in terms of skills and the recruits will come from the upper part of the skill distribution. Also, there are two special cases: when $\omega {\mu }_{\mathrm{c}}>{\mu }_{\mathrm{m}}$ and $\omega {\delta }_{\mathrm{c}}>{\delta }_{\mathrm{m}}$, no individual wishes to enlist, and when $\omega {\mu }_{\mathrm{c}}<{\mu }_{\mathrm{m}}$ and $\omega {\delta }_{\mathrm{c}}<{\delta }_{\mathrm{m}}$, all individuals wish to enlist.¹

Table 1. Four possible cases of skill sorting

	$\omega {\mu }_{\mathrm{c}}>{\mu }_{\mathrm{m}}$	$\omega {\mu }_{\mathrm{c}}<{\mu }_{\mathrm{m}}$
$\omega {\delta }_{\mathrm{c}}>{\delta }_{\mathrm{m}}$.	Nobody wishes to enlist.	Negative selection. Recruits will come from the lower part of the skill distribution
$\omega {\delta }_{\mathrm{c}}<{\delta }_{\mathrm{m}}$	Positive selection. Recruits will come from the upper part of the skill distribution.	Everyone wishes to enlist.

Figure 1 provides a graphical illustration of the model when the returns to skills are higher in the civilian labour market than in the military sector, whereas the opposite is true in the situation illustrated in Figure 2. In Figure 1, the recruit pool is negatively selected, and individuals with skills lower than ${\mathrm{s}}_{\mathrm{N}}$ will choose to enlist. In Figure 2, the recruit pool is positively selected, and individuals with skills higher than ${\mathrm{s}}_{\mathrm{P}}$ will choose to enlist. As such, negative selection occurs when enlistment incentives are generally stronger for those in the population with relatively low levels of skills, whereas positive selection is the case when enlistment incentives are generally stronger for those in the population with relatively high levels of skills.

Figure 1. This figure illustrates negative selection into the military. The lines represent the relationship between skill level and expected wage in the civilian and military sector, respectively. Individuals make their enlistment decision by comparing potential earnings in the two sectors. An individual enlists whenever the earnings that he or she can expect to get in the military are larger than those in the civilian sector. The return to skills is higher in the civilian sector than in the military sector. All individuals with a skill level below ${\mathrm{s}}_{\mathrm{N}}$ will choose the military sector.

Figure 2. This figure illustrates positive selection into the military. The lines represent the relationship between skill level and expected wage in the civilian and military sector, respectively. Individuals make their enlistment decision by comparing potential earnings in the two sectors. An individual enlists whenever the earnings that he or she can expect to get in the military are larger than those in the civilian sector. The return to skills is higher in the military sector than in the civilian sector. All individuals with a skill level above ${\mathrm{s}}_{\mathrm{P}}$ will choose the military sector.

Suppose now that the expected wage level in the civilian sector rises, because of a sudden boom in the civilian economy that increases the probability of finding a job. The economic boom shifts and turns the wage-skills line upwards in the civilian sector, as illustrated in Figures 3 and 4. If recruits are negatively selected, only individuals below the new and lower threshold ${\mathrm{s}}_{\mathrm{N}}^{\prime }$ will enlist. If recruits are positively selected, only individuals above the new and higher threshold ${\mathrm{s}}_{\mathrm{P}}^{\prime }$ will enlist. Hence, regardless of the type of selection that is the case, a boom in the civilian economy will mean that fewer individuals will find it optimal to enlist. The individuals who leave the military sector as a result of the economic boom will not be a random subset of the recruit pool, however. To see this, note that the skill level of the marginal individual, i.e. the individual who is indifferent about working in either the civilian sector or enlisting in the military, will decrease in both cases. Depending on the nature of selection into the military that is the case in the first place, this marginal individual will be drawn from the bottom, or the top, of the skill distribution of the recruit pool. The implication is that the enlistment rate drops, due to the withdrawal of the most skilled potential recruits. Consequently, if recruits are negatively selected, the average skill level of the recruit pool will decrease as a result of the boom in the civilian sector, whereas the opposite is the case if recruits are positively selected.

Figure 3. This figure illustrates the impact of a boom in the civilian economy when the recruit pool is negatively selected. Initially, all individuals with a skill level below ${\mathrm{s}}_{\mathrm{N}}$ will choose the military sector. The wage-skill line for the civilian sector shifts upwards and becomes steeper when the probability of finding a job in the civilian sector increases. Only individuals below the new and lower threshold ${\mathrm{s}}_{\mathrm{N}}^{\prime }$ will choose the military sector. Fewer individuals choose the military sector, and the average skill level of the remaining recruit pool decreases.

Figure 4. This figure illustrates the impact of a boom in the civilian economy when the recruit pool is positively selected. Initially, all individuals with a skill level above ${\mathrm{s}}_{\mathrm{P}}$ will choose the military sector. The wage-skill line for the civilian sector shifts upwards and becomes steeper when the probability of finding a job in the civilian sector increases. Only individuals above the new and higher threshold ${\mathrm{s}}_{\mathrm{P}}^{\prime }$ will choose the military sector. Fewer individuals choose the military sector, and the average skill level of the remaining recruit pool increases.

Summing up, this simple application of the Roy model provides some important insights on how the recruit flow is expected to react to a change in the civilian labour market environment. The relative return to skills in the civilian and the military sectors completely determines the nature of the sorting that occurs between the sectors, i.e. whether the recruit flow is positively, or negatively, selected. Regardless of the type of selection that prevails, we will expect to find a negative relationship between the size of the recruit flow and expected earnings in the civilian sector: the stronger the civilian economy, the lower the probability that a randomly chosen individual will prefer the military over the civilian sector. That is,

(5) $\frac{\mathrm{\partial }Pr\left[{\mathrm{w}}_{\mathrm{m}}>{\mathrm{w}}_{\mathrm{c}}\right]}{\mathrm{\partial }\omega }<0$(5)

However, due to the way in which different types of individuals select themselves into the military, the skill composition of those who are willing to join the military will not be neutral to these changes. That is,

(6) $\frac{\mathrm{\partial }\mathrm{E}\left(\mathrm{s}|{\mathrm{w}}_{\mathrm{m}}>{\mathrm{w}}_{\mathrm{c}}\right)}{\mathrm{\partial }\omega }\ne 0$(6)

As such, Equation (6)(6) $\frac{\mathrm{\partial }\mathrm{E}\left(\mathrm{s}|{\mathrm{w}}_{\mathrm{m}}>{\mathrm{w}}_{\mathrm{c}}\right)}{\mathrm{\partial }\omega }\ne 0$(6) represents the main hypothesis that is tested empirically in this paper. By examining the sign of Equation (6)(6) $\frac{\mathrm{\partial }\mathrm{E}\left(\mathrm{s}|{\mathrm{w}}_{\mathrm{m}}>{\mathrm{w}}_{\mathrm{c}}\right)}{\mathrm{\partial }\omega }\ne 0$(6) when applied to data, one can infer something about the nature of sorting that occurs. If enlistment incentives are stronger for those in the population with below-average levels of skills, so that recruits are negatively selected, the military will not only receive fewer recruits as a result of the strong civilian economy, but also find that the better ones are now missing. On the other hand, if incentives are stronger for those in the population with above-average levels of skills, so that recruits are positively selected, the military will still receive fewer recruits, but find that the recruits that remain are among the better ones.

Institutional Setting and Data

The Enlistment Test Procedure

The empirical analysis in this paper refers to the period when recruitment in the Swedish Armed Forces was voluntary. The compulsory peacetime draft was abandoned in 2010, and the relatively small proportion of each cohort needed for the Swedish Armed Forces was recruited on a voluntary basis until 2017, when the peacetime draft was re-instated. Consequently, all individuals who underwent the enlistment procedure during the period studied in this paper had chosen to participate of their own volition.

During the studied period, the enlistment procedure started with an application for basic military training. For those who fulfilled the basic requirements of being at least 18-years-old and holding Swedish citizenship, the first step of the application process was an online self-reporting questionnaire that asked background information on the individual, their physical and mental status, and areas of interest regarding military service. Individuals who met the requirements set by the military were invited to attend the enlistment test. The enlistment test procedure takes one or two days and involves a range of tests: medical status, cognitive ability, physical fitness, and a psychological evaluation.

This paper focuses on the test scores from the psychometric test, designed to measure individual differences in general cognitive ability, or general intelligence. Intelligence is typically understood as ‘a very general capability that, among other things, involves the ability to reason, plan, solve problems, think abstractly, comprehend complex ideas, learn quickly and learn from experience’ (Gottfredson 1997: 13). Psychometric tests of individual differences in general intelligence rest on the observation that even though people tend to have strengths and weaknesses in specific areas of cognitive ability, such as reasoning and spatial ability, individuals that perform well in one of the areas also tend to perform well in the others. This correlation is recognized as reflecting an underlying general cognitive ability (Deary, Penke, and Johnson 2010).

The test used in the Swedish military enlistment recognizes a hierarchical pattern in the components of general cognitive ability and consists of a battery of tests that require different kinds of cognitive performance, such as spatial ability, verbal ability and technical comprehension (Carlstedt and Annell 2010). All in all, the full test takes around 80 minutes to complete and is carried out at the Swedish Recruitment Agency’s facilities, under controlled conditions (Brandebo Fors et al. 2017).² Scores from ten sub-tests are used by the Recruitment Agency to construct a discrete full-scale score that reflects general cognitive ability, ranging from 1 (low) to 9 (high). This nine-point scale, often referred to as a stanine (standard nine) scale, approximates a normal distribution with a mean of five and a standard deviation of two (Lindqvist and Vestman 2011). The scaling is calibrated on a random sample from the population of (mostly) male conscripts who underwent enlistment testing in the year 2001 (Carlstedt and Annell 2010). Throughout this paper, this measure of general cognitive ability in stanine units is referred to as IQ, or IQ score.

From the military’s perspective, the testing procedure is intended to determine the individual’s ability to complete military training and to match individuals with a suitable position. For example, an IQ score of at least four (4) is required in order for the recruit to be admitted into basic military training; a test score of at least five (5) is required in order to be admitted into training as a squad leader; and a test score of at least six (6) is required in order to be admitted into the three-year commissioned officers training programme (Brandebo Fors et al. 2017; Jonsson 2014).

The outcome of the IQ test used in the Swedish military enlistment has been found to be predictive of individual performance during military service (Carlstedt 1999). Moreover, in line with a large body of literature that typically finds measures of general cognitive ability to be correlated with labour market outcomes and job performance (see Schmidt and Hunter 2004; Bowles, Gintis, and Osborne 2001), previous studies using data from the Swedish military enlistment tests have found the performance on the cognitive ability test to be predictive of wages, unemployment, and annual labour market earnings. For example, Lindqvist and Vestman (2011) found that a one (1) standard deviation increase in the test score was associated with an 8.9 percent increase in wages for a sample consisting of 14,703 men distributed evenly over the 1965–1974 birth cohorts. In particular, cognitive ability was found to be predictive of wages for skilled workers and earnings above the median.

Data

The empirical analysis in this paper is based on panel data for 290 Swedish municipalities over the years 2010 to 2016. The data set is created by averaging the IQ scores for 42,886 individuals who applied for military basic training during this period across their home municipalities.³ For each municipality, I calculated the mean of IQ scores for volunteers originating in that municipality in a specific year. This variable is labelled Mean IQ.⁴

In addition to mean IQ scores, I also collected information on the annual number of applicants who scored below four on the IQ test per 100,000 inhabitants in each municipality; the annual number of applicants who scored four or five on the IQ test per 100,000 inhabitants in each municipality; and the annual number of applicants who scored at least six on the IQ test per 100,000 inhabitants in each municipality. These variables are labelled: Application rate (Low ability); Application rate (Medium ability); and Application rate (High ability).

The panel data set containing these variables was merged with information on a range of regional attributes. The regional attribute of main interest is the employment rate in the municipality (Employment). In addition to this explanatory variable, the model includes five covariates on the municipality level. The purpose is to control for socio-demographic factors that might capture changes in the underlying population of a municipality, but also to control for factors that might influence the return to cognitive skills. For example, skilled workers might receive a relatively high wage premium from working in urban environments (Glaeser and Maré 2001). The regional covariates included in the models are: the median annual labour income in the municipality (Median income); the share of the municipality’s population with post-secondary education (Highly educated); the share of the municipality’s population aged 20–24 years that is enrolled in education (Study); the share of the municipality’s population that were born outside Sweden (Foreign); and the population density of the municipality (Population density). Data for all of these variables were collected from Statistics Sweden. Table 2 provides descriptive statistics of the variables included in the empirical analysis. Detailed definitions of the dependent and independent variables are presented in Table A1, in the Appendix.

Table 2. Summary statistics for dependent and independent variables

		Standard deviation
	Mean	Within municip.	Across municip.	Min	Max	N
Dependent variables
Mean IQ	5.58	0.68	0.40	2	9	1914
Application rate (Low ability)	4.38	5.59	2.58	0	53.16	2030
Application rate (Medium ability)	21.71	15.40	10.62	0	174.53	2030
Application rate (High ability)	26.49	17.90	15.42	0	223.45	2030
Independent variables
Employment (%)	56.86	0.86	4.24	43.26	70.94	2030
Median income (1000s SEK)	241.05	12.30	24.43	179.70	364.20	2030
Highly-educated (%)	17.84	0.84	7.41	8.27	56.88	2030
Study (%)	28.97	2.01	7.83	10.50	74.80	2030
Foreign (%)	12.32	1.37	5.67	3.83	41.04	2030
Population density (100 persons/km2)	1.44	0.26	5.09	0.002	54.96	2030

Notes: The sample used is a panel of 290 Swedish municipalities over the period 2010 to 2016. Application rates refers to the number of enlistment tests per 100,000 inhabitants in the municipality. Data from the Swedish Recruitment Agency and Statistics Sweden. Detailed definitions of the variables are presented in Table A1, in the Appendix. Unweighted values. Note that there are 116 missing values for the variable Mean IQ. These missing values relate to municipalities where no one performed the enlistment test in a year.

Empirical Specification

The purpose of this paper is to study how local labour market conditions, as measured by the rate of employment in the civilian labour market, influence the quality composition of those who volunteer for military service in Sweden. The empirical investigation uses panel data on 290 Swedish municipalities for the years 2010–2016. The outcome of interest is the mean IQ score of those, in a specific municipality, who volunteered for basic military training in a specific year. The starting point for the econometric analysis is the following regression model:

(7) $\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{I}{\mathrm{Q}}_{\mathrm{j}\mathrm{t}}={\alpha }_{\mathrm{j}}+{\tau }_{\mathrm{t}}+\beta {\mathrm{x}}_{\mathrm{j}\mathrm{t}}+{\mathbf{Z}}_{\mathbf{j}\mathbf{t}}\gamma +{\epsilon }_{\mathrm{j}\mathrm{t}}$(7)

where $\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{I}{\mathrm{Q}}_{\mathrm{j}\mathrm{t}}$ is the mean IQ score of those who volunteered for military basic training from municipality j in year t; ${\alpha }_{\mathrm{j}}$ and ${\tau }_{\mathrm{t}}$ are municipality and time-fixed effects; ${\mathrm{x}}_{\mathrm{j}\mathrm{t}}$ is the employment rate in municipality j in year t; ${\mathbf{Z}}_{\mathbf{j}\mathbf{t}}$ is a vector of covariates on the municipality level; and ${\epsilon }_{\mathrm{j}\mathrm{t}}$ is an error term. To account for the fact that the variation in mean IQ score is larger among less populated municipalities than among more populous ones, the regression is weighted by the size of the population in each municipality.⁵

To study how individuals select themselves into the military, one would like information on the cognitive abilities of both those who volunteer for military service and those who choose to remain in the civilian sector. Unfortunately, information on abilities is only available for those who actually volunteer for military service and perform the enlistment test. This is problematic, because it means that a naïve comparison of recruit quality across geographical regions with different labour market conditions will not only reflect a potential effect from self-selection of those residing in a municipality (the effect of main interest to this paper), but also unobserved differences between municipalities, which might confound the relationship of interest. For example, if skilled people tend to be unemployed to a lesser extent; and some municipalities, for some reason, have more skilled people living in them; a cross-sectional analysis might find a pattern where employment rates are positively associated with the skill level of those who volunteer for military service, when self-selection of those willing to serve, in fact works in the opposite direction within a municipality.⁶ The end result would be biased estimates of $\beta$ in Equation (7)(7) $\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{I}{\mathrm{Q}}_{\mathrm{j}\mathrm{t}}={\alpha }_{\mathrm{j}}+{\tau }_{\mathrm{t}}+\beta {\mathrm{x}}_{\mathrm{j}\mathrm{t}}+{\mathbf{Z}}_{\mathbf{j}\mathbf{t}}\gamma +{\epsilon }_{\mathrm{j}\mathrm{t}}$(7) .

The empirical strategy used in this paper in order to reduce the risk of bias due to such unobserved regional heterogeneity is to identify the impact of local labour market conditions on recruit quality via within-municipality variation, rather than cross-sectional variation. The major advantage of within-municipality estimation is that it avoids bias due to unobserved differences between municipalities that do not change over time. This is accomplished by the use of municipality fixed-effect terms (${\alpha }_{\mathrm{j}}$) that capture the effect from unobserved factors that vary across municipalities but are constant over time. However, if these unobserved differences do change over time in a way that is correlated with labour market conditions, there would still be bias arising from unobserved heterogeneity. By including time dummies (${\tau }_{\mathrm{t}}$) into the fixed-effects regression model, trends on the national level are differenced out from the within-municipality comparison. This means that the fixed-effects regression model controls for all factors that change over time and that affect all municipalities in the same manner, including changes in military wages and other forms of compensation, as well as recruiting policies and recruiting quotas on the national level. Moreover, the covariates (${\mathrm{Z}}_{\mathrm{j}\mathrm{t}}$) included into the regression model control for changes in observable characteristics of a municipality, including educational participation and attainment.

Lastly, the municipality fixed effects included in Equation (7)(7) $\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{I}{\mathrm{Q}}_{\mathrm{j}\mathrm{t}}={\alpha }_{\mathrm{j}}+{\tau }_{\mathrm{t}}+\beta {\mathrm{x}}_{\mathrm{j}\mathrm{t}}+{\mathbf{Z}}_{\mathbf{j}\mathbf{t}}\gamma +{\epsilon }_{\mathrm{j}\mathrm{t}}$(7) also help to solve a potential sample selection problem due to the fact that data on the variable Mean IQ is missing for certain years in some municipalities.⁷ The unbalanced panel can cause biased estimates if the reason a municipality has missing data is correlated with the error term. The fixed-effects analysis helps to solve this issue by allowing sample attrition to be correlated with the municipality fixed-effects term ${\alpha }_{\mathrm{j}}$ (Wooldridge 2012: 473–474).

Results

The results from regressing mean IQ scores of volunteers on civilian employment rates are reported in Table 3, which contains information on four model specifications. The first and second specifications are the regressions on the pooled data, with and without regional covariates. In the third and fourth specifications, time dummies and municipality fixed effects are consecutively added. All regressions are weighted by municipality population size averaged over the sample period. Robust standard errors are reported in parentheses.

Table 3. Estimates of the effect from regional attributes on mean IQ scores of volunteers

	Pooled data		w/Year fixed effects	w/Municipality and year fixed effects
	(1)	(2)	(3)	(4)
Employment (%)	0.0272***	0.0131**	0.0167**	−0.0815**
	(0.00313)	(0.00498)	(0.00526)	(0.0281)
Median income (1000s SEK)		−0.00107	−0.00204	0.00931
		(0.000801)	(0.00107)	(0.00828)
Highly-educated (%)		0.0154***	0.0176***	0.00540
		(0.00372)	(0.00409)	(0.0455)
Study (%)		0.00596*	0.00461	−0.000186
		(0.00272)	(0.00291)	(0.00738)
Foreign (%)		−0.0115***	−0.0120***	−0.00720
		(0.00202)	(0.00192)	(0.0287)
Population density (100 persons/km2)		0.000878	0.000714	−0.0552**
		(0.00129)	(0.00115)	(0.0198)
Year dummies?	No	No	Yes	Yes
Municipality fixed effects?	No	No	No	Yes
N	1914	1914	1914	1914
R^2	0.046	0.157	0.178	0.368

Notes: *** indicates that the coefficient is significant at the 0.1% level. ** indicates that the coefficient is significant at the 1% level. * indicates significance at the 5% level. Dependent variable is the mean IQ score of volunteers from a municipality. Annual values, 2010–2016. The individual measure of IQ ranges from 1 (low) to 9 (high). Regressions are weighted by municipality population size averaged over the sample period. Robust standard errors in parentheses.

Table 3 shows that an increase in the employment rate of a municipality is related to a decrease in the mean IQ score of volunteers originating in that municipality. This negative relationship, however, is only present when the model controls for unobserved heterogeneity across municipalities. Indeed, municipalities appear to differ with respect to unobserved characteristics that affect both employment rates and recruit quality in a systematic way. The first column of Table 3 is the simple regression of recruit quality on employment rates in the pooled data and indicates that a one percentage point increase in the civilian employment rate increases the mean IQ score of volunteers by 0.027 points, with a standard error of 0.003. Adding regional covariates in the second column lowers this estimate to around half its size, while adding time dummies in the third column has a minor effect on the size of the estimate. The fourth column adds municipality fixed effects and the estimated effect from civilian employment is dramatically lowered: a one percentage point increase in the civilian employment rate decreases the mean IQ score of volunteers by 0.08 points, with a standard error of 0.03. This estimate is statistically significant at the one percent level. In contrast, the effects from the other variables in the regression, with population density as the notable exception, are statistically insignificant. The results presented in Table 3 are consistent with the hypothesis that poor labour market opportunities at the local level tend to increase the quality composition of those who volunteer for military service.

The practical importance of the estimated relationship between the civilian employment rate and recruit quality is difficult to interpret, however. In order to gain a more comprehensive picture of the relationship, and to test the robustness of the results, Table 4 presents three regressions that relate the size of the recruit flow to the civilian employment rate, while controlling for regional covariates, year effects and unobserved heterogeneity across municipalities. In the first specification, the dependent variable is the number of individuals who scored below four on the IQ test per 100,000 inhabitants in the municipality. These individuals do not meet the basic requirement for military training and are referred to as low-ability individuals. In the second specification, the dependent variable is the number of individuals who scored four or five on the IQ test per 100,000 inhabitants in the municipality. These individuals score high enough to qualify for military training, but do not meet the requirement for commissioned officer training, and are referred to as medium-ability individuals. In the third specification, the dependent variable is the number of individuals who scored at least six on the IQ test per 100,000 inhabitants in the municipality. These individuals score high enough to qualify for commissioned officer training and are referred to as high-ability individuals. Semi-elasticities are reported alongside the coefficients in Table 4 to facilitate interpretation, and give the percentage change in the application rate from an absolute change in the independent variable, evaluated at mean values.

Table 4. Estimates of the effect from regional attributes on application rates

	Application rate from
	Low-ability individuals		Medium-ability individuals		High-ability individuals
	(1)	(2)	(3)	(4)	(5)	(6)
	Coefficient	Semi-elasticity	Coefficient	Semi-elasticity	Coefficient	Semi-elasticity
Employment (%)	−0.155	−0.0307	−3.551***	−0.142***	−6.069***	−0.182***
	(0.187)	(0.0370)	(0.546)	(0.0218)	(0.851)	(0.0255)
Median income (1000s SEK)	−0.109*	−0.0216*	−0.0135	−0.000540	−0.113	−0.00339
	(0.0549)	(0.0109)	(0.154)	(0.00616)	(0.281)	(0.00843)
Highly-educated (%)	0.317	0.0629	−1.086	−0.0433	1.831	0.0549
	(0.311)	(0.0616)	(1.124)	(0.0449)	(2.002)	(0.0600)
Study (%)	0.0987	0.0195	−0.0158	−0.000631	−0.759**	−0.0228**
	(0.0696)	(0.0138)	(0.196)	(0.00783)	(0.279)	(0.00837)
Foreign (%)	−0.0434	−0.00859	−0.893	−0.0356	−0.379	−0.0114
	(0.213)	(0.0421)	(0.654)	(0.0261)	(0.912)	(0.0273)
Population density (100 persons/km2)	0.197	0.0391	−0.193	−0.00771	−0.822	−0.0247
	(0.129)	(0.0255)	(0.379)	(0.0151)	(0.858)	(0.0257)
N	2030		2030		2030
R^2	0.425		0.655		0.674

Notes: *** indicates that the coefficient is significant at the 0.1% level. ** indicates that the coefficient is significant at the 1% level. * indicates significance at the 5% level. Dependent variables are the annual number of individuals who performed the enlistment test per 100,000 inhabitants in a municipality, across different subsets of the recruit pool. Annual values, 2010–2016. Low-ability individuals are those with a score below four on the IQ test; medium-ability individuals are those with a score of four or five on the IQ test; high-ability individuals are those who scored six or above on the IQ test. The semi-elasticity gives the percentage change in the dependent variable from an absolute change in the independent variable, evaluated at mean values. Regressions are weighted by municipality population size averaged over the sample period. Robust standard errors in parentheses. Municipality fixed effects and year dummies are included in all regressions.

Table 4 reveals that the application rate from individuals who score high on the IQ test are more responsive towards the employment rate in the municipality of origin, compared to the application rate from individuals who score low. The first column of Table 4 shows that a one percentage point increase in the employment rate decreases the number of low-ability enlistment tests per 100,000 inhabitants by 0.16 (with a standard error of 0.19). This effect corresponds to a 3 percent decrease in the application rate from low-ability individuals at mean values, but is not statistically different from zero at the five percent level. The third column shows that a one percentage point increase in the civilian employment rate decreases the number of medium-ability enlistment tests per 100,000 inhabitants by 3.6 (with a standard error of 0.5). This effect corresponds to a 14 percent decrease in the application rate from medium-ability individuals at mean values and is statistically significant at the 0.1 percent level. The fifth column shows that a one percentage point increase in the civilian employment rate decreases the number of high-ability enlistment tests per 100,000 inhabitants by 6.1 (with a standard error of 0.9). This effect corresponds to an 18 percent decrease in the application rate from high-ability individuals at mean values and is statistically significant at the 0.1 percent level. The implication from these disparities is that the quality composition of the recruit flow changes with the civilian employment rate.

To further address the practical implications of the results from Table 4, suppose that 100 individuals, originating in some municipality, perform the enlistment test annually. Assume also that the quality composition in this group of volunteers reflects the national averages over the period studied in this paper. The 100 individuals would then comprise 8 low-ability individuals; 39 medium-ability individuals; and 53 high-ability individuals. A one percentage point increase in the civilian employment rate at these recruitment volumes would, all else being equal, mean that around 10 fewer high-ability individuals and around 6 fewer medium-ability individuals perform the enlistment test every year, whereas the number of low-ability individuals stays largely unaffected. The 84 individuals who remain in the recruit pool after the increase in the civilian employment rate would now consist of 8 low-ability individuals; 33 medium-ability individuals; and 43 high-ability individuals. Hence, compared to the initial composition, the proportion of high-ability individuals in the pool of recruits who remain after the increase in the civilian employment rate decreases by around two percentage points, whereas the proportion of low-ability individuals increases by roughly the same amount.

Summing up, the results from Table 4 contribute to the understanding of the negative relationship between mean IQ scores and civilian employment rates reported in Table 3. As conditions in the civilian labour market improve, the number of individuals who volunteer for military service and perform the military enlistment tests goes down. However, the number of individuals with relatively high IQ scores decreases more than proportionally, while the number of individuals with relatively low IQ scores decreases less than proportionally. In other words, the number of volunteers declines, mainly due to the withdrawal of individuals who perform well on the IQ test. The end result is that the quality composition of the remaining recruit pool deteriorates.

Discussion

Between 2010 and 2017, the Swedish Armed Forces relied on voluntary recruitment to supply the enlisted ranks with personnel. The purpose of this paper was to study how local labour market conditions influence the quality composition of those who volunteer for military service. The empirical analysis was based on a panel data set containing IQ scores for those who applied for military basic training across Swedish municipalities during the period 2010 to 2016. A fixed-effects regression model was estimated in order to study the effect from labour market conditions on the quality composition of volunteers, while controlling for unobserved heterogeneity across municipalities.

The main finding is that that poor labour market opportunities at the local level tend to increase the mean IQ score of those who volunteer for military service, whereas the opposite is true if conditions in the civilian labour market move in a more favourable direction. The application rate from individuals that score high on the IQ test is more responsive towards the employment rate in the municipality of origin, compared to the application rate from individuals that score low: a one percentage point increase in the civilian employment rate is found to be associated with a two percentage point decrease in the share of volunteers who score high enough to qualify for commissioned officer training. Consistent with the view that a strong civilian economy favours negative self-selection into the military, the results from this paper suggest that the negative impact on recruitment volumes of a strong civilian economy is reinforced by a deterioration in recruit quality.

The re-instatement of the draft in Sweden might seem to be an ideal instrument for avoiding fluctuations in recruit quality by selecting only high-quality recruits for military service. However, such a policy is not likely to be optimal from a social welfare point of view, and it might not be desirable for the military organization, either. First, since the military does not have to pay a premium for drafting high-ability individuals, it is unlikely that the draft would balance a draftee’s productivity while in the military with the opportunity cost of pulling him or her out of the civilian labour market. The individuals selected for military service would be overly skilled (see Berck and Lipow 2011). Second, conscripts will form the main recruitment pool for the professional branch of the Swedish Armed Forces. Since enlistment incentives for individuals with relatively high earning capacity in the civilian sector can be expected to be relatively weak, putting too much emphasis on recruit quality might mean that individuals who are highly motivated to pursue a professional military career are crowded out of the selection process. The individuals selected for military service would have too little motivation to become professional soldiers and sailors. The difficult task for the military selection system is to balance the needs of the military with individual opportunity costs and preferences.

This paper is the first attempt to analyse the role of the civilian labour market for quality aspects of voluntary recruitment in Sweden. The empirical analysis is limited by the fact that measures of individual ability are only available for those who actually apply for military training in the first place. A more complete analysis of the nature of self-selection into the voluntary military force would also, in a more direct way, consider the distribution of ability among those who do not wish to enlist. Sweden’s re-instated military conscription, which allows for a mixture of conscripts and volunteers, will in the future surely provide opportunities to analyse this issue, including the causal nature of the relationships estimated in this paper, more deeply. Moreover, it should be kept in mind that the measure of cognitive ability used in this paper only represents a limited part of all those individual characteristics that together distinguish a high-quality recruit. Future research should thus consider recruit quality and self-selection in a broader context that includes both cognitive and non-cognitive skills.

Acknowledgments

The author would like to thank Stephanie Vincent Lyk-Jensen, Linn Karlsson, Magnus Wikström, Evelina Bonnier, Gauthier Lanot and Niklas Hanes, as well as two anonymous referees, for their valuable comments and suggestions. The paper has benefited from comments received from seminar participants at Umeå University and at the Swedish Defence Research Agency. Thanks, also, to Henrik Ströby, at the Swedish Armed Forces Human Resources Centre, for helping with the data.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1. In this simple application of the Roy model, self-selection into the military is based only on the pecuniary aspects of work in either sector and ensures that individuals are perfectly sorted by skill level. A more realistic model of self-selection into the military would also recognize that individual preference for military life matters for the enlistment decision. Introducing an independent and randomly distributed preference component into the model would imply that individuals are imperfectly sorted by skill level, but would not change the basic message of the model. For example, if skills and preference for military life are independent, so that individual preference for the non-pecuniary aspects of military life is unrelated to the state of the civilian economy, one would still find some highly-skilled individuals who prefer the military over the civilian sector, even though the recruit pool on average tends to be negatively selected in terms of skills, and vice versa.

2. See Carlstedt (2000) for a historical overview of psychometric testing in the Swedish military.

3. The full data consist mainly of young men: only 16 percent were women and the median age was 20 years. The age distribution of the applicants is compressed: 90 percent of the individuals in the full sample fall within the age range of 18–25 years. The distribution of IQ scores for the individuals in the full data is illustrated by Figure A1 in the Appendix. Note that the distribution of test scores appears to be skewed to the right, with a relatively heavy tail towards higher test scores. It is important to note that those who performed the enlistment tests did not comprise a random sample of all those willing to take the tests: there is active sorting before admittance into testing. All potential recruits were obliged to complete an online questionnaire before being allowed to participate in the enlistment test procedure. The questions were related to background information on the individual, educational attainment, physical and mental status, and areas of interest regarding military service. Only individuals who met the requirements set by the military were invited to attend the enlistment test procedure.

4. An alternative measure of the central tendency of the set of test scores is the median. Using the median, instead of the mean, in the empirical analysis does not alter the main results in terms of sign and statistical significance.

5. Table A2, in the Appendix, provides results from estimating the main model without population weights. Running the fixed-effects regression without population weights does not alter the size of the point estimate of the effect from employment on mean IQ scores. As expected, however, the precision of the point estimate decreases when the regression is run without weights, making the estimate statistically insignificant at the five-percent level (it is still statistically significant at the ten-percent level).

6. See, for example, Tano (2014) and Haapanen and Tervo (2012), on the topic of regional clustering of human capital in Sweden and Finland.

7. 116 observations drop out of the analysis entirely, because some municipalities do not have any individuals that performed the enlistment test in a specific year. These municipalities are among the least populated ones in the sample.

Appendix

Figure A1. The distribution of IQ scores in the full sample of 42,886 individuals who applied for basic military training over the period 2010–2016. Scores from ten sub-tests are used by the Swedish Recruitment Agency to construct a discrete full-scale score that reflects general cognitive ability, ranging from 1 (low) to 9 (high). The nine-point scale approximates a normal distribution with a mean of 5 and a standard deviation of 2. The scaling is calibrated on a random sample from the population of conscripts who underwent enlistment testing in the year 2001.

Appendix

Table A1. Definitions of dependent and independent variables

	Definition (Source in parentheses)
Dependent variables
Mean IQ	The mean of IQ scores for volunteers from the municipality. Individual test scores range from 1 (low) to 9 (high) (Swedish Recruitment Agency)
Application rate (Low ability)	The number of enlistment tests with an IQ score below four per 100,000 inhabitants in the municipality (Swedish Recruitment Agency and Statistics Sweden)
Application rate (Medium ability)	The number of enlistment tests with an IQ score of four or five per 100,000 inhabitants in the municipality (Swedish Recruitment Agency and Statistics Sweden)
Application rate (High ability)	The number of enlistment tests with an IQ score of six or above per 100,000 inhabitants in the municipality (Swedish Recruitment Agency and Statistics Sweden)
Independent variables
Employment	The employment rate, in percentage points, for persons above age 16 in the municipality (Statistics Sweden: Labour statistics based on administrative sources [RAMS])
Median income	The median annual labour income (1000s of SEK) of individuals in the municipality aged 20 and above (Statistics Sweden)
Highly-educated	The percentage of persons in the municipality aged 25–64 that have at least three years of post-secondary education (Statistics Sweden)
Study	The percentage of persons in the municipality aged 20–24 that are enrolled in education (Statistics Sweden)
Foreign	The percentage of persons in the municipality who are born outside Sweden (Statistics Sweden)
Population density	The population density (100 persons/km^2) of the municipality (Statistics Sweden)

Appendix

Table A2. Estimates of the effect from regional attributes on mean IQ scores of volunteers. Without population weights

	Pooled data		w/Year fixed effects	w/Municipality and year fixed effects
	(1)	(2)	(3)	(4)
Employment (%)	0.0217***	0.00869	0.0169*	−0.0765
	(0.00418)	(0.00664)	(0.00728)	(0.0431)
Median income (1000s SEK)		−0.000623	−0.00313*	0.00179
		(0.00119)	(0.00147)	(0.0116)
Highly-educated (%)		0.0187***	0.0228***	0.00557
		(0.00556)	(0.00565)	(0.0604)
Study (%)		−0.00326	−0.00474	−0.00250
		(0.00414)	(0.00411)	(0.0107)
Foreign (%)		−0.00479	−0.00607	−0.0103
		(0.00340)	(0.00340)	(0.0441)
Population density (100 persons/km2)		−0.000614	0.000145	−0.0414
		(0.00230)	(0.00232)	(0.0272)
Year dummies?	No	No	Yes	Yes
Municipality fixed effects?	No	No	No	Yes
N	1914	1914	1914	1914
R^2	0.014	0.029	0.045	0.255

Notes: *** indicates that the coefficient is significant at the 0.1% level. ** indicates that the coefficient is significant at the 1% level. * indicates significance at the 5% level. Dependent variable is the mean IQ score of volunteers from a municipality. Annual values, 2010–2016. The individual measure of IQ ranges from 1 (low) to 9 (high). Robust standard errors in parentheses.