This study made use of the 2014 World Bank Enterprise Survey data. The data is a stratified multi- stage random sampling, comprising geographically enumerated areas and cutting across all the sectors of the economy. After cleaning, the first analysis comprises of 2,238 samples: 252 of these are micro enterprises employing less than 5 persons, 1,180 are small enterprises with between 5 and 19 employees, 629 are medium-sized enterprises with between 20 to 99 employees, and 177 are large enterprises with 100 or more employees. The second analysis comprises of 2,249 sample size: 262 are micro enterprises, 1,228 are small enterprises, 609 are medium size firms and 150 are large size firms.
This survey contains information on each enterprise surveyed, the number of employees on establishment, in the 2012 fiscal year and in the 2009 fiscal year. This gave us the opportunity to identify for each firm size in the survey what happened in these two-points period in terms of employment generation and job loss. This enabled the identification of the enterprises (micro, small, medium and large enterprise) that were contributing more significantly to employment generation and less to job loss in the Nigerian economy. In identifying which firm size had been contributing to job creation and job destruction, it is worth noting that this data did not reflect the actual jobs created and lost in these firms throughout the period, it only reflected what the employment levels were in 2009 and 2012 and what obtained in the fiscal year that the enterprise started operation. This data is not capable of capturing the overall detail of what transpired between these two periods. Also, the
time period is a variable that will affect the data obtained. The economic depression occasioned by the 2007 financial crisis is still taking its toll on the world, especially with regard to access to credit for business enterprises, with its concomitant effect on demand and consumption. The resultant effect is a massive downsizing by firms in a bid to stay afloat.
To answer the question of which firm sizes create more jobs, it is crucial to examine all sectors of the economy, not just the manufacturing sector which was the focus of previous studies (Little, 1987; Davis et al., 1996b; and Biggs, Ramachandran, & Shah, 1998). This is a pioneer analysis in Nigeria in terms of identifying which firm size has net job creation ability in Nigeria, cutting across all regions and all sectors of the economy. According to the classification of the firm size by the survey, we examine the significant difference in net job creation ability across firm size categories. Secondly, we non-parametrically estimate the relationship between net employment growth and establishment size to avoid any potential bias or loss of information introduced by using arbitrary size boundaries by representing the information obtained in percentage form.
Our non-parametric analysis uses the locally-weighted scatterplot smoothing (LOWESS) method proposed by Cleveland (1979) and modified by Neumark et al. (2008), outlined as follows:
Step 1: Let yi be the employment growth rate of observation i (an establishment over a two-year period), xi the size of observation i measured using average size definition, and N the total number of observations. The standard implementation of locally-weighted mean smoothing would proceed as follows. Order the data such that xi β€ xi-1 for all i = 1, β¦, N-1. For each yi, choose the subset of the data that is indexed by i - = max(1, i-k) through i- = min(i+k, N), where k= [(N.h β 0.5)/2] and h is the pre-specified bandwidth that indicates the proportion of the data used in the calculation of the smoothed value π¦π¦π¦π¦οΏ½. Choose a function that assigns a weight wi to each observation j=i-, β¦, i-; observations outside of this range are given no weight. For example, one may choose a tri-cubic weight function (the kernel), in which case the smoothed value π¦π¦π¦π¦οΏ½ is calculated as:
π¦π¦π¦π¦οΏ½ = βππ+ππ=ππ+(π€π€π€π€.π¦π¦π€π€) βππ+ (π€π€π€π€) ππ=ππβ ,π€π€βπππππππ€π€π€π€ = οΏ½1β οΏ½ {π₯π₯π€π€βπ₯π₯π₯π₯} β οΏ½ 3 οΏ½3and β = 1.0001*max(xi- - xi, xi - xi-). 3.1
It can also be worked:
Given a repeated value for many observations, this first method is computationally non-feasible. It would involve calculating the repeated weighted average. Instead, this study utilises the following method where we first compute an average value y for each unique value of x and then calculate a smoothed value π¦π¦π¦π¦οΏ½ from the reduced dataset.
We use the following procedure. First, order the data such that xi β€ xi+1 for all I = 1, β¦, N-1. For each unique value of xi, create a zi= xi. Let the total number of z be M and order all of them such that zi< zi+1 for all I = 1, β¦, M-1. Then, let π¦π¦π¦π¦ = β π¦π¦πΆπΆππππ(πππποΏ½οΏ½ππππ)π¦π¦ππ for all I = 1, β¦, M, where πππ¦π¦οΏ½ = {(π¦π¦π¦π¦,π₯π₯π¦π¦): (π₯π₯π¦π¦=π§π§π¦π¦)} and
C(πππ¦π¦οΏ½) is the cardinality of πππ¦π¦οΏ½. Now apply the standard smoothing procedure to the observations (yi, zi), except that the weight function is adjusted using the frequency of yi. Again, using a tri-cubic weight function, this amounts to calculating the following smoothed value:
π¦π¦π¦π¦οΏ½ = βππ+ππ=ππβ(π€π€π€π€.π¦π¦π€π€) βππ+ππ=ππβ(π€π€π€π€) where wj =πΆπΆ(πππ¦π¦οΏ½).οΏ½1β οΏ½ {π₯π₯π€π€βπ₯π₯π₯π₯} β οΏ½ 3 οΏ½3. 3.2
These two methods essentially use the same information in the data although they usually assign slightly different smoothed values to different observations. Whereas the standard method gives multiple predicted values for each zi in cases where there is multiple xi such that xi = zi,, our method only returns one predicted value per unique value of x.
Non-parametric methods make fewer assumptions and this makes their applicability more acceptable and robust than the parametric methods. Another justification for non-parametric methods is its simplicity. It is easier to use non-parametric methods to parametric methods because of its robustness, and it leaves no room for improper use and misunderstanding. The essence of using non-parametric method here is to avoid any potential biases or loss of information introduced by using arbitrary size boundaries of firm sizes.
However, non-parametric methods have less power to parametric methods if the sample size is not large enough. With large sample size, a non-parametric method can draw the same degree of confidence has that of parametric methods.