«CSAE Working Paper WPS/2014-34 Aspire* Marcel Fafchamps Simon Quinn Stanford University, BREAD and NBER University of Oxford March 2015 Abstract We ...»
Since we rely on a comparison between competition winners and runners-up, we need to ensure that winners are not, ex ante, any better than runners-up. To this effect, we regress all descriptive statistics from the baseline questionnaire on winners and runners-up, and test whether the winner dummy is signiﬁcant. In Table 4 we regress characteristics collected at baseline on a winner dummy, using only winners and runners-up in the comparison. The results suggest that winners and runners-up are not different populations: only three of the 16 variables are signiﬁcant at the 10 percent level (winners are less likely to have been married, less likely to have spoken English, Extra cash could facilitate job search, as shown for Ethiopia by Franklin (2014). But the prize winnings far exceed the cost of searching for a wage job in Addis Ababa. Cash could also facilitate international migration, something that we cannot rule out but we suspect would only affect a small proportion of our study population. We do not observe international migrants in our data since, by design, they do not respond to the endline survey. We revisit this issue when we discuss attrition.
and more likely to have parents who run a busineses).15 In the impact analysis below we use these three variables as controls in one set of estimations, and show that this does not affect the results.
We have the scores given to each individual contestant by the committee judges. We perform the same balancedness analysis on the average scores and rankings given by individual judges. To recall, high scores to questions 1 to 8 indicate low performance while the opposite is true for questions 9 and 10. A low ranking means a better performance. The results are presented in Table
5. We see that winners are not signiﬁcantly different from runners-up in any of questions 1 to 8, nor are they ranked differently by judges as individuals. Winners perform better on the judges’ perception of the business growth potential, and the consequent recommendation to invest. In the analysis that follows, we use as controls the answers to questions 9 and 10, and the individual judge rankings.
Table 5 here.
We also conduct balancedness tests on time-invariant variables collected at endline. The reason for this additional check is to protect against possible misreporting in the self-reported information collected at baseline – notably age and education. We also include gender and an English speaking dummy collected at endline, as well as household size and the number of children. Balancedness test results are reported in Table 6. None of the variables is signiﬁcantly different between winners To verify how strong this pattern is, we estimate the same regression on the entire contestant population. We ﬁnd the same result for having been married, but the English speaking dummy is no longer signiﬁcant, nor is the result on parents having run a business. (Additionally, we ﬁnd that winners reported a signiﬁcantly larger planned investment of their own funds.)
Finally, we examine the data for signs of non-random attrition between winners and other contestants. We consider two types of attrition: (i) between baseline and the competition, and (ii) between the competition and the endline survey. The ﬁrst column of Table 7 regresses having competed on baseline characteristics. Individuals who ﬁlled at least the age question on the baseline questionnaire are regarded as part of the baseline population.16 There are 916 individuals at baseline, 442 of whom participated in the competition and were ranked by committee judges. Many baseline individuals did not fully complete the questionnaire, however, so that the results in column 1 should be interpreted as presenting correlates of attrition for those who ﬁlled most of the baseline questions.
We observe that individuals with children or who have travelled outside the country are less likely to participate to the competition, conditional on having ﬁlled the baseline questionnaire; individuals who speak English were more likely. This is not too surprising: the time cost of competing is more taxing for parents, and better travelled individuals probably have better outside options. Similarly, English-speakers may have anticipated having a higher chance in the competition. Column 2 adds answers to a question that was not ﬁlled in by many candidates: the percentage investment anticipated in a future business. This is highly signiﬁcant in predicting competition participation.
Variables such as age, gender and education are not signiﬁcant.
Next we examine whether contestants who ranked highly and won the competition are less likely to answer the enline questionnaire. Of the 442 contestants, 369 (83.4%) were interviewed at endline.
Column 3 regresses participating to the endline survey on a winner dummy and the contestant’s ﬁnal rank. We ﬁnd no evidence that top contestants are more or less likely to answer the endline questionnaire. In column 4 we repeat the analysis with baseline regressors.17 We ﬁnd that the nine contestants with children are, on average, less likely to answer the endline survey. Better educated contestants and those who do not speak English are more likely to answer. Winning status and committee ranking remain non-signiﬁcant.18 From this we conclude that there is not differential attrition by winning status, and more generally that attrition is not correlated with the performance of contestants in the competition.
4.2 Empirical results
We now turn to our main results. Coefﬁcients for equation (1) were estimated with the winner and two runners-up from each committee in the competition. There are 39 winners and 82 runnersup in all.19 Standard errors are clustered by judging committee throughout. Panel A of Table 8 presents estimation results using the winner dummy as sole regressor. Panel B adds unbalanced baseline variables and committee scores as controls.20 Panel C omits those controls, but instead includes judging committee ﬁxed effects.
As noted earlier, this leads to a loss of observations given that some baseline questions were not completed.
For completeness, we also run regressions of the kind reported in Table 8, where the outcome variable is whether the respondent was interviewed at endline. In every case, the coefﬁcient on winning is very small, and far from signiﬁcant: for the speciﬁcations in Panel A, Panel B and Panel C successively, the estimates are -0.044 (p = 0.510),
-0.058 (p = 0.412) and -0.007 (p = 0.908).
In Zambia, we had half as many contestants as in either of the other two countries; additionally, there were three committees in Zambia with only two contestants. For these reasons, there are only 21 runners-up for the seven winners in Zambia.
The complete list of controls is as follows: average marks for questions 9 and 10; the average of the ranks given to the contestant by individual judges; a dummy for whether the contestant is married; a dummy for whether the contestant’s parents have a business; and a dummy for whether the contestant speaks English. Because baseline information is missing for some contestants, including controls results in a loss of some observations.
The ﬁrst row of Table 8 presents the results for self-employment, which is our main dependent variable of interest. The dependent variable takes value 1 if the respondent answers ‘yes’ to the question “Do you derive an income from activities other than wage employment, i.e., are you selfemployed?”. We ﬁnd that winners are 33 percentage points more likely to be self-employed than runners-up six months after the competition; this is signiﬁcant at the 1% level. Of the 32 winners, 24 were self-employed at follow-up (75%); of the 72 runners-up, 30 were self-employed at followup (41.67%).
The second row presents a similar regression using data from time budget questions. The dependent variable is the number of hours in self-employment reported for the preceding day. The ﬁnding is similar: the point estimate is an extra 2.5 hours worked in self-employment, from a base of 1.7 hours; this effect, too, is signiﬁcant at the 1% level. Note that this is an average of 2.5 extra hours worked in self-employment, across all winners. This implies that winners starting their own business worked about 7.5 extra hours in that business on the preceding day; that is, it implies that those starting their own business treat it as a full-time occupation.21 Panel B of Table 8 presents coefﬁcient estimates obtained with unbalanced variables as controls;
Panel C includes judging committee ﬁxed effects. Results are similar – virtually identical for hours in self-employment, and slightly larger for the probability of being self-employed. From this we conclude that our results are driven by the effect of the cash grant, rather than by any inherent differences between winners and runners-up.
That is, 7.5 ≈ 2.5/0.33.
In rows 3 and 4 we present similar results for wage employment. Among our study population at endline, permanent wage employment is relatively common: of 380 respondents to the endline survey, 40% answered ‘yes’ to the question “Do you have a regular wage job?”. We do not expect a large lump sum transfer to increase the probability of being in wage employment – if anything, it may even reduce this probability if winners slack on search intensity. If anything, this is what we ﬁnd: winners and runners-up are not signiﬁcantly different in terms of the probability of have a permanent wage job at endline, and on average they work slightly fewer hours in wage employment (though the difference is not statistically signiﬁcant). Similarly, we ﬁnd no effect on search or on the monthly reservation wage.22 These ﬁndings are unaffected whether we include controls or committee ﬁxed effects.
Figure 2 illustrates these key outcomes against the committee ranking.23 Vertical lines denote the winner (ranking 1) and the respondent rankings used in the regressions as counter-factuals in Tanzania and Ethiopia (rankings 2 and 3). The graphs illustrate both the central idea behind the identiﬁcation strategy and the key results: outcomes are reasonably homogeneous for rankings 2 to 12, but are signiﬁcantly different for winners’ probability of self-employment and hours spent in self-employment.
Figure 2 here.
Next we investigate whether winning the prize affects ﬁrm performance. We examine ﬁve indicators of ﬁrm performance: average sales over the last month, average costs, self-reported proﬁts, We ﬁnd that winners spend an average of 0.85 hours fewer each day on leisure (including washing and grooming);
this is signiﬁcant with p 0.01. Aside from self-employment and leisure, we ﬁnd no signiﬁcant effect on any other category of time use (time in wage employment, time searching for work, time studying, time sleeping, time socialising and attending religious ceremonies, time doing chores and time on other activities).
Speciﬁcally, the ﬁgure shows coefﬁcients from a regression on dummy variables, with no controls. We cluster by judging committee and show 90% conﬁdence intervals. For clarity, we drop the eight Zambians who were ranked ‘1’ but who did not include a prize; results are robust to including them.
Marcel Fafchamps & Simon Quinn Aspire
proﬁts calculated as sales minus costs, and number of permanent employees. To avoid sample selection problems, we code each outcome as zero for respondents who are not self-employed; that is, we estimate the effect of winning on the unconditional expectation of ﬁrm performance.
We ﬁnd large and signiﬁcant effects. In the basic speciﬁcation (Panel A), four outcomes are signiﬁcant: total monthly sales, both measures of proﬁts, and the number of permanent employees.
The pattern repeats as we add controls (Banel B) and then committee ﬁxed effects (Panel C): coefﬁcients remain remarkably stable, though the addition of controls improves efﬁciency. In Panels B and C, we ﬁnd signiﬁcant positive effects on all measures of ﬁrm performance.
We then look at total income, expenditures, and assets. For total income, the point estimate is
positive — and is large in the Panel B and Panel C speciﬁcations — but it is not statistically significant, possibly because the variance of income is high.24 We get similar results for expenditures:
point estimates are positive, but not signiﬁcant. Finally, assets similarly move in the expected direction: winners have less debt and more savings in cash and in the bank by endline. The effect is statistically signiﬁcant at the 95% level for bank savings in the regression with controls and in the regression with committee ﬁxed effects. The magnitude of this effect is large: roughly a doubling of bank savings among winners six months after the competition. We estimate the same regression with combined personal wealth, deﬁned as bank and cash savings minus personal debt. As could be expected given our earlier results, point estimates are large; they are signiﬁcant at the 90% level in each speciﬁcation.
Finally, we investigate whether winners are more likely to be married six months after the competition, conditional on not being married at baseline. Getting married costs money, and winners may In separate regressions (available on request), we use log(income) as a dependent variable. We ﬁnd large estimated coefﬁcients, although not signiﬁcant: 0.581 for the Panel A speciﬁcation (p = 0.145), 0.801 for the Panel B speciﬁcation (p = 0.113), and 0.533 for the Panel C speciﬁcation (p = 0.200).
Marcel Fafchamps & Simon Quinn Aspire
have invested their winnings in paying for a wedding rather than investing in self-employment.
This is not what we ﬁnd: there is no signiﬁcant effect on the probability of being married. The proportion of married individuals at endline is quite small, however (7%), suggesting that our study population may be too young for an effect to be noticed.