Archival Tests of H2 and H3

21 argues that H2 and H3 occur because low SES college students follow slow strategies in benign conditions but adopt faster strategies in uncertain conditions. If our theory is valid, then SES and recession jointly affect a person’s likelihood of self-selection into accounting because they jointly affect the importance of job security to that person. That is, our interaction is likely to have an indirect effect on self-selection, via differences in the importance of job security. Following the recent statistics literature, we test the significance of the indirect effect using a bootstrapping technique Preacher and Hayes 2008; Hayes and Preacher 2013. Because our independent variable is an interaction term, we follow the guidance of Hayes and Preacher 2013 to create a multi-categorical independent variable using the linear weights for the interaction term. 13 We use 5000 bootstrap re-samples of the data to calculate bias-corrected confidence intervals for the total indirect effect. Significance is indicated by confidence intervals that do not include zero. In our analysis, there is a significant indirect effect of life history strategy on self- selection into accounting, which is mediated by the assessed importance of job security lower CI = 0.75, higher CI = 9.86. See Figure 1. Thus, our data support the conclusion that the SES by recession prime interaction affects self-selection into accounting because it affects the importance of job security.

5. Archival Tests of H2 and H3

– SES and Self-Selection into Accounting We now turn to tests of H2 and H3 using large-sample archival evidence to provide comfort about the generalizability of our findings. 13 We used the following weights: -0.25 for control prime, high SES; +0.25 for recession prime, high SES; +0.25 for control prime, low SES; -0.25 for recession prime, high SES . Following Hayes and Preacher’s 2013 guidance, we include two orthogonal control contrasts because fully representing the effects of a categorical variable with k categories requires k – 1 parameter estimates. 22 5.1. Sample We obtain data from the Higher Education Research Inst itute’s HERI Freshman Surveys, which contain data collected from millions of incoming college freshmen annually since 1971 UCLA 2013. The surveys describe how students choose which college to attend as well as their demographics, high school academic and extracurricular activities, opinions on a wide array of topics, values, and the field in which they intend to earn a degree. HERI data from 1971 to 1999 are available to all registered users of HERI’s website in an archive file. More recent data are available for purchase contingent on approval from HERI. In this study, we use subsets of HERI data from the archive as well as data purchased for 2000 and 2002. 14 We conduct our main tests using subsets of years from 1971-2002 for which our variables are available. 15 The sample size in our main tests is 125,125 accounting observations, 407,235 non-accounting business observations, and 2,993,954 non-accounting observations. 16 In addition, to test the robustness of H2 and to perform certain supplemental analyses described later, we also obtain a subset of the HERI database that collects responses both when individuals were freshman and when they were seniors. This dataset is only available from 1994 – 1999, thus tests using the senior data have lower statistical power than tests using the freshman dataset. Also, no recessions occurred during 1994 – 1999, thus we cannot use it to test H3. 14 The 2000 and 2002 samples differ somewhat from the earlier samples because HERI was unwilling to release the entire database to us. Rather, for 2000 and 2002 HERI provided us with all observations of students selecting any business field and a random sample of observations representing students selecting non-business fields. 15 See http:www.heri.ucla.eduabtcirp.php and Appendix B for more information about HERI variables and data availability. 16 Like many surveys, HERI is constructed using stratified sampling rather than strictly random sampling in order to ensure adequate coverage of the populations of interest to the study’s designers. Because their sampling is not strictly random, traditional statistical methods applied to HERI data will produce biased estimates if they are not adjusted to correct for HERI’s sampling methodology. Statistical corrections for nonrandom sampling are common in survey data analysis and involve weighting observations using their probability sampling weight e.g., Rosenbaum 1987; King et al. 2010. All of the means and statistical tests we report are corrected to account for survey weights. 23 5.2. Variables 5.2.1. Self-selection into accounting To test self-selection into accounting, we partition our sample into students planning to major in accounting and those not planning to major in accounting. This is a reliable predictor of selection into accounting jobs Madsen 2015. We use two comparison groups to test our hypothesis. We test self-selection into accounting against “non-accounting business” fields business administration, finance, international business, marketing, management, and other business and against “all non-accounting” fields the set of all non-accounting fields. 5.2.2. SES We manually search the HERI codebook and identify three measures of childhood SES that are both available throughout the whole sample period and relevant indicators of low SES. They are 1 whether or not the respondent indicated that cheap tuition was “very important” to them when selecting a college, as prior research has shown that low SES students are particularly sensitive to tuition costs Heller 1997, 638-642; 2 whether or not the respondent is a first generation college student , as first generation college students come from poorer families on average Terenzini et al. 1996, 8- 9, and whether or not the respondent’s estimate of their parental income is in the bottom third of the sample for that year, which is a direct measure of respondents’ perceptions about relative household income. 17 For each of these measures, we create a dummy variable equal to 1 for values representing low SES and 0 otherwise. 18 We then sum these three 17 Our results are robust to alternative specifications, such as the bottom quartile or the bottom half of the distributions. 18 Most of the HERI variables are ordinal measures with verbal anchors on each scale point, and the measures vary in the number of scale points and in the verbal anchors used. For example, parental income has 14 points, each associated with an income range e.g., 1 = “Less than 6,000,” 14 = “200,000 or more”, while cheap tuition has three 1 = “Not important,” 2 = “Somewhat important,” 3 = “Very important”. There is no consensus in the literature about how best to treat this type of ordinal variable statistically. Baum 2006, 161 and Campbell 2008 argue that dichotomization is justified in large samples such as ours, but that treating variables as continuous is also allowable. In our main tests, we transform ordinal variables into dichotomous variables around their medians unless otherwise noted in the text. We run robustness tests treating all ordinal predictors as continuous and our inferences do not change. 24 dummy variables to create a low SES index. Higher values of the low SES index indicate lower SES. Our main tests involve comparing values of the low SES index for students who have selected accounting degrees against values for students selecting other degrees. Table 3, Panel A sorts all fields in the HERI freshman database based on our low SES index in recession and non-recession years, with higher values indicating higher representation of low SES individuals. Accounting stands out relative to other fields and to business fields in particular, ranking 11 th overall out of 76 fields in low SES . These results suggest that accounting is by far the most appealing business field to low SES individuals, versus management 35 th , business administration 39 th , marketing 69 th , and finance 70 th . Table 3, Panel B shows that accounting has the 12 th largest decline in the low SES index in recession years. All other business fields experience smaller declines in the low SES index during recession years, as no other business field ranks higher than 38 th in the magnitude of the decline Finance. Indeed, accounting appears to occupy a unique niche among business fields and among most fields in general. 19 5.2.3. Uncertain Economic Conditions Our proxy for uncertain economic conditions is a year in which a recession occurred. We partition the sample period into recession years and non-recession years. Recession is an indicator for any year that includes at least one month classified by the National Bureau of Economic Research as a recession month. Recession years during our sample period are 1973-1975, 1980- 1982, and 1990-1991. 19 Of the fields with comparable low SES representation to accounting, only agriculture and nursing have similarly- sized declines in low SES representation in recession years. Thus, these fields may fit similar life history strategies to those fit by accounting. 25 5.3. Results 5.3.1. Hypothesis 2 – Univariate Tests of the Effect of SES on self-selection into accounting Table 4, Panel A presents univariate analyses. It provides mean values for each of our low SES indicators, both separately and when aggregated, in accounting, other business fields, and all non-accounting fields. H2 predicts that a low SES background is positively associated with self- selection into accounting. Consistent with H2, accounting students have higher values than non- accounting students for each of our low SES indicators. These differences are even larger when comparing accounting to other business fields. Table 4, Panel A also presents the univariate results among college seniors. Because people from disadvantaged backgrounds are more likely to drop out of college before they become seniors, all three low SES indicators are lower in the senior dataset than in the freshman dataset. The senior dataset also has substantially smaller sample sizes and lower statistical power. Notwithstanding these limitations that bias against our findings, H2 is robust among college seniors. Thus, our univariate analyses support H2. 5.3.2. Hypothesis 2 – Multivariate Tests of the Effect of SES on self-selection into accounting Table 4, Panel B presents results of a logit analysis that predicts the likelihood of selecting into accounting based on SES, which we estimate separately during the 1970s, 1980s, 1990s, and 2000s only the years 2000 and 2002 due to data limitations to illustrate the effect’s robustness over time. We estimate the following model: Select Accounting 0 or 1 = α + β 1 lowSES + β 2 Female + β 3 Minority + β 4 GPA + β 5 Drive + β 6 Self-confidence intelligence + β 7 Self-confidence social + β 8 Writing + ε 1 Our primary variable of interest is lowSES and we predict a positive effect for this variable, i.e., an odds ratio greater than one. We report coefficients in the form of odds ratios, which are exponentiated logit coefficients, to provide a clearer interpretation of the practical implications of 26 the effects of a given variable. In brief, odds ratios represent the number of people in our sample with a given attribute choosing accounting for every one person choosing a business degree under the heading “freshmen: accounting versus non-acctg business students” or any other degree under the heading “freshmen: accounting versus all non-acctg students”. Table 4, Panel B also shows similarly calculated results for college seniors under the “seniors” heading. To select control variables, we collected input from seven senior managers and partners in public accounting firms about desired attributes that lead to success in an accounting career. We presented each person with a list of 16 items from the HERI survey that we expected could be related to success and asked each person to select the five that are most likely to lead to success in accounting. We control for variables that were selected by a majority of respondents. Specifically, we include GPA on a four-point scale as a proxy for academic ability high school GPA for college freshmen and college GPA for college seniors, an indicator variable for above average self- assessed drive as a proxy for motivation, indicator variables for above average self-assessed intellectual and social self-confidence, and an indicator variable for above average self-assessed writing ability as a proxy for writing skills. We also include indicator variables for Female and for membership in a disadvantaged minority race or ethnicity, which we label Minority and includes Black, Hispanic, and American Indian people. We include Female and Minority to provide insight into the relation between our effect and the diversity of the accounting labor pool. See Appendix B for details about the variables. 20 The results in Table 4, Panel B support the idea that low childhood SES significantly and positively predicts the likelihood of selecting into accounting in each of the sample periods. Thus, 20 The 16 variables in our questionnaire were academic ability i.e. GPA, amount of time spent studying, amount of participation in clubs or student organizations, computer skills, competitiveness, creativity, drive to achieve, initiative, leadership ability, mathematical ability, public speaking ability, self-confidence, self-understanding, understanding of others, volunteer work, and writing ability. 27 multivariate analyses support H2. These analyses show that our effect is also robust to controlling for race, gender, academic ability, self-confidence, motivation, and writing ability. The results also indicate a few potentially problematic patterns for accounting. While GPA is positively associated with accounting in every sample period and relative to both business and non-business fields, individuals selecting accounting appear to possess lower social self-confidence and writing ability. We explore these issues in Section 6.1 to examine whether or not the low SES effect may help overcome some deficits in the entry-level accounting labor pool. Life history theory suggests this is possible, as those adopting slower life history strategies tend to be highly capable. Because low SES individuals have different life experiences than their high SES counterparts, they may also possess high levels of different capabilities and thus improve the quality of the entry-level accounting labor pool. 21 5.3.3. Hypothesis 3 – SES in recessions and non-recessions The HERI data also permit us to provide large-sample analysis of H3, which predicts that recessions decrease preferences for accounting when SES is low, and do so to a greater degree in accounting than in other fields. Our univariate analysis of H3 is presented in Table 5, Panel A. 21 We conduct untabulated tests to ensure that our results are robust to inclusion of all 16 HERI items that we identified as potentially important i.e., the 11 attributes from HERI that were not selected by our practitioner panel in addition to the five that were selected. However, some of the 11 variables are not available in all the years for which we conducted our primary tests, including four that are not available in any of the recession years. Thus, we run a series of robustness tests as follows: 1 we run models that include all 16 variables only in those years that all variables are available applicable only to H2, 2 we run models that include our five primary controls plus other controls that are available for all the years for which the five primary controls are available 7 of the 16 possible controls, and 3 we run models that include our five primary controls plus variables that are available in at least two recession years 9 of the 16 possible controls. In the reported univariate tests of H2 and H3, we use samples restricted to include the same sample years as those used in our multivariate tests of H2 and H3. We conduct robustness tests in which we use all sample years in our univariate difference in difference tests of H2 and H3. Univariate and multivariate tests of H2 effect of low SES and univariate tests of H3 effect of low SES by Recession are robust to all these alternate specifications for both the “non-accounting business” and “all non-accounting” comparison groups. Multivariate tests of H3 are robust to specifications 2 and 3 for the “all non-accounting” specification and to specification 2 for the “non-accounting business” comparison group. H3 is directionally consistent with our prediction, but insignificant for specification 3 for the “non-accounting business” comparison group. It is unclear whether H3’s lack of significance for this one specification for this one comparison group is due to the additional controls or to the restricted sample. 28 We conduct a difference in difference test that compares the decline in the low SES index in recession, as opposed to non-recession years in accounting to the decline in other fields. The results show that low SES representation on average drops in accounting, in other business fields, and in all other fields during recession years. This is consistent with recessions more negatively affecting rates of college attendance among people with fewer economic resources. However, the decline in low SES representation is larger in accounting than in other business fields p 0.01 and in the set of all other fields p 0.01. Thus, the results support H3. 22 We report multivariate analysis of H3 in Table 5, Panel B, which is based on the model: Select Accounting 0 or 1 = α + β 1 lowSES + β 2 lowSESRecession + β 3 Recession+ β 4 Female + β 5 Minority + β 6 GPA + β 7 Drive + β 8 Self-confidence intelligence + β 9 Self-confidence social + β 10 Writing + time period dummies + ε 2 Our primary variable of interest is the lowSESRecession interaction. We expect this interaction effect to be negative, i.e., an odds ratio lower than 1. The control variable definitions are the same as in equation 1. Because there are time trends that affect our variables of interest, including fewer first generation college students and increasing popularity of business fields in general , we also include dummy variables for each five-year time period in our sample to ensure that our analyses are not confounded by time effects. The reference time period to interpret each time period effect is 1971 – 1975, i.e., each time period effect tests the interest in accounting in that time period relative to interest in 1971 – 1975. 23 22 Ideally, we would also be able to explicitly identify the fields chosen by people who might otherwise select into accounting . However, we are unable to capture this data, even with our dataset of seniors that allows us to observe switches between fields. First, the senior dataset only includes years in which there were no recessions, thus we could not observe differences in substitutes between recession and non-recession years. Second, this dataset only captures alternatives to accounting for those who actually remained in college, and earning a degree is an inherently slower strategy than dropping out. Thus, the data will understate the life history speed of alternatives, and this understatement is likely stronger among low SES individuals, who are more likely to drop out of college. 23 We use time period dummies instead of year fixed effects because Recession is defined at the year level. In each model, the odds ratios for selecting accounting decrease over time. For example, for every person selecting accounting over another business discipline in 1971 – 1975, 0.372 selected accounting in the “boom years” of 1996 29 As predicted, the results show a negative interaction effect in which low SES interest in accounting is lower in recession years, and this decrease is greater in accounting than it is in business fields and in the set of all fields. This interaction is striking because, in both columns of Table 5 Panel B, Recession is associated with higher interest in accounting. That is, interest in accounting among low SES individuals is lower during times when overall interest in accounting is higher. We argue that this is due to people from poorer backgrounds adopting faster trajectories in their life history strategies when they observe cues of economic uncertainty. Importantly, if the recession interaction effect simply reflected fewer low SES individuals in college during recessions due to lower application rates or higher dropout rates, then lower SES representation in accounting would not differ from that of other fields. Thus, our multivariate analyses support H3. In sum, these archival analyses provide evidence both that our theorized relations generalize in a large-sample dataset of millions of real choices and that the effects are robust over time. Low SES individuals disproportionately prefer accounting because it is consistent with a slow life history strategy. However, recessions affect patterns of selection into accounting differently than other fields, as the preference for accounting by low SES individuals disproportionately weakens.

6. Supplemental Analyses