Do indirect questions reduce lying about corruption? Evidence from a quasi-field experiment.

Clarke, George R.G. ; Friesenbichler, Klaus S. ; Wong, Michael 等

INTRODUCTION

Managers who are concerned about being identified or looking bad have little reason to answer questions about corruption honestly. Managers who believe they might be identified if they admit to paying bribes could face legal problems, retaliation from the person they bribed, or might look bad to family, friends, and competitors. Even if managers think their responses will remain anonymous, they might not want to look bad to the interviewer. If they do not think the survey will change government behavior or they realize their individual responses are unlikely to affect average values in large surveys, they gain little from answering truthfully.

It would, therefore, not be surprising if some managers lied about paying bribes. Survey respondents often underreport socially frowned on and illegal behaviors (Coutts and Jann, 2011; Fisher and Tellis, 1998; Lee, 1993). How much they do this depends on how sensitive the topic is, how interviewers ask questions, how data are collected, and characteristics of the respondent and the interviewer (Coutts and Jann, 2011). People with something to hide are especially likely to lie when asked sensitive questions (Tourangeau et al, 2000).

Several recent papers have tried to estimate how much managers underreport bribes by identifying reticent respondents (Azfar and Murrell, 2009; Clausen et al, 2010; Jensen and Rahman, 2011). The interviewer asks the manager forced response questions (Warner, 1965) on sensitive topics unrelated to corruption. Researchers then use the managers' answers to the forced response questions to identify reticent managers. We describe this approach in detail in the paper's next section.

Underreporting is estimated by comparing reticent managers' answers to questions about corruption with less reticent managers' answers to the same questions. Azfar and Murrell (2009) found reticent managers in Romania were less likely to report paying bribes when registering their firm and during health inspections than were other respondents. They interpret this as suggesting that reticent managers answer questions on bribes strategically and, therefore, underreport corruption. Based on conservative estimates of reticence and underreporting, Azfar and Murrell (2009) estimated that about a third more firms paid bribes in Romania than reported they did on the survey.

This paper looks at the link between reticence and corruption using data from firm surveys in Sri Lanka and Bangladesh. As in the papers discussed above, we find reticent managers are less likely to report problems caused by corruption than are less reticent managers. This underreporting could lead us to underestimate how common bribes are in these countries. In the sample, for example, only about 2% of managers whose firms had not been inspected said tax inspectors typically requested or expected bribes. If no managers were reticent, however, we estimate at least 14% of managers would answer the question in this way. Similarly, only about 11 % of managers whose firms had been inspected said tax inspectors expected or asked for bribes during the inspection. If no managers were reticent, we estimate at least 22% of managers would answer the question in this way. The results confirm, using data from a different survey given in a different environment, that reticent managers answer questions about corruption differently than do other managers. Questions to identify reticent respondents would, therefore, be useful in many countries and surveys.

This paper also adds to the literature on reticence and corruption in several other ways. First, in contrast to previous studies (Azfar and Murrell, 2009; Jensen and Rahman, 2011), we find reticent managers are less, not more, likely to refuse to answer questions about corruption than are non-reticent managers. We estimate non-response rates would rise by between 2 and 8 percentage points on the four questions on corruption if managers were not reticent. A reticent manager might refuse to answer questions because he or she believes people will interpret the manager's refusals as a tacit admission of guilt. This belief would not be unreasonable; when the World Bank uses the Enterprise Survey data to calculate the percent of firms that pay bribes in a country, they treat firms that refuse to answer as if they paid bribes (World Bank, 2012a). Reticent managers might, therefore, prefer to answer 'no' rather than refuse to respond to the question.

It is not clear why our results on refusals are different from the results in previous studies. Although it is possible the differences are cultural, it is also possible that differences in the questionnaires cause the differences in results. In contrast to the survey in Azfar and Murrell (2009), our survey does not focus on corruption; it contains only a few questions on corruption at the end of the survey. Understanding why our results are different might be important for future work.

Second, we find reticence affects answers to both direct and indirect questions. Firm surveys often include indirect questions that ask the manager whether other firms pay bribes rather than asking direct questions about the manager's experiences. Indirect questions allow managers to answer without directly implicating themselves (Treisman, 2007, p. 214). It is hoped that managers will be more forthcoming when answering indirect questions. The results from this study suggest indirect questions are, at best, only partly successful. For two of the three indirect questions, reticent managers were less likely to report problems associated with corruption. If, as previous studies have suggested, reticent managers are less likely to report corruption because they answer strategically, this suggests indirect questions do not always encourage reticent managers to be candid.

Third, we control for the possibility that reticence is endogenous. We do this for two reasons. First, managers who face demands for bribes might become reticent, leading to reverse causation. Second, if a manager evades taxes or abuses his or her position, this will affect the paper's measure of reticence. If the same managers are also more likely to pay bribes, this could affect the correlation between the measures of reticence and corruption. Because of these concerns, we use instrumental variable (IV) estimation. We discuss the identification strategy below.

DATA

This study uses data from two World Bank management practice surveys. The questionnaire was based on the World Bank's Enterprise Survey questionnaire but was heavily modified to collect information on strategic behavior and industrial structure. It included the questions that Azfar and Murrell (2009) used to identify reticent respondents. Researchers in microeconomics, management, and psychology helped with questionnaire design. (1)

The World Bank conducted the survey of 1,001 firms in Bangladesh between May and June 2011 and the survey of 588 firms in Sri Lanka between June and November 2011. The sample included manufacturing; construction; services; transportation, storage and communications; and information technology. (2) World Bank (2012b) describes the sampling methodology. (3)

Because corruption is high in the two countries, these countries provide viable settings to study corruption. In 2013, Bangladesh ranked 136th out of 175 countries on Transparency International's Corruption Perceptions Index with a score of 27. (4) Although Sri Lanka ranked a little higher, 91st with a score of 37, it also performed poorly.

Corruption data

Because collecting information on corruption or related topics such as regulation was not the main goal of the survey, the survey includes only four questions on corruption. Three of these are identical to questions from the World Bank's Enterprise Surveys.

The first two questions are linked. Managers were first asked whether tax officials had visited or inspected their firm in the past year. Managers who said inspectors had visited their firms were then asked:

'J5. In any of these inspections or meetings with tax officials was a gift or informal payment expected or requested'.

[Tax inspections--inspected]

We refer to this question as 'tax inspections--inspected'. We include 'inspected' in the description to distinguish this question from a second question, which we refer to as 'tax inspections--not inspected'. The second question was only asked to managers who said tax officials had not inspected or visited their firm.

'J6. During inspections or meetings with tax officials, are gifts or payments typically expected or requested'.

[Tax inspections--not inspected]

Firm were only asked one question on tax inspections: the first was asked to managers of inspected firms and the second was asked to managers of non-inspected firms.

The third question on corruption is:

'37. It is said that establishments are sometimes required to make gifts or informal payments to public officials to "get things done" with regard to customs, taxes, licenses, regulations, services, etc. On average, how much do establishments like this one pay in informal payments or gifts to public officials for this purpose'.

[Any Bribes]

The manager could answer the question either as a percent of sales or in local currency. Because earlier studies have shown the way the manager answers the question affects the size of the reported bribe, we focus on whether the manager reported any bribes rather than on the amount of the bribe. (5) We refer to this question as 'any bribes' in Table 1 and the text.

The final question is:

'J30F. As I list some factors that can affect the current operations of a business, please look at this card and tell me if you think that each factor is No Obstacle, a Minor Obstacle, a Moderate Obstacle, a Major Obstacle, or a Very Severe Obstacle to the current operations of this establishment'.

[Perceptions]

The question then lists 'corruption' as one of five obstacles. (6) We refer to this question as 'perceptions' in Table 1 and the text.

The four questions differ in several important ways. One way they differ is whether they are asked directly or indirectly. Direct questions ask about the manager's own actions or experiences with corruption. In contrast, indirect questions ask about what the manager believes other firms do or about how common corruption is in general.

Surveys often use indirect questions about corruption so managers can answer questions without incriminating themselves. The World Bank's Enterprise Surveys, for example, ask managers whether 'firms like this one' pay bribes rather than asking the manager directly if he or she pays bribes. Similarly, Johnson et al. (2000) used a survey that asked about bribe payments by 'firms in your sector' and Svensson (2003) used a survey that asked how much a 'firm in your line of business and of similar size and characteristics' pays in bribes. Because indirect questions reduce underreporting of other sensitive behaviors (Dillman et al., 2008), they might do the same for questions about corruption.

Although indirect questions might reduce underreporting, they impose other costs. Fisher and Tellis (1998) find that indirect questions introduce variance unrelated to behavior and attitude because some respondents try to predict the behavior of the indirect question's subject. That is, some managers try to imagine whether other 'firms like this one' pay bribes--something they know less well than whether they pay bribes. The resulting measurement error will complicate econometric analysis. In addition, if managers answer questions based on what they think others do, it is difficult to link firm and manager characteristics with bribes.

Given the problems with indirect questions, it is important to know whether they reduce underreporting. If indirect questions encourage candor, we might expect reticent managers to be more candid when they answer indirect questions than when they answer direct questions. The evidence on this, however, is mixed. Azfar and Murrell (2009) find reticent managers in Romania are less likely than non-reticent managers to report bribes when registering and during health inspections whether questions are asked directly about their firm or indirectly about other firms. In contrast, Clausen et al. (2010) find reticent managers in Nigeria answered direct questions, but not indirect questions, differently than non-reticent managers.

A second way the questions differ is whether the questions are general or specific. Questions are specific if they ask about corruption during specific transactions or in specific agencies. In contrast, questions are 'general' if they ask about corruption without referring to a specific transaction or agency.

The first question ('tax inspections--inspected'), which was only asked to managers who said tax officials visited or inspected their firm, is direct and specific. Although the question does not ask whether bribes were paid, it asks whether bribes were requested or expected during the inspection. Because it asks about what happened to the manager's own firm, rather than about what happens to a typical firm, it is direct. It is specific because it asks about transactions with a specific agency, the tax authorities.

The second question ('tax inspections--not inspected), which was only asked to managers who said tax officials had not visited their firms, is indirect and specific. (7) The question asks whether tax officials typically request bribes during inspections. Because it does not directly ask about the manager's own experiences, it is indirect. (8) It is, however, specific because it asks about transactions with a specific agency.

The third question ('any bribes') is indirect and general. Because it asks what the manager thinks 'firms like this one' do rather than about what the manager's firm does, it is indirect. (9) It is general because it asks about many different transactions with many different agencies including getting licenses, getting government services, dealing with customs, and dealing with inspections.

The final question ('perceptions') is general; it does not specify an agency or transaction. Although it is harder to classify as direct or indirect, it appears to be indirect. First, it asks about how corruption affects 'the current operations of a business' not the current operations of the manager's business. Second, it does not ask specifically about the manager's experiences with bribes--it simply asks whether corruption is a problem. Even if the manager interpreted the question as if it referred to the manager's firm, the manager could report corruption is a serious problem even if he or she had never paid a bribe. A manager who lost a contract after refusing to pay a bribe, for example, might say corruption is a problem even if he or she did not pay a bribe. (10) Table 1 shows how we classify each of the four questions.

Managers can respond to questions on corruption in several ways: they can answer truthfully, they can lie, or they can refuse to answer. (11) Unfortunately, it is difficult to know how many respondents lie. Previous studies that have looked at how reticent managers answer questions on corruption have concluded reticent respondents sometimes lie about corruption (Azfar and Murrell, 2009; Clausen et al., 2010; Jensen and Rahman, 2011). It is likely, however, that some non-reticent respondents also lie.

Rather than answering truthfully or lying, many managers avoid the question. Because bribery is often illegal and is usually frowned on, most people assume managers who have paid bribes will be more likely to refuse to answer questions on corruption than managers who have not paid any bribes. Consistent with this, the World Bank treats refusals as yes responses when it calculates the percent of firms that pay bribes in a country (World Bank, 2012a). Similarly, Treisman (2007, p. 239) argues 'it might be respondents in some countries are more reluctant to admit paying bribes and so would reply "don't know" or "no answer" rather than "yes" to survey questions about [corruption]'.

Managers can avoid answering the question by refusing to answer or by saying they don't know. Both responses were common in the survey used in this study (see Table 2). Non-response rates varied from about 2.5% on the question on whether corruption was a serious problem (perceptions) to 33% on the question on how much firms need to pay to get things done (any bribes).

The non-response rate was higher on the indirect question on bribe payments during tax inspections (tax inspections--not inspected) than it was on the direct question (tax inspections--inspected). The non-response rate for managers of non-inspected firms was about 19% compared with about 9% for managers of inspected firms. This might be because managers whose firms had not been inspected did not know whether tax inspectors typically request bribes or not. Even when we exclude managers who said they did not know, however, managers whose firms had not been inspected were still more likely to refuse to answer the question than were managers of inspected firms.

Managers who refused to answer one question about corruption were often unwilling to answer other questions. Table 3 shows the correlation between dummy variables indicating whether the manager refused to answer each question. The positive and often statistically significant correlations show managers who refused to answer one question on corruption were also more likely to refuse to answer other questions.

Identifying reticent managers

We identify reticent managers--managers who are reluctant to answer questions candidly--using an approach developed by Azfar and Murrell (2009). (12) The approach uses the forced response version of random response questions (Warner, 1965). Although, as described below, random response questions were originally used to get people to answer sensitive questions truthfully, Azfar and Murrell (2009) propose using them instead to identify reticent managers.

In the forced response variant of the random response methodology, an interviewer asks a manager a sensitive question unrelated to corruption and asks them to toss a coin out of the interviewer's sight. If the coin shows heads, the manager answers 'yes'. If it shows tails, the manager answers the question. Forced response procedures are intended to encourage people to answer sensitive questions honestly. (13) If the manager answers 'yes' no one other than the manager, not even the interviewer, knows whether the manager is saying they committed the sensitive act or just that the coin showed heads.

Although forced response questions are meant to reduce underreporting of sensitive behaviors, they introduce a new bias. (14) The problem is some people ignore the instructions for the forced response questions and answer no even when the coin shows heads. (15) Azfar and Murrell (2009) note that less than half of respondents should answer no to the sensitive questions; even if no one has committed the sensitive act, the coin will show heads half the time. Azfar and Murrell (2009) and Clausen et al. (2010), however, find an implausible number of 'no' answers to the forced response questions. Following Azfar and Murrell (2009), we identify reticent respondents by looking at how they answer the forced response questions. It is important to note that the questions on corruption are not forced response questions, and the sensitive questions we use to identify reticent respondents are unrelated to corruption (see Table 4).

The list of forced response questions includes seven sensitive questions and three less sensitive questions. Azfar and Murrell (2009) include the three less sensitive questions so sophisticated respondents who realize long sequences of 'no's' are unlikely can sometimes answer yes. Following Azfar and Murrell (2009), we exclude the less sensitive questions when calculating the reticence index. Excluding them is reasonable because Edgell et al. (1982) show forced response questions work well with moderately sensitive questions; most people who are meant to say 'yes' to moderately sensitive forced response questions do so. As noted in the robustness checks, however, our results are similar if we include the three less sensitive questions when measuring reticence.

Table 5 shows the expected distribution of 'no' responses if no one had done any of the sensitive behaviors, the expected distribution if 30% of respondents had done each behavior, and the true distribution of 'no' answers. (16) If nobody had done any of the sensitive acts--that is, if everyone is an angel--we would expect the average respondent to answer no 3.5 times. The angel assumption, however, is unlikely to be true. If about 30% of respondents had done each act--and answered the questions truthfully that they had we would expect the average respondent to answer no 2.5 times. In fact, no responses were far more common than this--the average manager responded no about 4.2 times. This suggests many managers answer no even when the coin shows heads.

The high number of no's is not just because some managers always say no. Even if all managers are angels--that is, none have done any of the sensitive acts--too many answered 'no' five, six and seven times and too few answered 'no' zero, one, two or three times than we would expect if they followed the instructions for the forced response questions. If 30% of people have done each sensitive behavior, the distribution is even more skewed. Too many people responded 'no' four, five, six and seven times and too few people responded 'no' zero, one, two or three times. The high number of no's suggests reticence is not all-or-nothing; some people ignore the instructions to the forced response questions only some of the time. We, therefore, use the number of 'no' responses to measure reticence rather than using a simple measure that labels managers as reticent if they answer no seven times. Managers who answer 'no' more often are more reticent than managers who answer 'no' less often.

EMPIRICAL ESTIMATION

This section presents the empirical models that we estimate and the empirical results. We first estimate a model predicting whether the manager responds yes or no and then a second model predicting whether the managers responds to the question.

Estimating whether firms respond 'yes' or 'no'

Previous studies have looked at how reticent and non-reticent managers answer questions on corruption (Azfar and Murrell, 2009; Clausen et al., 2010). We, therefore, first look at whether reticent respondents who answer the questions on corruption are more or less likely to report bribes or that corruption is a serious problem than are non-reticent respondents. We assume that each manager's propensity to answer 'yes' depends on how reticent they are and various firm characteristics (FC). In all cases, a 'yes' implies corruption is more frequent or a greater problem.

Manager i's propensity to respond 'yes' is: Propensity to answer yes to corruption [question.sub.i] = [alpha] + [beta] reticence + [gamma] [FC.sub.i] + [[epsilon].sub.i]

We assume the error term, [[epsilon].sub.i], is normally distributed. We also assume managers will respond 'yes' if their propensity to answer 'yes' is greater than zero. Positive coefficients, therefore, mean managers with those characteristics are more likely to respond 'yes'.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

Because we assume c, has a normal distribution, the model is a Probit model. As a robustness check, we also estimate the model as a linear probability model. As discussed below, the linear probability model imposes less strict requirements on instruments than the probit model (Lewbel et al., 2012).

We are most interested in the coefficient on the measure of reticence. As discussed above, we use the number of no answers to the seven sensitive forced response questions to measure reticence. The results are similar, however, when we use all 10 forced response questions rather than just the seven sensitive questions. Because reticent managers are more likely to answer 'no' to the forced response questions, higher values mean greater reticence. A positive coefficient on the reticence variable would, therefore, imply reticent managers are more likely to say corruption is a problem than are less reticent managers. A negative coefficient would imply the reverse.

The control variables include other firm characteristics that might affect how the manager answers questions about corruption: that is, things that might affect either whether the firm has paid bribes or whether the manager refuses to answer or lies. These include characteristics of the firm, characteristics of the top manager, and characteristics of the respondent who is often, but not always, the top manager. The variables include the respondent's gender and age, whether the respondent is the firm's owner, the top manager's educational attainment, the firm's size and age, whether the firm exports, whether the firm is in Sri Lanka or Bangladesh, and the firm's sector of operations. (17)

Controlling for firm and manager characteristics is important if characteristics that affect whether the manager is reticent also affect how the manager answers questions about corruption. Managers of large firms, for example, might be more concerned about being identified based on the information in the survey than are managers of small firms. This might make them less willing to answer sensitive questions candidly than managers of small firms. But large firms also have more contact with regulators than small firms, meaning that they might be more likely to pay bribes than small firms. (18) Omitting relevant variables such as firm size could, therefore, result in a spurious correlation between reticence and responses to questions about corruption.

Even after controlling for observable firm and manager characteristics, there are two additional problems with the reticence variable. The first is the reticence variable does not only capture reticence. Managers who have committed fewer of the sensitive acts asked about in the forced response questions will answer 'no' to the forced response questions more frequently than other managers who have committed more of the sensitive acts. (19) The reticence measure is, therefore, also a measure of guilt. This could be a problem if managers who have committed more sensitive acts are more likely to pay bribes than managers who have committed fewer sensitive acts.

The second concern about the reticence variable is reverse causation. Firm managers who have had problems with corrupt officials might become reticent. They might, for example, become unwilling to admit to the sensitive acts that the forced response questions ask about if they worry doing so will give corrupt officials leverage over them. Similarly, unobserved characteristics of the firm or manager, such as how honest the manager is, might affect whether the firm pays bribes and whether the manager is reticent.

Because of these concerns, we use IV techniques throughout the analysis to control for the potential for endogeneity. To do this, we need an instrument for reticence--something that affects reticence but does not affect whether the manager has done any of the sensitive acts and does not affect whether the manager has paid bribes.

Because many things that might affect whether the respondent is reticent, such as the manager's gender or educational attainment, might also affect the manager's propensity to pay bribes, it is difficult to use firm or manager characteristics as instruments for reticence. We, therefore, adopt a different approach using an instrument based on the quality of the interaction between the manager and the interviewer. If the manager and the interviewer interact positively, the manager might become less reticent. Iarossi (2006, p. 157), for example, notes 'respondents are more willing to comply with requests from people who are similar to them, people who praise them, people who are familiar to them, and people with whom they like to be associated'. How well the manager and the interviewer interact depends not just on the manager; it also depends on the interviewer. Although we don't have any socio-demographic information on the interviewers, we can identify the firms interviewed by each interviewer. We can therefore make interviewer-level measures of reticence: the average reticence of other managers interviewed by the same interviewer. If successful interviewers have characteristics that appeal to many managers, interviewers who get truthful answers from other managers might be more successful with the next manager.

To see whether this is the case, we construct a variable indicating the average reticence of other managers interviewed by the same interviewer. When we do this, we omit the managers' own responses; that is, it is a 'leave-one-out' average. Because the leave-one-out average does not include the manager's own responses, it should only capture characteristics of the interviewer and not characteristics of the manager. As a result, it should be uncorrelated with the manager's decision about whether to pay bribes or not. In addition, it should not be affected by whether the manager has done any of the sensitive acts asked about in the forced response questions. As such, it should reduce both concerns about the reticence variable.

If some interviewers are better than others, we would expect the instrument to be positively correlated with the manager's reticence. The results are consistent with this. In the first-stage regression, the coefficient on the leave-one-out average is positive and statistically significant (F-stat = 112.3, p-value = 0.00). The first-stage regression results are shown in Appendix Tables A5 and A6 and are discussed in the robustness checks. Moreover, as noted in the robustness checks, the empirical results suggest that the reticence variable is endogenous.

Empirical results

Reticence

The coefficients on the indicator of reticence, the number of 'no' responses to the forced response questions, are negative in all four IV probit regressions and statistically significant in three of the four (see Table 6). The negative coefficients indicate reticent managers were less likely to say corruption is a serious problem, were less likely to say bribes were requested or expected during tax inspections if tax officials inspected their firm, and were less likely to say bribes were typically requested during tax inspections if tax officials did not inspect their firm. The negative but statistically insignificant coefficient in the final regression indicates reticent managers were also less likely to say bribes were needed to get things done.

The results suggest that reticent managers are less likely to report corruption in response to both general and specific questions about corruption and to both direct and indirect questions. The coefficients on the reticence variable are statistically significant in the direct question ('tax inspections inspected') and in two of the three indirect questions ('tax inspections--not inspected' and 'perceptions'). This suggests that indirect questions are not always effective in encouraging candor. Similarly, the coefficients are statistically significant in both specific questions and in one of the two general questions.

If reticent managers are lying about corruption--and this is how previous papers (Azfar and Murrell, 2009; Clausen et at, 2010) have interpreted similar results--then estimates of corruption that do not take reticence into account will underestimate the prevalence of corruption. How much reticence affects the estimates of corruption depends on how reticent managers are. But this depends partly on how likely it is that managers have done the sensitive acts. That is, the coefficients tell us how much an additional 'no' response to the sensitive randomized questions affects the likelihood that managers will answer 'yes' to the questions on corruption. The impact of reticence will therefore depend on how many 'no' responses we would expect if all managers answered the forced response questions honestly. This, in turn, depends on how many managers have done each sensitive act asked about in the forced response questions.

Table 7 shows the percent of managers who answer each question about corruption and different estimates of how many managers would answer the questions of corruption if reticence was not a problem. About 41% of managers in the sample said corruption was a major or very severe problem, about 10% said bribes were needed to get things done, about 11 % of managers whose firms had been inspected said bribes were requested or expected during the inspections and about 2% of managers whose firms had not been inspected said bribes were typically needed during inspections. If no one had done any of the sensitive acts, the average manager should answer 'no' to 3.5 of the 7 sensitive forced response questions. To see how reticence affects responses, we replace the managers' actual number of 'no' responses with 3.5, calculate the probability that each manager would answer each corruption question with a 'yes,' and then average the probabilities over all managers. When we do this, we calculate that about 50% of managers would say corruption was a serious problem. Similarly, if reticence was not affecting responses to the other questions, we calculate that about 11% would say that bribes were needed to get things done, that about 22% of managers whose firms had been inspected would say that bribes were requested or expected during tax inspections and that about 14% of managers whose firms had not been inspected would say that bribes are typically needed during tax inspections.

If some managers have done the sensitive acts, then we are underestimating the extent of reticence in the sample. If about 30% of managers have done each act, then we would expect that the average manager would respond 'no' to only 2.5 of the sensitive questions. In this case, we would expect that if no managers were reticent that about 65 % of managers would say that corruption was a serious problem, that about 12% would say that bribes were needed to get things done, that about 35% of managers whose firms had been inspected would say that bribes were requested or expected during tax inspections and that about 22% of managers whose firms had not been inspected would say that bribes are typically needed during tax inspections.

Other variables

There are few consistent results for the coefficients on the other control variables. The most consistent result is older respondents were more likely to complain about corruption than younger respondents. Older respondents were more likely to say corruption was a serious problem, more likely to say bribes were needed to get things done, and more likely to say bribes were expected during tax inspections. In the final regression, the coefficient on respondent's age was positive but was not statistically significant.

There is some evidence that better educated managers were less concerned about corruption than other managers and managers of larger firms and exporters were more concerned about corruption than other managers. Although the coefficients on these variables had consistent signs, they were mostly not statistically significant.

Model estimating response rates

The second question we ask is whether reticent respondents are more or less likely to refuse to respond to questions about corruption than less reticent respondents. We assume each manager's propensity to respond to questions about corruption depends on how reticent they are and various firm characteristics. Manager i's propensity to respond is therefore:

Propensity to refuse to [respond.sub.i] = [alpha] + [beta] [reticence.sub.i] + [gamma][FC.sub.i] + [[epsilon].sub.i],

We assume the error term, [[epsilon].sub.i], is normally distributed. We assume the manager will not respond if his or her propensity to refuse to answer the question is greater than zero

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

The dependent variable is, therefore, a dummy variable. Because e, has a normal distribution, we estimate the model as a Probit model.

The main variable of interest is the measure of reticence. We use the number of no answers to the seven sensitive forced response questions. Firm managers who more frequently respond 'no' to the forced response questions are more reticent than managers who respond 'no' less frequently. The variable, therefore, increases as the manager becomes more reticent. If reticent managers are more likely to refuse to respond than less reticent managers, the coefficient on the reticence variable would be positive. The reverse would be true if reticent managers were less likely to respond than less reticent managers. We control for endogeneity in the same way as in the previous section.

Empirical results

Reticence

The coefficient on the variable indicating reticence is statistically significant and negative in all four IV probit regressions (see Table 8). Because the dependent variable indicates the manager refused to answer the question and the reticence variable is larger when the manager is more reticent, the negative coefficient indicates managers who are more reticent are less likely to refuse to answer the questions on corruption than managers who are less reticent.

The coefficient is negative and statistically significant in all three regression with indirect dependent variables and in the only regression with a direct dependent variable. This suggests reticent respondents are less likely to refuse to respond to questions on corruption even when the question is asked indirectly. The coefficients are also significant in all general and specific questions.

Reticence, once again, has a large impact on the results (see Table 9). Even assuming no one has done any of the sensitive acts the forced response questions ask about, the estimates suggest that in the absence of reticence about 5% of managers would refuse to answer the question on whether corruption is a serious problem, compared with about 3 % of managers in the actual sample, and about 39% would refuse to answer the question on whether bribes are needed to get things done, compared with about 33 % in the sample. For firms that had been inspected by the tax authorities, the estimates suggest about 18% of managers would refuse to answer the question on whether bribes were requested during inspections if no managers were reticent compared with about 10% in the sample. Finally, for firms that had not been inspected, about 25 % would refuse to answer if none were reticent compared with about 19% in the sample. If any managers had done any of the sensitive acts, an even greater share of managers would refuse to answer if no managers were reticent.

Other variables

As with the previous questions, there were few other consistent patterns among the control variables. Firms in Sri Lanka were more likely to refuse to say whether corruption is a problem but were less likely to refuse to answer the other questions. Managers of large firms were less willing to answer the question on whether corruption was a serious problem but more likely to refuse to answer the other questions.

Robustness checks

Endogeneity

Throughout the analysis, we treat the reticence variable as if it were endogenous. This is appropriate given that Smith-Blundell exogeneity tests (Smith and Blundell, 1986) reject the null hypothesis of exogeneity at a 10% significance level or higher for seven of the eight regressions (see Table 6 and Table 8). Wu-Hausman tests give similar results for the linear probability model. (20)

Because we have only a single instrument, we cannot test over-identifying restrictions. We can, however, check instrument strength. This is useful because weak instruments can result in biased coefficients. As noted above, the leave-one-out average is highly correlated with the reticence variable in all regressions (see Tables A5 and A6 in the Appendix). Formal tests for instrument strength (Bound et al, 1995) in the linear probability model lead to similar conclusions (see Tables A3 and A4).

10 sensitive questions

As noted above, the forced response questions included three less sensitive questions. Azfar and Murrell (2009) included the three additional questions to allow sophisticated respondents to sometimes answer 'yes' if they understood tossing heads repeatedly was highly unlikely. Because forced response questions work well with moderately sensitive questions (Fisher and Tellis, 1998), this seems reasonable.

As a robustness check, we re-ran the regressions including all 10 questions when calculating the reticence measure. The coefficients on reticence are similar in size and statistical significance to the coefficients using only seven questions. The results suggest reticent respondents remain less likely to say corruption is a serious problem and bribes were either requested or typically needed during tax inspections (see Table A1 in Appendix). Reticent respondents were also less likely to refuse to answer all corruption questions (see Table A2 in Appendix).

Linear probability model

As a robustness check, we re-run the main regressions as a linear probit model using 2SLS to allow for endogeneity. Although, as discussed in Lewbel et al. (2012), the linear probability model has several problems, it puts fewer restrictions on the instruments than the probit model and is robust to heteroskedasticity. (21) In practice, the results from the linear probability model are similar to the results from the probit model; the coefficients on the reticence variable remain statistically significant and negative in the same regressions as in the Probit model.

CONCLUSION

As in previous studies (Azfar and Murrell, 2009; Clausen et al., 2010; Jensen and Rahman, 2011), but using a new data set, we find reticence affects estimates of corruption. Reticent managers were less likely to say corruption is a serious problem than non-reticent managers. Reticent managers of firms that tax inspectors visited were less likely to say the inspectors sought bribes than were non-reticent managers. And reticent managers of firms that tax inspectors did not visit were less likely to say bribes were typically needed during tax inspections than were non-reticent managers. Although the coefficient in the final regression was statistically insignificant, the point

estimate suggests reticent managers were also less likely to say bribes are needed than were non-reticent managers.

If reticent managers answer questions on corruption strategically, as most previous studies argue, raw survey data will understate how often firms pay bribes. We estimate that if no managers were reticent at least 22% of inspected firms would have reported inspectors sought bribes during tax inspections compared with only 11% in the sample (see Table 7). Similarly, we estimate at least 50% of managers would have said corruption was a serious problem if no managers were reticent compared with only 41 % in the sample. These estimates are lower bounds; if many people have done the sensitive acts the forced response questions ask about, reticence will affect the estimates even more.

The results also suggest asking questions indirectly does not always encourage reticent managers to answer questions on corruption candidly. The survey included three indirect and one direct question about corruption. For all four questions, including the three indirect questions, reticent managers were less likely to complain about corruption than non-reticent managers. For three of the four questions, including two of the three indirect questions, the differences were statistically significant. This is discouraging in that it suggests indirect questions do not consistently encourage greater candor from reticent managers.

Although reticent managers answer direct and some indirect questions differently than non-reticent managers, indirect questions have other problems. First, researchers find it difficult to interpret managers' answers to indirect questions; researchers do not know whether managers are referring to their experiences or to the managers' assumptions about other firms' experiences. Second, indirect questions can introduce measurement error if managers are not sure about other firms' experiences. This suggests more work is needed to assess whether indirect questions' benefits outweigh their costs.

We also find reticent managers in South Asia were less likely to refuse to answer questions than non-reticent managers, perhaps because they believe refusing to answer sounds like an admission of guilt. Reticent managers might, therefore, prefer to answer 'no' rather than refuse to answer to look less guilty. We estimate non-response rates would be between 2 and 8 percentage points higher if no managers were reticent and if no managers had done any of the sensitive acts asked about by the forced response questions. If some managers had done the sensitive acts, the non-response rate without reticence would be even higher.

Previous studies have found the opposite; they find reticent managers are more, not less, likely to refuse to answer questions on corruption than were non-reticent managers (Azfar and Murrell, 2009; Jensen and Rahman, 2011). It is not clear why our results are different from results in earlier studies. It might be because the cultural context is different in different countries. It might, however, be because of questionnaire or survey design. In contrast to Azfar and Murrell's (2009) survey, corruption was not the focus of our survey; the questionnaire only included four questions on bribes. It is also possible that different refusal rates in the different surveys cause the different results. More surveys, in different countries, are needed to resolve this issue.

As well as understanding why our results on refusal rates are different from earlier studies, it would also be useful to understand why reticent respondents answer questions differently. Although our results give more evidence on how reticent managers answer questions about corruption, we do not try to explain reticence psychologically.

APPENDIX

Additional tables

Table A1: IV Probit regressions for corruption
questions for all 10 forced response questions

                        Corruption is    Bribes are needed
                       serious problem   to get things done
                           [dummy]            [dummy]
Observation                 1,082               728

Reticence
  Number of no           -0.367 ***       -0.096
responses(10
questions)
  [high numbers        (-10.47)          (-0.99)
mean more likely to
be reticent]
Respondent
  Respondent is          -0.002           -0.576 ***
owner
  [dummy]               (-0.02)          (-2.70)
  Respondent is          -0.178            0.475 *
female
  [dummy]               (-0.84)           (1.72)
  Respondent's age        0.125 *          0.185
  [natural log]          (1.74)           (1.64)
Manager education
  Manager has            -0.236           -1.314 ***
secondary
education
  [dummy]               (-1.10)          (-4.28)
  Manager has            -0.318           -1.550 ***
vocational
education
  [dummy]               (-1.31)          (-3.99)
  Manager has            -0.133           -1.300 ***
tertiary education
  [dummy]               (-0.63)          (-4.23)
Firm characteristics
  Number of               0.053            0.030
workers
  [natural log]          (1.36)           (0.46)
  Age of firm            -0.142 **         0.168
  [natural log]         (-2.15)           (1.43)
  Firm exports            0.108            0.573 ***
  [dummy]                (1.00)           (3.30)
  Firm is in Sri         -1.039 ***       -0.265
Lanka
  [dummy]               (-6.88)          (-1.46)
  Retail firm             0.135           -0.243

  [dummy]                (0.60)          (-0.67)
  Manufacturing          -0.041           -0.223
firm
  [dummy]               (-0.42)          (-1.24)
Constant                  2.245 ***       -0.189
                         (6.19)          (-0.23)

                        Bribe was requested     Bribes typically
                       during tax inspection   needed during tax
                              [dummy]          inspections [dummy
Observation                     407                   336

Reticence
  Number of no          -0.317 ***              -0.372 ***
responses(10
questions)
  [high numbers        (-3.61)                 (-2.73)
mean more likely to
be reticent]
Respondent
  Respondent is          0.258                  -0.407
owner
  [dummy]               (1.13)                 (-1.46)
  Respondent is          0.125
female
  [dummy]               (0.31)
  Respondent's age       0.460 **               -0.017
  [natural log]         (2.46)                 (-0.07)
Manager education
  Manager has           -0.340                  -0.230
secondary
education
  [dummy]              (-0.66)                 (-0.52)
  Manager has           -0.314
vocational
education
  [dummy]              (-0.56)
  Manager has           -0.240                  -0.890
tertiary education
  [dummy]              (-0.45)                 (-1.56)
Firm characteristics
  Number of              0.285 ***               0.026
workers
  [natural log]         (3.27)                  (0.21)
  Age of firm           -0.494 ***               0.014
  [natural log]        (-3.16)                  (0.06)
  Firm exports           0.306                   0.755 **
  [dummy]               (1.48)                  (2.49)
  Firm is in Sri        -0.129
Lanka
  [dummy]              (-0.61)
  Retail firm           -0.035

  [dummy]              (-0.08)
  Manufacturing          0.225                  -0.484
firm
  [dummy]               (1.06)                 (-1.55)
Constant                 0.203                   1.562
                        (0.17)                  (1.22)

Note: T-statistics in parentheses.

Source: Authors' calculation based on data from the World Bank's
Enterprise Survey Bangladesh and Sri Lanka

* Statistically significant at 1%, 5% and 10% significance levels.

Table A2: IV Probit regressions for refusing to
answer for all 10 forced response questions

                       Corruption is    Bribes are needed
                      serious problem   to get things done
                          [dummy]            [dummy]
Observation                 997               1,092

Reticence
  Number of no         -0.340 ***        -0.192 ***
responses
  [high numbers       (-3.24)           (-3.38)
mean more likely to
be reticent]
Respondent
  Respondent is        -0.477 *          -0.162
owner
  [dummy]             (-1.65)           (-1.45)
  Respondent is         0.660 *           0.088
female
  [dummy]              (1.86)            (0.42)
  Respondent's age     -0.050            -0.165 **
  [natural log]       (-0.27)           (-2.28)
Manager education
  Manager has          -0.425            -0.316
secondary
education
  [dummy]             (-0.99)           (-1.43)
  Manager has                            -0.398
vocational
education
  [dummy]                               (-1.56)
                       -0.447            -0.321
  Manager has
tertiary education
  [dummy]             (-0.93)           (-1.46)
Firm characteristics
  Number of            -0.253             0.073 *
workers
  [natural log]       (-1.60)            (1.79)
  Age of firm           0.040             0.030
  [natural log]        (0.20)            (0.43)
  Firm exports          0.087            -0.091
  [dummy]              (0.23)           (-0.78)
  Firm is in Sri        0.791 **         -0.808 ***
Lanka
  [dummy]              (2.50)           (-6.31)
  Retail firm           0.086            -0.168
  [dummy]              (0.17)           (-0.69)
  Manufacturing         0.210            -0.138
firm
  [dummy]              (0.55)           (-1.31)
Constant                0.636             1.347 ***
                       (0.52)            (2.89)

                       Bribe was requested     Bribes typically
                      during tax inspection    needed during tax
                             [dummy]          inspections [dummy]
Observation                    460                    610

Reticence
  Number of no         -0.200 *                -0.211 **
responses
  [high numbers       (-1.94)                 (-2.46)
mean more likely to
be reticent]
Respondent
  Respondent is         0.110                  -0.344 **
owner
  [dummy]              (0.45)                 (-2.16)
  Respondent is        -0.239                   0.342
female
  [dummy]             (-0.43)                  (1.17)
  Respondent's age     -0.161                  -0.228 **
  [natural log]       (-1.09)                 (-2.11)
Manager education
  Manager has          -0.310                  -0.100
secondary
education
  [dummy]             (-0.51)                 (-0.37)
  Manager has          -0.594                   0.117
vocational
education
  [dummy]             (-0.92)                  (0.33)
                       -0.195                  -0.402
  Manager has
tertiary education
  [dummy]             (-0.32)                 (-1.51)
Firm characteristics
  Number of             0.207 ***              -0.011
workers
  [natural log]        (2.67)                 (-0.19)
  Age of firm          -0.002                   0.276 **
  [natural log]       (-0.02)                  (2.51)
  Firm exports         -0.107                  -0.026
  [dummy]             (-0.51)                 (-0.14)
  Firm is in Sri       -1.021 ***              -1.278 ***
Lanka
  [dummy]             (-3.37)                 (-4.56)
  Retail firm           0.514
  [dummy]              (1.18)
  Manufacturing        -0.005                  -0.357 **
firm
  [dummy]             (-0.03)                 (-2.39)
Constant                0.205                   0.949
                       (0.19)                  (1.38)

Note: T-statistics in parentheses.

Source: Authors' calculation based on data from the
World Bank's Enterprise Survey Bangladesh and Sri Lanka

***, **, * Statistically significant at 1%, 5% and 10%
significance levels.

Table A3: Linear probability estimation for corruption questions

                              Corruption is     Bribes are
                                 serious          needed
                             problem [dummy]   to get things
                                               done [dummy]

Observation                       1,092             736

Reticence
  Number of no responses       -0.143 ***       -0.017
  [high numbers mean more     (-6.07)          (-1.01)
likely to be reticent]
Respondent
  Respondent is owner           0.048           -0.063 ***
  [dummy]                      (1.28)          (-2.66)
  Respondent is female         -0.043            0.113 *
  [dummy]                     (-0.74)           (1.80)
  Respondent's age              0.060 **         0.026
  [natural log]                (2.33)           (1.23)
Manager education
  Manager has secondary        -0.045           -0.312 ***
education
  [dummy]                     (-0.57)          (-3.07)
  Manager has vocational       -0.096           -0.351 ***
education
  [dummy]                     (-1.04)          (-3.40)
  Manager has tertiary          0.025           -0.314 ***
education
  [dummy]                      (0.31)          (-3.06)
Firm characteristics
  Number of workers             0.028 **         0.003
  [natural loql                (1.98)           (0.29)
  Age of firm                  -0.049 *          0.019
  [natural log]               (-1.87)           (0.89)
  Firm exports                  0.012            0.113 ***
  [dummy]                      (0.27)           (2.99)
  Firm is in Sri Lanka         -0.395 ***       -0.042
  [dummy]                    (-10.93)          (-1.51)
  Retail firm                   0.022           -0.030
  [dummy]                     (0.36)           (-0.69)
  Manufacturing firm          -0.027            -0.035
  [dummy]                    (-0.68)           (-1.10)
Constant                       1.055 ***         0.406 ***
                               (7.43)           (2.98)
F-test for significance of    106.6            108.9
instruments in first-stage
p-value                        (0.00)           (0.00)

                                 Bribe was        Bribes typically
                                 requested         needed during
                                 during tax       tax inspections
                             inspection [dummy]       [dummy]

Observation                         417                 498

Reticence
  Number of no responses      -0.063 **            -0.014 *
  [high numbers mean more    (-2.13)              (-1.91)
likely to be reticent]
Respondent
  Respondent is owner          0.063               -0.005
  [dummy]                     (1.57)              (-0.24)
  Respondent is female         0.069               -0.008
  [dummy]                     (0.96)              (-0.87)
  Respondent's age             0.074 ***           -0.001
  [natural log]               (3.03)              (-0.10)
Manager education
  Manager has secondary       -0.035               -0.056
education
  [dummy]                    (-0.40)              (-0.86)
  Manager has vocational      -0.018               -0.080
education
  [dummy]                    (-0.19)              (-1.26)
  Manager has tertiary         0.005               -0.087
education
  [dummy]                     (0.06)              (-1.32)
Firm characteristics
  Number of workers            0.054 ***            0.004
  [natural loql               (3.09)               (0.59)
  Age of firm                 -0.074 ***            0.005
  [natural log]              (-3.14)               (0.42)
  Firm exports                 0.058                0.050
  [dummy]                     (1.38)               (1.62)
  Firm is in Sri Lanka        -0.012               -0.032 **
  [dummy]                    (-0.33)              (-2.35)
  Retail firm                 -0.001               -0.016
  [dummy]                    (-0.02)              (-0.92)
  Manufacturing firm           0.037               -0.023
  [dummy]                     (0.85)              (-1.23)
Constant                       0.214                0.149 **
                              (1.27)               (2.11)
F-test for significance of    29.2                 60.7
instruments in first-stage
p-value                       (0.00)               (0.00)

Note: T-statistics in parentheses.

Source: Authors' calculation based on data from the World
Bank's Enterprise Survey Bangladesh and Sri Lanka

***, **, * Statistically significant at 1%, 5% and 10%
significance levels.

Table A4: Linear probability estimation for
refusing to answer corruption questions

                           Corruption is       Bribes are
                              serious            needed
                              problem        to get things
                              [dummy]             done
                                                [dummy]
Observation                    1,104             1,104

Reticence
  Number of no             -0.016 **         -0.086 ***
responses
  [high numbers           (-1.97)           (-3.26)
mean more likely to be
reticent]
Respondent
  Respondent is            -0.009            -0.035
owner
  [dummy]                 (-0.92)           (-1.00)
  Respondent is             0.039             0.014
female
  [dummy]                  (1.37)            (0.24)
  Respondent's age          0.005            -0.045 *
  [natural log]            (0.77)           (-1.93)
Manager education
  Manager has              -0.013            -0.069
secondary education
  [dummy]                 (-0.47)           (-0.81)
  Manager has              -0.020            -0.100
vocational education
  [dummy]                 (-0.71)           (-1.07)
  Manager has              -0.012            -0.070
tertiary education
  [dummy]                 (-0.49)           (-0.83)
Firm characteristics
  Number of workers        -0.007 **          0.023 *
  [natural log]           (-2.12)            (1.67)
  Age of firm              -0.001             0.010
  [natural log]           (-0.33)            (0.45)
  Firm exports              0.004            -0.034
  ["dummy]                 (0.47)           (-0.87)
  Firm is in Sri Lanka      0.029 **         -0.244 ***
  [dummy]                  (2.58)           (-6.91)
  Retail firm               0.004            -0.056
  [dummy]                  (0.18)           (-0.90)
  Manufacturing firm        0.004            -0.058
  [dummy]                  (0.68)           (-1.55)
Constant                    0.090 **          0.889 ***
                           (2.12)            (5.72)
F-test for significance   107.7             107.7
of instruments in
first-stage
p-value                    (0.00)            (0.00)

                             Bribe was            Bribes
                             requested       typically needed
                             during tax         during tax
                             inspection        inspections
                              [dummy]            [dummy]
Observation                     470                634

Reticence
  Number of no             -0.061 **          -0.062 **
responses
  [high numbers           (-1.99)            (-1.99)
mean more likely to be
reticent]
Respondent
  Respondent is             0.023             -0.070 *
owner
  [dummy]                  (0.66)            (-1.90)
  Respondent is            -0.006              0.079
female
  [dummy]                 (-0.13)             (1.06)
  Respondent's age         -0.015             -0.057 **
  [natural log]           (-0.64)            (-1.98)
Manager education
  Manager has              -0.053              0.021
secondary education
  [dummy]                 (-0.54)             (0.20)
  Manager has              -0.084              0.027
vocational education
  [dummy]                 (-0.83)             (0.23)
  Manager has              -0.021             -0.061
tertiary education
  [dummy]                 (-0.21)            (-0.60)
Firm characteristics
  Number of workers         0.039 **          -0.004
  [natural log]            (2.43)            (-0.25)
  Age of firm              -0.006              0.076 ***
  [natural log]           (-0.24)             (2.61)
  Firm exports             -0.025             -0.011
  ["dummy]                (-0.64)            (-0.23)
  Firm is in Sri Lanka     -0.143 ***         -0.247 ***
  [dummy]                 (-4.53)            (-6.91)
  Retail firm               0.040             -0.106 **
  [dummy]                  (0.80)            (-2.11)
  Manufacturing firm       -0.009             -0.108 **
  [dummy]                 (-0.23)            (-2.41)
Constant                    0.407 *            0.573 ***
                           (1.94)             (3.45)
F-test for significance    37.0               65.6
of instruments in
first-stage
p-value                    (0.00)             (0.00)

Note: T-statistics in parentheses.

Source: Authors' calculation based on data from the
World Bank's Enterprise Survey Bangladesh and Sri Lanka

***, **, * Statistically significant at 1%, 5% and 10%
significance levels.

Table A5: First-stage regressions from IV Probit Model (see Table 6)

                                Corruption is    Bribes are needed
                               serious problem    get things done
                                   [dummy]            [dummy]

Observation                         1,092               736

Reticence
  Leave-one out                  0.688 ***         0.759 ***
average
  [high numbers                (10.64)           (10.38)
mean more likely to be
reticent]
Respondent
  Respondent is                 -0.036            -0.179
owner
  [dummy]                      (-0.28)           (-1.19)
  Respondent is                 -0.049             0.206
female
  [dummy]                      (-0.20)            (0.74)
  Respondent's age              -0.017             0.007
  [natural log]                (-0.19)            (0.07)
Manager education
  Manager has                    0.057             0.170
secondary education
  [dummy]                       (0.20)            (0.48)
  Manager has                   -0.075             0.034
vocational education
  [dummy]                      (-0.23)            (0.09)
  Manager has                   -0.045             0.105
tertiary education
  [dummy]                      (-0.16)            (0.29)
Firm characteristics
  Number of workers              0.034            -0.011
  [natural log]                 (0.67)           (-0.19)
  Age of firm                   -0.117            -0.134
  [natural log]                (-1.34)           (-1.30)
  Firm exports                   0.346 **          0.224
  [dummy]                       (2.42)            (1.32)
  Firm is in Sri Lanka           0.132            -0.189
  [dummy]                       (0.95)           (-1.23)
  Retail firm                   -0.048            -0.150
  [dummy]                      (-0.19)           (-0.54)
  Manufacturing firm            -0.217            -0.297 *
  [dummy]                      (-1.62)           (-1.82)
Constant                         1.574 ***         1.646 ***
                                (3.52)            (3.07)

                                Bribe was     Bribes typically
                                requested     needed during tax
                               during tax    inspections [dummy]
                               inspection
                                 [dummy]
Observation                        417               336

Reticence
  Leave-one out                  0.541 ***     0.824 ***
average
  [high numbers                 (5.43)        (5.45)
mean more likely to be
reticent]
Respondent
  Respondent is                  0.182         0.040
owner
  [dummy]                       (0.78)        (0.17)
  Respondent is                  0.302
female
  [dummy]                       (0.77)
  Respondent's age               0.115        -0.183
  [natural log]                 (0.78)       (-1.08)
Manager education
  Manager has                   -0.486         0.574
secondary education
  [dummy]                      (-0.83)        (1.32)
  Manager has                   -0.280
vocational education
  [dummy]                      (-0.45)
  Manager has                   -0.212         0.286
tertiary education
  [dummy]                      (-0.35)        (0.69)
Firm characteristics
  Number of workers              0.174 **     -0.116
  [natural log]                 (1.97)       (-1.25)
  Age of firm                   -0.340 **      0.083
  [natural log]                (-2.44)        (0.49)
  Firm exports                   0.087         0.470 *
  [dummy]                       (0.37)        (1.72)
  Firm is in Sri Lanka          -0.004
  [dummy]                      (-0.02)
  Retail firm                    0.116
  [dummy]                       (0.32)
  Manufacturing firm            -0.143        -0.465 *
  [dummy]                      (-0.62)       (-1.87)
Constant                         2.492 ***     0.974
                                (2.89)        (1.17)

Dependent variable is number of no responses
(ie, the reticence variable)

Note: T-statistics in parentheses.

Source: Authors' calculation based on data from the World
Bank's Enterprise Survey Bangladesh and Sri Lanka

***, **, * Statistically significant at 1%, 5% and 10%
significance levels.

Table A6: First-stage regressions from IV
Probit Model for refusals (see Table 8)

                             Corruption is     Bribes are
                            serious problem   needed to get
                                [dummy]        things done
                                                 [dummy]

Observation                      1,104            1,104

Reticence
  Leave-one out               0.680 ***         0.680 ***
average
  [high numbers mean        (10.67)           (10.67)
more likely to be
reticent]
Respondent
  Respondent is owner        -0.022            -0.022
  [dummy]                   (-0.16)           (-0.16)
  Respondent is female        0.011             0.011
  [dummy]                    (0.05)            (0.05)
  Respondent's age           -0.042            -0.042
  [natural log]             (-0.47)           (-0.47)
Manager education
  Manager has                 0.033             0.033
secondary education
  [dummy]                    (0.12)            (0.12)
  Manager has                -0.095            -0.095
vocational education
  [dummy]                   (-0.30)           (-0.30)
  Manager has tertiary       -0.049            -0.049
education
  [dummy]                   (-0.17)           (-0.17)
Firm characteristics
  Number of workers           0.039             0.039
  [natural log]              (0.76)            (0.76)
  Age of firm                -0.103            -0.103
  [natural log]             (-1.18)           (-1.18)
  Firm exports                0.335 **          0.335 **
  [dummy]                    (2.35)            (2.35)
  Firm is in Sri Lanka        0.13              0.13
  [dummy]                    (0.97)            (0.97)
  Retail firm                -0.066            -0.066
  [dummy]                   (-0.26)           (-0.26)
  Manufacturing firm         -0.224 *          -0.224 *
  [dummy]                   (-1.68)           (-1.68)
Constant                      1.616 ***         1.616 ***
                             (3.65)            (3.65)

                             Bribe was    Bribes typically
                             requested    needed during tax
                            during tax       inspections
                            inspection         [dummy]
                              [dummy]
Observation                     470              612

Reticence
  Leave-one out               0.575 ***     0.744 ***
average
  [high numbers mean         (6.06)        (8.38)
more likely to be
reticent]
Respondent
  Respondent is owner         0.140        -0.086
  [dummy]                    (0.62)       (-0.53)
  Respondent is female        0.329        -0.129
  [dummy]                    (0.85)       (-0.42)
  Respondent's age            0.152        -0.169
  [natural log]              (1.07)       (-1.48)
Manager education
  Manager has                -0.589         0.288
secondary education
  [dummy]                   (-1.05)        (0.90)
  Manager has                -0.430        -0.107
vocational education
  [dummy]                   (-0.72)       (-0.27)
  Manager has tertiary       -0.441         0.052
education
  [dummy]                   (-0.76)        (0.16)
Firm characteristics
  Number of workers           0.133        -0.012
  [natural log]              (1.61)       (-0.19)
  Age of firm                -0.353 ***     0.071
  [natural log]             (-2.72)        (0.60)
  Firm exports                0.194         0.327 *
  [dummy]                    (0.88)        (1.74)
  Firm is in Sri Lanka        0.04          0.13
  [dummy]                    (0.19)        (0.70)
  Retail firm                 0.163
  [dummy]                    (0.46)
  Manufacturing firm         -0.085        -0.311 *
  [dummy]                   (-0.39)       (-1.84)
Constant                      2.485 ***     1.212 **
                             (3.04)        (2.25)

Dependent variable is number of no responses
(ie, the reticence variable)

Note: T-statistics in parentheses.

Source: Authors' calculation based on data from the World
Bank's Enterprise Survey Bangladesh and Sri Lanka

***, **, * Statistically significant at 1%, 5% and 10%
significance levels.

Acknowledgements

The data used in this paper are from the Enterprise Surveys (http://www. enterprisesurveys.org), The World Bank. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Responsibility for all errors, omissions, and opinions rests solely with the authors. We would like to thank Jeff Nugent and Bilin Neyapti and participants at the Western Economic Association annual meetings, Public Choice Society annual meetings, and a seminar at Texas A&M International University for comments on earlier drafts. Finally, we would like to thank Josef Brada and anonymous referees for their helpful advice.

REFERENCES

Azfar, 0 and Murrell, P. 2009: Identifying reticent respondents: Assessing the quality of survey data on corruption and values. Economic Development and Cultural Change 57(2): 387-411.

Bound, J, Jaeger, DA and Baker, RM. 1995: Problems with instrumental variables when the correlation between the instruments and the endogenous explanatory variable is weak. Journal of the American Statistical Association 90(430): 443-450.

Clarke, GRG. 2011a: How petty is petty corruption? Evidence from firm surveys in Africa. World Development 39(7): 1122-1132.

Clarke, GRG. 2011b: Lying about firm performance: Evidence from a firm survey in Nigeria. Texas A&M International University: Laredo, TX.

Clarke, GRG. 2012a: Do reticent managers lie during firm surveys? Texas A&M International University: Laredo, TX.

Clarke, GRG. 2012b: What do managers mean when they say 'firms like theirs' pay bribes? International Journal of Economics and Finance 4(10): 147-160.

Clarke, GRG and Xu, LC. 2004: Privatization, competition and corruption: How characteristics of bribe takers and payers affect bribes to utilities. Journal of Public Economics 88(9-10): 2067-2097.

Clausen, B, Kraay, A and Murrell, P. 2010: Does respondent reticence affect the results of corruption surveys? Evidence from the World Bank Enterprise Survey for Nigeria. Policy Research Working Paper No. 5415. World Bank: Washington DC.

Coutts, E and Jann, B. 2011: Sensitive questions in online surveys: Experimental results for the randomized response technique (RRT) and the unmatched count technique (UCT). Sociological Methods and Research 40(1): 169-193.

Dillman, DA, Smith, JD and Christian, LM. 2008: Internet, mail and mixed-mode surveys: The tailored design method. John Wiley and Sons: Hoboken, NJ.

Djankov, S, La Porta, R, Lopez-de-Silanes, F and Shleifer, A. 2002: The regulation of entry. Quarterly Journal of Economics 117(1): 1-37.

Edgell, SE, Himmelfarb, S and Duchan, KL. 1982: Validity of forced responses in a randomized response model. Sociological Methods and Research 11(1): 89-100.

Fisher, RJ and Tellis, GJ. 1998: Removing social desirability bias with indirect questioning: Is the cure worse than the disease. Advances in Consumer Research 25(1): 563-567.

Fox, JA and Tracy, PE. 1986: Randomized response: A method for sensitive surveys. Sage Publications: Newbury Park, CA.

Friesenbichler, KS, Clarke, GRG and Wong, M. 2014: Price competition and market transparency: Evidence from a random response technique. Empirica 41(1): 5-21.

Iarossi, G. 2006: The power of survey design. World Bank: Washington DC.

Jensen, NM and Rahman, A. 2011: The silence of corruption: Identifying underreporting of business corruption through randomized response techniques. Policy Research Working Paper No. 5696. World Bank: Washington DC.

Johnson, S, Kaufmann, D, McMillan, J and Woodruff, C. 2000: Why do firms hide? Bribes and unofficial activity after communism. Journal of Public Economics 76(3): 495-520.

Johnson, S, McMillan, J and Woodruff, C. 2002: Property rights and finance. American Economic Review 92(5): 1335-1356.

Karlan, D and Zinman, J. 2012: List randomization for sensitive behavior: An application for measuring use of loan proceeds. Journal of Development Economics 98(1): 71-75.

Lee, RM. 1993: Doing research on sensitive topics. Sage Publications: London, UK.

Lensvelt-Mulders, GJLM, Hox, JJ, Van Der Heuden, PGM and Maas, CJM. 2005: Meta-analysis of randomized response research: Thirty-five years of validation. Sociological Methods and Research 35(3): 319-348.

Lewbel, A, Dong, Y and Yang, TT. 2012: Comparing features of convenient estimators for binary choice models with endogenous regressors. Canadian Journal of Economics 45(3): 810-829.

Recanatini, F, Wallsten, S and Xu, LC. 2000: Surveying surveys and questioning questions: Learning from World Bank experience. Policy Research Working Paper No. 2307. World Bank: Washington DC.

Safavian, MS, Graham, DH and Gonzalez-Vega, C. 2001: Corruption and microenterprises in Russia. World Development 29(7): 1215-1224.

Smith, RJ and Blundell, RW. 1986: An exogeneity test for a simultaneous equation Tobit model with an application to labor supply. Econometrica 54(4): 679-686.

Svensson, J. 2003: Who must pay bribes and how much? Evidence from a cross section of firms. Quarterly Journal of Economics 118(1): 207-230.

Tourangeau, R, Rips, U and Rasinski, K. 2000: The psychology of survey response. Cambridge University Press: Cambridge, UK.

Treisman, D. 2007: What have we learned about the causes and corruption from ten years of crossnational empirical research. Annual Review of Political Science 10(1): 211-244.

Warner, SL. 1965: Randomized response: A survey technique for eliminating evasive answer bias. Journal of the American Statistical Association 60(309): 63-66.

World Bank. 2012a: Enterprise surveys: Indicator descriptions. World Bank: Washington DC.

World Bank. 2012b: Sri Lanka 2011: Management practices surveys data set. World Bank: Washington DC.

GEORGE RG CLARKE [1], KLAUS S FRIESENBICHLER [2] & MICHAEL WONG [3]

[1] Division of International Banking and Finance Studies, A.R. Sanchez, Jr. School of Business, Texas A&M International University, 5201 University Boulevard, Laredo, TX 78041, USA.

[2] Austrian Institute of Economic Research (WIFO), 1030 Wien, Arsenal, Objekt 20, Vienna, Austria.

[3] The World Bank, 1818 H Street N.W., Washington D.C. 20433, USA.

(1) The data generated from the Sri Lanka survey and the questionnaire are available from the World Bank on the World Bank's Enterprise Survey website (http://www.enterprisesurveys.org).

(2) Classifications were based on ISIC rev 3.1.

(3) This document is available for registered users of World Bank Enterprise Survey data on the Enterprise Survey webpage (http://www.enterprisesurveys.org)

(4) Transparency International assigns scores on perceptions of public sector corruption ranging from 0 to 100, with 0 representing highly corrupt and 100 representing very clean.

(5) Clarke (2011a) shows managers who answer the question as a percentage of sales report paying far higher bribes than do managers who answer the question in local currency. He also shows this is true even after controlling for other things that might affect bribe payments.

(6) The other four are electricity, access to finance, access to land, and political instability.

(7) Managers who answered 'do not know' to whether they had met with or been inspected by tax officials were also asked the indirect question.

(8) In addition, it was asked only to managers who said tax inspectors had not inspected their firm, further suggesting it is not asking about the manager's own experiences.

(9) In practice, most studies assume managers answer question like this with their own experience in mind. See, for example, Clarke and Xu (2004, p. 2077), Johnson et al. (2000, p. 504), Johnson et al. (2002, pp. 1337-1338) and Svensson (2003, pp. 212-213).

(10) Consistent with this, Clarke (2012b) shows managers in Afghanistan who do not bid on government contracts are more likely to say corruption is a serious problem when bidding on government contracts than managers who do bid. Clarke (2012b) argues the most reasonable explanation is firms excluded from bidding due to corruption are more concerned about corruption than firms that bid.

(11) For the 'any bribes' question, although they cannot answer 'yes' or 'no', we will treat a positive amount as a 'yes' and zero as a 'no' in the discussion.

(12) The same, or a similar, approach has been used in Clarke (2011b, 2012a), Clausen et al. (2010), Friesenbichler et al. (2014) and Jensen and Rahman (2011). Azfar and Murrell (2009) and Clausen et al. (2010) provide greater detail on the methodology.

(13) See Fox and Tracy (1986) for a general discussion or Recanatini et al. (2000) for a discussion directly linked to the Enterprise Surveys. A related technique is list randomization where survey respondents are asked to report how many sensitive statements in a given list are true. Karlan and Zinman (2012) used list randomization in a study looking at loan use by microenterprises.

(14) Lensvelt-Mulders et al. (2005) suggest that random response questions reduce underreporting.

(15) Edgell et al. (1982), for example, found significant underreporting of sensitive behaviors when they used a version of the forced response methodology. They used a random number generator to generate random numbers. On some questions, however, the random number generator was not random and instead always forced the respondents to say yes. Despite this, about one-quarter of respondents answered some questions 'no'.

(16) The expected distribution assuming 30% have done each behavior is calculated assuming the probability an individual does each act is independent of the probability that they will do the other acts.

(17) We do not have any information on the educational attainment of the respondent (unless the respondent is also the top manager).

(18) Several studies have shown that corruption is greater when the burden of regulation is heavier (Djankov et al, 2002; Safavian et al., 2001).

(19) We would like to thank an anonymous referee for this observation.

(20) The one difference is that in the model for tax inspections--not inspected, the significance level falls to about 12%.

(21) In particular, the linear probability model can be applied with continuous, binary or categorical variables (Lewbel et al., 2012). The linear probability model is also more robust to heteroskedasticity. Because the error terms in the linear probability are heteroskedastic by construction, we use robust standard errors. Although robust standard errors are less efficient than using weighted least squares, they allow for more general heteroskedasticity.

Table 1: Classification of questions

           Specific                         General

Direct     Tax inspections--inspected
Indirect   Tax inspections--not inspected   Perceptions, any bribes

Table 2: Percent of firms that answered and
refused to answer questions on corruption

Sample                  Corruption   Bribes are
                        is serious   needed to
                         problem     get things
                                        done

                                  All

Answered                  1,548        1,070
Yes                         627          105
No                          921          965
Refused/Does not know        41          519
Refused                      28          253
Do not know                  13          266

Sample                      Bribe        Bribes typically
                         requested or      expected or
                           expected      requested during
                          during tax     tax inspections
                         inspections

                          Firms that     Firms that were
                        were inspected    not inspected

Answered                     599               736
Yes                           64                16
No                           535               721
Refused/Does not know         66               179
Refused                       52                95
Do not know                   14                35

Note: Counts are unweighted.

Source: Authors' calculation based on data from the
World Bank's Enterprise Survey Bangladesh and Sri Lanka

Table 3: Correlation between dummy variables indicating
that manager refused to answer corruption guestions

                       Bribes are  Corruption  Bribe was     Bribes
                       needed to   is serious  requested    typically
                       get things   problem    during tax    needed
                          done                 inspection  during tax
                                                           inspections
Bribes are needed         1.00
to get things done

Corruption is           0.06 **       1.00
serious problem          (0.03)

Bribe was requested     0.43 ***     -0.02        1.00
during tax inspection    (0.00)      (0.53)

Bribes typically        0.59 ***     -0.03       -- (a)       1.00
needed during tax        (0.00)      (0.44)
inspections

(a) Firms either answered guestion on whether bribe was requested
during tax inspection or question whether bribes are typically
needed.

Note: Dummy variables are coded '1' if firm refused to answer
question and 0 otherwise.

Source: Authors' calculation based on data from the World Bank's
Enterprise Survey Bangladesh and Sri Lanka

Table 4: Sensitive questions used to identify reticent respondents

Question   Forced response questions

1          Have you ever paid less in personal taxes than you should
           have under the law?

2          Have you ever paid less in business taxes than you should
           have under the law?

3          Have you ever made a misstatement on a job application?

           Have you ever used the office telephone for personal
           businesses?#

4          Have you ever inappropriately promoted an employee for
           personal reasons?

5          Have you ever deliberately not given your suppliers or
           clients what was due to them?

           Have you ever lied in your self-interest? #

6          Have you ever inappropriately hired a staff member for
           personal reasons?

           Have you ever been purposely late for work? #

7          Have you ever unfairly dismissed an employee for personal
           reasons

Note: The three bolded questions are less sensitive questions
that were included to allow sophisticated reticent respondents
to not have to give large numbers of 'no's' consecutively if
they realized that this would be unlikely.

Note: The three bolded questions are less sensitive questions
that were included to allow sophisticated reticent respondents
to not have to give large numbers of 'no's' consecutively if
they realized that this would be unlikely indicated with #

Source: Questionnaire for World Bank's Enterprise Survey

Table 5: Expected and actual distribution of 'no' responses

Number of          Expected     Expected if    Actual % of
no responses      if all are   30% have done   respondents
                   'angels'      each act       in survey

7                    0.8%          0.1%           12.9
6                    5.5%          0.8%           11.3
5                   16.4%          4.7%           18.2
4                   27.3%          14.4%          22.2
3                   27.3%          26.8%          18.8
2                   16.4%          29.8%           8.9
1                    5.5%          18.5%           4.9
0                    0.8%          4.9%            2.8
Expected number      3.5            2.5            4.2
of no responses

Note: Counts are unweighted. The 'angels' assumption assumes that
no one has done any of the sensitive acts. The '30% assumption'
assumes the probability of a respondent doing each act is 30% and
that the probability that any given agent does each act is
independent of the probability that they do other behaviors.

Source: Author's calculation based on data from the World Bank's
Enterprise Survey for Nigeria (2007 and 2009)

Table 6: IV Probit regressions for corruption questions

                       Corruption     Bribes are
                       is serious    needed to get
                         problem      things done
                         [dummy]        [dummy]
Observation               1,092           736
Reticence
  Number of no          -0.406 ***    -0.079
responses
  [high numbers        (-8.78)       (-0.71)
  mean more likely
  to be reticent]
Respondent
  Respondent is          0.118        -0.511 **
owner
  [dummy]               (1.13)       (-2.55)
  Respondent is         -0.199         0.544 **
female
  [dummy]              (-0.93)        (2.05)
  Respondent's age       0.148 **      0.188 *
  [natural log]         (2.05)        (1.68)
Manager education
  Manager has           -0.089        -1.275 ***
secondary
education
  [dummy]              (-0.41)       (-4.18)
  Manager has           -0.179        -1.506 ***
vocational
education
  [dummy]              (-0.73)       (-3.92)
Manager has              0.067        -1.225 ***
tertiary education
  [dummy]               (0.32)       (-4.01)
Firm characteristics
  Number of              0.057         0.027
workers
  [natural log]         (1.47)        (0.41)
  Age of firm           -0.129 *       0.167
  [natural log]        (-1.95)        (1.42)
  Firm exports           0.063         0.533 ***
  [dummy]               (0.58)        (3.12)
  Firm is in Sri        -1.053 ***    -0.200
Lanka
  [dummy]              (-6.79)       (-1.12)
  Retail firm            0.094        -0.280
  [dummy]               (0.42)       (-0.78)
  Manufacturing         -0.093        -0.187
firm
  [dummy]              (-0.94)       (-1.05)
Constant                 1.628 ***    -0.502
                        (4.68)       (-0.72)
Smith-Blundelltest      26.38 ***      0.90
of exogeneity
(p-value)               (0.00)        (0.76)

                          Bribe was      Bribes typically
                       requeste during    needed during
                        tax inspectic    tax inspections
                           [dummy]           [dummy]
Observation                  417               333
Reticence
  Number of no          -0.385 ***        -0.446 ***
responses
  [high numbers        (-3.59)           (-3.64)
  mean more likely
  to be reticent]
Respondent
  Respondent is          0.335            -0.287
owner
  [dummy]               (1.57)           (-1.01)
  Respondent is          0.368
female
  [dummy]               (1.04)
  Respondent's age       0.440 **          0.002
  [natural log]         (2.49)            (0.01)
Manager education
  Manager has           -0.274            -0.082
secondary
education
  [dummy]              (-0.54)           (-0.19)
  Manager has           -0.197
vocational
education
  [dummy]              (-0.36)
Manager has             -0.069            -0.719
tertiary education
  [dummy]              (-0.13)           (-1.39)
Firm characteristics
  Number of              0.269 ***         0.006
workers
  [natural log]         (3.26)            (0.04)
  Age of firm           -0.479 ***         0.047
  [natural log]        (-3.17)            (0.19)
  Firm exports           0.218             0.681 **
  [dummy]               (1.07)            (2.27)
  Firm is in Sri        -0.045
Lanka
  [dummy]              (-0.22)
  Retail firm           -0.189
  [dummy]              (-0.43)
  Manufacturing          0.181            -0.465
firm
  [dummy]               (0.87)           (-1.49)
Constant                -0.020             1.052
                       (-0.02)            (1.08)
Smith-Blundelltest       4.42 **           3.41 *
of exogeneity
(p-value)               (0.03)            (0.06)

Note: T-statistics in parentheses.

Source: Authors' calculation based on data from the
World Bank's Enterprise Survey Bangladesh and Sri Lanka

***, **, * Statistically significant at 1%, 5% and 10%
significance levels.

Table 7: Estimated impact of reticence on measures of corruption

Question                                Sample            Actual

Corruption is serious problem           All                41%
Bribes are needed to get things done                       10%
Bribe expected during tax inspections   Firms inspected    11%
Bribes typically expected during tax    Firms not           2%
inspections                             inspected

                                         All     30% engage in
Question                                angels   sensitive acts

Corruption is serious problem            50%          65%
Bribes are needed to get things done     11%          12%
Bribe expected during tax inspections    22%          35%
Bribes typically expected during tax     14%          26%
inspections

Note: Actualis the actual number of 'yes' responses to each of
the corruption questions. For the column 'all angels' is the
estimated number of 'yes' responses to the corruption questions
if all firms answered 3.5 sensitive questions with 'no'. For the
column '30% engage in sensitive acts' the probabilities are
calculated assuming that all firms answered 2.5 of the sensitive
questions as 'no'. Probabilities are calculated for all firms
with that firm's characteristics and the number of no answers in
the title.

Table 8: IV Probit regressions for refusing
to answer corruption questions

Observation             Corruption is    Bribes are needed
                       serious problem   to get things done
                           [dummy]            [dummy]
                            1,104              1,104

Reticence
  Number of no          -0.458 ***        -0.249 ***
responses
  [high numbers        (-4.18)           (-3.61)
mean more likely to
be reticent]
Respondent
  Respondent is         -0.169            -0.094
owner
  [dummy]              (-0.70)           (-0.88)
  Respondent is          0.543             0.037
female
  [dummy]               (1.57)            (0.18)
  Respondent's age       0.060            -0.142 **
  [natural log]         (0.34)           (-2.00)
Manager education
  Manager has           -0.169            -0.220
secondary education
  [dummy]              (-0.38)           (-0.99)
  Manager has           -0.269            -0.283
vocational
education
  [dummy]              (-0.50)           (-1.12)
  Manager has           -0.095            -0.207
tertiary education
  [dummy]              (-0.20)           (-0.95)
Firm characteristics
  Number of             -0.245 *           0.071 *
workers
  [natural log]        (-1.65)            (1.77)
  Age of firm            0.059             0.033
  [natural log]         (0.30)            (0.48)
  Firm exports          -0.022            -0.103
  [dummy]              (-0.06)           (-0.88)
  Firm is in Sri         0.776 **         -0.751 ***
Lanka
  [dummy]               (2.49)           (-5.86)
  Retail firm            0.133            -0.169
  [dummy]               (0.30)           (-0.72)
  Manufacturing          0.084            -0.170
firm
  [dummy]               (0.23)           (-1.64)
Constant                 0.065             1.162 ***
                        (0.06)            (2.81)
                         4.15 **           4.04 **
Smith-Blundell test
of exogeneity
(p-value)               (0.04)            (0.04)

Observation             Bribe was requested     Bribes typically
                       during tax inspection    needed during tax
                              [dummy]          inspections [dummy]
                                470                    612

Reticence
  Number of no          -0.291 **               -0.236 **
responses
  [high numbers        (-2.26)                 (-2.30)
mean more likely to
be reticent]
Respondent
  Respondent is          0.153                  -0.291 *
owner
  [dummy]               (0.66)                 (-1.85)
  Respondent is         -0.205                   0.353
female
  [dummy]              (-0.39)                  (1.20)
  Respondent's age      -0.147                  -0.209 *
  [natural log]        (-1.02)                 (-1.94)
Manager education
  Manager has           -0.278                  -0.023
secondary education
  [dummy]              (-0.48)                 (-0.08)
  Manager has           -0.499                   0.149
vocational
education
  [dummy]              (-0.80)                  (0.41)
  Manager has           -0.142                  -0.295
tertiary education
  [dummy]              (-0.24)                 (-1.10)
Firm characteristics
  Number of              0.196 **               -0.004
workers
  [natural log]         (2.57)                 (-0.06)
  Age of firm           -0.005                   0.267 **
  [natural log]        (-0.03)                  (2.41)
  Firm exports          -0.104                  -0.049
  [dummy]              (-0.52)                 (-0.27)
  Firm is in Sri        -0.969 ***              -1.307 ***
Lanka
  [dummy]              (-3.06)                 (-4.63)
  Retail firm            0.408
  [dummy]               (0.97)
  Manufacturing         -0.083                  -0.357 **
firm
  [dummy]              (-0.44)                 (-2.36)
Constant                 0.345                   0.599
                        (0.33)                  (1.01)
                         2.29                    3.65
Smith-Blundell test
of exogeneity
(p-value)               (0.12)                  (0.06)

Note: T-statistics in parentheses.

Source: Authors' calculation based on data from the World Bank's
Enterprise Survey Bangladesh and Sri Lanka

***, **, * Statistically significant at 1%, 5% and 10%
significance levels.

Table 9: Estimated impact of reticence on refusals to answer
questions on reticence
                                                      30% engage in
                                Actual   All angels   sensitive acts

Corruption is serious problem     3%         5%            11%
Bribes are needed to get         33%        39%            48%
  things done
Bribe expected during tax        10%        18%            26%
  inspections
Bribes typically expected        19%        25%            33%
  during tax inspections

Note: Actual is the actual number of people who refused to answer
each of the corruption questions. For the column 'all angels' is
the estimated number of 'yes' responses to the corruption
questions if all firms answered 3.5 sensitive questions designed
to identify reticent respondents with 'no'. For the column '30%
engage in sensitive acts' the probabilities are calculated
assuming that all firms answered 2.5 of the sensitive questions
with 'no'. Probabilities are calculated for all firms with that
firm's characteristics and the number of no answers in the title.