Innovation, financial reporting quality, and audit quality

He and Tian (2013) and Fang et al. (2014) document that analyst coverage and stock liquidity, respectively, could impede firms’ innovation. Plumlee et al. (2015) find that borrowers with more patent counts and citations receive lower interest rates from banks. Hirshleifer et al. (2012) provide evidence suggesting that overconfident CEOs in innovative industries achieve higher levels of innovation. Gu (2005) shows that the change of patent citation impact is positively associated with future earnings, but does not find evidence that market participants incorporate these implications in stock prices.

Information opacity of innovative firms may exacerbate the impact of managers’ opportunistic behavior on firms’ financial reporting quality.

Srivastava (2014) finds that successive cohorts of newly listed firms since 1970 exhibit progressively lower earnings quality measures (proxied by earnings volatility, matching between revenue and expense and earnings relevance), mainly because of increasing intangible intensity. In contrast to his time-series analysis of cohorts of newly list firms, our study focuses on cross-sectional difference in earnings quality (proxied by abnormal accruals and MBE) among public firms with patents.

Using a large sample, Lobo and Zhao (2013) already document that additional audit effort increases FRQ. In contrast, employing essentially a sub-sample empirically, we examine the moderating effect of auditor quality/auditor effort on the relationship between innovation proxies and FRQ.

Srivastava (2014) finds that the decline of earnings quality over time is due to the assimilation of successive cohorts of newly listed firms into the firm population and the increased intangible intensity of these firms. He examines the temporal association between intangible intensity, proxied by SG&A intensity, R&D intensity and market-to-book ratio, and earnings quality, proxied by volatility, relevance and matching. Our study differs from Srivastava (2014) in that we use a R&D output measure, patents, as the proxy for innovation, and abnormal accruals and MBE as proxies for earnings quality, and examine cross-sectional differences in the association between innovation and earnings quality for firms with different levels of audit quality.

Prior research (e.g., Hall et al. 2005; Gunny and Zhang 2014) documents that the value of a patent is best captured by citation counts subsequent to the patent grant date. In addition, citation count is shown to be positively associated with future earnings and market values and regarded as a superior measure for the economic value of the cited patents (Trajtenberg 1990; Hall et al. 2005).

As the market learns about a firm’s R&D activity, patents, and new product introductions, it responds to this information through valuations of the firm’s assets that are reflected in the market value of that firm (e.g., Tobin’s Q). Therefore, we use five measures of innovation that reflect this portrayal of the innovation process in the factor analysis to generate our alternative measure of innovation (InnvFactor). First, we use R&D intensity (i.e., R&D expenditure divided by total assets) as a measure of firms’ innovation activities in search of new product and process innovations. Second, we use the (log of) total number of patents applied by a firm as a measure of the outcome of the firm’s innovation activities. Third, we use the number of patent citations for each firm. Since self-citation and non-self-citation are different, we separate self-citation from external citation. Lastly, we use residual Tobin’s Q as the market level measure of firms’ innovation and exploration performance. Tobin’s Q (i.e., the ratio of a firm’s market value to its book value) is widely used as a proxy for firms’ intangible assets (Fama and French 1992; Chan et al. 2001). This intangible asset is attributed to the value created by firms’ innovation and exploration activities. However, a firm’s Tobin’s Q can be improved through operational efficiency (e.g., cost reductions) as well as by improving innovation and exploration activities. Therefore, we remove the variance in Tobin’s Q caused by efficiency improvements by regressing Tobin’s Q on inventory turnover, payables turnover, receivables turnover, and selling and administrative cost. Following Xue et al. (2012), we use the residual (TQInnv) from this regression as a measure of the firm’s innovation performance.

Google provides well-organized text files with information on the patents granted by the USPTO since 1976. The USPTO bulk downloadable patent files are available at http://​goo.​gl/​2qgb.

The data include both, the application year of the patent (i.e., the year the patent application was filed) and the grant year (i.e., the year the patent was granted). We use the patent application year in the data aggregation because the application year is regarded as a better representation of the actual time of the innovation than is the grant year (Griliches et al. 1988).

Please refer to Kogan et al. (2017) for detailed steps about patent data construction that we follow.

We classify an auditor as a specialist if the auditor has a national (city) market share of audit fees greater than 30% (50%) of the total audit fees in a two-digit SIC industry.

Prior studies suggest that firms in five R&D intensive industries (two-digit SIC codes: 28, 35, 36, 37 and 38) frequently apply for patents (e.g., Lev and Sougiannis 1996; Shi 2003; Plumlee et al. 2015).

In Table 2, we report the logs of patent count, citations, audit fees, and size (i.e., market value).

We also try the sensitivity test of using two-way clustering technique by auditor and year. Our inferences stay the same.

Our inferences remain consistent when we use the other three measures of innovation, i.e., Patent, CitePerPatent, InnvFactor.

The Tax Foundation provides state corporate tax rates from 2000 to 2013. We also check these data with the state corporate income tax rates reported in the Book of States to ensure consistency and accuracy. For each state-year observation, we compute the tax rate change from the prior year. For states with multiple tax brackets, we focus on changes in the top tax bracket.

There are several challenges associated with determining the relevant tax rate for a given firm. The approach taken by much of the previous research is to use the state in which a company has its headquarters. The underlying assumption is that most of the profits of the company are generated in that state. While this assumption is reasonable, it is often incorrect. In practice, state tax is assessed based on three main firm characteristics: percentage of sales, percentage of employees, and percentage of physical assets. As we do not have specific information on these three components, we attempt to approximate the state of taxation by deducing the state in which the firm conducts most of its business. We use the state count information from Garcia and Norli (2012), who compute the number of times a 10-K report mentions a U.S. state’s name in all 10-K filings from the SEC’s online database. The annual 10-K reports contain detailed information regarding a firm’s operations and financial performance during the year. More importantly, these reports can also contain information on the location of the firm’s properties and sales in different geographic areas. To construct the relevant state for our sample period we calculate the most frequently mentioned state for each firm and use that state to assess the corporate state income tax burden.

We also estimate Eq. (4) using PercentCredit in place of ExistCredit, where PercentCredit is the rate of state R&D tax credits. We do not tabulate these results as our inferences are unchanged when we use PercentCredit as the proxy for the exogenous shock.

As suggested by Duchin et al. (2010), and Hsu et al. (2015), our tests in Eq. (4) and later in Eq. (5) can be regarded as difference-in-differences tests because they compare the relation between the explanatory and dependent variables before and after an exogenous policy event.

Untabulated results show that our inferences are unchanged when we use the other patent proxies.