Body weight and Internet access: evidence from the rollout of broadband providers

Abstract

Obesity has become an increasingly important public health issue in the USA and many other countries. Hypothesized causes for this increase include declining relative cost of food and a decreasing share of the population working in labor-intensive occupations. In this paper, we suggest another factor: the Internet. Increasing Internet access could affect body weight through several channels. First, more time spent using the Internet, a sedentary activity, could lead to increases in body weight. Second, the prior literature has shown that economic activity (and income) increase with Internet access: given a positive health-income gradient, obesity rates could likewise increase, although the empirical evidence on the income-obesity gradient is mixed. Third, the Internet increases information and creates the possibility for online peer networks. Theoretically, increases in information should lead to more optimal consumer choices. At the same time, greater networking opportunities may result in peers having greater influence over positive or negative health behaviors. While we are unable to fully test these mechanisms, we are able to use the rollout of broadband Internet providers as a plausibly exogenous source of variation in Internet access to identify the reduced form effect of Internet use on body weight. We show that greater broadband coverage increases the body weight of white women and has both positive and negative effects on modifiable adult health behaviors including exercise, smoking, and drinking.

1 Introduction

Economists have long hypothesized that information is an important part of choice theory (Stigler 1961) and it is reasonable to assume that increases in access to information brought about by the Internet, which has transformed the way consumers acquire information, would improve health decisions and consequently health outcomes. However, this is not necessarily the case: consumers may substitute information from the Internet for visits to health professionals, resulting in exposure to lower quality health information (Wagner et al. 2001).¹ Beyond a direct information channel, the increasing prevalence of the Internet could influence health behaviors in other ways. Since it is a static activity, more time allocated toward surfing the Internet could increase body weight and this may lead individuals to experience the negative health consequences of a sedentary lifestyle (Owen et al. 2010). Additionally, social connections made through the Internet could influence weight gain and health through peer effects. Alternatively, increased Internet availability in an area may increase income which in turn could affect health given the positive health-income gradient, although the empirical evidence on the income-obesity gradient is mixed. In this paper, we explore the reduced form relationship between Internet access and adult body weight.

We use the rollout of broadband providers across counties in the USA during the 2000s as a plausibly exogenous proxy for increasing Internet access and use over this period. Similar identification strategies relying on the rollout of a policy or new technology have been used to estimate the effects of the food stamp program (Almond et al. 2011; Hoynes and Schanzenbach 2009), the Special Supplemental Nutrition Program for Women, Infants, and Children (WIC) (Hoynes et al. 2011), community health centers (Bailey and Goodman-Bacon 2015), and, of particular relevance to this work, electricity (Bailey and Collins 2011), broadcast television (Gentzkow 2006), and broadband Internet (Bhuller et al. 2013; Bellou 2015; Guldi and Herbst 2017; Kolko 2012). During the period of broadband rollout, the consumption of information on Web MD increased 20 fold, suggesting a relationship between broadband availability and the consumption of online health information. An existing body of work examines the relationship between increased use of the Internet to search for health information and the demand for health services,² but less research focuses on whether Internet availability directly affects health outcomes. The research that has been conducted is limited by time period, sample size, or is purely descriptive.³ Given that the Internet is increasingly used as a clearinghouse for health information⁴ and that this appears to have accelerated with the advent of widespread high-speed access, understanding whether broadband access affects health outcomes is of great importance to understanding aggregate public health trends.

To investigate the relationship between Internet and health, we focus on body weight, overweight status, and obesity as our primary health outcomes. Obesity and the rising share of overweight individuals is an increasing public health concern in many countries including the USA (Cawley and Meyerhoefer 2012). However, the social causes behind this increase in body weight are not entirely understood. Additionally, weight is a health outcome modifiable through behavior: time use, diet, exercise, and the utilization of weight loss products all directly influence BMI. Likewise, consumers’ decisions related to these products are all potentially affected by Internet availability. While much research has been devoted to understanding weight gain over time, little work has examined the relationship between technology and body weight.⁵ We also look at modifiable health behaviors including exercise, smoking, and drinking to provide evidence on the pathways by which Internet access affects weight gain. Ours is the first paper we are aware of to directly estimate the reduced form effect of the rollout of broadband service on health. We find that the expansion of broadband coverage is associated with increases in the average BMI and obesity of women and that these effects are particularly salient for white women. These results suggest technology is an important part of the obesity discussion.

2 Conceptual framework

We suggest several channels through which the Internet may influence an individual’s health behavior. Individuals are assumed to be making rational choices to optimize their utility subject to income and time constraints but with imperfect information. We posit that increases in access to the Internet may lead to increases in a consumer’s information set, expansion of the consumer’s peer network, changes in time use, and increases the consumer’s income. These may in turn influence the individual’s allocation of time for sedentary or dynamic activities, which may in turn affect measured body weight.

2.1 Information channel

Access to online health information has grown over time. Use of the Internet as a key source of this information has become increasingly common as consumers turn to websites, discussion boards, and social media. For example, the number of unique visitors to WebMD, an online publisher of health information, increased from 1.7 million in December 1999 to a monthly average of 40.8 million in the third quarter of 2007.⁶ Data from the Pew Research Center’s Internet and American Life Project shows Internet use among adults increased from 46% in March 2000 to 75% in December of 2007. Over the same period, the share of Internet users who ever looked for health information online increased from 55 to 75%. Broken down by gender, 61% of women and 47% of men reported using the Internet to look for health information in 2000 and this increased to 81% of women and 68% of men by 2007, suggesting that women may engage with online health information more readily than men.⁷ Quality of the health information discovered online, however, is variable. At one end of the quality spectrum are sources that provide correct and timely information. At the other end, information may be inaccurate, misleading, and potentially dangerous (Impicciatore et al. 1997; Akatsu and Kuffner 1998; Donald et al. 1998; McLellan 1998; Biermann et al. 1999; Purcell, Wilson, and Delamothe, Purcell et al. 2002).^8,9 Poor health information, or overwhelming amounts of information,¹⁰ may lead consumers to forgo visits to professionals when more reliable advice is in fact needed. With this in mind, there are opposing viewpoints regarding how online information influences the consumption of health services. Some researchers find online health information is a complement to health services (Suiziedelyte 2012), and others find it serves as a substitute (Wagner et al. 2001). Although individuals may use online resources with the intent to make informed health-focused lifestyle changes, the potential difficulty in assessing the quality of this information may curb their ability to improve own health.

2.2 Social networks and time use

Beyond direct health information effects, the use of the Internet may also affect health through social networks. Networking through the Internet could alter health-related behaviors that have been shown to be influenced by peer effects including: positive health behaviors such as exercising (Carrell et al. 2011); or negative behaviors such as drinking alcohol, smoking, or using illegal drugs (Kremer and Levy 2008, b; Lundborg 2006).¹¹

Additionally, when deciding what activities to engage in, individuals face a time constraint. Using the Internet is a fairly sedentary activity. Whether increasing time spent on the Internet leads to weight gain is in part a question of what type of activity it displaces. If more time is spent online at the expense of time spent reading a book, then we might not expect much change in body weight since both activities are similarly sedentary. If, however, an individual previously walked around a library looking for information and subsequently searches for this information online, we expect increases in body weight, all other things equal.

2.3 Income channel

An earlier literature has shown that Internet access improves wages, productivity, and growth (Kolko 2012; Akerman et al. 2015; Atasoy 2013). An expansive literature in economics has documented a positive relationship between income and health (e.g., Case and Paxon 2008). Increases in income is another mechanism by which Internet access could affect body weight. While the income-health gradient is generally positive, a number of studies show that the relationship between obesity and economic conditions is quite complicated, suggesting that the income effects of the Internet on health could be positive or negative.¹²

3 Background

3.1 Media, Internet access, and socioeconomic outcomes

There is a rich literature documenting the influence of media on socioeconomic outcomes. Researchers have found evidence that the introduction of broadcast television to a market leads to drops in voter turnout (Gentzkow 2006), and improvements in test scores (Gentzkow and Shapiro 2008). Others have found that the variety of television programming can influence political outcomes (DellaVigna and Kaplan 2007), fertility (Jensen and Oster 2009; La Ferrara, Chong, and Duryea 2012; Kearney and Levine 2015a; and Trudeau 2015), and can affect child outcomes (Kearney and Levine 2015b). Other researchers have examined the increasing availability of high-speed Internet and found it to be associated with improvements in wages and labor market opportunities (Akerman et al. 2015; Atasoy 2013; Dettling 2017; Kolko 2012). Additionally, the rollout of broadband has been linked with a wide variety of other outcomes including increases in voter turn-out (Poy and Schuller 2016); marriage market matching (Bellou 2015; Potarca 2017); reductions in teen fertility (Guldi and Herbst 2017); and increased incidence and reporting of sex crimes (Bhuller et al. 2013). Complementing this other work, some researchers have explored the relationship between the Internet and health outcomes (Bessière et al. 2010) or the demand for health care services (Baker et al. 2003; Suziedelyte 2012; Wagner et al. 2001). Yet, no paper has examined the potentially causal relationship between broadband expansion and obesity, a prominent health concern in many countries.

3.2 Obesity

Since 1980, the world obesity rate has doubled and today most of the world’s population live in countries where being obese is more likely to cause death than being underweight.¹³ Figure 1 demonstrates the trends in the rates of obesity and overweight for US adults aged 18 to 64, separately by race and gender, from 1990 to 2007. There are several patterns evident in the raw data. First, rates of obesity and overweight are lower for white women than non-white women while they are more similar for men. Second, the obesity and overweight rates rose sharply across all race and gender subgroups, affecting a large fraction of the population. By 2007, half or more of each group is classified as overweight. In response to these alarming trends, obesity has been declared as one of the leading problems in public health in the USA and other developed countries. The health costs of obesity are estimated to be at least 9.1%, or as high as 20.6% of total health costs in the USA, suggesting substantial room for cost savings through interventions that stem the cause of weight gain (Cawley and Meyerhoefer 2012; Finkelstein et al. 2009).¹⁴

Fig. 1

Trends in fraction obese and overweight: 1990–2007. Source: author's calculations using the 1990–2007 Behavioral Risk Factor Surveillance System surveys

Central explanations for the increase in American obesity are changes in food consumption or calorie expenditures. Prior work suggests that important factors in explaining the rise include the decreasing relative cost of food, the shift away from manual labor and to more sedentary work, increasing maternal labor supply, and the shift to a more sedentary lifestyle (Cawley 2011).¹⁵ Additional work suggests that a peer’s body weight in an individual’s social network may influence their own body weight, suggesting obesity may be contagious (Christakis and Fowler 2007; Cunningham et al. 2012; Fletcher 2011). Last, other work suggests that technological change may be underlying these other proposed causes.¹⁶

Theoretically, as we describe in our Conceptual Framework section, the effect of Internet access on behavioral and environmental factors related to obesity remains ambiguous. Improved information on the negative health consequences of obesity, means to achieve a healthy weight, and access to social networks that promote healthy lifestyles could decrease obesity. At the same time, false information and access to social networks promoting negative health choices along with the potential for increased sedentary lifestyle suggest greater access may increase obesity. Last, increases in income could improve body weight measures if the income-health gradient is positive for body weight. This suggests that an empirical examination of the effects could provide useful information regarding the overall effect of the increasing availability of broadband Internet on obesity.

4 Data

To proxy for Internet access, we use data on the broadband providers in a county over time. Data on broadband providers comes from the Federal Communication Commission’s Form 477. This form documents the number of providers in each zip code in each year from 1999 to 2008, and the information is consolidated into a dataset available from the Federal Communications Commission. For the purposes of our analysis, we group zip codes to county and create population-weighted variables representing the fraction of the county with at least one broadband provider.¹⁷ Although this is not a measure of individual use, it is correlated with use and serves as a good proxy for use (Guldi and Herbst 2017).

For data on health outcomes, we use the Behavioral Risk Factor Surveillance System (BRFSS) surveys from 1999 to 2007. The BRFSS is one of the largest data sets in the USA that provides information on adult health and health-related behaviors for a representative sample of non-institutionalized adults who are at least 18 years old.¹⁸ Interviews are conducted by state health departments, assisted by the US Centers for Disease Control and Prevention, through monthly telephone interviews to collect data on health and health-related behaviors. The surveys consist of a set of standard core questions, optional modules, and state-specific questions.¹⁹ We use the BRFSS to analyze the effects of county-level broadband availability on six outcomes covering weight and modifiable health behaviors that may affect weight and may change as a result of broadband availability. Our main weight measures are body mass index (BMI), an indicator for overweight status (BMI ≥ 25), and an indicator for obese status (BMI ≥ 30). In some specifications, we also look at extreme obesity (BMI ≥ 40). Health behaviors include three indicator variables for any exercise activity in the last 30 days, any binge drinking events (five or more drinks in one occasion) in the last 30 days, and whether an individual currently smokes. Figure 2 demonstrates that the proportion of the population overweight rises over the same period as the proportion of the population with access to a broadband Internet provider.

Fig. 2

Broadband coverage and share overweight: 1999–2007. Source: author’s calculations using the 1999–2007 Behavioral Risk Factor Surveillance System surveys and the 1999–2007 Federal Communications Commission Form 447 data

Using the county geographic identifiers in the BRFSS, we match individuals to our county-level broadband availability measure in each year. We limit the sample to adults age 18-64. We believe it is important to look at heterogeneity by race and gender due to cultural and behavioral differences in responses to Internet access and health interventions. To look at this heterogeneity, we further stratify the sample into white men, white women, non-white men, and non-white women. We drop observations from LA due to changes in infrastructure associated with Hurricane Katrina and observations from VA due to a large number of unmatched zip codes in the FCC data. Finally, we drop all observations without county identifiers and those missing any demographic control variables (age, gender, race, marital status, and education). This results in an unbalanced panel of counties. Because some counties are not consistently in the sample from 1999 to 2007, we restrict the sample to a balanced panel of counties as a robustness check and find that our baseline results do not substantially change (see Appendix Table 13).

Table 1 shows summary statistics for the variables used in the analysis. While the average county broadband availability increased from 68.9% in 1999 to 98.2% in 2007 in the USA as a whole, respondents in the BRFSS tend to be in counties with higher levels of broadband coverage with an average county-level broadband coverage of 95.6% in 1999 and 99.5% in 2007. The average age in the sample is approximately 40 years old, 67.7% are white, 58.7% are married, and 34.1% have at least a bachelor’s degree.

Table 1 Summary statistics for full sample, 1999–2007

We augment our BRFSS data with additional county-level covariates including the unemployment rate and real gross domestic product per capita. Table 2 shows the means of socio-economic characteristics for US counties, the subset of counties in our sample (“BRFSS counties”), and the subset of counties not in our sample (“non-BRFSS counties”). Overall, the BRFSS counties in our sample are wealthier, more urban, more educated, and spend more on social welfare than the counties not in our sample. The counties in the BRFSS, however, represent 93.9% of the US population, which suggests that the omitted counties are largely sparsely populated and our estimates represent the majority of individuals living in the USA.²⁰

Table 2 County-level socio-economic variables

5 Methods

We use within-county changes in broadband providers to identify the impact of Internet access on health outcomes in adults. We perform the analysis separately by race and gender for several reasons. First, as shown in the raw data (Fig. 1), for both levels and trends there are clear differences by race in the rates of obesity and overweight by race and gender. Second, as we mention in the conceptual framework section, during our period of study, women were more likely than men to seek out health information via the Internet. This suggests that there may be important differences by gender in the behavioral health responses to Internet access. Third, adult interactions with the health system vary by race and gender. For example, data collected by the Center for Disease Control during 1997 to 1998, contemporaneous with the beginning of our study period, demonstrates that the number of doctor visits varies by gender and race.²¹ Lastly, when attempting to lose weight, clinical evidence suggests different weight loss strategies are more effective for different gender and race groups (Jerome et al. 2015). Due to these observable differences by race and gender in the health outcomes we consider, the use of the Internet to seek out health information, and the utilization of healthcare, we expect the effects of Internet access on body weight to vary on these dimensions as well. Specifically, for each race by gender subgroup, we estimate the following reduced form model:

$$ {Y}_{\mathrm{icmt}}=\beta +{\beta}_1{\mathrm{Broadband}}_{\mathrm{c}\mathrm{t}}+{\beta}_2{X}_{\mathrm{icmt}}+{\gamma}_{\mathrm{c}}+{\lambda}_{\mathrm{m}}+{\tau}_{\mathrm{t}}+{\delta}_{\mathrm{c}\mathrm{t}}+{\epsilon}_{\mathrm{icmt}} $$

(1)

Here i indexes individuals, c indexes counties, m indexes months, and t indexes years. Broadband_ct is the percentage of zip codes in a county with at least one broadband provider, and β₁ is the coefficient of interest. County, month of the year, and year fixed effects are represented by γ_c, λ_m, and τ_t, respectively. Including these fixed effects absorbs time invariant differences in health across counties, national differences in health specific to months of the year, and national differences in health across years. δ_ct is a county-specific linear time trend, which we include in many of our specifications so that identification comes from deviations from trends within counties and over time. Y_icmt is either BMI, an indicator for overweight, an indicator for obese, or one of our other adult health outcomes of interest.²²

In Eq. 1, β₁ is identified from within county changes in health that coincide with within county changes in coverage, holding national average health within a year constant. A key assumption behind the identification of β₁ is that there are no trends in health prior to the entrance of broadband providers into counties relative to those who have not yet had an entrant. In other words, this model assumes that adult health was not improving (or declining) before a county experienced a change in broadband availability (relative to counties where at that time there was no change in providers).

To examine the validity of our assumption that the timing and degree of county broadband expansion is exogenous to county adult health, the differential trends assumption, we first turn to the careful work of other researchers who have used the same policy instrument, some of whom have examined pre-treatment trends directly for other outcomes (Atasoy 2013; Bellou 2015; Guldi and Herbst 2017; and Kolko 2012). Second, we directly test the robustness of our results to county-specific linear time trends for all of our estimates. We also test whether future broadband adoption influences past adult weight outcomes by including both leads and lags (and find that it does not). Finally, for counties we can observe in the years leading up to the rollout, we perform an analysis where we assign a “zero” for broadband in years before 1999 and find similar effects of broadband on body weight even once we add county trends. Unfortunately, due to data limitations we are unable to do a graphical analysis showing differences in treated versus untreated counties before the expansion of broadband with the data we have. While our measure of broadband availability begins in 1999, broadband providers are present to some degree in the majority of counties observable in the BRFSS data in 1999, making it difficult to distinguish the exact starting date of treatment in these counties. Relatedly, because our broadband Internet access proxy is a continuous measure of penetration, it is unclear how to define a discrete time in which broadband entrance into a county was initiated. Compounding these issues, not all BRFSS counties are observable in the pre-1999 era; and many that are available are not continuously in the data during the pre-1999 years due to the balance issues discussed in the Data section.²³ Even with this data limitation, our other, non-graphical tests, offer support of our parallel trends assumption.

Causal identification of β₁ also assumes that there are no contemporaneous unobserved changes in county policies, demographic composition, or characteristics that jointly induce broadband entrance and directly impact adult health. This assumption would be violated if, for example, county level policies designed to improve health also led to (or coincided with) the entrance of broadband providers. It is impossible to be entirely sure that this assumption holds, though we test it as rigorously as possible. Specifically, we add relevant time-varying county and individual-level observable characteristics (represented by X_icmt), to see if our estimates are sensitive to controlling for variables that would likely be correlated with unobservable changes. Individual-level controls include indicators for single year of age, education (high school, some college, and 4+ years of college), and marital status. In the non-white samples, we include an indicator for Hispanic and an indicator for all other reported races/ethnicities, using non-Hispanic black as our base group; and we control for economic conditions by including the county unemployment rate and real per capita income. In addition, we reviewed the literature on obesity prevention programs and related interventions that were plausibly changing during the time and found no reason to think these would be correlated with broadband penetration or cause a spurious relationship between broadband penetration and weight gain.²⁴ Finally, we note that any obesity prevention programs we are unaware of whose effectiveness is facilitated by Internet access will place a downward bias on our results. Therefore, even if this bias is a concern, our estimates would still rule out a zero or negative effect of broadband on weight gain in white women.

6 Results

6.1 Body weight

Table 3 shows our core results from the regressions of weight-related outcomes for the samples of white men and white women. Across the different outcomes, for white men the sign on our measure of broadband access are positive, but relatively small in magnitude and statistically insignificant. Overall, this suggests no consistent effect of Internet availability on health for white men. On the other hand, an increase in Internet availability has robust and moderately sized effects on a variety of measures of body weight for white women.²⁵ Starting at column 5, for white women a 10 percentage point increase in the fraction of the population in a county with at least one Internet provider would increase BMI by 0.1026 and the probability of being obese by 0.006. These represent effects that are 0.39 and 3.00% of the mean, respectively.

Table 3 Estimates of the effect of broadband availability on weight, white samples

Moving across the columns of Table 3 shows that the results are generally robust to a variety of alternate specifications such as adding demographic and county level controls. One concern in difference-in-differences type models is that there may be differential trends in health between counties that expand broadband access and those that do not. However, as can be seen in moving from column 7 to 8 of Table 3, there is little change in the coefficients after adding county linear trends.²⁶ Overall, we consider these estimates to be reasonably robust. We also estimated similar models for the non-white samples of men and women. These results, reported in Table 4, follow a similar pattern, though most coefficients are not statistically significant.²⁷ We also explored the estimates by age and found that the effects for white women are strongest for the youngest group (age 18 to 34), the ages with the highest rates of reported Internet use during this period.²⁸

Table 4 Estimates of the effect of broadband availability on weight, non-white samples

We explored the robustness of the body weight estimates in two additional analyses that involved extending our pre-period beyond the time for which we have broadband data available. First, we expanded our sample to also include years back to 1990. Since broadband Internet providers were virtually non-existent at this time, we set our county broadband penetration measure to zero for years 1990 to 1993 and omitted years 1994 to 1996, when broadband rollout started. This allows us to more convincingly test for robustness to county-linear trends by extending a pre-period. We show that estimated effects on body weight (see Appendix Table 16) are similar to our baseline estimates in Tables 3 and 4. Next, for white women only, we estimate a similar model but where we include the years 1994 to 1996 and assign all counties zero broadband coverage in these years (Appendix Table 17). We see that including these years leads to a slightly smaller coefficient estimate in column 2 (which includes 1994 to 1996) than in column 1 (which excludes 1994 to 1996), which is exactly what we expect due to measurement error since some of these counties are expanding broadband during this early period but are assigned zero broadband coverage. Finally, again for white women only, we estimate the effects separately for the 1999 to 2003 period, the time when the greatest rollout of broadband occurred, and for the 2004 to 2007 period, when broadband rollout was nearly completed. These estimates are reported in Appendix Table 18 and show that, indeed, the effects are concentrated in the earlier period. While Appendix Tables 17 and 18 show results for white women only, we ran these tests on the other demographic groups and found consistent results to our main estimates for these groups.

6.2 Health behaviors

If the Internet increases weight gain through increased sedentary activity, we would possibly see decreased reported exercise with Internet rollout. Similarly, exposure to lower quality health information, or the influence of expanded social networks, could increase obesity through worsening health behaviors. Though it is difficult to completely isolate any one of these mechanisms, in this section we explore changes in health behaviors as a pathway through which increased obesity occurs. During our sample period, the BRFSS consistently collects information on a number of health behaviors of interest: exercise, binge drinking, and smoking. The estimates of our model with these health behaviors as outcomes are in Table 5 (whites) and Table 6 (non-whites). Our estimates show that Internet access increases harmful health behaviors for white men and women. Specifically, we see that for white women, Internet access increases exercise, binge drinking, and smoking, although the estimated effects on binge drinking and smoking are not robust to including county linear trends.²⁹ For non-whites, however, except for binge drinking, our estimates do not provide statistically meaningful evidence on modifiable health behaviors.

Table 5 Estimates of the effect of broadband availability on health behaviors, white samples

Table 6 Estimates of the effect of broadband availability on health behaviors, non-white samples

Although our main estimates reveal that Internet access leads to weight gain, our Table 5 estimates also show that for white women broadband coverage increases exercise activity, which is generally thought to lower body weight. The exercise variable we use, however, does not capture exercise frequency or intensity. It may be that Internet access fails to improve exercise behaviors to a level of intensity that improves body weight. Another possible explanation for increasing exercise and weight gain is that white women who begin a new exercise regime in response to information about exercise on the Internet over-estimate how many calories they burn and in turn over-compensate with the calories they eat and drink. This is consistent with anecdotal evidence that some people gain weight when they start an exercise program and epidemiological work which attempts to explain why people who exercise lose less weight than expected (Miller et al. 1997; Thomas et al. 2012; Melanson et al. 2013; Dhurandhar et al. 2015).³⁰ Given the high calorie content of alcohol, increases in binge drinking are also consistent with increased weight in white women. While increases in exercise and binge drinking are present for white men (in models without county linear time trends), the estimated effect of exercise is larger than the effect of binge drinking, which may mean that any weight gain from drinking is overcome with the additional exercise.

To better understand the exercise results, we look at exercise intensity as an outcome. These results are in Table 7 (whites) and Table 8 (non-whites). Here the outcome variable is either an indicator for moderate exercise (as opposed to no exercise or vigorous exercise) or vigorous exercise (relative to no exercise or only moderate exercise). For white men, the standard errors are large making it difficult to draw a firm conclusion on how Internet changes exercise intensity. For white women, we see that for those who exercised, the broadband coverage increased moderate exercise and not vigorous exercise, which is consistent with a story of Internet access failing to improve health behaviors at a level of intensity that offsets the increased weight gain from drinking. However, a caveat to these results for white women is that they are not robust to including linear time trends. Estimated effects for non-white men and women are generally not statistically significant, except when we include linear time trends.

Table 7 Estimates of the effect of broadband availability on health behaviors, white samples

Table 8 Estimates of the effect of broadband availability on health behaviors, non-white samples

6.3 Role of income

We explore the relationship between income and broadband access by adding an interaction between our broadband variable and whether the observation was in a relatively high or low-income county (based on pre-period 1999 income levels). This allows for heterogeneous effects of broadband Internet access for those who are already living in affluent areas versus poorer areas.³¹ The results in Table 9 are not statistically significant, although there is a pattern showing that for richer counties (as measured by either being above the median, or at the 80th percentile), broadband access increases weight gain and obesity; with a larger relative effect compared to lower income counties. Potentially, this relatively lower negative effect on health for low income counties is because lower income areas had more to gain from the economic activity associated with broadband. This may suggest that there are important differences in how income interacts with broadband, but income itself is not necessarily the principal channel explaining declines in health. We are hesitant to draw strong conclusions from these estimates, however, since they lack statistical significance.

Table 9 Effect of broadband on weight interacted with county income indicators in 1999, white samples

6.4 Other potential mechanisms

In addition to health behaviors and income, we explore several other potential mechanisms. Using the BRFSS data, we were able to estimate how broadband Internet access affects the following outcomes (all indicator variables): (1) employed at the time of survey, (2) currently trying to lose weight, (3) has any type of health insurance, (4) had a routine check-up in the last 12 months, and (5) had a cholesterol check in the last 12 months. These results are in Appendix Table 19. Unfortunately, due to a lack of statistical precision, these estimates do not provide clear evidence on a single mechanism.

6.5 Falsification tests

As Fig. 2 shows, obesity rates were trending up since the 1990s and 2000s during the rollout of broadband Internet. While Figs. 1 and 2 show national trends, it is possible that there is a spurious pre-trend in body weight for counties that later increase their broadband Internet coverage. We conduct a falsification test by check the timing of broadband availability and changes in body weight by replacing our broadband measure with a three year lead of our broadband measure. More specifically, we estimate the following model for each of our three weight measures:

$$ {Y}_{\mathrm{icmt}}=\beta +{\beta}_1{\mathrm{Broadband}}_{\mathrm{c}\mathrm{t}+3}+{\beta}_2{X}_{\mathrm{icmt}}+{\gamma}_{\mathrm{c}}+{\lambda}_{\mathrm{m}}+{\tau}_{\mathrm{t}}+{\delta}_{\mathrm{c}\mathrm{t}}+{\epsilon}_{\mathrm{icmt}} $$

(2)

For this analysis, we match individual-level data from the 1996–2004 BRFSS to their corresponding county-level broadband measure 3 years in the future.³² We include the same control variables in X_icmt, fixed effects, and county-specific time trends as in Eq. 1. Results from these regressions are reported in Table 10 (whites) and Table 11 (non-whites). Our estimates are not statistically significant for any regression in either table. Furthermore, the estimates are of mixed signs and when the estimates are positive, the magnitudes are generally smaller than our main estimates (Tables 3 and 4).

Table 10 Estimates of the effect of future broadband availability (t + 3) on weight, white samples

Table 11 Estimates of the effect of future broadband availability (t + 3) on weight, non-white samples

Next, we examine this in a slightly different way by estimating a model that includes a set of 3 lags and 3 leads of the coverage variable as follows:

$$ {Y}_{\mathrm{icmt}}=\beta +\sum \limits_{n=-3}^3{\rho}_n{\mathrm{Broadband}}_{\mathrm{c}\mathrm{t}+n}+{\beta}_3{X}_{\mathrm{icmt}}+{\gamma}_{\mathrm{c}}+{\lambda}_{\mathrm{m}}+{\tau}_{\mathrm{t}}+{\delta}_{\mathrm{c}\mathrm{t}}+{\epsilon}_{\mathrm{icmt}} $$

(3)

Here, ${\sum}_{n=-3}^3{\rho}_n{\mathrm{Broadband}}_{\mathrm{ct}+n}:$ reflects the current period, as well as 1, 2, and 3 years before and after the current period. This model tests whether the current change in weight is occurring with the current change in broadband or with a change in broadband in earlier or later years. Because Broadband_ct + n, n ∈ [−3..3], are not dichotomous variables, there is not a reference/excluded group. In all of our models, effects are identified off of the relative changes in levels of penetration relative to changes in levels of obesity across both counties and years. If there were pre-trends in the model, then we would expect current changes in weight to be correlated with future levels of broadband. The results of the model for white women are reported in Table 12.³³ Across these three models, the coefficient on current broadband (Broadband_ct) is similar in magnitude to our main estimates in Table 3 and statistically significant for BMI and overweight. Furthermore, the estimates for broadband leads and lags are not statistically significant and small relative to our main estimates.

Table 12 Falsification test using 3-year leads and lags of broadband, white women

Taken together, these tests suggest that our results are not driven by a spurious correlation between broadband availability and body weight.

6.6 Multiple hypothesis testing

Another concern is that with so many outcomes tested, we might have estimated coefficients that are, by chance, statistically significant. To address the issue of multiple hypothesis testing (MHT), we use a reporting index test akin to those implemented by Anderson (2008) and Kling et al. (2007).³⁴ When dealing with multiple outcomes, the procedure to construct an index is to (1) separate outcomes into “families” of similar outcomes, (2) normalize the outcomes for “sign agreement” such that all beneficial (harmful) outcomes have a positive (negative) sign, and (3) construct a weighted average through demeaning and dividing by the control group standard deviation. This addresses MHT concerns because for each family of outcomes there is now only a single hypothesis test. Below we discuss MHT within the context of subgroups, weight-related outcomes, and health health-behavior related outcomes.

First, while we would like to do a similar correction for subgroups, the current applied microeconometric literature has only developed subgroup MHT for the experimental setting (List et al. 2016). One problem with applying a correction (such as an index) to subgroups is that it would implicitly assume that the reactions across subgroups should be the same, something we believe is unlikely to be true since there are known health and health behavior differences by gender and race, which we describe in the Conceptual Framework and Background sections. Along these lines, we believe the best evidence against concerns about MHT is that we see statistically significant effects for white women: the group that is the most likely to use the Internet to access health information and therefore the group who arguably has the highest propensity to be treated.

Second, while we could apply such an index to our “family” of weight-related outcomes (BMI, overweight, obese, and extremely obese), each of these are (essentially) already different transformations of the underlying measure of BMI. Therefore, it seems to make more sense to simply use BMI as our “summary index” in this case. As reported in Table 3, BMI is statistically significant for white women across all our models. Likewise, BMI follows the same patterns as the other weight outcomes for white women as well as the other subgroups we examine as shown in Tables 3 and 4, suggesting that BMI is a good summary measure of the weight outcomes we explore in the paper.

Last, we consider the other health outcomes we examine: exercise, drinking, and smoking. We create a single “risky health behavior” index using these variables to address MHT issues. We normalize these variables to have a mean zero and standard deviation of one in the overall sample, such that a large value for these variables represents worse health. Then we create the index by averaging these three composite variables together into a single index of negative health behaviors such that a positive coefficient reflects more drinking and smoking and less exercise. These results are reported in Appendix Table 20 and show similar inference when using the health behavior summary index of negative health behaviors as with models with each of the individually measured health outcomes where we also control for county linear trends (see Tables 5 and 6, columns 4 and 8), though we note that for white women the index measure, while positive, is no longer statistically significant and that for non-white women is now statistically significant. Taken together, our results suggest that while we provide strong evidence that broadband coverage leads to increases in white women’s body weight, we are unable to unequivocally demonstrate that this happens through changes in an index of health behaviors. However, when creating our index we have assumed that lower exercise is a worse health behavior. Yet, as we discuss above within the context of our main results, some anecdotal and epidemiological evidence suggests women gain weight with increased exercise. Therefore, it may be that our exercise estimates are, in part, responsible for the weight gains we observe.

7 Conclusion

Obesity has become an increasingly important public health issue in the USA and many other countries. Hypothesized causes for this uptick include declining relative cost of food and decreasing share of the population working in labor-intensive occupations. In this paper, we hypothesize that the Internet may also influence the obesity rate.

We use the rollout of broadband Internet providers as a plausible source of exogenous variation in Internet use to identify the effects of Internet use on obesity and body weight. We show that greater broadband coverage increases body weight and has both positive and negative effects on modifiable adult health behaviors including exercise, smoking, and drinking. A 10% increase in broadband availability increases obesity among white women by 0.0035, which represents an effect of 1.8% of the mean. Referring to Fig. 1, between 1990 and 2007, obesity increased by 15% for white women; suggesting that while the Internet is by no means the driving force behind this increase, it is at least part of the story. Our back of the envelope calculation suggests that increased medical costs for white women from obesity due to broadband Internet coverage over this time came to approximately $2.2 billion.³⁵ On the other hand, our results show that broadband coverage was not a substantial cause of increased body weight for non-white women or for men over this time. This discrepancy is likely due to differences in behavioral responses to broadband Internet availability and differences in how Internet use influences health behaviors.

How can we explain the mechanisms behind our findings that broadband Internet availability increases obesity in white women? Theoretically, if increases in broadband availability improve information, this should lead to more optimal consumer choices. However, as we show, such choices do not necessarily mean health improves: greater networking opportunities available through the Internet may result in peers having greater influence over positive or negative health behaviors. Indeed, a number of papers have linked the Internet to expanding social circles (Wellman et al. 1996; Wellman and Gulia 1999; Wellman et al.2001; Zhao 2006). While a pure information effect should decrease the likelihood of obesity, peer effects on health behaviors may work in either direction. Our estimates suggest broadband Internet availability increases drinking for men and women. While Internet use is a sedentary activity, counter to this explanation, we find some evidence of increases in exercise for men and women. While it is possible that increased exercise can lead to weight gain as appears to be occurring for white women, the large effects on exercise for white men appear to cancel out any weight increase. On the other hand, there are relatively small effects of exercise for white women. Taken together, broadband Internet coverage appears to provide both positive and negative health benefits with a net effect of increasing obesity rates for white women.

Appendix

Table 13 Effect of broadband on weight using “balanced counties” samples

Table 14 Estimates of the effect of broadband availability on weight using month by year fixed effects specifications, white samples

Table 15 Estimates of the effect of broadband availability on weight using month by year fixed effects specifications, non-white samples

Table 16 Robustness test, 1990–2007 BRFSS with broadband set to zero before 1999, excludes 1994–1996

Table 17 Robustness test including pre-broadband era BRFSS samples, white women

Table 18 Estimates of the effect of broadband availability on weight, white women samples by period

Table 19 Other potential mechanisms

Table 20 Effect of broadband coverage on negative health behavior index