Jump to ContentJump to Main Navigation
The Sympathetic StateDisaster Relief and the Origins of the American Welfare State$

Michele Landis Dauber

Print publication date: 2012

Print ISBN-13: 9780226923482

Published to Chicago Scholarship Online: January 2014

DOI: 10.7208/chicago/9780226923505.001.0001

Show Summary Details
Page of

PRINTED FROM CHICAGO SCHOLARSHIP ONLINE (www.chicago.universitypressscholarship.com). (c) Copyright University of Chicago Press, 2017. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a monograph in CHSO for personal use (for details see http://www.chicago.universitypressscholarship.com/page/privacy-policy). Subscriber: null; date: 20 November 2017

(p.231) Appendix Data, Methods, and Supplementary Tables for Chapter Seven

(p.231) Appendix Data, Methods, and Supplementary Tables for Chapter Seven

Source:
The Sympathetic State
Publisher:
University of Chicago Press

Chapter 7 is based primarily on a sample of letters to Eleanor Roosevelt from the collection Eleanor Roosevelt: Papers 1884–1964, Series 150.1: Material Assistance Requested, housed at the Franklin D, Roosevelt Library in Hyde Park, New York. The records span the period 1934–45 and contain approximately 40,000 letters. There are approximately 25,000 letters in the section from which I systematically sampled 529, essentially at random. Mrs. Roosevelt's mail was “colossal” during the 1930s and was exceeded in volume only by that addressed to her husband.1 There is no similar surviving collection of letters to President Roosevelt requesting financial assistance.2 The letters to Mrs. Roosevelt sampled for this chapter are requests for financial or other material assistance, most commonly for loans and gifts, but also for in-kind donations of clothing or other items.

I used the following sampling method: The letters were grouped by year in boxes, with the general number of letters fairly consistent across time (the archives staff has since reboxed them). I selected every fourth box, which yielded several boxes per year. This weighted the sample appropriately for variation in the number of letters per year. From each box sampled, I selected twenty-five letters by taking every tenth letter, beginning with the tenth so that if there were a few letters that had been placed on the top of a file by a prior researcher or archivist because they were thought to be especially interesting or otherwise anomalous they would be not be included. I also excluded one letter that was totally illegible and three in foreign languages. One sampled letter was excluded from most statistical analysis because it was an extreme outlier in its length. I did not sample letters written after 1940 because my preliminary investigation revealed that beginning in 1941 the letters were related to World War II rather than to requests for economic aid. Mrs. Roosevelt confirmed this; she wrote in her (p.232) autobiography that the mail received during the war years was of “an entirely different character” than it had been during the Depression.3

These 529 letters were coded for hundreds of aspects of their content, including basic demographic information such as the writer's name, gender, race, address, region, date of writing, and whether the 1930 census classified the writer's home as urban or rural.4 Using Ancestry.com, a commercial genealogy website, I located the 1930 manuscript census forms for over half (267) of the letters and was able to code the letters for aspects of the writers' households that were included in the census, including age, race, marital status, number of children, and various economic indicators.

As shown in table 7.1, the 267 writers in the census subset were largely similar to both the general population in 1930 and to known features of the writers of the overall sample. However, there were some differences that appear to be primarily attributable to the ease or difficulty of locating writers in the census database. I found a census form for the writers of 56 percent of letters written between 1933 and 1936 and 45 percent of those written between 1937 and 1940. Because of the time lag between the 1930 census and the time of writing of the letter to Eleanor Roosevelt, in some cases the census form pertains to a different household than the one in which the writer was living at the time he or she wrote the letter (for example, the writer's family of origin or, in a few cases, the writer's husband's family of origin).

Married women were overrepresented in the census subset (79 percent of female writers age fifteen and over) compared with the 1930 census enumeration (61.1 percent) and with the number of adult women in the overall sample who self-reported that they were married (58 percent), suggesting that women who were already married in 1930, and hence did not change their names between the 1930 census and the writing of their letter, were easier to find.5 Similarly, 14.2 percent of the writers were male in the overall sample of 529 letters, compared with 20 percent of the census subset, again suggesting that men may have been easier to locate than women because they did not change their names between 1930 and the date that they wrote Mrs. Roosevelt. A logistic regression, reported in table A.2 below, shows that among writers in the census subset, men were only 38 percent as likely to be married as women even after controlling for age, adding support for the theory that women who were already married in 1930 were easier to locate than those who were not.

In addition I was somewhat more likely to find the census forms of writers from rural than from urban areas. Nationally, I located a census form for 60 percent of rural writers and 47 percent of those from urban (p.233) areas, and this pattern was consistent across regions, with the exception of the West, where it was reversed. This could be because rural dwellers were less geographically mobile in the 1930s than those living in cities. While there was extensive rural-urban migration during the 1920s, the population movement out of rural areas slowed during the 1930s, presumably due to the lack of employment opportunities in cities during that decade, and then accelerated during the 1940s.6

The writers in both the full sample of 529 letters and the census subset of 267 were disproportionately female (85 percent and 80 percent, respectively). This overrepresentation of women was stable over time and did not vary by year. As noted above, 58 percent of adult female writers stated that they were married. Only 47 percent of men said that they were married (compared with 60 percent in the general population);7 however, it is not certain that a lower proportion of male writers were married because men were nearly twice as likely as women to remain silent about their marital status (27 percent versus 14 percent). Male writers in the census subset were younger than female writers,8 but as table A.2 shows, women in the census subset were much more likely to be married, even taking age into account.

Two-thirds of married women writers in the full sample made no mention of their own employment or lack of it, while 13.3 percent described themselves as unemployed for various reasons such as the Depression, pregnancy, or disability. Only 12 percent said they were working in the waged labor force. The remainder included women who said they worked together with a husband in a family farm or business. The number of employed married women writers in the census subset was also very small (5.2 percent), and not significantly different (p 〉 0.1) from the comparable proportion of working married women in the general population (9.3 percent of married native white women age fifteen and over).9

Both the full sample of letters and the census subset originated from areas of the country that closely approximated the US population distribution in 1930. The geographic distribution of the full sample of letters and the census subset tracks the regional divisions of the 1930 census distribution so closely that a chi-square goodness of fit test shows that a hypothetical random sample of this size from the 1930 US population with the geographical characteristics of the letters would not be statistically significantly different from the distribution in the 1930 census enumeration.10 Because the large number of regions used in the 1930 and 1940 censuses may lead to misleadingly large chi-square values, I also performed these tests using the four-region geographic division from the 1950 census, which yielded (p.234) the same results. This four-region division was used for the analyses in chapter seven unless otherwise noted.11

As table 7.2 shows, writers from urban, as opposed to rural, areas were somewhat overrepresented in both the full sample and the census subset, particularly in the South and West. Similarly, rural writers in the census subset were less likely to live on a farm in 1930 than the overall rural population in their regions, again particularly in the South and West. For instance, in the South Atlantic region that included Georgia, the Carolinas, and Virginia, there were very few letters in the census subset from rural-farm dwellers, despite the fact that over 37 percent of the population in that region was classified as rural-farm.12 This gap, like the general over-representation of urban writers, is probably due to the almost total lack of letters from rural blacks in those regions, many of whom lived on farms as tenant farmers, sharecroppers, and field hands. It could be that rural southern blacks did write but were so transient they were impossible to locate in the census; however, it is unlikely that if they had written in proportion to their numbers in the population that there would be virtually no such writers in the entire census subset. Moreover, the likelihood that half of black writers did not identify their race in their letters makes it impossible to test this proposition.13

It seems more plausible that rural southern blacks did not write very many of the letters. Although literacy was nearly universal among whites (both urban and rural) by 1930, among rural blacks, the rate of illiteracy was over 20 percent (26.5 percent for those living on a farm). In the south, black illiteracy was between six and ten times higher than that in other regions. Illiteracy was a problem chiefly but not only because it prevented people from writing letters; the illiterate population also did not read Mrs. Roosevelt's various newspaper columns and magazine articles about her mail and her philanthropic activities and was likely less motivated to write as a result. Rural southern blacks in 1930 also had the lowest rates of radio ownership in the country and were more isolated and poorer than others. The rural south in general lagged far behind the rest of the country in radio adoption due to the fact that sets and batteries were expensive and electricity was unavailable. The lack of a radio set, like literacy, may have impacted the relative rate of writing to Mrs. Roosevelt since she frequently mentioned both her mail and her charitable activities in her radio addresses.14

The occupational distribution of the heads of households in writers' families in the census subset was very similar to that in 1930 for the general population, except for the absence of farm laborers. The 1930 census did not use socioeconomic occupational categories, but instead divided US (p.235) workers by occupations within industries. Beginning with the 1937 unemployment census, the government began to develop socioeconomic categories (e.g., professionals, semiskilled workers, clerks and kindred workers, farmers, farm laborers, etc.) for workers across industries. In 1938 a Census Bureau researcher, Dr. Alba Edwards, revisited earlier censuses, including 1930, and recalculated occupational statistics using the new categories. Using Edwards's methodology for assigning occupational status, I applied the 1937 categories to the occupations of the heads of household of writers in the census subset, and then used Dr. Edwards's statistics to make comparisons between the occupational distribution in the census subset and the general population in 1930. These categories were further refined as part of the 1940 census, which explained the adoption of the new occupational categories and provided a guide to the occupations included in each major occupational group, similar to that used in 1937 and 1938.15

One difference between the writers and the country as a whole in terms of occupation is that the heads of writers' households in the census subset had an unemployment rate of 7.5 percent in 1930 compared with a national rate at that time of 4.7 percent, and this difference is significant (p 〈 0.05). These rates are for Class A and B unemployment, which, according to the Census Bureau, includes practically all those who would be considered unemployed “in the practically accepted meaning of the term.” Figures for head of household unemployment for classes C–G are not available; however, these groups were extremely small and would be unlikely to change the overall calculation much if at all.16

Writers in the census subset were similar to the general population with respect to home and radio ownership. Although the national radio ownership rates among writers were similar to those in the general population, urban writers were somewhat less likely to own a radio in 1930 than urban dwellers generally, while writers from the South were much more likely to own a radio in 1930 than the general southern population (28 percent of southern writers versus approximately 5 percent of southern families). This probably again reflects the lack of writers from the southern rural-farm population generally and African Americans in particular, who had the lowest rates of radio ownership in the country.17 As discussed in chapter seven, nonfarm writers' home values and rents were somewhat higher than writers' county average. It was not possible to make the same calculation for farm values because the Census Bureau lost this data.18 Writers also had a larger median family size than that in the 1930 census (4.8 versus 3.4);19 however, the difference in nonfarm housing cost was independent of family size.20

(p.236) Although the 1930 census did not contain income information, it did include other socioeconomic indicators from which an estimate of class status can be derived. An index was constructed out of the four socioeconomic variables, in which writers received one point for each of four designations: (1) writers whose 1930 head of household was in one of the top three occupations according to the classifications used in the 1937 unemployment census and in Alba Edwardss 1938 reanalysis of the 1930 occupational categories (professionals, skilled workers, and proprietors); (2) writers whose 1930 head of household owned rather than rented their homes; (3) writers whose 1930 head of household owned a radio; and (4) writers who lived in households with homes that ranked in the upper third of the scale created by the ratio of their home value or rental to the median county value or rental.

Writers who scored three or four total points on this index (37/207, or 18 percent of those for whom census data on these variables was available) were coded as “high index.” The index for low status was similar, except that it used the bottom three occupations (laborers, servants, farm laborers). Those who rented rather than owned, those who lacked a radio, and those who ranked in the bottom third on the home or rental value scale each received one point, yielding forty-six writers (22 percent) who were coded as “low index.” As discussed chapter seven, chi-square tests revealed no relationship between being “high index” and making middle-class arguments (p = 0.15) or between being “low index” and making middle-class arguments (p = 0.14).

In addition to the demographic information given by the letter itself and the census, the letters were also extensively coded for elements of their content, including the writer's description of her life and family, the type of aid requested, the various reasons the writer gave for the request, the number of words used, whether she mentioned the Depression, whether someone in the household was a veteran, or how she described Mrs. Roosevelt.

In performing the analyses reported in this chapter, I constructed twenty-nine dummy variables, coding each letter for the presence or absence of one of twenty-nine different excuses or mitigating factors. These excuses were identified and coded through an inductive process of reading the letters in order to determine how the writers “defined the categories in practice.”21 Examples of various excuses are given in chapter seven. Each letter received a code of 0 or 1 for each of the excuse dummy variables. The “excuse count” used as a dependent variable described in tables A.4A.8 is the sum of the values on these twenty-nine dummy variables. Theoretically, the range for this variable is from 0 to 29; however, in the real distribution (p.237) the range is from 0 to 12 with a mean of 4.3 and a median of 4. The mean rate of excuses per one hundred words is 1.3. Only fourteen letters out of 529 (2.6 percent) received a score of 0 on this variable. The list of excuses in the count is as follows:

  1. 1 health

  2. 2 middle-class status

  3. 3 children

  4. 4 pride

  5. 5 sudden reversal of fortunes

  6. 6 macroeconomic conditions

  7. 7 death

  8. 8 agedness

  9. 9 divorce

  10. 10 writer's prior successful work history

  11. 11 husband's prior successful work history

  12. 12 writer employed but salary too low

  13. 13 husband employed but salary too low

  14. 14 writer unemployed

  15. 15 husband unemployed

  16. 16 writer's history of unemployment

  17. 17 husband's history of unemployment

  18. 18 unemployment of others in the household

  19. 19 previous efforts to avoid relief

  20. 20 hardworking

  21. 21 inclusion of supporting documentation or references

  22. 22 writer has only part time work

  23. 23 husband has only part time work

  24. 24 husband's pay cut

  25. 25 writer's pay cut

  26. 26 writer or husband is a student

  27. 27 aged parents

  28. 28 other dependents

  29. 29 writer or husband just recently reemployed

For the logistic regressions described in tables 7.3 and A.1, some of the variables referring to a single excuse (e.g., unemployment) were collapsed into a single variable regardless of which member of the household was referenced. The three variables for unemployment (writer, husband, and other member of household) were combined into “unemployment,” as (p.238)

Table A.1. Log odds from logistic regression on various excuses (N = 502)

Health

Middle class

Children

Workin but not enogh salary

Pride

Unemployment

Macroecnomic conditios

Old age

Hardworking

Previous efforts to avoid relief

Prior successful work history

History of unemployment

Death

Sudden reversal of fortunes

Politics

Vetera status

Citize nship

Class

Entitlement

A.1

A.2

B.1

B.2

C.1

C.2

D.1

D.2

E.1

E.2

F.1

F.2

G.1

G.2

H.1

H.2

I.1

I.2

J.1

J.2

K.1

K.2

L.1

L.2

M.1

M.2

N.1

N.2

O.1

O.2

P.1

P.2

Q.1

Q.2

R.1

R.2

S.1

S.2

Urban

0.35+

0.32

0.17

0.18

−0.16

−0.16

0.33

0.26

0.33

0.25

0.28

0.30

0.26

0.14

0.08

0.03

−0.02

−0.07

0.30

0.13

0.17

0.08

0.61*

0.55*

0.67*

0.62*

0.23

0.09

−0.06

−0.06

0.26

0.25

1.23

1.22

0.08

−0.06

0.06

0.06

Female

−0.08

−0.18

−0.40

−0.30

0.36

0.39

−0.00

0.07

0.31

0.40

0.09

0.10

0.47

0.32

0.94+

0.80

−0.25

−0.15

−0.25

−0.33

0.01

0.01

−0.60*

−0.68*

0.33

0.32

−0.09

−0.17

0.29

0.23

−0.44

−0.65

−1.09*

−1.25*

0.18

0.00

0.06

−0.01

Region1

North Central

−0.05

−0.03

−0.15

−0.17

0.12

0.11

0.54

0.54*

0.54*

0.55*

0.28

0.28

0.23

0.26

0.74+

0.76+

0.36

0.35

0.03

0.07

0.42

0.43

0.52+

0.54+

0.21

0.21

0.25

0.28

0.33

0.34

−1.02+

−0.99

−0.02

0.02

0.53

0.57

0.09

0.10

South

0.38

0.40

−0.24

−0.26

0.26

0.26

0.44

0.44+

0.36

0.36

−0.04

−0.04

0.15

0.17

0.34

0.35

0.01

0.01

−0.02

− 0.01

0.05

0.04

0.63*

0.64*

0.34

0.34

0.11

0.11

0.22

0.23

−0.03

0.01

−0.65

−0.60

0.55

0.57

0.04

0.05

West

−0.12

−0.10

−0.24

−0.26

0.07

0.06

0.15

0.16

0.39

0.42

0.02

0.02

−0.03

0.02

0.34

0.35

0.01

0.00

−0.40

−0.37

0.36

0.38

0.01

0.03

−0.52

−0.50

−0.35

−0.32

−0.26

−0.25

−0.07

0.01

−0.81

−0.70

−0.09

−0.05

−0.14

−0.13

1937–40 Period2

−0.01

−0.01

−0.11

−0.11

−0.01

0.00

0.20

0.21

−0.03

−0.01

−0.24

−0.24

−0.83**

−0.87**

0.51+

0.50+

−0.25

−0.24

0.34

0.36

−0.83***

− 0.84***

−0.68**

−0.69***

0.05

0.05

−1.05***

−1.07***

0.01

0.00

−0.41

−0.43

0.03

0.04

0.27

0.27

−0.16

−0.16

Pay back debt

0.43+

−0.24

−0.06

0.34

0.35

−0.14

1.07***

0.54

0.16

1.05***

0.59*

0.53*

0.31

0.99**

0.14

0.79+

0.42

0.92+

0.18

Loan

−0.26

0.34+

−0.09

0.52**

0.58**

−0.02

−0.17

−0.32

0.61*

0.42+

0.36

−0.06

0.14

0.29

−0.19

−0.73

−0.64

−0.20

−0.25

Constant

−0.42

−0.32

0.29

0.13

−0.61+

−0.65+

−1.01**

−1.32***

−1.79***

−2.16***

−0.84***

−0.82*

−2.06***

2.08***

−3.66***

−3.52***

−1.68***

−2.03***

−1.56***

− 1.84***

−1.43***

−1.67***

−1.02**

−1.02**

−2.22***

−2.31***

−1.57***

−1.80***

−1.78***

−1.68***

−2.32**

−2.12**

−3.19***

−2.96***

−4.13***

−4.08***

−1.15**

−1.03**

LR Chi−Square

6.92

11.55

5.77

9.62

4.78

5.03

9.62

20.84**

7.02

18.35**

6.00

6.38

13.34*

27.20***

10.25

13.33

3.16

8.89

6.17

27.86***

15.55*

23.44**

25.50***

29.89***

12.16+

14.21+

16.69+

30.08***

3.45

4.26

5.51

9.96

11.67+

13.61+

1.42

4.01

1.03

2.58

D.F.

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

6

8

LR Model 1 vs.Model (df2)

4.63+

3.84

.25

11.22**

11.33**

.38

13.87**

3.07

5.73+

21.69***

7.89*

4.39

2.05

**

.8

4.45

1.94

2.59

1.55

(+) p 〈 0.1,

(*) p 〈 0.1,

(**) p 〈 0.1,

(***) p 〈 0.1,

(1) The reference group is “North East.”

(2) The reference group is “1933–36 Period”

(p.239) (p.240) (p.241) (p.242) were the variables for “working but salary not enough,” “prior successful work history” and “history of unemployment.” For those analyses I was less interested in the number of excuses in a particular letter than in testing whether writers with particular characteristics might be more or less likely to discuss unemployment, perhaps due to gender or to regional or temporal fluctuations in the unemployment rate.

Claims concerning farm conditions were coded as follows: claims to having owned or operated a farm in the past were coded as “prior successful work history,” farm and crop failures were generally coded as “sudden reversal of fortune,” past farm losses were coded as and claims that an operating farm no longer produced sufficient income were coded as “working but salary not enough.”

Table A.1 is an expanded version of table 7.3 (table A.1, above, displays coefficients rather than odds ratios as shown in table 7.3). Findings from this table are discussed in chapter seven. The dependent variables are eighteen of the various excuses that writers use in their letters and are displayed across the top of the table. Included are the fourteen most frequently used excuses, as well as justifications for aid that rely on politics, veteran status, citizenship, and appeals to class, and one additional dependent variable, “entitlement” (S.l, S.2 in table A.1), which is a combination of the dependent variables veteran, citizenship, and politics such that a letter was coded as “1” on “entitlement” if it contained any of these three excuses.22 For each dependent variable, the first model (A.1, B.1, etc.) is a baseline model including variables for whether the writer was from an urban or a rural area according to the 1930 census, the writer's gender, the writer's geographic region,23 and the time period in which the letter was written.24 The second model (A2, B2, etc.) for each dependent variable adds two explanatory variables: (1) whether a writer was requesting a loan and (2) whether a writer was requesting money to get out of debt.25 Logistic regression was used because the dependent variable (odds of using a particular justification in a given letter) is a binary variable with a binomial distribution. These logistic regression models take the form:

logit(Y)=log(π/(1π))=α+β1x1+β2x2+...+βixi

Where log represents the natural logarithm, π is the probability that the dichotomous outcome variable Y = 1 (e.g., that a writer used the justification for aid that her husband was working but his wages were too low to support the family); α is the Y intercept; βs are regression coefficients; and x1 to x1 are vectors of the independent variables (e.g., gender, urban, region, debt) included in the model.

(p.243) As table A.1 indicates, the models are nested. Tests of the goodness of fit of these models compared a relatively more complex model with a simpler model to see if the more complex model fit the dataset significantly better. In most cases, neither model offered an improvement over the null model. For example, neither model fit the data well in predicting whether a writer would discuss health, middle-class status, or children, possibly because these justifications appeared nearly universally in the letters.

In many cases the expanded model with the two additional variables (e.g., D.2) fit the data significantly better than the baseline model (e.g., D.l). In these cases both the significant coefficients for the explanatory variables and the significant likelihood ratio chi-square statistics supported the conclusion that the model containing the explanatory variables better fit the data than the baseline model. This was the case in predicting whether a writer would discuss: working but salary not enough (D. 1 v. D.2, p 〈 0.01); pride (E.l v. E.2, p 〈 0.01); macroeconomic conditions (G.l v. G.2, p 〈 0.01); previous efforts to avoid relief (J.l v. J.2, p 〈 0.001); prior successful work history (K.l v. K.2, p 〈 0.05), sudden reversal of fortunes (N.l v. N.2, p 〈 0.01).

In the case of history of unemployment, both models fit the data well, and the explanatory variable for loan was significant (p 〈 0.05); however, comparing model L.l and L.2, the likelihood ratio is 4.39 with 2 degrees of freedom, and shows that the addition of the explanatory variables was not a significant improvement (p 〉 0.1) over the baseline model.

The model containing the additional two explanatory variables (e.g., D.2) tests the effects of known demographic information about writers as well as features of the letter (asking for loans or debt relief) as predictors of whether a particular excuse was used to seek aid in a given letter. Generally, coefficients for the variables representing demographic characteristics were not significant. However, there were a few exceptions. Men were more likely to mention a history of unemployment than women, and writers living in urban areas were more likely to discuss a history of unemployment than writers from rural areas. As discussed in chapter seven, letters written in the second half of the decade were less likely to mention macroeconomic conditions, prior successful work history, a history of unemployment, and a sudden reversal of fortune.

The dependent variable in table A.2 is the odds of being married in 1930 for writers in the census sample who were over age fifteen in 1930. The explanatory variable is the writer's gender, and the control variables are the writer's age (in years) in 1930 and an age squared term to control for the possibility that the relationship between age and marriage is curvilinear. (p.244)

Table A.2. Odds ratios from logistic regression of gender and age on marital status (N = 228)

Model 1

Model 2

Male

0.41*

Age

1.69***

1.66***

Age Squared

0.99***

0.99***

LR Chi-Square

57.91***

62.21***

Degrees of freedom

2

3

(*) p 〈 0.05

** p 〈 0.05

(***) p 〈 0.05

Logistic regression was used because the dependent variable (odds of being married) is a binary variable with a binomial distribution. The logistic regression models in table A.2 take the form described in the discussion for table A.1 above:

Where log represents the natural logarithm, π is the probability that the di-chotomous outcome variable Y = 1 (e.g., that a writer is married); α is the Y intercept; βs are regression coefficients; x1 is the vector of the independent variable gender, and x2 is the vector of the independent variable age, and x3 is the vector of the independent variable age squared.

The chi-square statistics suggest that both models are significant improvements over the null model. As the table indicates, the models are nested. Comparing model 2 with model 1 resulted in a LR chi-square statistic of 4.30 with 1 degree of freedom indicating that model 2 fits the data significantly better than model 1 (p 〈 0.05).

Model 1 is a baseline model and includes age as a predictor for marital status. The log likelihood of this model is −99.257481. Not surprisingly, the coefficient for age is positive and significant and the coefficient for age squared is negative and significant, indicating that the relationship of age to marriage is curvilinear. Model 2 tests the effect of gender on marital status and adds the writers sex to the baseline model. All three coefficients in the model are significant (not shown). The log likelihood of model 2 is −97.105302. In this model, being older and being female both significantly increased the probability of a writer being married. Holding age constant, being male decreased the odds that a writer would be married by a factor of 0.41, or 59 percent. Conversely, being female more than doubled the odds that a writer would be married.

In table A.3, logistic regression was also used to test whether features (p.245) of the writers' circumstances gleaned from the census data were related to making certain arguments. Here I tested whether writers within the census subset who had some indicators that they were better off financially were more likely to make arguments and excuses tied to being middle class. The logistic regression models in table A.3 take the form described in the discussion for table A.1 above:

Where log represents the natural logarithm, π is the probability that the dichotomous outcome variable Y = 1 (i.e., that a writer will claim to be a member of the middle class); α is the Y intercept; βs are regression coefficients; and x1 to xi are vectors of the independent variables (e.g., gender, urban, region, socioeconomic factors) included in the model.

The four socioeconomic variables available from the 1930 census used as proxies for class status are: (1) writers whose 1930 head of household was in one of the top three occupations according to the classifications used in the 1937 unemployment census and in Alba Edwards's 1938 reanalysis of the 1930 occupational categories (professionals, skilled workers, and proprietors); (2) writers whose 1930 head of household owned rather than rented their homes; (3) writers whose 1930 head of household owned a radio; and (4) writers who in 1930 lived in households with homes with a value or monthly rent above the median for their county of residence, i.e., those living in a better neighborhood.

The dependent variable in table A.3 is the odds that a writer will make an argument or excuse related to being middle or upper class. Examples of such excuses are given in chapter seven and include such things as stating that “we are one of the South's best families” or referencing objects and practices associated with class status such as professional standing, servants, and expensive furnishings. The explanatory variables are the four socioeconomic factors relating to financial circumstances of the writer's 1930 household. The control variables are for the writer living in an urban, as opposed to rural, area at the time of writing (as designated by the Census Bureau), the writer's gender, the region in which they resided at the time of writing, and whether the letter was written in the first or second half of the decade.

The two models are nested. Model 1 is a baseline model, containing the control variables for urban/rural residence, sex, region, and time of writing. None of the coefficients were significant. Model 2 tested the effect of the four dummy variables for socioeconomic status on making excuses and arguments connected to middle-class status. Only the coefficient for having a high relative home or rental value was significant (p 〈 0.05). None of (p.246)

Table A.3. Odds ratio from logistic regression on the making of middle-class arguments and excuses (N = 203)

Model 1

Model 2

Urban

1.38

1.27

Female

0.82

0.78

Region1

North Central

0.66

0.65

South

0.68

0.54

West

0.89

0.83

1937–40 Period2

0.83

0.87

Socioeconomic Factors

High Status Occupation

1.26

Owning a Home

0.92

Owning a Radio

1.16

Above Median Home or Rental Value

2.04*

LR Chi-Square

3.89

9.35

Degrees of Freedom

6

10

(*) p 〈 0.05

(1) The reference group is “North East.” The four-region system established in the 1950 census was used rather than the 7-region format in use in 1930 in order to reduce the degrees of freedom used in the analysis. The reference group for region is “North East.” The overall effect of the four dummy variable set for region was tested using a LR chi-square test between a model containing the region dummies and one that did not. The inclusion of the variables for region did not significantly improve the model fit.

(2) The reference group is “1933–36 Period”

the other variables that are indicators of being better off, such as owning a home or having a professional occupation, were associated with claiming to be middle class.

As the nonsignificant LR chi-square tests in table A.3 indicate, neither model fit the data well. The difference in LR chi-square between model 1 and model 2 is 7.72 which was not significant (p 〉 0.1), showing that adding the four socioeconomic factors did not improve the model fit. Overall, writers who were objectively better off (based on the available information about writers' class status) were no more likely than other, less well-off, writers to mention their class status as a basis for seeking assistance. However, as indicated by the significant coefficient for high relative home or rental value, the odds that writers who lived in a better neighborhood compared to the rest of their county of residence would make middle-class claims were twice as high as those of other writers (p 〈 0.05). This may be due to the fact that the nature of the request required writers to disclose (p.247) their address to Mrs. Roosevelt so that she could send a check. Because the writer's address necessarily revealed the neighborhood in which she lived, it carried with it the potential to refute a claim to be well-to-do, particularly if the letter were to be forwarded to local relief officials for investigation. For certain writers, such as those who resided in slums or poor rural districts, this may have inhibited them from claiming to belong to the middle or upper class.

In table A.4, the dependent variable is a count variable consisting of the number of excuses used in each letter. This variable consisted of twentynine excuses, constructed as described above. The control variables include the same demographic variables used in the logistic regressions discussed above: residing in an urban as opposed to rural area, the writer's gender, the region of residence, and the time period in which the letter was written. The explanatory variables related to debts and loans are also included in order to test their effect on the number of excuses.

Because the outcome variable, the number of excuses in a given letter, is a count variable, event count models rather than linear regression models were used. The Poisson distribution is a discrete probability distribution that expresses the probability of a number of events occurring in a fixed period of time. In this distribution the mean number of occurrences is equal to the variance. Where, as here, the variance is greater than the mean, then the data suffer from overdispersion, and the negative binomial distribution is preferred because it includes an ancillary parameter a that reflects unobserved heterogeneity among observations.26 Where α is zero, negative binomial regression is simply a Poisson model. Where α is greater than zero, the negative binomial model is more suitable. The negative binomial distribution is also a discrete probability distribution:

Pr(y|x)=Γ(γ+α1)γ|Γ(α1)(α1α1+λ)(λα1+λ)γ

For each of the models displayed in table A.4, the LR test of α suggests an alpha greater than 0 (p 〈 0.001). Therefore, the negative binomial model is preferred over the Poisson.

As table A.4 shows, the models are nested such that model 1 is nested in model 2, in order to compare a relatively more complex model to a simpler model to see if the more complex model fits the dataset significantly better. Between model 1 and model 2, the likelihood ratio is 30.54 with 2 degrees of freedom, and is statistically significant (p 〈 0.001). This means that adding the two explanatory variables (debt and loan) significantly improved (p.248)

Table A.4. Incident-rate ratios from negative binomial model on the number of excuses (N = 502)

Model 1

Model 2

Urban

1.16*

1.11

Female

1.02

1.01

Region1

North Central

1.17*

1.17*

South

1.14

1.14

West

1.03

1.04

1937–40 Period2

0.88*

0.88**

Pay back debt

1.31***

Loan

1.15**

α

0.11***

0.08***

LR Chi-Square

17.86**

48.41***

Degrees of Freedom

6

8

LR chi-square test Model 1 vs.

30.54(2)***

Model 2

(*) p 〈 0.05,

(**) p 〈 0.01,

(***) p 〈 0.001,

(1) The reference group is “North East.” Writers from the Midwest may have used more excuses than those from the northeast, as the coefficient for the control variable “north central” was significant (p 〈 0.05). The overall effect of the four-dummy variable set for region was tested using a LR chi-square test between a model containing the region dummies and one that did not. The inclusion of the variables for region did not significantly improve the model fitting. In that case, the likelihood ratio is 6.67, with 3 degrees of freedom and is not statistically significant. The coefficients for the explanatory variables of debt and loan were essentially identical in the model without the region dummies and in the model that included them.

(2) The reference group is “1933–36 Period”

the model fitting. Both the significant coefficients for these variables and the significant likelihood ratio statistic support the conclusion that model 2 is the model that best fits the data.

As table A.4 indicates, both of the explanatory variables are strong predictors of the number of excuses even when the control variables are included in the model. Asking for debt relief and asking for a loan rather than another form of aid increased the number of excuses by 31 percent and 15 percent, respectively, holding the control variables constant. Consistent with the findings reported in tables 7.3 and A.1 showing that writers in the latter half of the decade were less prone to use certain excuses, writers in the second half of the decade used 12 percent fewer excuses than those who wrote earlier.

Table A.5 uses letters from female writers only, and tests (in addition to the demographic and explanatory variables used in table A.4) the effect (p.249)

Table A.5. Incident-rate ratios from negative binomial model on the number of excuses, for female writers only (N = 415)

Model 1

Model 2

Model 3

Urban

1.14*

1.09

1.10

Region1

North Central

1.13

1.14

1.13

South

1.12

1.13

1.14

West

1.04

1.05

1.04

1937–40 Period2

0.90

0.90

0.90

Pay back debt

1.34***

1.32***

Loan

1.12*

1.13*

Married women

1.18**

α

0.12***

0.09***

0.09***

LR Chi-Square

10.41

38.51***

46.93***

Degree of Freedom

5

7

8

LR chi-sq (df)

28.10(2)***

8.42(1)**

(*) p 〈 0.05,

(**) p 〈 0.01,

(***) p 〈 0.001,

(1) The reference group is “North East.” The overall effect of the four-dummy variable set for region was tested using a LR chi-square test between a model containing the region dummies and one that did not. The inclusion of the variables for region did not significantly improve the model fitting. In that case, the likelihood ratio is 3.99, with 3 degrees of freedom and is not statistically significant. The coefficients for the explanatory variables of debt, loan, and marriage were essentially identical in the model without the region dummies and in the model that included them.

(2) The reference group is “1933–36 Period”

of being married on the number of excuses deployed in each letter. Married women have a mean number of excuses of 4.69, compared with 3.95 for unmarried women, and that difference was significant using a t-test for samples of unequal variance (p 〈 0.01). Since Levene's test suggests unequal variances, Satterthwaite's estimates that do not assume equal variance were used. The dependent variable is a count variable consisting of the number of excuses used in each letter. The control variables include the same demographic variables used in the logistic and negative binomial regressions discussed above: residing in an urban as opposed to rural area, the region of residence, and the time period in which the letter was written. The explanatory variables related to debts and loans were included. In addition, because table A.5 tests the effect of marriage on the excuse count of female writers, a dummy variable for “married” was added as an additional explanatory variable.

As in table A.4 above, the outcome variable, the number of excuses in a given letter, is a count variable. Therefore, event count models rather than (p.250) linear regression models are used. For each of the models displayed in table A.5, the LR test of α suggests an alpha greater than 0 (p 〈 0.001). Therefore, the negative binomial model is preferred over the Poisson.

As table A.5 shows, the three models are nested. The chi-square statistics suggest that model 2 and model 3 are significant improvements over the null model (p 〈 0.001). Model 1 is the baseline model, containing the control variables for urban/rural residence, region, and time of writing. Model 2 adds the explanatory variables of asking for debt relief or a loan, both of which are positive and significant. Model 3 tests the effect of marriage on the number of excuses deployed by female writers and adds to model 2 a dummy variable for marriage, which along with the variables for debt and loan, is also positive and significant (p 〈 0.01). Comparing model 2 with baseline model 1 resulted in a LR chi-square statistic of 28.10 with 2 degrees of freedom (p 〈 0.001). Comparing model 3, which added marital status, with model 2 resulted in a LR chi-square statistic of 8.42 with one degree of freedom (p 〈 0.01). As shown in table A.5 above, the results of the LR chi-squared tests show that model 3 fits the data significantly better than the other two models.

Table A.5 displays the incident rate ratios for all three models. The results of model 3 show that for female writers, describing themselves as indebted, asking for loans, and being married all predicted an increase in the number of excuses. Writers requesting debt relief used an average of 32 percent more excuses than others, while those asking for loans (as opposed to other forms of aid, such as gifts) used 13 percent more excuses. Holding constant all of the control variables and other explanatory variables, being married increased the number of excuses deployed by female writers by 18 percent.

Table A.6 uses letters from the census subset because these are the only writers for whom the 1930 census occupational data are available. As in tables A.4 and A.5, the dependent variable is a count variable consisting of the number of excuses used in each letter. The control variables include the same demographic variables used in the logistic and negative binomial regressions discussed above: residing in an urban as opposed to rural area, the region of residence, and the time period in which the letter was written. Table A.6 tests the effect of the explanatory variables loan and debt on the excuse count of these writers when considering the effect of covariates for both the length of the letter and having a high occupational status as proxies for literacy. A letter received a “1” for high occupational status if the writer's 1930 head of household was in one of the top three occupations according to the classifications used in the 1937 unemployment census (p.251)

Table A.6. Incident-rate ratios from negative binomial model on the number of excuses for census subset only (N = 203)

Model 1

Model 2

Female

1.08

1.10

Urban

1.11

1.09

Region1

North Central

1.34**

1.33**

South

1.22*

1.22*

West

1.06

1.06

1937–40 Period2

0.86*

0.87*

Word count

1.00***

1.00***

High status occupation

1.14+

1.12

Pay back debt

1.21**

Loan

1.12+

α

0.05**

0.04*

LR Chi-Square

50.34***

60.76***

Degrees of Freedom

8

10

LR chi-sq test M.1 vs. M.2

10.42(2)**

(+) p 〈 0.10,

(*) p 〈 0.05,

(**) p 〈 0.01,

(***) p 〈 0.001,

(1) The reference group is “North East.” The overall effect of the four-dummy variable set for region was tested using a LR chi-square test between a model containing the region dummies and one that did not. The inclusion of the variables for region significantly improved the model fit (p 〈 0.01). In that case, the likelihood ratio is 11.72, with 3 degrees of freedom. The coefficients for the explanatory variables of debt and loan were essentially identical in the model without the region dummies and in the model that included them.

(2) The reference group is “1933–36 Period”

and in Alba Edwards's 1938 reanalysis of the 1930 occupational categories (professionals, skilled workers, and proprietors).

Because the outcome variable is a count variable, event count models rather than linear regression models are used to analyze the relationship between the covariates. For both model 1 and model 2, the LR test of α suggests an alpha greater than 0 (p 〈 0.01). Therefore, the negative binomial model is preferred over the Poisson.

As table A.6 shows, the two models are nested. The chi-square statistics suggest that both models are significant improvements over the null model (p 〈 0.001). Model 1 is the baseline model, containing the control variables for urban/rural residence, region, and time of writing, as well as the additional controls for high status occupation and word count. Model 2 adds the explanatory variables of asking for debt relief or a loan. Comparing model 2 with baseline model 1 resulted in a LR chi-square statistic of (p.252) 10.42 with 2 degrees of freedom (p 〈 0.01). Thus adding the explanatory variables (debt and loan) significantly improved the model fit.

Table A.6 displays the incident rate ratios for both models. The results of model 2 show that for the smaller group of writers in the census subset, having a high status occupation is not significant, while the length of the letter is positively and significantly associated with the number of excuses. But even when all of the controls intended to approximate literacy, as well as the other controls, are held constant, writers requesting debt relief used an average of 21 percent more excuses (p 〈 0.01) than other writers, and writers asking for a loan used 12 percent more excuses on average than those requesting aid in the form of a gift, though this result was only marginally significant (p 〈 0.1).

Table A.7 tests the effects of literacy on the number of excuses for the full sample. As in tables A.4A.6 above, the dependent variable is a count variable consisting of the number of excuses used in each letter. The control variables include the same demographic variables used in the logistic and negative binomial regressions discussed above: gender, residing in an urban as opposed to rural area, the region of residence, and the time period in which the letter was written. An additional control variable for the length of the letter (word count) is included.

Because the outcome variable is a count variable, event count models rather than linear regression models are used. For both models, the LR test of α suggests an alpha greater than 0 (p 〈 0.001). Therefore, the negative binomial model is preferred over the Poisson.

As table A.7 shows, the two models are nested. The chi-square statistics suggest that both models are significant improvements over the null model (p 〈 0.001). Model 1 is the baseline model, containing the control variables for urban/rural residence, gender, region, and time of writing as well as an additional control for word count. Model 2 adds the explanatory variables of asking for debt relief or a loan. Comparing model 2 with model 1 resulted in a LR chi-square statistic of 21.62 with 2 degrees of freedom (p 〈 0.001). Thus adding the explanatory variables (debt and loan) significantly improved the model fit.

Table A.7 displays the incident rate ratios for both models. The results of model 2 show that the length of the letter was positively and significantly associated with the number of excuses, though other possible predictors for literacy such as being from an urban area or from the Northeast were not associated with an increase in the number of excuses. But even when the controls intended to approximate literacy are held constant, the explanatory variables are still strongly predictive and significant. Writers asking (p.253)

Table A.7. Incident-rate ratios from negative binomial model on the number of excuses controlling for length of letter (N = 502)

Model 1

Model 2

Urban

1.11

1.08

Female

1.03

1.04

Region1

North Central

1.22**

1.21**

South

1.16*

1.16*

West

0.97

0.97

1937–40 Period2

0.88**

0.88**

Word count

1.00***

1.00***

Pay back debt

1.20**

Loan

1.16**

α

0.07***

0.05***

LR Chi-Square

97.39***

119.01***

Degrees of Freedom

7

9

LR chi-sq test Ml. ys. M2

21.62(2)***

(*) p 〈 0.05,

(**) p 〈 0.01,

(***) p 〈 0.001,

(1) The reference group is “North East.” The overall effect of the four-dummy variable set for region was tested using a LR chi-square test between a model containing the region dummies and one that did not. The inclusion of the variables for region significantly improved the model fitting. In that case, the likelihood ratio is 13.28 with 3 degrees of freedom and is statistically significant (p 〈 0.01). However, the coefficients for the explanatory variables of debt and loan (as well as for the other controls) were essentially identical in the model without the region dummies and in the model that included them.

(2) The reference group is “1933–36 Period”

for debt relief and loans used 20 percent and 16 percent more excuses, respectively, than others regardless of the length of the letter. Therefore, even taking into account the effects of literacy (to the extent possible with the data available) the explanatory variables continued to exert an effect on the number of excuses.

Table A.8 repeats the analysis reported in table A.7 but for female writers only. An additional explanatory variable for marriage is added in order to gauge the effect of controlling for literacy on the finding from table A.5 that married women writers use more excuses in their letters than unmarried women. For all three models, the LR test of α suggests an alpha greater than 0 (p 〈 0.001). Therefore, the negative binomial model is preferred over the Poisson.

As table A.8 shows, the three models are nested. The chi-square statistics suggest that all three models are significant improvements over the null (p.254)

Table A.8. Incident-rate ratios from negative binomial model on the number of excuses controlling for length of letter for female writers only (N = 427)

Model 1

Model 2

Model 3

Urban

1.08

1.05

1.06

Region1

North Central

1.19*

1.19*

1.18*

South

1.16*

1.16*

1.17*

West

0.96

0.98

0.97

1937–40 Period2

0.90

0.90

0.91

Word count

1.00***

1.00***

1.00***

Pay back debt

1.22**

1.20**

Loan

1.14*

1.15**

Married women

1.18**

a

0.08***

0.06***

0.06***

LR Chi-Sauare

73.36***

92.15***

101.78***

Degrees of Freedom

6

8

9

LR chi-sq test (df)

18.79(2)***

9.63(1)***

(*) p 〈 0.05,

(**) p 〈 0.01,

(***) p 〈 0.001,

(1) The reference group is “North East.” The overall effect of the four-dummy variable set for region was tested using a LR chi-square test between a model containing the region dummies and one that did not. The inclusion of the variables for region significantly improved the model fit. In that case, the likelihood ratio is 9.35 with 3 degrees of freedom and is statistically significant (p 〈 0.05). However, the coefficients for the explanatory variables of debt, loan, and marriage (as well as the other controls) were essentially identical in the model without the region dummies and in the model that included them.

(2) The reference group is “1933–36 Period”

model (p 〈 0.001). Model 1 is the baseline model, containing the control variables for urban/rural residence, region, and time of writing, and including an additional control for word count. Model 2 adds the explanatory variables of asking for debt relief or a loan, both of which are positive and significant. Model 3 tests the effect of marriage on the number of excuses deployed by female writers and adds to model 2 a dummy variable for marriage, which along with the variables for debt and loan, is also positive and significant (p 〈 0.01).

As shown in table A.8, adding the explanatory variables debt and loan significantly improved the model fit (p 〈 0.001). Model 3, which includes marital status, fit the data better than the other two models (p 〈 0.01). Model 3 shows that even when possible controls for literacy such as Northeast region, urban, and word count are held constant, the effect of asking for debt relief or a loan, and of being married for women writers, continued to be positive and significant. Holding constant all of the control variables and other explanatory variables, being married increased the number of (p.255)

Table A.9. Coefficients from linear regression on logged word count (N = 502)

Model 1

Model 2

Urban

0.06

0.05

Female

−0.15*

−0.16*

Region1

North Central

−0.12

−0.11

South

−0.05

−.005

West

0.07

0.08

1937–40 Period2

0.01

0.01

Number of Excuses

0.13***

0.13***

Debt

0.11

Loan

.00

Constant

5.33***

5.34***

F Score

28.97***

22.96***

Degrees of Freedom

7

9

R Square

0.29

.30

Incremental F Test M1 vs. M2 (df)

1.66(2)

(*) p 〈 0.05,

** p 〈 0.01,

(***) p 〈 0.001 〈!set angle brackets as less-than signs!〉

(1) The reference group is “North East.” The overall effect of the four-dummy variable set for region was tested using an incremental F test for the null-hypothesis that the set of region variables had no effect on the dependent variable of log (word count). According to this test, the inclusion of the variables for region did not significantly improve the model fit.

(2) The reference group is “1933–36 Period”

excuses deployed by female writers by 18 percent regardless of the length of the letter.

Table A.9 reports the results of a linear regression on the log of the word count for each letter in order to estimate the effect of the writer's gender on the length of the letter, holding constant a number of demographic and letter content variables, including the number of excuses. The dependent variable was logged both to address outliers in the data and for interpret-ability in terms of percent change. The regression has the form:

Y=a+b1X1+b2X2+...biXi

Where Y is the natural log of the word count, a is the constant, x is a vector of the independent variable (e.g., urban, female, region, number of excuses), and b is a vector of estimated parameters.

Model 1 includes a set of demographic variables such as the writer's sex and region, as well as a variable for the number of excuses in each letter. Model 2 adds two additional content variables for debt relief and loans (p.256)

Table A.10. Coefficients from linear regression on number of excuses per 100 words (N = 502)Urban

Urban

0.06

Female

0.27*

Region1

North Central

0.35***

South

−0.16

West

0.08

1937–40 Period2

−0.08

Number of Excuses

0.13***

Pay back debt

0.07

Loan

0.19***

Constant

0.86***

F Score

3.76***

Degrees of Freedom

8

R Square

0.06

(*) p 〈 0.05,

** p 〈 0.01,

(***) p 〈 0.001,

(1) The reference group is “North East.” The overall effect of the fourdummy variable set for region was tested using an incremental F test for the null-hypothesis that the set of region variables had no effect on the dependent variable of density 100 (number of excuses per 100 words). This test yielded an F of 5.89 with 3 degrees of freedom. According to this test, the inclusion of the variables for region signifi-cantly improved the model fit. The significant coefficient (p 〈 0.001) for North Central suggests that writers from the Midwest region had a greater rate of excuses per 100 words than those from the northeast.

(2) The reference group is “1933–36 Period”

in order to test the effect of these variables on the length of the letter and to ensure that the effect of gender is independent of the effect of both the number of excuses and of the use of particular excuses. The F score of 28.97 for model 1 indicates that this model is a significant improvement over the null model (p 〈 0.001). Overall the independent variables explain 29 percent of the variation in the dependent variable. The incremental F test shows that model 2 is not a significant improvement over model 1. As discussed in chapter seven, the results of model 1 show that women wrote letters that were shorter than those of male writers, holding constant the number of excuses and the other variables.

In table A.10, the dependent variable is the number of excuses per one hundred words. As discussed in chapter seven, men had a mean rate of excuses per one hundred words of 1.14, while for women the mean rate was 1.36, and this difference was significant (p 〈 0.01). (Since Levene's test suggests unequal variance, Satterthwaite's estimates that do not assume equal variance were used). Table A.10 reports the results of a linear regression on the rate of excuses per 100 words to test whether a difference between (p.257) men and women on this measure persists when the effect of other covariates such as the number of excuses and the use of certain excuses is considered.

The overall fit of the model is good. The F score of 3.76 indicates that this model is a significant improvement over the null model. Overall, the independent variables explain only about 6 percent of the variation in the dependent variable. An incremental F test on the joint significance of the two content variables for debt relief and requests for loans indicates that the addition of these variables significantly improved the model fit (p 〈 0.05). The positive and significant coefficient for “female” indicates that letters from women had a higher rate of excuses per one hundred words than did those written by men holding constant the other covariates in the model, including those that might be related to literacy, such as being from an urban area or from the Northeast rather than the South, as well as those related to the type and purpose of the request. (p.258)

Notes:

(1) . Hellman, “Mrs. Roosevelt,” 70.

(2) . Most letters to FDR of that type were not filed by subject by the White House mail-room staff and were not retained by archivists. However, a number of observers have estimated that the bulk of FDR's mail was comprised of individual requests for assistance like those reported on in this chapter. See Jasper Mayer and Mary Roos, Analysis of General Run of All Mail Addressed to President Roosevelt, by Correspondence Division, Federal Emergency Relief Administration, received at White House between March 23–27, 1934, President's Official File (OF 50), Box 4, Franklin and Eleanor Roosevelt Library, Hyde Park, New York; Whitehurst, “Dear Mr. President,” 12; Smith and Morris, “DearMr. President,” 213–14; Hellman, “Mrs. Roosevelt,” 70; Sussman, Dear FDR, 123–24; Howe, “President's Mailbag,” 23. Lila Sussman estimates that of the five to eight thousand letters the president received every day, as many as 80 percent were requests from individuals for financial assistance. Sussman, Dear FDR, 132, n. 23. A selection of letters to FDR requesting aid in McElvaine appears identical in content and tone to those discussed in this chapter. McElvaine, Down and Out, 72–73, 75, 86–87, 106–9, 116, 158, 161–62, 166–70. During 1933 before Mrs. Roosevelt had sufficient staff to handle the mail herself, some letters were sent to Ellen Woodward at the Federal Emergency Relief Administration for response, and from there such letters went to state relief offices and were often not retained. Some of these can be found in FERA Old General Subject Series, Correspondence with White House, NARA; Whitehurst, “Dear Mr. President,” 12; Wick-enden, “Relief Programs,” 181. Those letters that remain in the FERA files appear to be indistinguishable from the letters analyzed in this chapter, however the fact that forwarded mail was often not preserved rules out a systematic comparison. Those letters to the Roosevelts that concerned eligibility for particular government programs such as work relief under the Civil Works Administration or Works Progress Administration were generally forwarded to the relevant agencies for responses. Some of these letters can still be found in the files of certain agencies, chiefly FERA. (p.318) These letters primarily deal with complaints about unfairness, incompetence, or corruption in relief administration. Some are described in Cohen, Making a New Deal, 283.

(3) . Roosevelt, This I Remember, 100.

(4) . The 1930 Census with a few exceptions defined as urban any city or incorporated municipality with a population of 2,500 or more. In addition, a few townships and other political subdivisions mostly located on the eastern seaboard having a total population of 10,000 or more, and a population density of 1,000 or more per square mile, were considered to be urban. All other areas were defined as rural. The Census Bureau also modified for New England its practice of classifying all towns of 2,500 or more inhabitants as urban, because it resulted in classifying as urban a considerable number of places that were mainly rural in their general characteristics, which resulted in some minor changes of classification. Steven Rug-gles, Matthew Sobek, Trent Alexander, Catherine A. Fitch, Ronald Goeken, Patricia Kelly Hall, Miriam King, and Chad Ronender, Integrated Public Use Microdata Series: version 3.0 (Machine Readable Database), Minneapolis: Minnesota Population Center (producer and distributer), 2004, http:/ipums.umn/usa/ (accessed May 30, 2007).

(5) . US Bureau of the Census, Fifteenth Decennial Census of the United States: 1930, Population, vol. 2, 842, Table 4 (“Marital Condition of the Population 15 Years Old and Over, By Sex, Color, and Nativity, for the United States, 1890–1930”).

(6) . Pearson, “Significance of Urban History,” 231; Maisel, “Variables Commonly Ignored,” 266; Thornthwaite, Internal Migration.

(7) . US Census Bureau, Fifteenth Decennial Census of the United States: 1930, Population, vol. 2, General Report, 842, Table 4 (“Marital Condition of the Population 15 Years Old and Over, By Sex, Color, and Nativity, for the United States, 1890–1930”).

(8) . The average age of male writers was twenty-nine, compared with 35.4 for women, and this difference was significant (p 〈 0.01).

(9) . US Census Bureau, Fifteenth Decennial Census of the United States: 1930, Population, vol. 4, Occupations by States, 68, Table 24 (“Number and Proportion of Women 15 Years Old and Over Gainfully Occupied”).

(10) . For the full sample of letters, chisq (6) = 10.21 (p = 0.1161). The same test on the census subset produced the same result, chisq (6) = 7.95 (p = 0.2419). Furthermore, the letters for which a 1930 manuscript census form was found and those for which a census form was not found were not significantly different in their regional distribution, chisq (6) = 8.1905 (p = 0.224).

(11) . US Census Bureau, Fifteenth Decennial Census of the United States, Population, vol. 2, General Report, vi, Map of the United States, Showing Geographic Divisions; US Bureau of the Census, Seventeenth Decennial Census of the United States, Characteristics of the Population, vol. 2, xi, Figure 1 (“Regions and Geographic Divisions of the United States”).

(12) . US Census Bureau, Fifteenth Decennial Census of the United States:1930, Population, vol. 2, General Report, 12, Table 8 (“Rural-Farm and Rural Nonfarm (Village) Population, By Divisions and States: 1930 and 1920”).

(13) . Although the 1930 census tabulated information on other racial and ethnic groups, such as Mexicans and Japanese, only a few letters were received from members of these groups, which comprised 1.2 percent and 0.1 percent of the US population, respectively. US Census Bureau, Fifteenth Decennial Census of the United States: 1930. (p.319) Population, vol. 2, General Report, Statistics by Subjects, 32, Table 4 (“Total Population by Color or Race for the United States, 1790–1930”).

(14) . US Census Bureau, Fifteenth Decennial Census of the United States: 1930, Population, vol. 2, General Report, Table 3 (“Illiteracy in the Urban, Rural-Farm, and Rural-Non-farm Population, by Color and Nativity, for the United States: 1930”) and Table 10 (“Illiteracy in the Population Ten Years Old and Over, by Color and Nativity, by Divisions and States: 1930”). US Census Bureau, Fifteenth Census of the United States: 1930, Population, vol. 6, Families, 53, Table 60 (“Families Having Radio Set, in Urban, Rural-Farm, and Rural-Nonfarm Areas, by Divisions and States: 1930”). Craig, “How America Adopted Radio,” 189; Joint Committee on Radio Research, “Rural Radio Ownership,” 6–10 (detailing how southern and rural households lagged in radio adoption); Podber, “Early Radio in Appalachia,” 394–96; Beasley, White House Press Conferences, 9, 57, 127, 280, 310, 333; Cohen, Dear Mrs. Roosevelt, 5; Cook, Eleanor Roosevelt, 13; Hoff-Wilson and Lightman, Without Precedent, 10; Knepper, Dear Mrs. Roosevelt, 23; Kearney, Anna Eleanor Roosevelt, 113.

(15) . See US Census Bureau and Edwards, Socio-Economic Grouping; US Census Bureau, Sixteenth Decennial Census of the United States: 1940, Population, vol. 2, Characteristics of the Population, United States Summary, 15–17, Table XI (“Employed Workers, by Major Occupation Group and Sex, for the United States: 1940”); US Census Bureau and Edwards, Comparative Occupation Statistics, 183–89. The occupational categories used were: professional, skilled worker, proprietor, clerk, semiskilled worker, farmer, other laborer, servant class, and farm laborer (as well as unknown and silent).

(16) . US Census Bureau, Fifteenth Decennial Census of the United States: 1930, Unemployment, vol. 2, pp. 2–3, 51–52, Table 19 (“Unemployment Returns-Class A, By Sex and Family Relationship, by Geographic Divisions and States: 1930”) and Table 20 (“Unemployment Returns-Class B, By Sex and Family Relationship, by Geographic Divisions and States: 1930”); US Census Bureau, Fifteenth Decennial Census of the United States: 1930, Population, vol. 6, Families, 49, Table 56 (“Families Classified by Sex of Head and by Age of Man Head, By Divisions and States: 1930.”

(17) . US Census Bureau, Fifteenth Census of the United States: 1930, Population, vol. 6, Families, 53, Table 60 (“Families Having Radio Set, in Urban, Rural-Farm, and Rural-Nonfarm Areas, by Divisions and States: 1930”).

(18) . See http://1930census.archives.gov/FAQ.html (accessed May 30, 2007).

(19) . Using a conventional calculation for median, the median family size for writers in the census subset is five. The 1930 Census Bureau used a formula for computing the median number of people in a family replicated here for the census subset in order to make it comparable to the census figure rather than using a more conventional calculation. See US Census Bureau, Fifteenth Decennial Census of the United States: 1930, Population, vol. 6, “Families,” 7n3.

(20) . The proportion of median housing value and the number of people in the family are essentially uncorrelated (Pearson's r = 0.2).

(21) . Gerteis, “Possession of Civic Virtue,” 593.

(22) . As discussed in chapter 7, writers only rarely deployed these excuses, which are analytically related as non-fault-based claims that have been emphasized as important in the existing welfare state literature. In order to guard against the possibility that the nonsignificant results were due to the low occurrence of these variables in (p.320) the data, I combined them into a composite variable “entitlement,” which occurred in 151 letters.

(23) . The reference group is “North East.” The four-region system established in the 1950 census was used rather than the seven-region format in use in 1930 in order to reduce the degrees of freedom used in the analysis. The overall effect of the four dummy variable set for region was tested using a LR chi-square test between a model containing the region dummies and one that did not. The inclusion of the variables for region did not significantly improve the model fitting for any of the reported models on any of the nineteen dependent variables.

(24) . The reference group is “1933–36 Period.” The letters cover the period 1933–40, coinciding roughly with Roosevelt's first and second terms. There are two justifications for splitting the decade into halves: first, just as there were regional differences and differences between the city and the country that might have affected the content of the letters, there were changes in the economic circumstances of the nation over the course of the 1930s that might have affected the strategies writers adopted. Splitting the decade into two parts rather than ten allowed for the capture of some of that variation with a minimum of added variables. Second, claims of entitlement might be expected to lag the development of institutions during the first term, so that it would be reasonable to expect that claims based on fault would diminish and those based on entitlement would increase in the second half of the decade.

(25) . One hundred seven writers requested assistance with unpaid debts, and 198 requested aid in the form of a loan; fifty-six writers were in both categories.

(26) . Long and Freese, Regression Models for Categorical Dependent Variables Using Stata, 372.