Prevalence of mental health problems
Figure 1 Panel A reports the prevalence of depression among PhD students, other Master’s graduates, two population samples representing different age groups, and various PhD student subsamples. In a given year, 6.8% of the PhD students receive medication or diagnosis for depression, while the corresponding prevalence for other Master’s graduates is 5.3%. Sample splits among the PhD student population suggest depression is more prevalent among women, students who were older at the time of enrolling in a PhD program, students of soft sciences, and students who have a parent with a prior depression or anxiety diagnosis; high school GPA appears to have little predictive power for depression prevalence. The highest prevalence of depression among the above sample splits, recorded for PhD students with parents struggling with mental health, is about the same as that for the general population of 18–70-year-olds (9.4% vs. 9.0%). Table 1 reports full details.
Table 1
Annual prevalence of depression (%), anxiety (%), and suicides (‱) among PhD students, their subgroups, and benchmarks. Panel B reports on student age in the year of starting PhD studies and excludes statistics on suicide to preserve privacy. “Meta-analysis” refers to estimates from a meta-analysis on PhD students’ mental health5, whereas “Derived estimate” adjusts our base estimate for PhD students by multiplying it with two factors: the ratio of clinically significant screened depression and medically validated depression in a representative sample of the Swedish population of 18–70-year-olds9, and the ratio of medically validated depression among 18–70-year-olds and 20–39-year-olds in the Swedish population. Standard errors are reported below the means. N refers to the number of individual-year observations.
| Depression (%) | Anxiety (%) | Suicide (‱) | Age | Female (%) | N |
Panel A: PhD students and benchmarks | | | | | | |
PhD students | 6.84 | 5.06 | 0.21 | 29.90 | 46.86 | 48,186 |
| 0.12 | 0.10 | 0.21 | 0.02 | 0.23 | |
Other Master’s graduates | 5.34 | 4.01 | 0.27 | 29.64 | 55.47 | 448,500 |
| 0.03 | 0.03 | 0.08 | 0.005 | 0.07 | |
Population, 20–39 years | 6.96 | 6.09 | 1.11 | 29.53 | 48.92 | 26,689,985 |
| 0.00 | 0.00 | 0.02 | 0.001 | 0.01 | |
Population, 18–70 years | 8.99 | 10.02 | 1.36 | 43.49 | 49.35 | 69,822,141 |
| 0.00 | 0.00 | 0.01 | 0.002 | 0.01 | |
Panel B: PhD student subgroups | | | | | | |
Men | 5.22 | 3.87 | Not rep. | 29.73 | 0.00 | 25,606 |
| 0.14 | 0.12 | Not rep. | 0.02 | 0.00 | |
Women | 8.68 | 6.42 | Not rep. | 30.11 | 1.00 | 22,580 |
| 0.19 | 0.16 | Not rep. | 0.02 | 0.00 | |
Mentally healthy parents | 6.65 | 4.97 | Not rep. | 29.85 | 46.95 | 45,085 |
| 0.12 | 0.10 | Not rep. | 0.02 | 0.24 | |
Mentally unhealthy parents | 9.45 | 5.86 | Not rep. | 29.99 | 42.70 | 836 |
| 1.01 | 0.81 | Not rep. | 0.11 | 1.71 | |
GPA ≤ median | 6.89 | 5.52 | Not rep. | 30.61 | 41.51 | 17,010 |
| 0.19 | 0.18 | Not rep. | 0.03 | 0.38 | |
GPA > median | 6.82 | 4.81 | Not rep. | 29.52 | 49.78 | 31,176 |
| 0.14 | 0.12 | Not rep. | 0.02 | 0.28 | |
Age ≤ 26 | 5.13 | 3.84 | Not rep. | 27.43 | 44.49 | 24,648 |
| 0.14 | 0.12 | Not rep. | 0.01 | 0.32 | |
Age > 26 | 8.64 | 6.34 | Not rep. | 32.50 | 49.35 | 23,538 |
| 0.18 | 0.16 | Not rep. | 0.02 | 0.33 | |
Hard sciences | 5.51 | 3.92 | Not rep. | 28.63 | 39.57 | 21,628 |
| 0.16 | 0.13 | Not rep. | 0.02 | 0.33 | |
Soft sciences | 7.87 | 6.06 | Not rep. | 29.98 | 55.17 | 13,704 |
| 0.23 | 0.20 | Not rep. | 0.03 | 0.42 | |
Panel C: Clinically significant | | | | | | |
Meta-analysis | 24 | 17 | | | | |
| 3 | 3 | | | | |
Derived estimate | 10.59 | 12.20 | | | | |
The prevalence of depression reported in prior work is much larger than that for the PhD students in our sample. A meta-analysis estimates the prevalence of clinically significant symptoms of depression at 24%, i.e. at 3.5 times as high as in our sample5. Because prior estimates are based on screening instruments and thus are not necessarily directly comparable with estimates obtained from medical data, we report a derived estimate that aims to make our results more comparable with existing evidence. The derived estimate, 10.6%, corresponds to 44% of the meta-analysis estimate and it is well below the lower bound of the meta-analysis estimate’s 95% confidence interval, 18%.
The derived estimate multiplies our base estimate of 6.8% with two factors. The first factor, 1.2, captures differences between the measures used in this study and the meta-analysis. It is obtained by dividing screened prevalence of clinically significant depression in a representative sample of 18–70-year-olds in the Swedish population6 with prevalence of medically validated depression in the population of the same age range in our data. The second factor, 1.29, eliminates the possible effects of age on treatment behavior, including the likelihood to seek treatment10–11, the type of treatment received12, and the healthcare setting utilized. The factor is computed as the ratio of medically validated depression among 18–70-year-olds and 20–39-year-olds in the Swedish population. The latter group has the same average age, 30, as PhD students and other Master’s graduates in our sample, and its age range also reflects that of these populations.
Given realistic assumptions and bounds for unobserved parameters, the derived estimate cannot have any significant negative bias (see supplement for details). Rather, if anything, our derived estimate is inflated, making comparisons between it and the meta-analysis estimate conservative.
Figure 1 Panel B reports on the prevalence of anxiety. In a given year, 5.1% of the PhD students have anxiety, while the corresponding prevalence for Master’s graduates (general population of 20–39-year-olds) is 4.0% (6.1%). The sample split results are qualitatively similar for anxiety and depression except that students with lower-than-median GPA have larger prevalence of anxiety than those with higher-than average GPAs.
A meta-analysis estimates the prevalence of clinically significant symptoms of anxiety at 17%, i.e. at 3.4 times the incidence in our sample. By using the same method as for depression to compute the derived estimate for PhD students’ prevalence of anxiety, we arrive at 5.1% × 1.47 × 1.64 = 12.2%. This figure is 72% of the corresponding meta-analysis estimate and narrowly within its 95% confidence interval. As for depression, our estimate for the difference between the derived estimate and the meta-analysis estimate for anxiety is conservative (see supplement for details).
Past research finds prevalence of suicidal ideation among PhD students can exceed 10%5. Our results suggest these suicide ideations are highly unlikely to lead to completed suicides: there are 48,186 individual-PhD year observations in our sample but only one suicide. The propensity of suicides for PhD students is about the same as that for Master’s graduates, and 19% of the corresponding propensity in the population of 20–39-year-olds.
How does mental health change after starting PhD studies?
Next, we turn to an analysis designed to study the association between mental health and entering a PhD program. Figure 2 Panel A plots the prevalence of depression for PhD students and the control group. There are three noteworthy patterns in the figure. First, depression prevalence on average doubles during the seven years the subjects are being followed, likely reflecting the effect of age on treatment behavior10–12 and being ill13–14. Second, PhD students have consistently higher prevalence rates, indicating average differences in vulnerability to mental health problems among the two groups. Third, the prevalence rates increase more for PhD students upon entering the doctoral program than for the control group upon graduating with a Master’s degree. Before entering or graduation, the two groups show a similar pre-trend. Panel B shows anxiety follows a similar pattern.
Panel C reports the coefficients from a difference-in-differences linear probability model explaining annual prevalence of depression with a constant, the treatment indicator (being a PhD student), the event time indicators (for treated t = 1 equals first year in the PhD program; for control t = 1 equals year of graduation with a Master’s degree), and their interactions. Controls include year, gender, age, parental mental health indicators, and high-school GPA.
The interaction term is positive for all years in the PhD program, and it is significantly different from zero at the 5% level from the third year onwards. The point estimate for the interaction term in the fifth year suggests a 1.57 percentage point (t = 3.78) increase in depression for graduate students compared to their peers. Table 2 shows these results are not sensitive to whether we include GPA and parental mental health controls in the regression. Given that there are no discernible differences in trends in depression prior to year 0 and its prevalence among the PhD students in their fifth year of studies is 8.21%, the estimate suggests 19% of the prevalence of their depression can be attributed to PhD studies.
Table 2
A linear probability model is estimated on depression and anxiety. The independent variables are the treatment indicator (being a PhD student), the event time indicators (for treated t = 1 equals first year in PhD program; for control t = 1 equals year of graduation with Master’s degree), and their interactions. The baseline models also control for year, gender, and age, and additional controls include high school GPA and indicators for parental anxiety and depression. The t-values reported in parentheses below coefficients assume clustering at the individual level.
Dependent variable | | Depression | | Anxiety |
Specification | | Baseline | | Additional controls | | Baseline | | Additional controls |
Reported coefficient | | Main effect | Inte-raction with PhD student | | Main effect | Inte-raction with PhD student | | Main effect | Inte-raction with PhD student | | Main effect | Inte-raction with PhD student |
PhD student | | 0.58 | | | 0.50 | | | 0.54 | | | 0.55 | |
| | (1.87) | | | (1.59) | | | (1.96) | | | (1.99) | |
Event year = − 2 | | -0.61 | -0.20 | | -0.61 | -0.20 | | -0.65 | -0.43 | | -0.65 | -0.43 |
| | (-6.87) | (-0.69) | | (-6.87) | (-0.69) | | (-6.98) | (-1.42) | | (-6.98) | (-1.42) |
Event year = − 1 | | -0.23 | -0.10 | | -0.23 | -0.10 | | -0.28 | 0.03 | | -0.28 | 0.03 |
| | (-3.10) | (-0.46) | | (-3.10) | (-0.46) | | (-3.12) | (0.10) | | (-3.12) | (0.10) |
Event year = + 1 | | 0.24 | 0.45 | | 0.24 | 0.45 | | 0.38 | 0.11 | | 0.38 | 0.11 |
| | (3.06) | (1.74) | | (3.06) | (1.74) | | (4.11) | (0.37) | | (4.11) | (0.37) |
Event year = + 2 | | 0.80 | 0.41 | | 0.80 | 0.40 | | 0.63 | 0.16 | | 0.63 | 0.16 |
| | (8.30) | (1.26) | | (8.31) | (1.26) | | (6.16) | (0.48) | | (6.16) | (0.48) |
Event year = + 3 | | 1.39 | 0.74 | | 1.39 | 0.74 | | 1.20 | 0.21 | | 1.20 | 0.21 |
| | (12.99) | (2.06) | | (13.00) | (2.05) | | (10.82) | (0.58) | | (10.82) | (0.58) |
Event year = + 4 | | 1.89 | 1.09 | | 1.89 | 1.09 | | 1.50 | 0.94 | | 1.50 | 0.94 |
| | (16.52) | (2.87) | | (16.52) | (2.87) | | (13.00) | (2.41) | | (13.00) | (2.41) |
Event year = + 5 | | 2.61 | 1.57 | | 2.61 | 1.57 | | 1.74 | 0.58 | | 1.74 | 0.58 |
| | (21.30) | (3.78) | | (21.31) | (3.78) | | (14.75) | (1.43) | | (14.75) | (1.43) |
High-school GPA | | | | | 0.28 | | | | | | -0.01 | |
| | | | | (2.90) | | | | | | (-0.19) | |
Parental anxiety | | | | | 2.16 | | | | | | 1.22 | |
| | | | | (1.94) | | | | | | (1.79) | |
Parental depression | | | | | 2.93 | | | | | | 1.18 | |
| | | | | (3.59) | | | | | | (2.30) | |
Mean dependent variable | | 4.89 | | 4.89 | | 3.61 | | 3.61 |
Adjusted R2 | | 0.01 | | 0.02 | | 0.01 | | 0.01 |
Number of observations | | 374,097 | | 374,097 | | 374,097 | | 374,097 |
Panel D repeats the difference-in-differences analysis for anxiety. The results are broadly consistent with those for depression, although weaker. The interaction term is the largest in the fourth year of PhD studies, when it constitutes a 0.94 percentage point (t = 2.41) increase in anxiety compared to the control group and accounts for 16% of PhD student anxiety diagnoses. The interaction term is clearly positive (0.58%) also in the fifth year, but not statistically significant at conventional levels (t = 1.43).