The effects of screening for gestational diabetes

Abstract

I estimate the effects of screening “low-risk” women for gestational diabetes using a regression discontinuity design and exploiting exogenous variation in testing at the overweight threshold of the body mass index in Finland. I find that screening low-risk mothers just above the overweight threshold increases the number of mothers diagnosed with gestational diabetes. There is a 1.5 percentage point, or 27%, increase in the probability of being diagnosed with gestational diabetes at the threshold, which translates into a 10.7 percentage point local average treatment effect given the 14.0 percentage point jump in the screening rates. The estimates on the effect on insulin treatment are, however, small and imprecise, suggesting that screening low-risk mothers did not result in diagnoses needing insulin treatment. The cost estimates in the existing literature suggest that the policy is cost-effective. The results also suggest that universal screening could decrease health disparities between mothers with low and high levels of education, given that gestational diabetes is treated if diagnosed. The effect on the probability of having an abnormal test result is over twice as great for the less educated mothers compared to the more educated mothers. Large effects of around 10–20% on adverse birth outcomes (low birth weight, macrosomia, metabolic testing, and C-section) can be ruled out.

Appendices

Appendix A Additional tables and figures

Table 3 Mothers’ covariates and outcome variables

Table 4 Main results from local linear regressions by mother’s education

Table 5 Main results from local linear regressions by birth parity

Table 6 Local linear regressions: Donut hole

Table 7 Main regressions with quadratic trends

Table 8 Effect of screening for gestational diabetes on outcomes: Results from using the age cutoff of 25 years

Fig. 4

Histogram: Mother’s BMI. All mothers in the main analysis sample are included. The unrestricted means are in bins of 0.1 BMI units, and the trend is linear fit with a bandwidth of 1.0 BMI units

Fig. 5

Mother’s characteristics. See Fig. 2 for figure notes

Fig. 6

Mother’s height and weight. See Fig. 2 for figure notes

Fig. 7

Bandwidth plots for abnormal test result and insulin treatment. Regression estimates using different bandwidths and triangular kernel and controlling for mothers’ covariates. The vertical line is the optimal bandwidth (CCT)

Fig. 8

Age RD setup. The unrestricted means are in bins of 1 month, and the trend is linear fit with a bandwidth of 12 months

Fig. 9

First-time mothers over age 25. See Fig. 2 for figure notes. First-time mothers over age 25 should be screened regardless of their BMI

Fig. 10

Age RD setup for overweight mothers. The unrestricted means are in bins of one month, and the trend is linear fit with a bandwidth of 12 months

Appendix B Effect of screening on birth outcomes

The results suggest that the expansion of screening increased the number of cases that could be treated with dietary advice. A natural question is whether screening affects birth outcomes. However, it is not clear how screening might affect birth outcomes and the mechanism behind any such effects. Mothers are already given dietary advice at the beginning of their pregnancy, as Fig. 1 shows, and I could not find an effect on insulin treatment. Furthermore, there might be some unanticipated consequences if a mother has normal test results while having a “bad lifestyle” or diet or in the case of abnormal test results causing stress, both of which might cancel out the positive effects of treatment. It has been shown that maternal psychological factors like stress may significantly contribute to pregnancy complications and unfavorable development of the (unborn) child (see, e.g., Mulder et al. 2002, for a review).

A larger problem concerns issues of power. As I do not expect to see large effects of screening on many birth outcomes, and as adverse birth outcomes are rather rare in today’s Finland, the sample sizes needed to detect the effects would be enormous. Table 9 shows the population sizes needed to detect different effect sizes for multiple birth outcomes with the power of 0.8. The estimated sample sizes are calculated using Pearson’s chi-squared test and are estimations for a two-sample proportions test. In order to detect a 1% effect on low birth weight, I would need 10 million observations. With the current sample size of roughly 70,000 births varying with the bandwidth used, I am able to detect with high certainty an effect size of 7% to 20% depending on the outcome variable. Hence, it is highly unlikely that significant estimates would be found, given the small sample and expected effect sizes.

Table 10 provides estimates from the local linear regressions using a triangular kernel for birth outcomes. The estimates from the first-stage regression (i.e., the jump in the treatment) varies from 12.1 to 13.3 percentage points, depending on the bandwidth used. The optimal bandwidths obtained with the CCT bandwidth selector are now considerably lower (1.40 to 1.97 BMI units) than in the previous analysis. The intention-to-treat effects for adverse birth outcomes are small, ranging from −0.002 to 0.007 and imprecise due to the low fraction of adverse birth outcomes and insufficient mass around the cutoff. The IV estimates range from −0.012 to 0.054. Hence, I cannot rule out that there is no effect or a larger effect. For most of the outcomes, however, I can rule out effects larger than 10–20%. However, for C-section, which is the most frequent outcome, I am able to detect an effect of 7% with a power of 0.8 and a sample size of roughly 55,000, or 27,500 on both sides,¹³ as shown in Table 9. Hence, the chances that the effect is larger than 7% are very small because the sample size is around 52,000.

Table 9 Power calculations: Estimated sample sizes for a two-sample proportions test for different outcome variables and effect sizes with a power of 0.8

Table 10 Effect of screening for gestational diabetes on birth outcomes: Main results from the local linear regressions